Remove rocm6.3 workaround flags and macro

This commit is contained in:
OscarXu
2025-05-28 21:05:21 -05:00
parent 772debdf8f
commit 653bc83f8a
2 changed files with 3 additions and 3 deletions

View File

@@ -42,8 +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 --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")
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 -mllvm --misched-bottomup=1")
# 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})

View File

@@ -170,7 +170,7 @@
#define CK_USE_PK4_LAYOUT_SHUFFLE 1
// using .co compiled shader for moe_stage2_blockscale
#define CK_USE_ASM_MOE_BLOCKSCALE 1
#define CK_USE_ASM_MOE_BLOCKSCALE 0
// block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
#define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1