CANNBot 算子测试报告模拟数据【免费下载链接】cann-outreach项目地址: https://gitcode.com/cann/cann-outreach生成时间: 2026-07-01 17:01:06 | 工具: op_tester | 数据来源: 模拟未实跑1. 总览指标数值算子数8测试用例总数49通过47失败2通过率95.9%2. 算子级汇总算子种类通路dtype用例数通过失败状态abselementwisePyTorch(.so)fp3212120✅ PASSaddelementwisePyTorch(triton)fp16761⚠ PARTIALsoftmaxreductionPyTorch(.so)fp32660✅ PASSsparse_gemmmatmul(2:4 sparse)Bin(exec)fp32550✅ PASStrilelementwisePyTorch(.so)fp32660✅ PASSmamba_selective_scanscanPyTorch(.so)fp32541⚠ PARTIALsumreductionPyTorch(.so)fp32440✅ PASSmatmulmatmulPyTorch(.so)fp16440✅ PASS3. 详细测试用例abs (elementwise, fp32)用例shape结果max_diffmean_diff耗时(ms)备注C1[256]✅ PASS0.000e000.000e000.18C2[512]✅ PASS0.000e000.000e000.16C3[1024]✅ PASS0.000e000.000e000.17C4[4096]✅ PASS0.000e000.000e000.19C5[32768]✅ PASS0.000e000.000e000.24C6[1000]✅ PASS0.000e000.000e000.17非对齐C7[13]✅ PASS0.000e000.000e000.15极小 shapeC8[512,512]✅ PASS0.000e000.000e000.22C9[1024,1024]✅ PASS0.000e000.000e000.28C10[128,128,128]✅ PASS0.000e000.000e000.31C11[64,32,16,8]✅ PASS0.000e000.000e000.204DC12[65536]✅ PASS0.000e000.000e000.35add (elementwise, fp16)用例shape结果max_diffmean_diff耗时(ms)备注C1[1024,1024]✅ PASS0.000e000.000e000.42C2[8192]✅ PASS0.000e000.000e000.19C3[256,256,256]✅ PASS0.000e000.000e000.383DC4[4096,4096]✅ PASS0.000e000.000e001.12C5[128,128]✅ PASS0.000e000.000e000.14C6[65536]✅ PASS0.000e000.000e000.21C7[10000]❌ FAIL1.953e-034.882e-040.18fp16 非对齐尾部精度softmax (reduction, fp32)用例shape结果max_diffmean_diff耗时(ms)备注C1[128,128] dim-1✅ PASS1.192e-072.980e-080.33C2[16,256,32] dim1✅ PASS1.490e-073.725e-080.41中间维归约C3[1024,1024] dim-1✅ PASS2.384e-075.960e-080.88C4[256] dim0✅ PASS5.960e-081.490e-080.121DC5[64,64,64] dim1✅ PASS1.788e-074.470e-080.46C6[2048,512] dim-1✅ PASS2.682e-076.705e-081.34sparse_gemm (matmul(2:4 sparse), fp32)用例shape结果max_diffmean_diff耗时(ms)备注C1[128,128]x[128,128]✅ PASS0.000e000.000e000.74C2[2,128,128] batched✅ PASS0.000e000.000e001.18batchedC3[256,256]x[256,256]✅ PASS0.000e000.000e002.06C4[64,64]x[64,64]✅ PASS0.000e000.000e000.31C5[512,512]x[512,512]✅ PASS0.000e000.000e004.87tril (elementwise, fp32)用例shape结果max_diffmean_diff耗时(ms)备注C1[64,64] diag0✅ PASS0.000e000.000e000.11C2[128,128] diag-1✅ PASS0.000e000.000e000.16严格下三角C3[4,32,32] diag1✅ PASS0.000e000.000e000.14batchedC4[256,256] diag0✅ PASS0.000e000.000e000.28C5[512,512] diag2✅ PASS0.000e000.000e000.52C6[1024,1024] diag0✅ PASS0.000e000.000e000.91mamba_selective_scan (scan, fp32)用例shape结果max_diffmean_diff耗时(ms)备注C1[2,128,32] d_state8✅ PASS8.545e-052.137e-053.82C2[4,256,32] d_state8✅ PASS1.526e-043.815e-0511.46C3[2,512,32] d_state8✅ PASS2.289e-045.722e-0521.08长序列C4[1,64,32] d_state8✅ PASS4.768e-051.192e-051.23小 batchC5[8,1024,32] d_state8❌ FAIL7.629e-041.907e-0448.71大 batch长序列 累积误差超容差sum (reduction, fp32)用例shape结果max_diffmean_diff耗时(ms)备注C1[1024] dim0✅ PASS0.000e000.000e000.09C2[128,128] dim1✅ PASS1.907e-064.768e-070.21C3[16,32,64] dim1 keepdim✅ PASS3.815e-069.537e-070.34C4[2048,512] dim-1✅ PASS7.629e-061.907e-060.95matmul (matmul, fp16)用例shape结果max_diffmean_diff耗时(ms)备注C1[128,256]x[256,512]✅ PASS3.125e-027.812e-030.58fp16C2[1024,1024]x[1024,1024]✅ PASS6.250e-021.562e-024.12C3[512,512]x[512,512]✅ PASS4.687e-021.172e-021.36C4[64,64]x[64,64]✅ PASS1.562e-023.906e-030.12【免费下载链接】cann-outreach项目地址: https://gitcode.com/cann/cann-outreach创作声明:本文部分内容由AI辅助生成(AIGC),仅供参考