[ { "op": "Conv2dFwdOp", "label": "bottleneck-expand-1x1", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 22.5, "cannbot_us": 28.5, "ratio": 1.2666666666666666, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 2, "ours_us": 22.5, "cannbot_us": 28.5, "ratio": 1.2666666666666666, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 3, "ours_us": 22.5, "cannbot_us": 28.25, "ratio": 1.2555555555555555, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 4, "ours_us": 22.5, "cannbot_us": 28.75, "ratio": 1.2777777777777777, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 5, "ours_us": 22.5, "cannbot_us": 28.0, "ratio": 1.2444444444444445, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" } ], "errors": [], "ours_us": 22.5, "cannbot_us": 28.5, "ratio": 1.2666666666666666, "ratio_min": 1.2444444444444445, "ratio_max": 1.2777777777777777 }, { "op": "Conv2dFwdOp", "label": "bottleneck-expand-1x1", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 35.375, "cannbot_us": 35.5, "ratio": 1.0035335689045937, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 2, "ours_us": 34.75, "cannbot_us": 35.5, "ratio": 1.0215827338129497, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 3, "ours_us": 35.25, "cannbot_us": 35.75, "ratio": 1.0141843971631206, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 4, "ours_us": 35.0, "cannbot_us": 35.75, "ratio": 1.0214285714285714, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" }, { "rep": 5, "ours_us": 35.25, "cannbot_us": 35.5, "ratio": 1.0070921985815602, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-10/result.json" } ], "errors": [], "ours_us": 35.25, "cannbot_us": 35.5, "ratio": 1.0141843971631206, "ratio_min": 1.0035335689045937, "ratio_max": 1.0215827338129497 }, { "op": "Conv2dFwdOp", "label": "bottleneck-expand-1x1-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-27/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "bottleneck-expand-1x1-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-27/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "bottleneck-reduce-1x1", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 33.625, "cannbot_us": 29.0, "ratio": 0.862453531598513, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 2, "ours_us": 33.375, "cannbot_us": 28.5, "ratio": 0.8539325842696629, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 3, "ours_us": 33.75, "cannbot_us": 28.75, "ratio": 0.8518518518518519, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 4, "ours_us": 33.75, "cannbot_us": 29.0, "ratio": 0.8592592592592593, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 5, "ours_us": 33.625, "cannbot_us": 28.75, "ratio": 0.8550185873605948, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" } ], "errors": [], "ours_us": 33.625, "cannbot_us": 28.75, "ratio": 0.8550185873605948, "ratio_min": 0.8518518518518519, "ratio_max": 0.862453531598513 }, { "op": "Conv2dFwdOp", "label": "bottleneck-reduce-1x1", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 40.0, "cannbot_us": 39.5, "ratio": 0.9875, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 2, "ours_us": 39.75, "cannbot_us": 40.0, "ratio": 1.0062893081761006, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 3, "ours_us": 40.0, "cannbot_us": 39.75, "ratio": 0.99375, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 4, "ours_us": 39.75, "cannbot_us": 39.25, "ratio": 0.9874213836477987, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" }, { "rep": 5, "ours_us": 39.75, "cannbot_us": 39.625, "ratio": 0.9968553459119497, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-11/result.json" } ], "errors": [], "ours_us": 39.75, "cannbot_us": 39.625, "ratio": 0.99375, "ratio_min": 0.9874213836477987, "ratio_max": 1.0062893081761006 }, { "op": "Conv2dFwdOp", "label": "bottleneck-reduce-1x1-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-28/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "bottleneck-reduce-1x1-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-28/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "classifier-1x1", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 27.625, "cannbot_us": 74.5, "ratio": 2.6968325791855206, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 2, "ours_us": 33.0, "cannbot_us": 76.0, "ratio": 2.303030303030303, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 3, "ours_us": 28.0, "cannbot_us": 76.75, "ratio": 2.7410714285714284, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 4, "ours_us": 32.25, "cannbot_us": 76.125, "ratio": 2.36046511627907, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 5, "ours_us": 28.125, "cannbot_us": 76.625, "ratio": 2.7244444444444444, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" } ], "errors": [], "ours_us": 28.125, "cannbot_us": 76.125, "ratio": 2.6968325791855206, "ratio_min": 2.303030303030303, "ratio_max": 2.7410714285714284 }, { "op": "Conv2dFwdOp", "label": "classifier-1x1", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 32.5, "cannbot_us": 74.625, "ratio": 2.296153846153846, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 2, "ours_us": 38.125, "cannbot_us": 77.125, "ratio": 2.0229508196721313, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 3, "ours_us": 34.125, "cannbot_us": 77.0, "ratio": 2.2564102564102564, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 4, "ours_us": 38.625, "cannbot_us": 77.125, "ratio": 1.9967637540453074, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" }, { "rep": 5, "ours_us": 35.0, "cannbot_us": 77.125, "ratio": 2.2035714285714287, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-13/result.json" } ], "errors": [], "ours_us": 35.0, "cannbot_us": 77.125, "ratio": 2.2035714285714287, "ratio_min": 1.9967637540453074, "ratio_max": 2.296153846153846 }, { "op": "Conv2dFwdOp", "label": "classifier-1x1-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-30/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "classifier-1x1-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-30/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "deeplabv3-aspp-3x3-rate12", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 2048, 32, 32), (256, 2048, 3, 3), torch.float16) refused: NotImplementedError: conv2d dilation=12 is not supported; this kernel implements dilation=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-14/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "deeplabv3-aspp-3x3-rate12", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 2048, 32, 32), (256, 2048, 3, 3), torch.float16) refused: NotImplementedError: conv2d dilation=12 is not supported; this kernel implements dilation=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-14/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "default-dense-cube-fp16", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 87.5, "cannbot_us": 26.0, "ratio": 0.29714285714285715, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 2, "ours_us": 87.5, "cannbot_us": 28.375, "ratio": 0.3242857142857143, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 3, "ours_us": 87.5, "cannbot_us": 26.0, "ratio": 0.29714285714285715, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 4, "ours_us": 87.5, "cannbot_us": 27.625, "ratio": 0.3157142857142857, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 5, "ours_us": 87.5, "cannbot_us": 25.75, "ratio": 0.29428571428571426, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" } ], "errors": [], "ours_us": 87.5, "cannbot_us": 26.0, "ratio": 0.29714285714285715, "ratio_min": 0.29428571428571426, "ratio_max": 0.3242857142857143 }, { "op": "Conv2dFwdOp", "label": "default-dense-cube-fp16", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 87.75, "cannbot_us": 26.25, "ratio": 0.29914529914529914, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 2, "ours_us": 88.0, "cannbot_us": 30.25, "ratio": 0.34375, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 3, "ours_us": 87.75, "cannbot_us": 26.5, "ratio": 0.301994301994302, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 4, "ours_us": 88.0, "cannbot_us": 28.125, "ratio": 0.3196022727272727, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" }, { "rep": 5, "ours_us": 87.75, "cannbot_us": 26.0, "ratio": 0.2962962962962963, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-31/result.json" } ], "errors": [], "ours_us": 87.75, "cannbot_us": 26.5, "ratio": 0.301994301994302, "ratio_min": 0.2962962962962963, "ratio_max": 0.34375 }, { "op": "Conv2dFwdOp", "label": "dense-cube-bf16", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 53.75, "cannbot_us": 36.625, "ratio": 0.6813953488372093, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 2, "ours_us": 53.875, "cannbot_us": 39.5, "ratio": 0.7331786542923434, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 3, "ours_us": 53.875, "cannbot_us": 39.75, "ratio": 0.7378190255220418, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 4, "ours_us": 54.0, "cannbot_us": 45.375, "ratio": 0.8402777777777778, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 5, "ours_us": 53.75, "cannbot_us": 39.75, "ratio": 0.7395348837209302, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" } ], "errors": [], "ours_us": 53.875, "cannbot_us": 39.75, "ratio": 0.7378190255220418, "ratio_min": 0.6813953488372093, "ratio_max": 0.8402777777777778 }, { "op": "Conv2dFwdOp", "label": "dense-cube-bf16", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 54.0, "cannbot_us": 36.0, "ratio": 0.6666666666666666, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 2, "ours_us": 54.0, "cannbot_us": 38.375, "ratio": 0.7106481481481481, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 3, "ours_us": 54.0, "cannbot_us": 45.25, "ratio": 0.8379629629629629, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 4, "ours_us": 54.0, "cannbot_us": 42.875, "ratio": 0.7939814814814815, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" }, { "rep": 5, "ours_us": 54.0, "cannbot_us": 37.625, "ratio": 0.6967592592592593, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-36/result.json" } ], "errors": [], "ours_us": 54.0, "cannbot_us": 38.375, "ratio": 0.7106481481481481, "ratio_min": 0.6666666666666666, "ratio_max": 0.8379629629629629 }, { "op": "Conv2dFwdOp", "label": "depthwise-direct-fp16", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 4, 8, 9), (4, 1, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=4 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-35/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "depthwise-direct-fp16", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 4, 8, 9), (4, 1, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=4 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-35/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "dilation-cube-fp16", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 16, 9, 10), (16, 16, 3, 3), torch.float16) refused: NotImplementedError: conv2d dilation=2 is not supported; this kernel implements dilation=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-33/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "dilation-cube-fp16", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 16, 9, 10), (16, 16, 3, 3), torch.float16) refused: NotImplementedError: conv2d dilation=2 is not supported; this kernel implements dilation=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-33/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "fp32-direct", "dtype": "float32", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: cannbot's conv2d.py::build_conv2d_kernel declares support for ('float16', 'bfloat16'); float32 is not one of them\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-37/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "fp32-direct", "dtype": "float32", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: cannbot's conv2d.py::build_conv2d_kernel declares support for ('float16', 'bfloat16'); float32 is not one of them\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-37/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "grouped-direct-fp16", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 8, 9, 10), (8, 2, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=4 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-34/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "grouped-direct-fp16", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 8, 9, 10), (8, 2, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=4 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-34/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "highres-3x3-s1", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 29082.5, "cannbot_us": 3656.125, "ratio": 0.12571563655119058, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 2, "ours_us": 29077.25, "cannbot_us": 3654.0, "ratio": 0.12566525376368123, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 3, "ours_us": 29079.125, "cannbot_us": 3657.875, "ratio": 0.12579040806764302, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 4, "ours_us": 29082.625, "cannbot_us": 3657.375, "ratio": 0.12575807720245336, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 5, "ours_us": 29078.25, "cannbot_us": 3655.875, "ratio": 0.12572541332439194, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" } ], "errors": [], "ours_us": 29079.125, "cannbot_us": 3656.125, "ratio": 0.12572541332439194, "ratio_min": 0.12566525376368123, "ratio_max": 0.12579040806764302 }, { "op": "Conv2dFwdOp", "label": "highres-3x3-s1", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 29084.875, "cannbot_us": 3657.875, "ratio": 0.12576553964904438, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 2, "ours_us": 29081.375, "cannbot_us": 3661.875, "ratio": 0.12591822085441284, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 3, "ours_us": 29086.875, "cannbot_us": 3662.0, "ratio": 0.12589870860998303, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 4, "ours_us": 29081.0, "cannbot_us": 3660.125, "ratio": 0.12585966782435268, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" }, { "rep": 5, "ours_us": 29085.125, "cannbot_us": 3664.375, "ratio": 0.12598794057099635, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-4/result.json" } ], "errors": [], "ours_us": 29084.875, "cannbot_us": 3661.875, "ratio": 0.12589870860998303, "ratio_min": 0.12576553964904438, "ratio_max": 0.12598794057099635 }, { "op": "Conv2dFwdOp", "label": "highres-3x3-s1-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-21/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "highres-3x3-s1-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-21/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "large-grid-depthwise-fp16", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((16, 256, 56, 56), (256, 1, 1, 1), torch.float16) refused: NotImplementedError: conv2d groups=256 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-38/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "large-grid-depthwise-fp16", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((16, 256, 56, 56), (256, 1, 1, 1), torch.float16) refused: NotImplementedError: conv2d groups=256 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-38/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "late-stage-1x1", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 12.0, "cannbot_us": 20.75, "ratio": 1.7291666666666667, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 2, "ours_us": 12.625, "cannbot_us": 20.75, "ratio": 1.6435643564356435, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 3, "ours_us": 13.125, "cannbot_us": 21.0, "ratio": 1.6, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 4, "ours_us": 12.5, "cannbot_us": 21.0, "ratio": 1.68, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 5, "ours_us": 12.5, "cannbot_us": 