mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-01 12:11:19 +00:00
This reverts commit 2cbbf5dcb3.
This commit is contained in:
@@ -57,7 +57,7 @@ struct SmoothquantPipelineTwoPass
|
||||
|
||||
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));
|
||||
|
||||
auto reduce_absmax_func = ReduceOp::AbsMax{};
|
||||
auto reduce_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
|
||||
@@ -77,7 +77,7 @@ struct SmoothquantPipelineTwoPass
|
||||
auto absmax = block_reduce2d.template MakeYBlockTile<XTensorType>();
|
||||
set_tile(absmax, reduce_absmax_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 x = load_tile(x_window);
|
||||
const auto smscale = load_tile(smscale_window);
|
||||
@@ -121,7 +121,7 @@ struct SmoothquantPipelineTwoPass
|
||||
move_tile_window(qy_window, {0, stride_to_right_most_window});
|
||||
|
||||
// recompute y and quantize y to qy
|
||||
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 x = load_tile(x_window);
|
||||
const auto smscale = load_tile(smscale_window);
|
||||
|
||||
Reference in New Issue
Block a user