From 31fe7704cad9eb1feb0c3fe4ebe2653d5ad2e373 Mon Sep 17 00:00:00 2001 From: Si Yudong Date: Thu, 28 Nov 2024 11:07:02 +0800 Subject: [PATCH] Improve performance of shape 1024x1024x1024 out of box (#2839) Close https://github.com/intel/intel-xpu-backend-for-triton/issues/2822 After: 90% -> 103% --- benchmarks/triton_kernels_benchmark/gemm_benchmark.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/benchmarks/triton_kernels_benchmark/gemm_benchmark.py b/benchmarks/triton_kernels_benchmark/gemm_benchmark.py index 9941b0c5f0..2ad8636414 100644 --- a/benchmarks/triton_kernels_benchmark/gemm_benchmark.py +++ b/benchmarks/triton_kernels_benchmark/gemm_benchmark.py @@ -32,7 +32,7 @@ ] + [ triton.Config( {'BLOCK_SIZE_M': 256, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 4, 'grf_mode': 'large'}, - num_stages=s, num_warps=32) for s in [2, 3] + num_stages=s, num_warps=32) for s in [2, 3, 4] ] + [ triton.Config( {'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 4, 'grf_mode': 'large'},