20.75, "ratio": 1.66, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" } ], "errors": [], "ours_us": 12.5, "cannbot_us": 20.75, "ratio": 1.66, "ratio_min": 1.6, "ratio_max": 1.7291666666666667 }, { "op": "Conv2dFwdOp", "label": "late-stage-1x1", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 21.25, "cannbot_us": 24.75, "ratio": 1.1647058823529413, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 2, "ours_us": 21.25, "cannbot_us": 25.25, "ratio": 1.188235294117647, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 3, "ours_us": 21.625, "cannbot_us": 25.5, "ratio": 1.1791907514450868, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 4, "ours_us": 21.5, "cannbot_us": 25.25, "ratio": 1.1744186046511629, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" }, { "rep": 5, "ours_us": 21.0, "cannbot_us": 25.25, "ratio": 1.2023809523809523, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-12/result.json" } ], "errors": [], "ours_us": 21.25, "cannbot_us": 25.25, "ratio": 1.1791907514450868, "ratio_min": 1.1647058823529413, "ratio_max": 1.2023809523809523 }, { "op": "Conv2dFwdOp", "label": "late-stage-1x1-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-29/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "late-stage-1x1-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-29/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "midres-5x5-s1", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 4193.625, "cannbot_us": 794.625, "ratio": 0.18948403827237773, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 2, "ours_us": 4195.875, "cannbot_us": 795.5, "ratio": 0.18959096731909317, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 3, "ours_us": 4193.625, "cannbot_us": 800.625, "ratio": 0.1909147813645712, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 4, "ours_us": 4192.875, "cannbot_us": 793.125, "ratio": 0.18916018245237456, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 5, "ours_us": 4193.875, "cannbot_us": 793.5, "ratio": 0.1892044946499359, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" } ], "errors": [], "ours_us": 4193.625, "cannbot_us": 794.625, "ratio": 0.18948403827237773, "ratio_min": 0.18916018245237456, "ratio_max": 0.1909147813645712 }, { "op": "Conv2dFwdOp", "label": "midres-5x5-s1", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 4199.5, "cannbot_us": 804.375, "ratio": 0.1915406596023336, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 2, "ours_us": 4201.75, "cannbot_us": 806.125, "ratio": 0.1918545843993574, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 3, "ours_us": 4199.875, "cannbot_us": 807.625, "ratio": 0.19229738980326797, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 4, "ours_us": 4198.25, "cannbot_us": 806.375, "ratio": 0.19207407848508307, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" }, { "rep": 5, "ours_us": 4198.625, "cannbot_us": 804.875, "ratio": 0.1916996635803388, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-5/result.json" } ], "errors": [], "ours_us": 4199.5, "cannbot_us": 806.125, "ratio": 0.1918545843993574, "ratio_min": 0.1915406596023336, "ratio_max": 0.19229738980326797 }, { "op": "Conv2dFwdOp", "label": "midres-5x5-s1-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-22/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "midres-5x5-s1-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-22/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "mobilenetv2-depthwise", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 32, 56, 56), (32, 1, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=32 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-15/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "mobilenetv2-depthwise", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 32, 56, 56), (32, 1, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=32 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-15/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-1x1", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 15.25, "cannbot_us": 18.25, "ratio": 1.1967213114754098, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 2, "ours_us": 15.25, "cannbot_us": 18.25, "ratio": 1.1967213114754098, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 3, "ours_us": 15.5, "cannbot_us": 18.25, "ratio": 1.1774193548387097, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 4, "ours_us": 15.25, "cannbot_us": 18.5, "ratio": 1.2131147540983607, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 5, "ours_us": 15.25, "cannbot_us": 18.25, "ratio": 1.1967213114754098, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" } ], "errors": [], "ours_us": 15.25, "cannbot_us": 18.25, "ratio": 1.1967213114754098, "ratio_min": 1.1774193548387097, "ratio_max": 1.2131147540983607 }, { "op": "Conv2dFwdOp", "label": "resnet-1x1", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 37.0, "cannbot_us": 26.0, "ratio": 0.7027027027027027, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 2, "ours_us": 36.75, "cannbot_us": 25.75, "ratio": 0.7006802721088435, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 3, "ours_us": 36.75, "cannbot_us": 25.75, "ratio": 0.7006802721088435, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 4, "ours_us": 36.75, "cannbot_us": 26.0, "ratio": 0.7074829931972789, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" }, { "rep": 5, "ours_us": 36.75, "cannbot_us": 25.75, "ratio": 0.7006802721088435, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-9/result.json" } ], "errors": [], "ours_us": 36.75, "cannbot_us": 25.75, "ratio": 0.7006802721088435, "ratio_min": 0.7006802721088435, "ratio_max": 0.7074829931972789 }, { "op": "Conv2dFwdOp", "label": "resnet-1x1", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 15.625, "cannbot_us": 18.5, "ratio": 1.184, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 2, "ours_us": 15.75, "cannbot_us": 18.0, "ratio": 1.1428571428571428, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 3, "ours_us": 15.75, "cannbot_us": 18.5, "ratio": 1.1746031746031746, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 4, "ours_us": 15.75, "cannbot_us": 18.125, "ratio": 1.1507936507936507, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 5, "ours_us": 15.625, "cannbot_us": 18.375, "ratio": 1.176, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" } ], "errors": [], "ours_us": 15.75, "cannbot_us": 18.375, "ratio": 1.1746031746031746, "ratio_min": 1.1428571428571428, "ratio_max": 1.184 }, { "op": "Conv2dFwdOp", "label": "resnet-1x1", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 37.75, "cannbot_us": 25.875, "ratio": 0.6854304635761589, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 2, "ours_us": 37.75, "cannbot_us": 26.0, "ratio": 0.6887417218543046, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 3, "ours_us": 37.5, "cannbot_us": 25.75, "ratio": 0.6866666666666666, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 4, "ours_us": 37.75, "cannbot_us": 26.0, "ratio": 0.6887417218543046, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" }, { "rep": 5, "ours_us": 37.5, "cannbot_us": 25.875, "ratio": 0.69, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-8/result.json" } ], "errors": [], "ours_us": 37.75, "cannbot_us": 25.875, "ratio": 0.6887417218543046, "ratio_min": 0.6854304635761589, "ratio_max": 0.69 }, { "op": "Conv2dFwdOp", "label": "resnet-1x1-bias", "dtype": "bfloat16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-26/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-1x1-bias", "dtype": "bfloat16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-26/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-1x1-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-25/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-1x1-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-25/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-3x3", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1199.375, "cannbot_us": 398.0, "ratio": 0.33183949973944765, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 2, "ours_us": 1198.25, "cannbot_us": 397.625, "ratio": 0.33183809722512, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 3, "ours_us": 1197.5, "cannbot_us": 400.625, "ratio": 0.3345511482254697, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 4, "ours_us": 1199.5, "cannbot_us": 396.375, "ratio": 0.3304501875781576, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 5, "ours_us": 1198.0, "cannbot_us": 398.75, "ratio": 0.3328464106844741, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" } ], "errors": [], "ours_us": 1198.25, "cannbot_us": 398.0, "ratio": 0.33183949973944765, "ratio_min": 0.3304501875781576, "ratio_max": 0.3345511482254697 }, { "op": "Conv2dFwdOp", "label": "resnet-3x3", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1203.5, "cannbot_us": 406.5, "ratio": 0.3377648525135023, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 2, "ours_us": 1204.375, "cannbot_us": 405.25, "ratio": 0.33648157758173325, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 3, "ours_us": 1206.875, "cannbot_us": 403.0, "ratio": 0.33392024857586744, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 4, "ours_us": 1207.5, "cannbot_us": 404.75, "ratio": 0.33519668737060043, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" }, { "rep": 5, "ours_us": 1206.0, "cannbot_us": 406.625, "ratio": 0.3371683250414594, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-1/result.json" } ], "errors": [], "ours_us": 1206.0, "cannbot_us": 405.25, "ratio": 0.33648157758173325, "ratio_min": 0.33392024857586744, "ratio_max": 0.3377648525135023 }, { "op": "Conv2dFwdOp", "label": "resnet-3x3", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1198.75, "cannbot_us": 399.0, "ratio": 0.33284671532846716, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 2, "ours_us": 1198.75, "cannbot_us": 399.625, "ratio": 0.33336809176225235, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 3, "ours_us": 1198.875, "cannbot_us": 397.75, "ratio": 0.33176936711500365, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 4, "ours_us": 1199.875, "cannbot_us": 398.5, "ratio": 0.3321179289509324, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 5, "ours_us": 1198.875, "cannbot_us": 397.625, "ratio": 0.33166510270044836, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" } ], "errors": [], "ours_us": 1198.875, "cannbot_us": 398.5, "ratio": 0.3321179289509324, "ratio_min": 0.33166510270044836, "ratio_max": 0.33336809176225235 }, { "op": "Conv2dFwdOp", "label": "resnet-3x3", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1203.125, "cannbot_us": 407.625, "ratio": 0.33880519480519483, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 2, "ours_us": 1202.375, "cannbot_us": 408.25, "ratio": 0.3395363343382888, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 3, "ours_us": 1201.5, "cannbot_us": 406.0, "ratio": 0.3379109446525177, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 4, "ours_us": 1201.625, "cannbot_us": 407.125, "ratio": 0.3388120253822948, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" }, { "rep": 5, "ours_us": 1202.125, "cannbot_us": 404.875, "ratio": 0.3367994176978268, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-0/result.json" } ], "errors": [], "ours_us": 1202.125, "cannbot_us": 407.125, "ratio": 0.33880519480519483, "ratio_min": 0.3367994176978268, "ratio_max": 0.3395363343382888 }, { "op": "Conv2dFwdOp", "label": "resnet-3x3-bias", "dtype": "bfloat16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-18/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-3x3-bias", "dtype": "bfloat16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-18/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-3x3-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-17/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnet-3x3-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-17/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnext-grouped-3x3", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 128, 28, 28), (256, 4, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=32 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-16/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "resnext-grouped-3x3", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: ['cannbot: docs/reports/R298-data/cannbot-kernels/conv2d.py::build_conv2d_kernel((1, 128, 28, 28), (256, 4, 3, 3), torch.float16) refused: NotImplementedError: conv2d groups=32 is not supported; this kernel implements groups=1 only']", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-16/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stage-transition-3x3-s2", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 8358.625, "cannbot_us": 1514.5, "ratio": 0.1811900880826691, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 2, "ours_us": 8354.0, "cannbot_us": 1532.0, "ratio": 0.18338520469236294, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 3, "ours_us": 8355.375, "cannbot_us": 1514.875, "ratio": 0.1813054470924405, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 4, "ours_us": 8358.25, "cannbot_us": 1525.5, "ratio": 0.18251428229593517, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 5, "ours_us": 8354.875, "cannbot_us": 1506.375, "ratio": 0.1802989272729993, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" } ], "errors": [], "ours_us": 8355.375, "cannbot_us": 1514.875, "ratio": 0.1813054470924405, "ratio_min": 0.1802989272729993, "ratio_max": 0.18338520469236294 }, { "op": "Conv2dFwdOp", "label": "stage-transition-3x3-s2", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 8369.125, "cannbot_us": 1523.25, "ratio": 0.1820082744611892, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 2, "ours_us": 8363.75, "cannbot_us": 1542.375, "ratio": 0.18441189657749216, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 3, "ours_us": 8365.75, "cannbot_us": 1523.5, "ratio": 0.18211158593072946, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 4, "ours_us": 8361.625, "cannbot_us": 1538.25, "ratio": 0.1839654373402299, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" }, { "rep": 5, "ours_us": 8366.5, "cannbot_us": 1521.625, "ratio": 0.181871152811809, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-3/result.json" } ], "errors": [], "ours_us": 8365.75, "cannbot_us": 1523.5, "ratio": 0.18211158593072946, "ratio_min": 0.181871152811809, "ratio_max": 0.18441189657749216 }, { "op": "Conv2dFwdOp", "label": "stage-transition-3x3-s2-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-20/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stage-transition-3x3-s2-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-20/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stage-transition-5x5-s2", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 24026.375, "cannbot_us": 3914.625, "ratio": 0.1629303213655826, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 2, "ours_us": 24033.25, "cannbot_us": 3927.25, "ratio": 0.1634090270770703, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 3, "ours_us": 24026.375, "cannbot_us": 3940.625, "ratio": 0.16401246546763712, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 4, "ours_us": 24029.625, "cannbot_us": 3911.5, "ratio": 0.16277823728002414, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 5, "ours_us": 24029.375, "cannbot_us": 3933.125, "ratio": 0.16367987099123468, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" } ], "errors": [], "ours_us": 24029.375, "cannbot_us": 3927.25, "ratio": 0.1634090270770703, "ratio_min": 0.16277823728002414, "ratio_max": 0.16401246546763712 }, { "op": "Conv2dFwdOp", "label": "stage-transition-5x5-s2", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 