ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

prune debug message

Browse Source
This commit is contained in:
Feng Shijie
2025-07-30 06:20:30 +00:00
parent 2e5d4c74cd
commit 3f43b841d4
2 changed files with 32 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
example/ck_tile/18_flatmm/run_flatmm_example.inc
View File

@@ -114,6 +114,35 @@ int run_flatmm_example_with_layouts(int argc,
auto per_channel_scale_dev_ptr = ck_tile::FlatmmScalePointer<ScaleGranularityN>{
static_cast<float*>(per_channel_scale_dev_buf.GetDeviceBuffer())};
invoke_flatmm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
decltype(per_token_scale_dev_ptr),
decltype(per_channel_scale_dev_ptr),
UsePersistentKernel>(a_dev_buf,
b_shuffle_dev_buf,
c_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
kbatch,
per_token_scale_dev_ptr,
per_channel_scale_dev_ptr,
n_warmup,
n_repeat);
c_dev_buf.FromDevice(c_rslt_host.data());
bool pass = true;
if(arg_parser.get_int("v") == 1)
@@ -212,35 +241,6 @@ int run_flatmm_example_with_layouts(int argc,
ck_tile::hip_check_error(hipFree(d_B));
ck_tile::hip_check_error(hipFree(d_C));
invoke_flatmm<FlatmmConfig,
ADataType,
BDataType,
ck_tile::tuple<>,
AccDataType,
CDataType,
ALayout,
BLayout,
ck_tile::tuple<>,
CLayout,
decltype(per_token_scale_dev_ptr),
decltype(per_channel_scale_dev_ptr),
UsePersistentKernel>(a_dev_buf,
b_shuffle_dev_buf,
c_dev_buf,
M,
N,
K,
stride_A,
stride_B,
stride_C,
kbatch,
per_token_scale_dev_ptr,
per_channel_scale_dev_ptr,
n_warmup,
n_repeat);
c_dev_buf.FromDevice(c_rslt_host.data());
c_gpu_ref_dev_buf.FromDevice(c_gpu_ref_host.data());
const float max_accumulated_value =
*std::max_element(c_gpu_ref_host.mData.begin(), c_gpu_ref_host.mData.end());

6
include/ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp
View File

@@ -287,9 +287,9 @@ struct FlatmmKernel
const int persistent_block_size = prop.multiProcessorCount * maxActiveBlocksPerCU;
const int total_work_tile_cnt = TilePartitioner::GridSize(kargs.M, kargs.N);
std::cout << "maxActiveBlocksPerCU: " << maxActiveBlocksPerCU
<< ", persistent_block_size: " << persistent_block_size
<< ", total_work_tile_cnt: " << total_work_tile_cnt << std::endl;
// std::cout << "maxActiveBlocksPerCU: " << maxActiveBlocksPerCU
// << ", persistent_block_size: " << persistent_block_size
// << ", total_work_tile_cnt: " << total_work_tile_cnt << std::endl;
assert(kargs.k_batch == 1);
return dim3(min(persistent_block_size, total_work_tile_cnt), 1, kargs.k_batch);