ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-18 17:48:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

optimze small N case using vec io and using rcp div

Browse Source
This commit is contained in:
dummycoderfe
2024-11-04 08:50:09 +00:00
parent cb6c5d39dc
commit 27ff3dec9f
4 changed files with 18 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp
View File

@@ -125,7 +125,8 @@ struct Layernorm2dFwdPipelineOnePass
// compute inv-std
auto inv_std = tile_elementwise_in(
[&](const auto& v_) {
return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ + epsilon));
return type_convert<ComputeDataType>(1.0f) *
__builtin_amdgcn_rcpf(sqrt(v_ + epsilon));
},
var);

5
include/ck_tile/ops/welford/block/block_welford.hpp
View File

@@ -356,7 +356,8 @@ CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTe
int count)
{
using DataType = typename VarDistributedTensor_::DataType;
tile_elementwise_inout([&count](auto& x) { x = x / type_convert<DataType>(count); },
var_tensor);
tile_elementwise_inout(
[&count](auto& x) { x = x * __builtin_amdgcn_rcpf(type_convert<DataType>(count)); },
var_tensor);
}
} // namespace ck_tile

13
include/ck_tile/ops/welford/thread/thread_welford.hpp
View File

@@ -12,7 +12,7 @@ CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count)
{
// TODO: check nan? maybe no
T delta = x - mean;
mean += delta / count;
mean += delta * __builtin_amdgcn_rcpf(count);
T delta2 = x - mean;
var += delta * delta2;
}
@@ -21,11 +21,12 @@ template <typename T>
CK_TILE_DEVICE static void
welford_merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
{
int count = count_a + count_b;
T count_ = type_convert<T>(count);
T count_a_ = type_convert<T>(count_a);
T count_b_ = type_convert<T>(count_b);
T count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
int count = count_a + count_b;
T count_ = type_convert<T>(count);
T count_a_ = type_convert<T>(count_a);
T count_b_ = type_convert<T>(count_b);
T count_b_over_count =
count == 0 ? type_convert<T>(0) : count_b_ * __builtin_amdgcn_rcpf(count_);
T delta = mean_b - mean_a;
mean_a += delta * count_b_over_count;