diff --git a/benchmarks/config/omp/dnn-bf16.json b/benchmarks/config/omp/dnn-bf16.json index f03655586..d96532de1 100644 --- a/benchmarks/config/omp/dnn-bf16.json +++ b/benchmarks/config/omp/dnn-bf16.json @@ -4,29 +4,29 @@ "gemm_bf16_omp_2_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] }, "gemm_bf16_omp_4_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] }, "gemm_bf16_omp_8_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] }, "gemm_bf16_omp_16_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] } }}, @@ -35,29 +35,29 @@ "mlp_bf16_omp_2_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] }, "mlp_bf16_omp_4_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] }, "mlp_bf16_omp_8_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] }, "mlp_bf16_omp_16_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "1", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "1", "2", "1024", "1024", "1024", "1024" ], "extensions": [] } }} diff --git a/benchmarks/config/omp/dnn-fp32.json b/benchmarks/config/omp/dnn-fp32.json index c68de1b5f..d8eedbcab 100644 --- a/benchmarks/config/omp/dnn-fp32.json +++ b/benchmarks/config/omp/dnn-fp32.json @@ -4,29 +4,29 @@ "gemm_fp32_omp_2_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] }, "gemm_fp32_omp_4_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] }, "gemm_fp32_omp_8_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] }, "gemm_fp32_omp_16_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "0", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] } }}, @@ -35,29 +35,29 @@ "mlp_fp32_omp_2_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "2", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] }, "mlp_fp32_omp_4_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "4", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] }, "mlp_fp32_omp_8_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "8", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] }, "mlp_fp32_omp_16_dnn": { "type": "XSMM-DNN", "benchmark": "xsmm_dnn_mlp", - "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" }, - "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "0", "1024", "1024", "1024", "1024" ], + "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0", "LIBXSMM_DNN_DISABLE_SFC": "1" }, + "flags": [ "100", "256", "5", "F", "32", "32", "32", "0", "1", "1024", "1024", "1024", "1024" ], "extensions": [] } }} diff --git a/cmake/modules/xsmm-dnn.cmake b/cmake/modules/xsmm-dnn.cmake index c72c8d512..41ccdec9a 100644 --- a/cmake/modules/xsmm-dnn.cmake +++ b/cmake/modules/xsmm-dnn.cmake @@ -16,8 +16,8 @@ else() FetchContent_Declare( xsmm_dnn - URL https://github.com/libxsmm/libxsmm-dnn/archive/6e042517be4b975f431fb92d803ddc5d9dafcc64.tar.gz - URL_HASH SHA256=2d92577b865a823ec223a675bea1537a61b717a2b8fb9f06dd7e46a19183df60 + URL https://github.com/libxsmm/libxsmm-dnn/archive/a6305b41322a0b39be541e09443de424b4285cf7.tar.gz + URL_HASH SHA256=dccc4292991361786a61bd00c020d6916e746051c693a3873920cbab05b538c3 ) FetchContent_GetProperties(xsmm_dnn) diff --git a/cmake/modules/xsmm.cmake b/cmake/modules/xsmm.cmake index 4b683a22a..d1d845afc 100644 --- a/cmake/modules/xsmm.cmake +++ b/cmake/modules/xsmm.cmake @@ -13,8 +13,8 @@ else() FetchContent_Declare( xsmm - URL https://github.com/libxsmm/libxsmm/archive/c5fe65139050d7460d6929b2b507e3e349687477.tar.gz - URL_HASH SHA256=c55467f2440e5d894907a6297e02c1f5c9f96fefa27a807fca3ad60bf7f4c5ab + URL https://github.com/libxsmm/libxsmm/archive/85851d4368f730069086e5acf65eaa3ae3e80852.tar.gz + URL_HASH SHA256=5b63f968ca0c618e0edf6760399a72c07cabed88b103e0ced11e30b321a824b7 ) FetchContent_GetProperties(xsmm)