24037.875, "cannbot_us": 3918.625, "ratio": 0.16301877765817485, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 2, "ours_us": 24036.5, "cannbot_us": 3932.25, "ratio": 0.16359494934786678, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 3, "ours_us": 24036.125, "cannbot_us": 3945.75, "ratio": 0.16415915626998945, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 4, "ours_us": 24036.25, "cannbot_us": 3916.625, "ratio": 0.16294659108637996, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" }, { "rep": 5, "ours_us": 24037.0, "cannbot_us": 3953.75, "ratio": 0.16448600074884553, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-6/result.json" } ], "errors": [], "ours_us": 24036.5, "cannbot_us": 3932.25, "ratio": 0.16359494934786678, "ratio_min": 0.16294659108637996, "ratio_max": 0.16448600074884553 }, { "op": "Conv2dFwdOp", "label": "stage-transition-5x5-s2-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-23/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stage-transition-5x5-s2-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-23/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stem-3x3-s2", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 389.875, "cannbot_us": 255.75, "ratio": 0.6559794806027573, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 2, "ours_us": 387.75, "cannbot_us": 254.25, "ratio": 0.655705996131528, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 3, "ours_us": 389.5, "cannbot_us": 258.625, "ratio": 0.6639922978177151, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 4, "ours_us": 387.875, "cannbot_us": 248.5, "ratio": 0.6406703190460844, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 5, "ours_us": 389.25, "cannbot_us": 258.375, "ratio": 0.6637764932562621, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" } ], "errors": [], "ours_us": 389.25, "cannbot_us": 255.75, "ratio": 0.6559794806027573, "ratio_min": 0.6406703190460844, "ratio_max": 0.6639922978177151 }, { "op": "Conv2dFwdOp", "label": "stem-3x3-s2", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 394.25, "cannbot_us": 260.375, "ratio": 0.660431198478123, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 2, "ours_us": 395.375, "cannbot_us": 260.0, "ratio": 0.6576035409421436, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 3, "ours_us": 394.5, "cannbot_us": 262.5, "ratio": 0.6653992395437263, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 4, "ours_us": 394.5, "cannbot_us": 253.375, "ratio": 0.6422686945500634, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" }, { "rep": 5, "ours_us": 394.75, "cannbot_us": 262.875, "ratio": 0.6659278024065864, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-2/result.json" } ], "errors": [], "ours_us": 394.5, "cannbot_us": 260.375, "ratio": 0.660431198478123, "ratio_min": 0.6422686945500634, "ratio_max": 0.6659278024065864 }, { "op": "Conv2dFwdOp", "label": "stem-3x3-s2-bias", "dtype": "float16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-19/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stem-3x3-s2-bias", "dtype": "float16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-19/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stride-padding-nondiv-cube-fp16", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 36.5, "cannbot_us": 21.75, "ratio": 0.5958904109589042, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 2, "ours_us": 36.25, "cannbot_us": 24.25, "ratio": 0.6689655172413793, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 3, "ours_us": 36.5, "cannbot_us": 22.0, "ratio": 0.6027397260273972, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 4, "ours_us": 36.5, "cannbot_us": 24.25, "ratio": 0.6643835616438356, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 5, "ours_us": 36.5, "cannbot_us": 22.0, "ratio": 0.6027397260273972, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" } ], "errors": [], "ours_us": 36.5, "cannbot_us": 22.0, "ratio": 0.6027397260273972, "ratio_min": 0.5958904109589042, "ratio_max": 0.6689655172413793 }, { "op": "Conv2dFwdOp", "label": "stride-padding-nondiv-cube-fp16", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 36.75, "cannbot_us": 21.5, "ratio": 0.5850340136054422, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 2, "ours_us": 36.625, "cannbot_us": 24.0, "ratio": 0.6552901023890785, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 3, "ours_us": 36.75, "cannbot_us": 21.25, "ratio": 0.5782312925170068, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 4, "ours_us": 36.75, "cannbot_us": 24.0, "ratio": 0.6530612244897959, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" }, { "rep": 5, "ours_us": 36.75, "cannbot_us": 21.5, "ratio": 0.5850340136054422, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-32/result.json" } ], "errors": [], "ours_us": 36.75, "cannbot_us": 21.5, "ratio": 0.5850340136054422, "ratio_min": 0.5782312925170068, "ratio_max": 0.6552901023890785 }, { "op": "Conv2dFwdOp", "label": "stride2", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1811.5, "cannbot_us": 362.0, "ratio": 0.19983439138835218, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 2, "ours_us": 1811.75, "cannbot_us": 359.25, "ratio": 0.19828894715054504, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 3, "ours_us": 1811.75, "cannbot_us": 366.125, "ratio": 0.20208362080861045, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 4, "ours_us": 1812.25, "cannbot_us": 362.125, "ratio": 0.1998206649192992, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 5, "ours_us": 1812.5, "cannbot_us": 364.375, "ratio": 0.20103448275862068, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" } ], "errors": [], "ours_us": 1811.75, "cannbot_us": 362.125, "ratio": 0.19983439138835218, "ratio_min": 0.19828894715054504, "ratio_max": 0.20208362080861045 }, { "op": "Conv2dFwdOp", "label": "stride2", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1817.0, "cannbot_us": 374.625, "ratio": 0.20617776554760595, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 2, "ours_us": 1819.75, "cannbot_us": 375.0, "ratio": 0.20607226267344417, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 3, "ours_us": 1816.875, "cannbot_us": 372.75, "ratio": 0.20515995872033024, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 4, "ours_us": 1816.25, "cannbot_us": 375.625, "ratio": 0.2068134893324157, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" }, { "rep": 5, "ours_us": 1818.125, "cannbot_us": 375.75, "ratio": 0.20666895840495014, "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-7/result.json" } ], "errors": [], "ours_us": 1817.0, "cannbot_us": 375.0, "ratio": 0.20617776554760595, "ratio_min": 0.20515995872033024, "ratio_max": 0.2068134893324157 }, { "op": "Conv2dFwdOp", "label": "stride2-bias", "dtype": "bfloat16", "regime": "eager", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-24/result.json" } ] }, { "op": "Conv2dFwdOp", "label": "stride2-bias", "dtype": "bfloat16", "regime": "graph", "n": 0, "required_n": 5, "publishable": false, "trials": [], "errors": [ { "rep": 1, "reason": "RuntimeError: no admissible cannbot candidate: [\"cannbot: Conv2dFwdOp: 3 operands -- cannbot's build_conv2d_kernel has no bias parameter, so a case that declares bias_shape is a different operator expression and is refused rather than compared without the bias\"]", "source": "/home/dyq/workspace/docs/reports/R322-data/conv/case-24/result.json" } ] }, { "op": "CountNonzeroFwdOp", "label": "3d-multidim-reduce", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 114.75, "cannbot_us": 10.75, "ratio": 0.09368191721132897, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 116.0, "cannbot_us": 11.25, "ratio": 0.09698275862068965, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 115.75, "cannbot_us": 11.5, "ratio": 0.09935205183585313, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 114.25, "cannbot_us": 10.75, "ratio": 0.09409190371991247, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 113.375, "cannbot_us": 10.5, "ratio": 0.0926130099228225, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 114.75, "cannbot_us": 10.75, "ratio": 0.09409190371991247, "ratio_min": 0.0926130099228225, "ratio_max": 0.09935205183585313 }, { "op": "CountNonzeroFwdOp", "label": "3d-multidim-reduce", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 116.0, "cannbot_us": 18.25, "ratio": 0.15732758620689655, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 114.625, "cannbot_us": 17.5, "ratio": 0.15267175572519084, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 114.5, "cannbot_us": 18.0, "ratio": 0.1572052401746725, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 115.25, "cannbot_us": 16.875, "ratio": 0.14642082429501085, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 114.25, "cannbot_us": 17.75, "ratio": 0.15536105032822758, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 114.625, "cannbot_us": 17.75, "ratio": 0.15536105032822758, "ratio_min": 0.14642082429501085, "ratio_max": 0.15732758620689655 }, { "op": "CountNonzeroFwdOp", "label": "sparsity-hidden", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 93.875, "cannbot_us": 29.0, "ratio": 0.30892143808255657, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 93.625, "cannbot_us": 28.5, "ratio": 0.30440587449933243, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 94.25, "cannbot_us": 29.75, "ratio": 0.3156498673740053, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 93.5, "cannbot_us": 28.75, "ratio": 0.3074866310160428, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 93.875, "cannbot_us": 29.5, "ratio": 0.31424766977363516, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 93.875, "cannbot_us": 29.0, "ratio": 0.30892143808255657, "ratio_min": 0.30440587449933243, "ratio_max": 0.3156498673740053 }, { "op": "CountNonzeroFwdOp", "label": "sparsity-hidden", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 103.75, "cannbot_us": 37.0, "ratio": 0.3566265060240964, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 103.0, "cannbot_us": 36.75, "ratio": 0.3567961165048544, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 103.625, "cannbot_us": 35.875, "ratio": 0.34620024125452353, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 102.75, "cannbot_us": 34.75, "ratio": 0.3381995133819951, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 103.5, "cannbot_us": 37.875, "ratio": 0.36594202898550726, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 103.5, "cannbot_us": 36.75, "ratio": 0.3566265060240964, "ratio_min": 0.3381995133819951, "ratio_max": 0.36594202898550726 }, { "op": "CountNonzeroFwdOp", "label": "sparsity-hidden", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 93.25, "cannbot_us": 28.75, "ratio": 0.30831099195710454, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 94.375, "cannbot_us": 28.0, "ratio": 0.29668874172185433, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 93.5, "cannbot_us": 27.875, "ratio": 0.29812834224598933, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 93.75, "cannbot_us": 28.0, "ratio": 0.2986666666666667, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 93.25, "cannbot_us": 28.0, "ratio": 0.3002680965147453, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 93.5, "cannbot_us": 28.0, "ratio": 0.2986666666666667, "ratio_min": 0.29668874172185433, "ratio_max": 0.30831099195710454 }, { "op": "CountNonzeroFwdOp", "label": "sparsity-hidden", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 104.0, "cannbot_us": 36.5, "ratio": 0.35096153846153844, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 104.125, "cannbot_us": 36.375, "ratio": 0.34933973589435774, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 101.75, "cannbot_us": 34.25, "ratio": 0.3366093366093366, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 101.5, "cannbot_us": 34.25, "ratio": 0.3374384236453202, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 101.875, "cannbot_us": 35.5, "ratio": 0.34846625766871164, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 101.875, "cannbot_us": 35.5, "ratio": 0.34846625766871164, "ratio_min": 0.3366093366093366, "ratio_max": 0.35096153846153844 }, { "op": "CountNonzeroFwdOp", "label": "sparsity-seq", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 40.25, "cannbot_us": 8.5, "ratio": 0.2111801242236025, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 41.0, "cannbot_us": 8.5, "ratio": 0.2073170731707317, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 40.875, "cannbot_us": 8.75, "ratio": 0.21406727828746178, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 41.0, "cannbot_us": 9.0, "ratio": 0.21951219512195122, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 40.75, "cannbot_us": 8.75, "ratio": 0.2147239263803681, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 40.875, "cannbot_us": 8.75, "ratio": 0.21406727828746178, "ratio_min": 0.2073170731707317, "ratio_max": 0.21951219512195122 }, { "op": "CountNonzeroFwdOp", "label": "sparsity-seq", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 53.0, "cannbot_us": 13.0, "ratio": 0.24528301886792453, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 2, "ours_us": 52.5, "cannbot_us": 12.25, "ratio": 0.23333333333333334, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 3, "ours_us": 55.5, "cannbot_us": 13.25, "ratio": 0.23873873873873874, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 4, "ours_us": 55.25, "cannbot_us": 12.5, "ratio": 0.22624434389140272, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" }, { "rep": 5, "ours_us": 55.5, "cannbot_us": 12.875, "ratio": 0.23198198198198197, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/CountNonzeroFwdOp/coverage/CountNonzeroFwdOp.json" } ], "errors": [], "ours_us": 55.25, "cannbot_us": 12.875, "ratio": 0.23333333333333334, "ratio_min": 0.22624434389140272, "ratio_max": 0.24528301886792453 }, { "op": "EluFwdOp", "label": "mlp-hidden", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 36.0, "cannbot_us": 30.5, "ratio": 0.8472222222222222, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 34.875, "cannbot_us": 29.25, "ratio": 0.8387096774193549, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 36.0, "cannbot_us": 29.125, "ratio": 0.8090277777777778, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 34.375, "cannbot_us": 28.5, "ratio": 0.8290909090909091, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 36.25, "cannbot_us": 29.75, "ratio": 0.8206896551724138, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 36.0, "cannbot_us": 29.25, "ratio": 0.8290909090909091, "ratio_min": 0.8090277777777778, "ratio_max": 0.8472222222222222 }, { "op": "EluFwdOp", "label": "mlp-hidden", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 43.75, "cannbot_us": 36.0, "ratio": 0.8228571428571428, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 45.25, "cannbot_us": 36.75, "ratio": 0.8121546961325967, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 43.75, "cannbot_us": 36.0, "ratio": 0.8228571428571428, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 44.625, "cannbot_us": 33.5, "ratio": 0.7507002801120448, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 43.75, "cannbot_us": 35.75, "ratio": 0.8171428571428572, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 43.75, "cannbot_us": 36.0, "ratio": 0.8171428571428572, "ratio_min": 0.7507002801120448, "ratio_max": 0.8228571428571428 }, { "op": "EluFwdOp", "label": "mlp-hidden", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 34.875, "cannbot_us": 28.75, "ratio": 0.8243727598566308, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 35.25, "cannbot_us": 29.5, "ratio": 0.8368794326241135, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 35.0, "cannbot_us": 29.875, "ratio": 