From 19e43cb175a99342ca87a307241bf28d356df681 Mon Sep 17 00:00:00 2001 From: OscarXu Date: Sat, 17 May 2025 00:51:44 -0500 Subject: [PATCH] Cmakelist diable compiler-bound flags --- example/65_gemm_multiply_multiply/CMakeLists.txt | 14 ++++---------- 1 file changed, 4 insertions(+), 10 deletions(-) diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt index f26e332ef5..4ac588e255 100644 --- a/example/65_gemm_multiply_multiply/CMakeLists.txt +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -17,6 +17,7 @@ target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVAT add_example_executable(example_moe_gemm1_xdl_fp8 moe_gemm1_xdl_fp8.cpp) add_example_executable(example_moe_gemm2_xdl_fp8 moe_gemm2_xdl_fp8.cpp) add_example_executable(example_moe_gemm2_xdl_fp8_blockscale moe_gemm2_xdl_fp8_blockscale.cpp) +add_example_executable(example_moe_gemm1_xdl_fp8_blockscale moe_gemm1_xdl_fp8_blockscale.cpp) list(APPEND gpu_list gfx942 gfx950) set(target 0) @@ -41,16 +42,8 @@ set(GEMM_OPTIONS) list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32") list(APPEND GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker) set(BLOCKSCALE_GEMM_OPTIONS) -list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --schedmodel=0 -mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental") -list(APPEND BLOCKSCALE_GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker) -target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS}) -target_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS}) -target_compile_options(example_moe_gemm2_xdl_fp8 PRIVATE ${GEMM_OPTIONS}) -target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS}) -target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${BLOCKSCALE_GEMM_OPTIONS}) - # list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --disable-schedmodel-in-sched-mi=1 -mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental -mllvm --misched-bottomup=1") -list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32 -mllvm --schedmodel=0 -mllvm --amdgpu-sched-strategy=gcn-iterative-max-occupancy-experimental") +list(APPEND BLOCKSCALE_GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32") list(APPEND BLOCKSCALE_GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker) target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS}) target_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS}) @@ -59,6 +52,7 @@ target_compile_options(example_gemm_multiply_multiply_xdl_fp8_ab_scale PRIVATE $ target_compile_options(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle PRIVATE ${BLOCKSCALE_GEMM_OPTIONS}) target_compile_options(example_moe_gemm2_xdl_fp8_blockscale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS}) +target_compile_options(example_moe_gemm1_xdl_fp8_blockscale PRIVATE ${BLOCKSCALE_GEMM_OPTIONS}) #hacky fix for bs_moe_stage2 with rocm < 6.4 -target_compile_definitions(example_moe_gemm2_xdl_fp8_blockscale PRIVATE MOE_STAGE2_ASM_DIR="${CMAKE_CURRENT_SOURCE_DIR}/hsa/") +target_compile_definitions(example_moe_gemm2_xdl_fp8_blockscale PRIVATE MOE_STAGE2_ASM_DIR="${CMAKE_CURRENT_SOURCE_DIR}/hsa/") \ No newline at end of file