Revert "[CK-Tile] Add the API to load SGPR (#2878)" (#2904)

This reverts commit 2cbbf5dcb3.
This commit is contained in:
asleepzzz
2025-09-24 05:33:51 +08:00
committed by GitHub
parent 959df2a155
commit f161b5b738
40 changed files with 167 additions and 273 deletions

View File

@@ -92,13 +92,13 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
index_t num_n_tile_iteration =
amd_wave_read_first_lane(integer_divide_ceil(row_size, Block_N));
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
using XTensorType = decltype(cast_tile<ComputeDataType>(load_tile(a_window)));
auto square_sum = block_reduce2d.template MakeYBlockTile<XTensorType>();
set_tile(square_sum, reduce_square_sum_func.GetIdentityValue<ComputeDataType>());
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto a = load_tile(a_window);
const auto b = load_tile(b_window);
@@ -149,7 +149,7 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
if constexpr(kSaveX)
__syncthreads();
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = [&]() {
if constexpr(kSaveX)
@@ -226,7 +226,7 @@ struct AddRmsnorm2dRdquantFwdPipelineThreePass
}
move_tile_window(gamma_window, {Block_N});
for(int iN = amd_wave_read_first_lane(0); iN < num_n_tile_iteration; ++iN)
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
auto x = [&]() {
if constexpr(kSaveX)