0.8535714285714285, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 35.25, "cannbot_us": 28.875, "ratio": 0.8191489361702128, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 35.0, "cannbot_us": 29.25, "ratio": 0.8357142857142857, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 35.0, "cannbot_us": 29.25, "ratio": 0.8357142857142857, "ratio_min": 0.8191489361702128, "ratio_max": 0.8535714285714285 }, { "op": "EluFwdOp", "label": "mlp-hidden", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 42.0, "cannbot_us": 34.25, "ratio": 0.8154761904761905, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 44.75, "cannbot_us": 35.5, "ratio": 0.7932960893854749, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 43.0, "cannbot_us": 34.75, "ratio": 0.8081395348837209, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 43.5, "cannbot_us": 36.0, "ratio": 0.8275862068965517, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 42.25, "cannbot_us": 34.75, "ratio": 0.8224852071005917, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 43.0, "cannbot_us": 34.75, "ratio": 0.8154761904761905, "ratio_min": 0.7932960893854749, "ratio_max": 0.8275862068965517 }, { "op": "EluFwdOp", "label": "mlp-hidden-wide", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 68.25, "cannbot_us": 57.25, "ratio": 0.8388278388278388, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 68.75, "cannbot_us": 56.75, "ratio": 0.8254545454545454, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 68.25, "cannbot_us": 57.0, "ratio": 0.8351648351648352, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 68.25, "cannbot_us": 56.0, "ratio": 0.8205128205128205, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 68.5, "cannbot_us": 57.0, "ratio": 0.8321167883211679, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 68.25, "cannbot_us": 57.0, "ratio": 0.8321167883211679, "ratio_min": 0.8205128205128205, "ratio_max": 0.8388278388278388 }, { "op": "EluFwdOp", "label": "mlp-hidden-wide", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 78.0, "cannbot_us": 61.75, "ratio": 0.7916666666666666, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 80.25, "cannbot_us": 62.5, "ratio": 0.778816199376947, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 79.25, "cannbot_us": 62.125, "ratio": 0.7839116719242902, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 76.75, "cannbot_us": 61.5, "ratio": 0.8013029315960912, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 78.0, "cannbot_us": 62.25, "ratio": 0.7980769230769231, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 78.0, "cannbot_us": 62.125, "ratio": 0.7916666666666666, "ratio_min": 0.778816199376947, "ratio_max": 0.8013029315960912 }, { "op": "EluFwdOp", "label": "mlp-hidden-wide", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 68.25, "cannbot_us": 56.0, "ratio": 0.8205128205128205, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 66.25, "cannbot_us": 55.25, "ratio": 0.8339622641509434, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 67.625, "cannbot_us": 55.0, "ratio": 0.8133086876155268, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 67.0, "cannbot_us": 55.0, "ratio": 0.8208955223880597, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 68.5, "cannbot_us": 56.75, "ratio": 0.8284671532846716, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 67.625, "cannbot_us": 55.25, "ratio": 0.8208955223880597, "ratio_min": 0.8133086876155268, "ratio_max": 0.8339622641509434 }, { "op": "EluFwdOp", "label": "mlp-hidden-wide", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 75.75, "cannbot_us": 61.25, "ratio": 0.8085808580858086, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 2, "ours_us": 76.25, "cannbot_us": 61.0, "ratio": 0.8, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 3, "ours_us": 75.375, "cannbot_us": 60.5, "ratio": 0.802653399668325, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 4, "ours_us": 76.875, "cannbot_us": 60.75, "ratio": 0.7902439024390244, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/EluFwdOp/coverage/EluFwdOp.json" }, { "rep": 5, "ours_us": 76.125, "cannbot_us": 62.25, "ratio": 0.8177339901477833, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/EluFwdOp/coverage/EluFwdOp.json" } ], "errors": [], "ours_us": 76.125, "cannbot_us": 61.0, "ratio": 0.802653399668325, "ratio_min": 0.7902439024390244, "ratio_max": 0.8177339901477833 }, { "op": "GeluFwdOp", "label": "llama-8b-ffn-decode", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 3.5, "cannbot_us": 2.0, "ratio": 0.5714285714285714, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 2, "ours_us": 3.5, "cannbot_us": 1.75, "ratio": 0.5, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 3, "ours_us": 3.5, "cannbot_us": 2.0, "ratio": 0.5714285714285714, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 4, "ours_us": 3.5, "cannbot_us": 1.75, "ratio": 0.5, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 5, "ours_us": 3.5, "cannbot_us": 1.75, "ratio": 0.5, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GeluFwdOp/coverage/GeluFwdOp.json" } ], "errors": [], "ours_us": 3.5, "cannbot_us": 1.75, "ratio": 0.5, "ratio_min": 0.5, "ratio_max": 0.5714285714285714 }, { "op": "GeluFwdOp", "label": "llama-8b-ffn-decode", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 3.5, "cannbot_us": 3.75, "ratio": 1.0714285714285714, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 2, "ours_us": 3.5, "cannbot_us": 3.5, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 3, "ours_us": 3.5, "cannbot_us": 4.0, "ratio": 1.1428571428571428, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 4, "ours_us": 3.5, "cannbot_us": 3.5, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 5, "ours_us": 3.5, "cannbot_us": 3.75, "ratio": 1.0714285714285714, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GeluFwdOp/coverage/GeluFwdOp.json" } ], "errors": [], "ours_us": 3.5, "cannbot_us": 3.75, "ratio": 1.0714285714285714, "ratio_min": 1.0, "ratio_max": 1.1428571428571428 }, { "op": "GeluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 184.5, "cannbot_us": 128.25, "ratio": 0.6951219512195121, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 2, "ours_us": 184.0, "cannbot_us": 128.5, "ratio": 0.6983695652173914, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 3, "ours_us": 184.375, "cannbot_us": 128.25, "ratio": 0.6955932203389831, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 4, "ours_us": 184.0, "cannbot_us": 128.25, "ratio": 0.6970108695652174, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 5, "ours_us": 184.0, "cannbot_us": 128.0, "ratio": 0.6956521739130435, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GeluFwdOp/coverage/GeluFwdOp.json" } ], "errors": [], "ours_us": 184.0, "cannbot_us": 128.25, "ratio": 0.6956521739130435, "ratio_min": 0.6951219512195121, "ratio_max": 0.6983695652173914 }, { "op": "GeluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 194.25, "cannbot_us": 137.5, "ratio": 0.7078507078507078, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 2, "ours_us": 192.375, "cannbot_us": 135.25, "ratio": 0.7030539311241065, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 3, "ours_us": 195.0, "cannbot_us": 135.0, "ratio": 0.6923076923076923, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 4, "ours_us": 192.75, "cannbot_us": 137.5, "ratio": 0.7133592736705577, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 5, "ours_us": 193.75, "cannbot_us": 135.5, "ratio": 0.6993548387096774, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GeluFwdOp/coverage/GeluFwdOp.json" } ], "errors": [], "ours_us": 193.75, "cannbot_us": 135.5, "ratio": 0.7030539311241065, "ratio_min": 0.6923076923076923, "ratio_max": 0.7133592736705577 }, { "op": "GeluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 181.75, "cannbot_us": 125.25, "ratio": 0.6891334250343879, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 2, "ours_us": 181.875, "cannbot_us": 125.5, "ratio": 0.6900343642611684, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 3, "ours_us": 182.0, "cannbot_us": 125.75, "ratio": 0.6909340659340659, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 4, "ours_us": 181.75, "cannbot_us": 125.25, "ratio": 0.6891334250343879, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 5, "ours_us": 181.25, "cannbot_us": 125.25, "ratio": 0.6910344827586207, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GeluFwdOp/coverage/GeluFwdOp.json" } ], "errors": [], "ours_us": 181.75, "cannbot_us": 125.25, "ratio": 0.6900343642611684, "ratio_min": 0.6891334250343879, "ratio_max": 0.6910344827586207 }, { "op": "GeluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 192.375, "cannbot_us": 134.5, "ratio": 0.6991552956465237, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 2, "ours_us": 189.5, "cannbot_us": 132.375, "ratio": 0.6985488126649076, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 3, "ours_us": 193.125, "cannbot_us": 134.625, "ratio": 0.6970873786407767, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 4, "ours_us": 189.75, "cannbot_us": 132.375, "ratio": 0.6976284584980237, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GeluFwdOp/coverage/GeluFwdOp.json" }, { "rep": 5, "ours_us": 190.75, "cannbot_us": 134.375, "ratio": 0.7044560943643512, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GeluFwdOp/coverage/GeluFwdOp.json" } ], "errors": [], "ours_us": 190.75, "cannbot_us": 134.375, "ratio": 0.6985488126649076, "ratio_min": 0.6970873786407767, "ratio_max": 0.7044560943643512 }, { "op": "GemmFwdOp", "label": "ds-v3-decode-down", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 57.0, "cannbot_us": 60.125, "ratio": 1.0548245614035088, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 57.25, "cannbot_us": 61.5, "ratio": 1.074235807860262, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 57.75, "cannbot_us": 61.0, "ratio": 1.0562770562770563, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 58.5, "cannbot_us": 62.75, "ratio": 1.0726495726495726, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 59.75, "cannbot_us": 62.0, "ratio": 1.0376569037656904, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 57.75, "cannbot_us": 61.5, "ratio": 1.0562770562770563, "ratio_min": 1.0376569037656904, "ratio_max": 1.074235807860262 }, { "op": "GemmFwdOp", "label": "ds-v3-decode-down", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 66.75, "cannbot_us": 69.0, "ratio": 1.0337078651685394, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 67.375, "cannbot_us": 70.25, "ratio": 1.0426716141001855, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 66.125, "cannbot_us": 68.375, "ratio": 1.0340264650283555, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 67.75, "cannbot_us": 72.0, "ratio": 1.062730627306273, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 67.75, "cannbot_us": 70.75, "ratio": 1.044280442804428, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 67.375, "cannbot_us": 70.25, "ratio": 1.0426716141001855, "ratio_min": 1.0337078651685394, "ratio_max": 1.062730627306273 }, { "op": "GemmFwdOp", "label": "ds-v3-decode-gate-up", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 70.25, "cannbot_us": 77.375, "ratio": 1.101423487544484, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 71.125, "cannbot_us": 75.625, "ratio": 1.063268892794376, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 72.5, "cannbot_us": 78.125, "ratio": 1.0775862068965518, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 71.5, "cannbot_us": 75.875, "ratio": 1.0611888111888113, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 73.125, "cannbot_us": 78.125, "ratio": 1.0683760683760684, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 71.5, "cannbot_us": 77.375, "ratio": 1.0683760683760684, "ratio_min": 1.0611888111888113, "ratio_max": 1.101423487544484 }, { "op": "GemmFwdOp", "label": "ds-v3-decode-gate-up", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 72.5, "cannbot_us": 87.0, "ratio": 1.2, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 73.75, "cannbot_us": 84.875, "ratio": 1.1508474576271186, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 74.5, "cannbot_us": 86.5, "ratio": 1.1610738255033557, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 73.0, "cannbot_us": 84.375, "ratio": 1.1558219178082192, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 75.25, "cannbot_us": 87.375, "ratio": 1.1611295681063123, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 73.75, "cannbot_us": 86.5, "ratio": 1.1610738255033557, "ratio_min": 1.1508474576271186, "ratio_max": 1.2 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-attn-proj", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 806.625, "cannbot_us": 794.75, "ratio": 0.9852781651944832, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 805.875, "cannbot_us": 797.125, "ratio": 0.98914223669924, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 807.5, "cannbot_us": 794.125, "ratio": 0.98343653250774, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 807.0, "cannbot_us": 795.25, "ratio": 0.9854399008674102, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 809.375, "cannbot_us": 796.625, "ratio": 0.9842471042471043, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 807.0, "cannbot_us": 795.25, "ratio": 0.9852781651944832, "ratio_min": 0.98343653250774, "ratio_max": 0.98914223669924 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-attn-proj", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 816.5, "cannbot_us": 805.375, "ratio": 0.9863747703612982, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 816.375, "cannbot_us": 805.625, "ratio": 0.986832031848109, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 817.25, "cannbot_us": 804.5, "ratio": 0.9843988987457938, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 814.125, "cannbot_us": 803.125, "ratio": 0.9864885613388608, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 819.375, "cannbot_us": 804.625, "ratio": 0.9819984744469871, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 816.5, "cannbot_us": 804.625, "ratio": 0.9863747703612982, "ratio_min": 0.9819984744469871, "ratio_max": 0.986832031848109 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-attn-proj", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 826.25, "cannbot_us": 796.125, "ratio": 0.9635400907715582, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 825.75, "cannbot_us": 794.375, "ratio": 0.962004238570996, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 827.0, "cannbot_us": 794.875, "ratio": 0.9611547762998791, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 828.75, "cannbot_us": 800.0, "ratio": 0.9653092006033183, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 826.5, "cannbot_us": 794.25, "ratio": 0.9609800362976406, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 826.5, "cannbot_us": 794.875, "ratio": 0.962004238570996, "ratio_min": 0.9609800362976406, "ratio_max": 0.9653092006033183 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-attn-proj", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 835.5, "cannbot_us": 804.125, "ratio": 0.9624476361460204, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 836.25, "cannbot_us": 804.25, "ratio": 0.9617339312406576, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 834.25, "cannbot_us": 803.5, "ratio": 0.9631405454000599, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 836.0, "cannbot_us": 808.5, "ratio": 0.9671052631578947, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 834.625, "cannbot_us": 801.5, "ratio": 0.96031151714842, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 835.5, "cannbot_us": 804.125, "ratio": 0.9624476361460204, "ratio_min": 0.96031151714842, "ratio_max": 0.9671052631578947 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-down", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 474.625, "cannbot_us": 396.5, "ratio": 0.8353963655517513, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 474.125, "cannbot_us": 396.75, "ratio": 0.8368046401265489, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 474.125, "cannbot_us": 396.5, "ratio": 0.8362773530187186, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 472.125, "cannbot_us": 395.75, "ratio": 0.8382314005824728, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 474.75, "cannbot_us": 395.75, "ratio": 0.8335966298051606, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 474.125, "cannbot_us": 396.5, "ratio": 0.8362773530187186, "ratio_min": 0.8335966298051606, "ratio_max": 0.8382314005824728 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-down", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 486.375, "cannbot_us": 405.75, "ratio": 0.8342328450269854, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 485.0, "cannbot_us": 405.25, "ratio": 0.8355670103092784, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 483.875, "cannbot_us": 403.75, "ratio": 0.8344097132523896, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 480.5, "cannbot_us": 404.125, "ratio": 0.8410509885535901, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 483.125, "cannbot_us": 403.25, "ratio": 0.8346701164294955, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 483.875, "cannbot_us": 404.125, "ratio": 0.8346701164294955, "ratio_min": 0.8342328450269854, "ratio_max": 0.8410509885535901 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-gate-up", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 482.375, "cannbot_us": 454.75, "ratio": 0.9427312775330396, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 483.25, "cannbot_us": 455.5, "ratio": 0.9425763062596999, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 484.375, "cannbot_us": 456.0, "ratio": 0.9414193548387096, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 485.75, "cannbot_us": 457.0, "ratio": 0.9408131755018013, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 485.0, "cannbot_us": 455.5, "ratio": 0.9391752577319588, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 484.375, "cannbot_us": 455.5, "ratio": 0.9414193548387096, "ratio_min": 0.9391752577319588, "ratio_max": 0.9427312775330396 }, { "op": "GemmFwdOp", "label": "ds-v3-prefill-gate-up", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 491.5, "cannbot_us": 465.125, "ratio": 0.9463377416073245, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 493.5, "cannbot_us": 464.625, "ratio": 0.9414893617021277, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 491.5, "cannbot_us": 464.75, "ratio": 0.9455747711088505, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 495.5, "cannbot_us": 463.125, "ratio": 0.9346619576185671, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 493.75, "cannbot_us": 463.625, "ratio": 0.9389873417721519, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 493.5, "cannbot_us": 464.625, "ratio": 0.9414893617021277, "ratio_min": 0.9346619576185671, "ratio_max": 0.9463377416073245 }, { "op": "GemmFwdOp", "label": "k-dominant-7168x16384", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 4800.625, "cannbot_us": 4452.375, "ratio": 0.9274573623226142, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 4983.625, "cannbot_us": 4466.0, "ratio": 0.8961348416062606, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 4690.25, "cannbot_us": 4485.0, "ratio": 0.9562390064495496, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 4697.875, "cannbot_us": 4329.5, "ratio": 0.9215868876885827, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 4891.25, "cannbot_us": 4401.375, "ratio": 0.8998466649629441, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 4800.625, "cannbot_us": 4452.375, "ratio": 0.9215868876885827, "ratio_min": 0.8961348416062606, "ratio_max": 0.9562390064495496 }, { "op": "GemmFwdOp", "label": "k-dominant-7168x16384", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 4777.375, "cannbot_us": 4465.75, "ratio": 0.9347706638059604, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 4829.75, "cannbot_us": 4445.5, "ratio": 0.9204410166157668, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 4716.0, "cannbot_us": 4487.125, "ratio": 0.9514684054283291, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 4745.0, "cannbot_us": 4343.375, "ratio": 0.9153582718651212, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 4820.5, "cannbot_us": 4410.125, "ratio": 0.914868789544653, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 4777.375, "cannbot_us": 4445.5, "ratio": 0.9204410166157668, "ratio_min": 0.914868789544653, "ratio_max": 0.9514684054283291 }, { "op": "GemmFwdOp", "label": "mid-m16-attn", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 112.75, "cannbot_us": 109.625, "ratio": 0.9722838137472284, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 109.875, "cannbot_us": 107.0, "ratio": 0.9738339021615472, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 110.0, "cannbot_us": 106.875, "ratio": 0.9715909090909091, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 102.25, "cannbot_us": 101.75, "ratio": 0.9951100244498777, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 113.5, "cannbot_us": 110.0, "ratio": 0.9691629955947136, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 110.0, "cannbot_us": 107.0, "ratio": 0.9722838137472284, "ratio_min": 0.9691629955947136, "ratio_max": 0.9951100244498777 }, { "op": "GemmFwdOp", "label": "mid-m16-attn", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 117.5, "cannbot_us": 115.75, "ratio": 0.9851063829787234, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 116.0, "cannbot_us": 114.125, "ratio": 0.9838362068965517, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 113.75, "cannbot_us": 111.75, "ratio": 0.9824175824175824, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 106.5, "cannbot_us": 109.0, "ratio": 1.0234741784037558, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 117.375, "cannbot_us": 116.75, "ratio": 0.9946751863684771, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 116.0, "cannbot_us": 114.125, "ratio": 0.9851063829787234, "ratio_min": 0.9824175824175824, "ratio_max": 1.0234741784037558 }, { "op": "GemmFwdOp", "label": "mid-m32-attn", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 103.25, "cannbot_us": 102.25, "ratio": 0.9903147699757869, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 102.0, "cannbot_us": 100.5, "ratio": 0.9852941176470589, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 100.5, "cannbot_us": 99.5, "ratio": 0.9900497512437811, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 109.75, "cannbot_us": 106.5, "ratio": 0.9703872437357631, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 104.0, "cannbot_us": 101.625, "ratio": 0.9771634615384616, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 103.25, "cannbot_us": 101.625, "ratio": 0.9852941176470589, "ratio_min": 0.9703872437357631, "ratio_max": 0.9903147699757869 }, { "op": "GemmFwdOp", "label": "mid-m32-attn", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 109.125, "cannbot_us": 111.25, "ratio": 1.0194730813287514, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 108.125, "cannbot_us": 109.375, "ratio": 1.0115606936416186, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 105.5, "cannbot_us": 107.0, "ratio": 1.014218009478673, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 112.25, "cannbot_us": 114.0, "ratio": 1.0155902004454342, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 108.875, "cannbot_us": 110.25, "ratio": 1.0126291618828933, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 108.875, "cannbot_us": 110.25, "ratio": 1.014218009478673, "ratio_min": 1.0115606936416186, "ratio_max": 1.0194730813287514 }, { "op": "GemmFwdOp", "label": "mid-m64-down", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 53.25, "cannbot_us": 50.75, "ratio": 0.9530516431924883, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 53.0, "cannbot_us": 52.0, "ratio": 0.9811320754716981, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 54.25, "cannbot_us": 52.5, "ratio": 0.967741935483871, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 54.0, "cannbot_us": 52.5, "ratio": 0.9722222222222222, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 54.75, "cannbot_us": 52.25, "ratio": 0.954337899543379, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 54.0, "cannbot_us": 52.25, "ratio": 0.967741935483871, "ratio_min": 0.9530516431924883, "ratio_max": 0.9811320754716981 }, { "op": "GemmFwdOp", "label": "mid-m64-down", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 63.75, "cannbot_us": 60.0, "ratio": 0.9411764705882353, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 64.25, "cannbot_us": 61.375, "ratio": 0.9552529182879378, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 63.5, "cannbot_us": 60.375, "ratio": 0.9507874015748031, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 61.5, "cannbot_us": 62.25, "ratio": 1.0121951219512195, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 62.875, "cannbot_us": 60.75, "ratio": 0.9662027833001988, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 63.5, "cannbot_us": 60.75, "ratio": 0.9552529182879378, "ratio_min": 0.9411764705882353, "ratio_max": 1.0121951219512195 }, { "op": "GemmFwdOp", "label": "mid-m96-gate-up", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 71.625, "cannbot_us": 70.0, "ratio": 0.9773123909249564, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 67.0, "cannbot_us": 68.375, "ratio": 1.0205223880597014, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 66.375, "cannbot_us": 68.25, "ratio": 1.0282485875706215, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 68.25, "cannbot_us": 68.5, "ratio": 1.0036630036630036, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 70.625, "cannbot_us": 69.75, "ratio": 0.9876106194690265, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 68.25, "cannbot_us": 68.5, "ratio": 1.0036630036630036, "ratio_min": 0.9773123909249564, "ratio_max": 1.0282485875706215 }, { "op": "GemmFwdOp", "label": "mid-m96-gate-up", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 81.0, "cannbot_us": 79.25, "ratio": 0.9783950617283951, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 76.5, "cannbot_us": 77.0, "ratio": 1.0065359477124183, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 74.5, "cannbot_us": 76.0, "ratio": 1.0201342281879195, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 77.0, "cannbot_us": 75.25, "ratio": 0.9772727272727273, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 78.25, "cannbot_us": 78.125, "ratio": 0.9984025559105432, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 77.0, "cannbot_us": 77.0, "ratio": 0.9984025559105432, "ratio_min": 0.9772727272727273, "ratio_max": 1.0201342281879195 }, { "op": "GemmFwdOp", "label": "square-1k-nn", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 19.5, "cannbot_us": 27.25, "ratio": 1.3974358974358974, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 19.75, "cannbot_us": 27.25, "ratio": 1.379746835443038, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 19.5, "cannbot_us": 27.0, "ratio": 1.3846153846153846, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 20.0, "cannbot_us": 27.75, "ratio": 1.3875, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 19.5, "cannbot_us": 27.0, "ratio": 1.3846153846153846, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 19.5, "cannbot_us": 27.25, "ratio": 1.3846153846153846, "ratio_min": 1.379746835443038, "ratio_max": 1.3974358974358974 }, { "op": "GemmFwdOp", "label": "square-1k-nn", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 28.75, "cannbot_us": 37.0, "ratio": 1.2869565217391303, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 29.0, "cannbot_us": 37.25, "ratio": 1.2844827586206897, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 27.25, "cannbot_us": 35.5, "ratio": 1.3027522935779816, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 29.5, "cannbot_us": 37.0, "ratio": 1.2542372881355932, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 27.5, "cannbot_us": 34.875, "ratio": 1.268181818181818, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 28.75, "cannbot_us": 37.0, "ratio": 1.2844827586206897, "ratio_min": 1.2542372881355932, "ratio_max": 1.3027522935779816 }, { "op": "GemmFwdOp", "label": "square-1k-nn", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 19.5, "cannbot_us": 27.25, "ratio": 1.3974358974358974, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 19.75, "cannbot_us": 27.0, "ratio": 1.3670886075949367, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 19.5, "cannbot_us": 27.5, "ratio": 1.4102564102564104, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 19.75, "cannbot_us": 27.5, "ratio": 1.3924050632911393, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 20.0, "cannbot_us": 27.5, "ratio": 1.375, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 19.75, "cannbot_us": 27.5, "ratio": 1.3924050632911393, "ratio_min": 1.3670886075949367, "ratio_max": 1.4102564102564104 }, { "op": "GemmFwdOp", "label": "square-1k-nn", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 30.25, "cannbot_us": 37.0, "ratio": 1.2231404958677685, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 30.125, "cannbot_us": 36.75, "ratio": 1.2199170124481329, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 28.375, "cannbot_us": 35.5, "ratio": 1.251101321585903, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 28.0, "cannbot_us": 36.875, "ratio": 1.3169642857142858, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 29.5, "cannbot_us": 35.75, "ratio": 1.2118644067796611, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 29.5, "cannbot_us": 36.75, "ratio": 1.2231404958677685, "ratio_min": 1.2118644067796611, "ratio_max": 1.3169642857142858 }, { "op": "GemmFwdOp", "label": "wide-n-24576", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1244.625, "cannbot_us": 1014.125, "ratio": 0.8148036557195942, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 1238.5, "cannbot_us": 1014.5, "ratio": 0.8191360516754138, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 1256.875, "cannbot_us": 1015.75, "ratio": 0.8081551466931874, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 1260.5, "cannbot_us": 1015.125, "ratio": 0.8053351844506148, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 1242.75, "cannbot_us": 1014.125, "ratio": 0.816032991349829, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 1244.625, "cannbot_us": 1014.5, "ratio": 0.8148036557195942, "ratio_min": 0.8053351844506148, "ratio_max": 0.8191360516754138 }, { "op": "GemmFwdOp", "label": "wide-n-24576", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1253.5, "cannbot_us": 1023.875, "ratio": 0.8168129238133227, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 2, "ours_us": 1246.5, "cannbot_us": 1024.25, "ratio": 0.8217007621339751, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 3, "ours_us": 1263.5, "cannbot_us": 1023.25, "ratio": 0.8098535813217254, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 4, "ours_us": 1268.5, "cannbot_us": 1023.375, "ratio": 0.8067599527000394, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/GemmFwdOp/coverage/GemmFwdOp.json" }, { "rep": 5, "ours_us": 1249.125, "cannbot_us": 1021.5, "ratio": 0.817772440708496, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/GemmFwdOp/coverage/GemmFwdOp.json" } ], "errors": [], "ours_us": 1253.5, "cannbot_us": 1023.375, "ratio": 0.8168129238133227, "ratio_min": 0.8067599527000394, "ratio_max": 0.8217007621339751 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "ratio_min": 0.8571428571428571, "ratio_max": 0.8571428571428571 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "ratio_min": 1.25, "ratio_max": 1.25 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1.75, "cannbot_us": 1.375, "ratio": 0.7857142857142857, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 2.0, "cannbot_us": 1.5, "ratio": 0.75, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 2.0, "cannbot_us": 1.5, "ratio": 0.75, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.8571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 1.75, "cannbot_us": 1.5, "ratio": 0.7857142857142857, "ratio_min": 0.75, "ratio_max": 0.8571428571428571 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 2.0, "cannbot_us": 2.625, "ratio": 1.3125, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 2.0, "cannbot_us": 2.75, "ratio": 1.375, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 2.0, "cannbot_us": 2.5, "ratio": 1.25, "ratio_min": 1.25, "ratio_max": 1.375 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate-deep", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 2.25, "cannbot_us": 2.125, "ratio": 0.9444444444444444, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "ratio_min": 0.8888888888888888, "ratio_max": 0.9444444444444444 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate-deep", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.75, "cannbot_us": 2.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 2.625, "cannbot_us": 3.0, "ratio": 1.1428571428571428, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 2.75, "cannbot_us": 3.0, "ratio": 1.0909090909090908, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 2.75, "cannbot_us": 2.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 2.75, "cannbot_us": 2.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 2.75, "cannbot_us": 2.75, "ratio": 1.0, "ratio_min": 1.0, "ratio_max": 1.1428571428571428 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate-deep", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 2.25, "cannbot_us": 2.0, "ratio": 0.8888888888888888, "ratio_min": 0.8888888888888888, "ratio_max": 0.8888888888888888 }, { "op": "HardsigmoidFwdOp", "label": "mbv3-se-gate-deep", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.625, "cannbot_us": 2.75, "ratio": 1.0476190476190477, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 2, "ours_us": 2.625, "cannbot_us": 2.75, "ratio": 1.0476190476190477, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 3, "ours_us": 2.625, "cannbot_us": 2.75, "ratio": 1.0476190476190477, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 4, "ours_us": 2.5, "cannbot_us": 3.0, "ratio": 1.2, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" }, { "rep": 5, "ours_us": 2.5, "cannbot_us": 2.75, "ratio": 1.1, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardsigmoidFwdOp/coverage/HardsigmoidFwdOp.json" } ], "errors": [], "ours_us": 2.625, "cannbot_us": 2.75, "ratio": 1.0476190476190477, "ratio_min": 1.0476190476190477, "ratio_max": 1.2 }, { "op": "HardswishFwdOp", "label": "mbv3-stage2", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 37.5, "cannbot_us": 35.0, "ratio": 0.9333333333333333, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 37.0, "cannbot_us": 34.0, "ratio": 0.918918918918919, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 37.25, "cannbot_us": 34.5, "ratio": 0.9261744966442953, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 38.0, "cannbot_us": 35.75, "ratio": 0.9407894736842105, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 37.75, "cannbot_us": 35.0, "ratio": 0.9271523178807947, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 37.5, "cannbot_us": 35.0, "ratio": 0.9271523178807947, "ratio_min": 0.918918918918919, "ratio_max": 0.9407894736842105 }, { "op": "HardswishFwdOp", "label": "mbv3-stage2", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 45.25, "cannbot_us": 40.75, "ratio": 0.9005524861878453, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 47.0, "cannbot_us": 40.125, "ratio": 0.8537234042553191, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 46.5, "cannbot_us": 39.25, "ratio": 0.8440860215053764, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 45.75, "cannbot_us": 42.0, "ratio": 0.9180327868852459, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 45.25, "cannbot_us": 42.375, "ratio": 0.93646408839779, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 45.75, "cannbot_us": 40.75, "ratio": 0.9005524861878453, "ratio_min": 0.8440860215053764, "ratio_max": 0.93646408839779 }, { "op": "HardswishFwdOp", "label": "mbv3-stage2", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 37.25, "cannbot_us": 34.75, "ratio": 0.9328859060402684, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 37.125, "cannbot_us": 34.375, "ratio": 0.9259259259259259, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 37.0, "cannbot_us": 34.0, "ratio": 0.918918918918919, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 36.625, "cannbot_us": 34.0, "ratio": 0.9283276450511946, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 37.25, "cannbot_us": 34.75, "ratio": 0.9328859060402684, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 37.125, "cannbot_us": 34.375, "ratio": 0.9283276450511946, "ratio_min": 0.918918918918919, "ratio_max": 0.9328859060402684 }, { "op": "HardswishFwdOp", "label": "mbv3-stage2", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 44.0, "cannbot_us": 39.25, "ratio": 0.8920454545454546, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 47.75, "cannbot_us": 40.5, "ratio": 0.8481675392670157, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 46.0, "cannbot_us": 40.0, "ratio": 0.8695652173913043, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 44.5, "cannbot_us": 39.0, "ratio": 0.8764044943820225, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 44.5, "cannbot_us": 39.5, "ratio": 0.8876404494382022, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 44.5, "cannbot_us": 39.5, "ratio": 0.8764044943820225, "ratio_min": 0.8481675392670157, "ratio_max": 0.8920454545454546 }, { "op": "HardswishFwdOp", "label": "mbv3-stage3", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 24.0, "cannbot_us": 23.0, "ratio": 0.9583333333333334, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 24.25, "cannbot_us": 22.5, "ratio": 0.9278350515463918, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 23.625, "cannbot_us": 21.5, "ratio": 0.91005291005291, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 24.25, "cannbot_us": 22.25, "ratio": 0.9175257731958762, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 24.25, "cannbot_us": 22.5, "ratio": 0.9278350515463918, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 24.25, "cannbot_us": 22.5, "ratio": 0.9278350515463918, "ratio_min": 0.91005291005291, "ratio_max": 0.9583333333333334 }, { "op": "HardswishFwdOp", "label": "mbv3-stage3", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 32.75, "cannbot_us": 27.5, "ratio": 0.8396946564885496, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 33.5, "cannbot_us": 29.0, "ratio": 0.8656716417910447, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 30.75, "cannbot_us": 26.75, "ratio": 0.8699186991869918, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 32.5, "cannbot_us": 26.5, "ratio": 0.8153846153846154, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 33.25, "cannbot_us": 27.5, "ratio": 0.8270676691729323, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 32.75, "cannbot_us": 27.5, "ratio": 0.8396946564885496, "ratio_min": 0.8153846153846154, "ratio_max": 0.8699186991869918 }, { "op": "HardswishFwdOp", "label": "mbv3-stage3", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 24.0, "cannbot_us": 23.0, "ratio": 0.9583333333333334, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 23.25, "cannbot_us": 22.25, "ratio": 0.956989247311828, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 23.75, "cannbot_us": 22.5, "ratio": 0.9473684210526315, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 23.5, "cannbot_us": 23.0, "ratio": 0.9787234042553191, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 23.875, "cannbot_us": 22.875, "ratio": 0.9581151832460733, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 23.75, "cannbot_us": 22.875, "ratio": 0.9581151832460733, "ratio_min": 0.9473684210526315, "ratio_max": 0.9787234042553191 }, { "op": "HardswishFwdOp", "label": "mbv3-stage3", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 31.5, "cannbot_us": 27.5, "ratio": 0.873015873015873, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 2, "ours_us": 33.5, "cannbot_us": 30.25, "ratio": 0.9029850746268657, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 3, "ours_us": 32.25, "cannbot_us": 29.0, "ratio": 0.8992248062015504, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 4, "ours_us": 30.75, "cannbot_us": 27.5, "ratio": 0.8943089430894309, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/HardswishFwdOp/coverage/HardswishFwdOp.json" }, { "rep": 5, "ours_us": 30.75, "cannbot_us": 27.5, "ratio": 0.8943089430894309, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/HardswishFwdOp/coverage/HardswishFwdOp.json" } ], "errors": [], "ours_us": 31.5, "cannbot_us": 27.5, "ratio": 0.8943089430894309, "ratio_min": 0.873015873015873, "ratio_max": 0.9029850746268657 }, { "op": "LeakyReluFwdOp", "label": "gan-feat", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 63.125, "cannbot_us": 55.125, "ratio": 0.8732673267326733, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 62.75, "cannbot_us": 54.75, "ratio": 0.8725099601593626, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 62.75, "cannbot_us": 54.0, "ratio": 0.8605577689243028, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 63.5, "cannbot_us": 54.5, "ratio": 0.8582677165354331, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 62.75, "cannbot_us": 54.625, "ratio": 0.8705179282868526, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 62.75, "cannbot_us": 54.625, "ratio": 0.8705179282868526, "ratio_min": 0.8582677165354331, "ratio_max": 0.8732673267326733 }, { "op": "LeakyReluFwdOp", "label": "gan-feat", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 74.5, "cannbot_us": 60.5, "ratio": 0.8120805369127517, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 70.25, "cannbot_us": 64.25, "ratio": 0.9145907473309609, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 73.25, "cannbot_us": 59.5, "ratio": 0.8122866894197952, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 71.25, "cannbot_us": 62.625, "ratio": 0.8789473684210526, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 74.5, "cannbot_us": 60.5, "ratio": 0.8120805369127517, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 73.25, "cannbot_us": 60.5, "ratio": 0.8122866894197952, "ratio_min": 0.8120805369127517, "ratio_max": 0.9145907473309609 }, { "op": "LeakyReluFwdOp", "label": "gan-feat", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 63.0, "cannbot_us": 54.5, "ratio": 0.8650793650793651, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 62.5, "cannbot_us": 54.75, "ratio": 0.876, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 62.5, "cannbot_us": 54.875, "ratio": 0.878, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 62.0, "cannbot_us": 54.5, "ratio": 0.8790322580645161, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 62.0, "cannbot_us": 53.75, "ratio": 0.8669354838709677, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 62.5, "cannbot_us": 54.5, "ratio": 0.876, "ratio_min": 0.8650793650793651, "ratio_max": 0.8790322580645161 }, { "op": "LeakyReluFwdOp", "label": "gan-feat", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 74.0, "cannbot_us": 61.25, "ratio": 0.8277027027027027, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 70.0, "cannbot_us": 59.5, "ratio": 0.85, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 73.0, "cannbot_us": 62.25, "ratio": 0.8527397260273972, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 71.5, "cannbot_us": 60.75, "ratio": 0.8496503496503497, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 73.5, "cannbot_us": 59.875, "ratio": 0.814625850340136, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 73.0, "cannbot_us": 60.75, "ratio": 0.8496503496503497, "ratio_min": 0.814625850340136, "ratio_max": 0.8527397260273972 }, { "op": "LeakyReluFwdOp", "label": "gan-feat-deep", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 32.0, "cannbot_us": 27.875, "ratio": 0.87109375, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 33.0, "cannbot_us": 29.0, "ratio": 0.8787878787878788, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 32.25, "cannbot_us": 27.25, "ratio": 0.8449612403100775, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 33.0, "cannbot_us": 28.75, "ratio": 0.8712121212121212, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 33.0, "cannbot_us": 28.25, "ratio": 0.8560606060606061, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 33.0, "cannbot_us": 28.25, "ratio": 0.87109375, "ratio_min": 0.8449612403100775, "ratio_max": 0.8787878787878788 }, { "op": "LeakyReluFwdOp", "label": "gan-feat-deep", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 42.0, "cannbot_us": 34.75, "ratio": 0.8273809523809523, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 42.5, "cannbot_us": 34.5, "ratio": 0.8117647058823529, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 42.0, "cannbot_us": 35.0, "ratio": 0.8333333333333334, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 42.25, "cannbot_us": 34.625, "ratio": 0.8195266272189349, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 43.0, "cannbot_us": 36.25, "ratio": 0.8430232558139535, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 42.25, "cannbot_us": 34.75, "ratio": 0.8273809523809523, "ratio_min": 0.8117647058823529, "ratio_max": 0.8430232558139535 }, { "op": "LeakyReluFwdOp", "label": "gan-feat-deep", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 31.75, "cannbot_us": 27.5, "ratio": 0.8661417322834646, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 32.0, "cannbot_us": 28.25, "ratio": 0.8828125, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 31.25, "cannbot_us": 26.5, "ratio": 0.848, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 32.5, "cannbot_us": 28.75, "ratio": 0.8846153846153846, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 32.25, "cannbot_us": 28.5, "ratio": 0.8837209302325582, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 32.0, "cannbot_us": 28.25, "ratio": 0.8828125, "ratio_min": 0.848, "ratio_max": 0.8846153846153846 }, { "op": "LeakyReluFwdOp", "label": "gan-feat-deep", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 42.75, "cannbot_us": 35.0, "ratio": 0.8187134502923976, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 2, "ours_us": 39.0, "cannbot_us": 33.625, "ratio": 0.8621794871794872, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 3, "ours_us": 42.125, "cannbot_us": 34.0, "ratio": 0.8071216617210683, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 4, "ours_us": 40.25, "cannbot_us": 34.0, "ratio": 0.84472049689441, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" }, { "rep": 5, "ours_us": 42.5, "cannbot_us": 36.75, "ratio": 0.8647058823529412, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/LeakyReluFwdOp/coverage/LeakyReluFwdOp.json" } ], "errors": [], "ours_us": 42.125, "cannbot_us": 34.0, "ratio": 0.84472049689441, "ratio_min": 0.8071216617210683, "ratio_max": 0.8647058823529412 }, { "op": "MaskedFillFwdOp", "label": "elementwise-16M", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 76.5, "cannbot_us": 76.5, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 73.75, "cannbot_us": 73.5, "ratio": 0.9966101694915255, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 75.5, "cannbot_us": 76.25, "ratio": 1.009933774834437, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 74.875, "cannbot_us": 74.75, "ratio": 0.998330550918197, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 76.0, "cannbot_us": 75.25, "ratio": 0.9901315789473685, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 75.5, "cannbot_us": 75.25, "ratio": 0.998330550918197, "ratio_min": 0.9901315789473685, "ratio_max": 1.009933774834437 }, { "op": "MaskedFillFwdOp", "label": "elementwise-16M", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 82.5, "cannbot_us": 81.875, "ratio": 0.9924242424242424, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 78.75, "cannbot_us": 78.0, "ratio": 0.9904761904761905, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 81.75, "cannbot_us": 82.0, "ratio": 1.003058103975535, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 80.5, "cannbot_us": 80.75, "ratio": 1.0031055900621118, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 79.5, "cannbot_us": 79.5, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 80.5, "cannbot_us": 80.75, "ratio": 1.0, "ratio_min": 0.9904761904761905, "ratio_max": 1.0031055900621118 }, { "op": "MaskedFillFwdOp", "label": "elementwise-16M", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 76.625, "cannbot_us": 77.25, "ratio": 1.0081566068515497, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 74.125, "cannbot_us": 73.75, "ratio": 0.9949409780775716, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 76.25, "cannbot_us": 75.75, "ratio": 0.9934426229508196, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 75.5, "cannbot_us": 74.75, "ratio": 0.9900662251655629, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 76.75, "cannbot_us": 77.75, "ratio": 1.013029315960912, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 76.25, "cannbot_us": 75.75, "ratio": 0.9949409780775716, "ratio_min": 0.9900662251655629, "ratio_max": 1.013029315960912 }, { "op": "MaskedFillFwdOp", "label": "elementwise-16M", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 80.5, "cannbot_us": 82.5, "ratio": 1.0248447204968945, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 79.25, "cannbot_us": 78.5, "ratio": 0.9905362776025236, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 80.0, "cannbot_us": 80.5, "ratio": 1.00625, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 80.75, "cannbot_us": 80.5, "ratio": 0.9969040247678018, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 81.125, "cannbot_us": 83.25, "ratio": 1.0261941448382126, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 80.5, "cannbot_us": 80.5, "ratio": 1.00625, "ratio_min": 0.9905362776025236, "ratio_max": 1.0261941448382126 }, { "op": "MaskedFillFwdOp", "label": "elementwise-16M", "dtype": "float32", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 130.25, "cannbot_us": 130.75, "ratio": 1.0038387715930903, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 127.0, "cannbot_us": 126.75, "ratio": 0.9980314960629921, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 129.625, "cannbot_us": 129.5, "ratio": 0.9990356798457087, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 126.25, "cannbot_us": 126.75, "ratio": 1.003960396039604, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 131.25, "cannbot_us": 132.375, "ratio": 1.0085714285714287, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 129.625, "cannbot_us": 129.5, "ratio": 1.0038387715930903, "ratio_min": 0.9980314960629921, "ratio_max": 1.0085714285714287 }, { "op": "MaskedFillFwdOp", "label": "elementwise-16M", "dtype": "float32", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 136.0, "cannbot_us": 137.125, "ratio": 1.0082720588235294, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 132.0, "cannbot_us": 136.0, "ratio": 1.0303030303030303, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 138.0, "cannbot_us": 134.5, "ratio": 0.9746376811594203, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 132.75, "cannbot_us": 134.75, "ratio": 1.015065913370998, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 137.875, "cannbot_us": 138.0, "ratio": 1.0009066183136899, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 136.0, "cannbot_us": 136.0, "ratio": 1.0082720588235294, "ratio_min": 0.9746376811594203, "ratio_max": 1.0303030303030303 }, { "op": "MaskedFillFwdOp", "label": "elementwise-256M", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1115.25, "cannbot_us": 1170.125, "ratio": 1.0492042143017262, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 1115.25, "cannbot_us": 1167.25, "ratio": 1.0466263169692893, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 1115.875, "cannbot_us": 1166.125, "ratio": 1.045031925618909, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 1114.5, "cannbot_us": 1165.125, "ratio": 1.0454239569313593, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 1117.0, "cannbot_us": 1166.625, "ratio": 1.044427036705461, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 1115.25, "cannbot_us": 1166.625, "ratio": 1.0454239569313593, "ratio_min": 1.044427036705461, "ratio_max": 1.0492042143017262 }, { "op": "MaskedFillFwdOp", "label": "elementwise-256M", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1211.0, "cannbot_us": 1179.75, "ratio": 0.9741948802642444, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 1209.25, "cannbot_us": 1177.5, "ratio": 0.9737440562332024, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 1211.875, "cannbot_us": 1184.125, "ratio": 0.9771015987622486, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 1206.125, "cannbot_us": 1179.5, "ratio": 0.9779251735931185, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 1203.75, "cannbot_us": 1179.375, "ratio": 0.9797507788161994, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 1209.25, "cannbot_us": 1179.5, "ratio": 0.9771015987622486, "ratio_min": 0.9737440562332024, "ratio_max": 0.9797507788161994 }, { "op": "MaskedFillFwdOp", "label": "elementwise-256M", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1176.375, "cannbot_us": 1133.5, "ratio": 0.9635532887047072, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 1176.125, "cannbot_us": 1135.25, "ratio": 0.9652460410245509, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 1174.75, "cannbot_us": 1135.625, "ratio": 0.9666950414981911, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 1171.625, "cannbot_us": 1133.875, "ratio": 0.9677797930225115, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 1175.125, "cannbot_us": 1136.0, "ratio": 0.966705669609616, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 1175.125, "cannbot_us": 1135.25, "ratio": 0.9666950414981911, "ratio_min": 0.9635532887047072, "ratio_max": 0.9677797930225115 }, { "op": "MaskedFillFwdOp", "label": "elementwise-256M", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 1183.5, "cannbot_us": 1142.375, "ratio": 0.9652513730460498, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 2, "ours_us": 1183.625, "cannbot_us": 1139.625, "ratio": 0.9628260639983103, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 3, "ours_us": 1182.0, "cannbot_us": 1140.75, "ratio": 0.9651015228426396, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 4, "ours_us": 1181.0, "cannbot_us": 1140.625, "ratio": 0.9658128704487722, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" }, { "rep": 5, "ours_us": 1181.5, "cannbot_us": 1141.375, "ratio": 0.9660389335590351, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MaskedFillFwdOp/coverage/MaskedFillFwdOp.json" } ], "errors": [], "ours_us": 1182.0, "cannbot_us": 1140.75, "ratio": 0.9652513730460498, "ratio_min": 0.9628260639983103, "ratio_max": 0.9660389335590351 }, { "op": "MishFwdOp", "label": "yolo-p3", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 190.0, "cannbot_us": 88.25, "ratio": 0.4644736842105263, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 189.5, "cannbot_us": 87.0, "ratio": 0.45910290237467016, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 189.75, "cannbot_us": 87.25, "ratio": 0.4598155467720685, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 189.5, "cannbot_us": 85.125, "ratio": 0.4492084432717678, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 190.0, "cannbot_us": 89.875, "ratio": 0.4730263157894737, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 189.75, "cannbot_us": 87.25, "ratio": 0.4598155467720685, "ratio_min": 0.4492084432717678, "ratio_max": 0.4730263157894737 }, { "op": "MishFwdOp", "label": "yolo-p3", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 197.0, "cannbot_us": 93.375, "ratio": 0.4739847715736041, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 197.75, "cannbot_us": 94.125, "ratio": 0.47597977243994943, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 199.75, "cannbot_us": 94.25, "ratio": 0.4718397997496871, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 199.5, "cannbot_us": 93.0, "ratio": 0.46616541353383456, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 197.0, "cannbot_us": 95.75, "ratio": 0.4860406091370558, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 197.75, "cannbot_us": 94.125, "ratio": 0.4739847715736041, "ratio_min": 0.46616541353383456, "ratio_max": 0.4860406091370558 }, { "op": "MishFwdOp", "label": "yolo-p3", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 187.0, "cannbot_us": 89.75, "ratio": 0.4799465240641711, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 187.0, "cannbot_us": 88.0, "ratio": 0.47058823529411764, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 186.875, "cannbot_us": 88.25, "ratio": 0.47224080267558527, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 187.0, "cannbot_us": 86.0, "ratio": 0.45989304812834225, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 187.25, "cannbot_us": 90.125, "ratio": 0.48130841121495327, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 187.0, "cannbot_us": 88.25, "ratio": 0.47224080267558527, "ratio_min": 0.45989304812834225, "ratio_max": 0.48130841121495327 }, { "op": "MishFwdOp", "label": "yolo-p3", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 194.5, "cannbot_us": 97.125, "ratio": 0.4993573264781491, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 195.625, "cannbot_us": 94.0, "ratio": 0.4805111821086262, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 197.25, "cannbot_us": 94.375, "ratio": 0.4784537389100127, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 196.5, "cannbot_us": 93.375, "ratio": 0.4751908396946565, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 194.5, "cannbot_us": 97.5, "ratio": 0.5012853470437018, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 195.625, "cannbot_us": 94.375, "ratio": 0.4805111821086262, "ratio_min": 0.4751908396946565, "ratio_max": 0.5012853470437018 }, { "op": "MishFwdOp", "label": "yolo-p4", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 97.0, "cannbot_us": 46.25, "ratio": 0.47680412371134023, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 96.75, "cannbot_us": 45.125, "ratio": 0.4664082687338501, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 96.75, "cannbot_us": 45.75, "ratio": 0.4728682170542636, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 96.75, "cannbot_us": 46.5, "ratio": 0.4806201550387597, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 97.0, "cannbot_us": 49.0, "ratio": 0.5051546391752577, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 96.75, "cannbot_us": 46.25, "ratio": 0.47680412371134023, "ratio_min": 0.4664082687338501, "ratio_max": 0.5051546391752577 }, { "op": "MishFwdOp", "label": "yolo-p4", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 105.25, "cannbot_us": 52.0, "ratio": 0.49406175771971494, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 104.25, "cannbot_us": 52.75, "ratio": 0.5059952038369304, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 106.0, "cannbot_us": 53.75, "ratio": 0.5070754716981132, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 106.0, "cannbot_us": 54.25, "ratio": 0.5117924528301887, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 106.0, "cannbot_us": 55.125, "ratio": 0.5200471698113207, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 106.0, "cannbot_us": 53.75, "ratio": 0.5070754716981132, "ratio_min": 0.49406175771971494, "ratio_max": 0.5200471698113207 }, { "op": "MishFwdOp", "label": "yolo-p4", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 95.75, "cannbot_us": 47.0, "ratio": 0.4908616187989556, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 95.5, "cannbot_us": 47.5, "ratio": 0.4973821989528796, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 95.5, "cannbot_us": 45.75, "ratio": 0.4790575916230366, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 95.5, "cannbot_us": 45.5, "ratio": 0.47643979057591623, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 95.75, "cannbot_us": 49.25, "ratio": 0.5143603133159269, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 95.5, "cannbot_us": 47.0, "ratio": 0.4908616187989556, "ratio_min": 0.47643979057591623, "ratio_max": 0.5143603133159269 }, { "op": "MishFwdOp", "label": "yolo-p4", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 103.25, "cannbot_us": 54.375, "ratio": 0.5266343825665859, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 2, "ours_us": 104.375, "cannbot_us": 53.625, "ratio": 0.5137724550898204, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 3, "ours_us": 106.25, "cannbot_us": 52.0, "ratio": 0.4894117647058824, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 4, "ours_us": 106.25, "cannbot_us": 51.625, "ratio": 0.4858823529411765, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/MishFwdOp/coverage/MishFwdOp.json" }, { "rep": 5, "ours_us": 103.0, "cannbot_us": 57.0, "ratio": 0.5533980582524272, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/MishFwdOp/coverage/MishFwdOp.json" } ], "errors": [], "ours_us": 104.375, "cannbot_us": 53.625, "ratio": 0.5137724550898204, "ratio_min": 0.4858823529411765, "ratio_max": 0.5533980582524272 }, { "op": "ReluFwdOp", "label": "hidden-state-decode", "dtype": "bfloat16", "regime": "eager", "n": 4, "required_n": 5, "publishable": false, "trials": [ { "rep": 1, "ours_us": 1.75, "cannbot_us": 1.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 2, "ours_us": 1.75, "cannbot_us": 1.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 3, "ours_us": 1.75, "cannbot_us": 1.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 4, "ours_us": 1.75, "cannbot_us": 1.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/ReluFwdOp/coverage/ReluFwdOp.json" } ], "errors": [ { "rep": 5, "reason": "device timing failed: RuntimeError: expected 1 kernel_details.csv under /home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/profiler/ReluFwdOp_0006_case, got []", "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/coverage/ReluFwdOp.json" } ] }, { "op": "ReluFwdOp", "label": "hidden-state-decode", "dtype": "bfloat16", "regime": "graph", "n": 4, "required_n": 5, "publishable": false, "trials": [ { "rep": 1, "ours_us": 1.75, "cannbot_us": 1.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 2, "ours_us": 1.75, "cannbot_us": 1.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 3, "ours_us": 2.0, "cannbot_us": 2.0, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 4, "ours_us": 1.75, "cannbot_us": 1.75, "ratio": 1.0, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/ReluFwdOp/coverage/ReluFwdOp.json" } ], "errors": [ { "rep": 5, "reason": "device timing failed: RuntimeError: expected 1 kernel_details.csv under /home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/profiler/ReluFwdOp_0006_case, got []", "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/coverage/ReluFwdOp.json" } ] }, { "op": "ReluFwdOp", "label": "hidden-state-prefill", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 30.5, "cannbot_us": 28.5, "ratio": 0.9344262295081968, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 2, "ours_us": 32.0, "cannbot_us": 30.25, "ratio": 0.9453125, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 3, "ours_us": 30.0, "cannbot_us": 27.875, "ratio": 0.9291666666666667, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 4, "ours_us": 30.25, "cannbot_us": 27.375, "ratio": 0.9049586776859504, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 5, "ours_us": 29.875, "cannbot_us": 27.25, "ratio": 0.9121338912133892, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/coverage/ReluFwdOp.json" } ], "errors": [], "ours_us": 30.25, "cannbot_us": 27.875, "ratio": 0.9291666666666667, "ratio_min": 0.9049586776859504, "ratio_max": 0.9453125 }, { "op": "ReluFwdOp", "label": "hidden-state-prefill", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 37.75, "cannbot_us": 34.5, "ratio": 0.9139072847682119, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 2, "ours_us": 41.0, "cannbot_us": 36.75, "ratio": 0.8963414634146342, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 3, "ours_us": 39.5, "cannbot_us": 37.25, "ratio": 0.9430379746835443, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 4, "ours_us": 39.25, "cannbot_us": 34.0, "ratio": 0.8662420382165605, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 5, "ours_us": 39.25, "cannbot_us": 34.25, "ratio": 0.8726114649681529, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/coverage/ReluFwdOp.json" } ], "errors": [], "ours_us": 39.25, "cannbot_us": 34.5, "ratio": 0.8963414634146342, "ratio_min": 0.8662420382165605, "ratio_max": 0.9430379746835443 }, { "op": "ReluFwdOp", "label": "hidden-state-prefill", "dtype": "float16", "regime": "eager", "n": 4, "required_n": 5, "publishable": false, "trials": [ { "rep": 1, "ours_us": 30.375, "cannbot_us": 29.5, "ratio": 0.9711934156378601, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 2, "ours_us": 31.125, "cannbot_us": 29.25, "ratio": 0.9397590361445783, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 3, "ours_us": 30.0, "cannbot_us": 28.0, "ratio": 0.9333333333333333, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 5, "ours_us": 29.375, "cannbot_us": 27.5, "ratio": 0.9361702127659575, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/coverage/ReluFwdOp.json" } ], "errors": [ { "rep": 4, "reason": "D036 selection could not be measured (RuntimeError: expected 1 kernel_details.csv under /home/dyq/workspace/docs/reports/R322-data/runs/pass-4/ReluFwdOp/profiler/ReluFwdOp_0001_d036_ReluFwdOp_hidden-state-prefill_float16, got []); fell back to the pre-D036 first-match order, which this record says so that the number is not read as best-of-pool", "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/ReluFwdOp/coverage/ReluFwdOp.json" } ] }, { "op": "ReluFwdOp", "label": "hidden-state-prefill", "dtype": "float16", "regime": "graph", "n": 4, "required_n": 5, "publishable": false, "trials": [ { "rep": 1, "ours_us": 38.25, "cannbot_us": 35.0, "ratio": 0.9150326797385621, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 2, "ours_us": 40.5, "cannbot_us": 35.75, "ratio": 0.8827160493827161, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 3, "ours_us": 39.25, "cannbot_us": 34.5, "ratio": 0.8789808917197452, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/ReluFwdOp/coverage/ReluFwdOp.json" }, { "rep": 5, "ours_us": 39.25, "cannbot_us": 34.25, "ratio": 0.8726114649681529, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/ReluFwdOp/coverage/ReluFwdOp.json" } ], "errors": [ { "rep": 4, "reason": "graph fallback is not a cannbot comparison", "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/ReluFwdOp/coverage/ReluFwdOp.json" } ] }, { "op": "SiluFwdOp", "label": "llama-8b-ffn-decode", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.5, "cannbot_us": 1.75, "ratio": 0.7, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 2, "ours_us": 2.5, "cannbot_us": 1.75, "ratio": 0.7, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 3, "ours_us": 2.5, "cannbot_us": 1.75, "ratio": 0.7, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 4, "ours_us": 2.75, "cannbot_us": 1.75, "ratio": 0.6363636363636364, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 5, "ours_us": 2.5, "cannbot_us": 1.75, "ratio": 0.7, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SiluFwdOp/coverage/SiluFwdOp.json" } ], "errors": [], "ours_us": 2.5, "cannbot_us": 1.75, "ratio": 0.7, "ratio_min": 0.6363636363636364, "ratio_max": 0.7 }, { "op": "SiluFwdOp", "label": "llama-8b-ffn-decode", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 2.75, "cannbot_us": 3.5, "ratio": 1.2727272727272727, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 2, "ours_us": 2.75, "cannbot_us": 3.25, "ratio": 1.1818181818181819, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 3, "ours_us": 2.75, "cannbot_us": 3.25, "ratio": 1.1818181818181819, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 4, "ours_us": 2.75, "cannbot_us": 3.75, "ratio": 1.3636363636363635, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 5, "ours_us": 2.75, "cannbot_us": 3.25, "ratio": 1.1818181818181819, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SiluFwdOp/coverage/SiluFwdOp.json" } ], "errors": [], "ours_us": 2.75, "cannbot_us": 3.25, "ratio": 1.1818181818181819, "ratio_min": 1.1818181818181819, "ratio_max": 1.3636363636363635 }, { "op": "SiluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 136.25, "cannbot_us": 104.25, "ratio": 0.7651376146788991, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 2, "ours_us": 134.5, "cannbot_us": 106.75, "ratio": 0.7936802973977695, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 3, "ours_us": 132.75, "cannbot_us": 99.875, "ratio": 0.7523540489642184, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 4, "ours_us": 135.75, "cannbot_us": 103.375, "ratio": 0.7615101289134438, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 5, "ours_us": 133.75, "cannbot_us": 103.0, "ratio": 0.7700934579439253, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SiluFwdOp/coverage/SiluFwdOp.json" } ], "errors": [], "ours_us": 134.5, "cannbot_us": 103.375, "ratio": 0.7651376146788991, "ratio_min": 0.7523540489642184, "ratio_max": 0.7936802973977695 }, { "op": "SiluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 145.0, "cannbot_us": 109.125, "ratio": 0.7525862068965518, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 2, "ours_us": 142.75, "cannbot_us": 111.5, "ratio": 0.7810858143607706, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 3, "ours_us": 142.5, "cannbot_us": 105.0, "ratio": 0.7368421052631579, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 4, "ours_us": 149.0, "cannbot_us": 108.625, "ratio": 0.7290268456375839, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 5, "ours_us": 143.5, "cannbot_us": 109.125, "ratio": 0.7604529616724739, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SiluFwdOp/coverage/SiluFwdOp.json" } ], "errors": [], "ours_us": 143.5, "cannbot_us": 109.125, "ratio": 0.7525862068965518, "ratio_min": 0.7290268456375839, "ratio_max": 0.7810858143607706 }, { "op": "SiluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 134.25, "cannbot_us": 106.25, "ratio": 0.7914338919925512, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 2, "ours_us": 134.625, "cannbot_us": 104.5, "ratio": 0.776230269266481, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 3, "ours_us": 131.25, "cannbot_us": 101.375, "ratio": 0.7723809523809524, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 4, "ours_us": 136.0, "cannbot_us": 106.25, "ratio": 0.78125, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 5, "ours_us": 133.0, "cannbot_us": 100.125, "ratio": 0.7528195488721805, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SiluFwdOp/coverage/SiluFwdOp.json" } ], "errors": [], "ours_us": 134.25, "cannbot_us": 104.5, "ratio": 0.776230269266481, "ratio_min": 0.7528195488721805, "ratio_max": 0.7914338919925512 }, { "op": "SiluFwdOp", "label": "llama-8b-ffn-prefill", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 143.625, "cannbot_us": 111.25, "ratio": 0.7745865970409052, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 2, "ours_us": 142.25, "cannbot_us": 109.75, "ratio": 0.7715289982425307, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 3, "ours_us": 138.75, "cannbot_us": 107.625, "ratio": 0.7756756756756756, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 4, "ours_us": 147.0, "cannbot_us": 111.0, "ratio": 0.7551020408163265, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SiluFwdOp/coverage/SiluFwdOp.json" }, { "rep": 5, "ours_us": 140.0, "cannbot_us": 106.0, "ratio": 0.7571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SiluFwdOp/coverage/SiluFwdOp.json" } ], "errors": [], "ours_us": 142.25, "cannbot_us": 109.75, "ratio": 0.7715289982425307, "ratio_min": 0.7551020408163265, "ratio_max": 0.7756756756756756 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 50.5, "cannbot_us": 30.875, "ratio": 0.6113861386138614, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 50.5, "cannbot_us": 29.625, "ratio": 0.5866336633663366, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 51.0, "cannbot_us": 31.375, "ratio": 0.6151960784313726, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 50.25, "cannbot_us": 29.0, "ratio": 0.5771144278606966, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 50.5, "cannbot_us": 30.25, "ratio": 0.599009900990099, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 50.5, "cannbot_us": 30.25, "ratio": 0.599009900990099, "ratio_min": 0.5771144278606966, "ratio_max": 0.6151960784313726 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 61.75, "cannbot_us": 39.5, "ratio": 0.6396761133603239, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 61.25, "cannbot_us": 38.5, "ratio": 0.6285714285714286, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 58.25, "cannbot_us": 37.75, "ratio": 0.648068669527897, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 58.25, "cannbot_us": 35.875, "ratio": 0.6158798283261803, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 59.25, "cannbot_us": 37.75, "ratio": 0.6371308016877637, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 59.25, "cannbot_us": 37.75, "ratio": 0.6371308016877637, "ratio_min": 0.6158798283261803, "ratio_max": 0.648068669527897 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 50.25, "cannbot_us": 32.125, "ratio": 0.6393034825870647, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 49.5, "cannbot_us": 29.75, "ratio": 0.601010101010101, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 50.0, "cannbot_us": 31.25, "ratio": 0.625, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 50.0, "cannbot_us": 31.25, "ratio": 0.625, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 50.0, "cannbot_us": 30.5, "ratio": 0.61, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 50.0, "cannbot_us": 31.25, "ratio": 0.625, "ratio_min": 0.601010101010101, "ratio_max": 0.6393034825870647 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 61.25, "cannbot_us": 40.25, "ratio": 0.6571428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 60.375, "cannbot_us": 36.75, "ratio": 0.6086956521739131, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 57.5, "cannbot_us": 39.5, "ratio": 0.6869565217391305, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 57.75, "cannbot_us": 38.0, "ratio": 0.658008658008658, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 58.625, "cannbot_us": 35.75, "ratio": 0.6098081023454158, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 58.625, "cannbot_us": 38.0, "ratio": 0.6571428571428571, "ratio_min": 0.6086956521739131, "ratio_max": 0.6869565217391305 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden-wide", "dtype": "bfloat16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 98.0, "cannbot_us": 59.5, "ratio": 0.6071428571428571, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 97.75, "cannbot_us": 58.375, "ratio": 0.5971867007672634, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 98.5, "cannbot_us": 60.75, "ratio": 0.616751269035533, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 97.25, "cannbot_us": 58.5, "ratio": 0.6015424164524421, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 98.0, "cannbot_us": 58.25, "ratio": 0.5943877551020408, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 98.0, "cannbot_us": 58.5, "ratio": 0.6015424164524421, "ratio_min": 0.5943877551020408, "ratio_max": 0.616751269035533 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden-wide", "dtype": "bfloat16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 109.75, "cannbot_us": 67.875, "ratio": 0.6184510250569476, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 108.75, "cannbot_us": 65.75, "ratio": 0.6045977011494252, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 106.125, "cannbot_us": 66.0, "ratio": 0.6219081272084805, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 105.75, "cannbot_us": 65.0, "ratio": 0.6146572104018913, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 106.75, "cannbot_us": 65.125, "ratio": 0.6100702576112412, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 106.75, "cannbot_us": 65.75, "ratio": 0.6146572104018913, "ratio_min": 0.6045977011494252, "ratio_max": 0.6219081272084805 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden-wide", "dtype": "float16", "regime": "eager", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 95.875, "cannbot_us": 59.5, "ratio": 0.6205997392438071, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 95.75, "cannbot_us": 57.0, "ratio": 0.5953002610966057, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 96.25, "cannbot_us": 60.0, "ratio": 0.6233766233766234, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 95.875, "cannbot_us": 56.75, "ratio": 0.5919165580182529, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 95.875, "cannbot_us": 57.25, "ratio": 0.5971316818774446, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 95.875, "cannbot_us": 57.25, "ratio": 0.5971316818774446, "ratio_min": 0.5919165580182529, "ratio_max": 0.6233766233766234 }, { "op": "SoftplusFwdOp", "label": "mlp-hidden-wide", "dtype": "float16", "regime": "graph", "n": 5, "required_n": 5, "publishable": true, "trials": [ { "rep": 1, "ours_us": 106.25, "cannbot_us": 67.75, "ratio": 0.6376470588235295, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-1/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 2, "ours_us": 105.5, "cannbot_us": 64.5, "ratio": 0.6113744075829384, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-2/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 3, "ours_us": 105.625, "cannbot_us": 66.0, "ratio": 0.6248520710059171, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-3/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 4, "ours_us": 105.5, "cannbot_us": 63.5, "ratio": 0.6018957345971564, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-4/SoftplusFwdOp/coverage/SoftplusFwdOp.json" }, { "rep": 5, "ours_us": 104.0, "cannbot_us": 63.25, "ratio": 0.6081730769230769, "source": "/home/dyq/workspace/docs/reports/R322-data/runs/pass-5/SoftplusFwdOp/coverage/SoftplusFwdOp.json" } ], "errors": [], "ours_us": 105.5, "cannbot_us": 64.5, "ratio": 0.6113744075829384, "ratio_min": 0.6018957345971564, "ratio_max": 0.6376470588235295 } ]