ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 17:26:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

added implicit gemm v4 (nchw, kcyx)

Browse Source
This commit is contained in:
Chao Liu
2019-05-30 20:28:01 -05:00
parent b2439ec9dd
commit cb78cc74b7
5 changed files with 40 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/include/ConstantTensorDescriptor.hip.hpp
View File

@@ -215,7 +215,7 @@ struct ConstantTensorDescriptor
// do carry check in reversed order, starting from lowest dimension
// don't check the highest dimension
static_for<0, nDim - 1, 1>{}([&](auto IDimReverse) {
static_for<0, nDim, 1>{}([&](auto IDimReverse) {
constexpr index_t idim = nDim - 1 - IDimReverse.Get();
constexpr auto IDim = Number<idim>{};
@@ -241,7 +241,7 @@ struct ConstantTensorDescriptor
// do borrow check in reversed order, starting from lowest dimension
// don't check the highest dimension
static_for<0, nDim - 1, 1>{}([&](auto IDimReverse) {
static_for<0, nDim, 1>{}([&](auto IDimReverse) {
constexpr index_t idim = nDim - 1 - IDimReverse.Get();
constexpr auto IDim = Number<idim>{};

30
src/include/blockwise_generic_tensor_slice_op.hip.hpp
View File

@@ -3,7 +3,7 @@
// slice a (normal or merged) tensor, and copy it into another (normal or merged) tensor
// memory layout (ordering of dimensions) can be different between src and dst
// For now, only support SubLengths == 1 on a merged dimension
// For now, only support SubLengths[...] == 1 on a merged dimension
template <index_t BlockSize,
class Float,
class SrcDesc,
@@ -84,8 +84,8 @@ struct BlockwiseGenericTensorSliceCopy_v1
constexpr auto repeat_lengths = SliceLengths{} / data_per_cluster_per_dims;
// for now, only support SubLengths.Get() == 1 on a merged dimension that is merge from
// multiple dimensions
// for now, only support SubLengths.Get() == 1 on a merged dimension that constains
// multiple original dimensions
static_for<0, nDim, 1>{}([&](auto IDim_) {
constexpr auto IDim = decltype(IDim_){};
@@ -292,7 +292,8 @@ struct BlockwiseGenericTensorSliceCopy_v1
static_if<SrcDesc::ContainMultipleOriginalDimensions(IDim)>{}([&](auto fwd) {
// logic for a merged dimension, also works for non-merged dimension, but its logic may
// be unncessarily complicated for compiler to remove uselss calculations
// be unncessarily complicated for compiler to remove calculations that are useless for
// a non-merged dimension
// extract partial original dimensions
constexpr auto src_partial_original_dims =
@@ -309,6 +310,27 @@ struct BlockwiseGenericTensorSliceCopy_v1
src_partial_original_desc.UpdateMultiIndexGivenStepSizeOf1dIndex(
old_src_partial_original_multi_id, StepSize, direction);
#if 0
{
if(debug_flag && get_block_1d_id() == 0)
{
printf("id %5u %5u: "
"old_src_partial_original_multi_id %u %u %u, "
"new_src_partial_original_multi_id %u %u %u, "
"mThreadSrcOffset %u, mThreadDstOffset %u \n",
get_block_1d_id(),
get_thread_local_1d_id(),
old_src_partial_original_multi_id[0],
old_src_partial_original_multi_id[1],
old_src_partial_original_multi_id[2],
new_src_partial_original_multi_id[0],
new_src_partial_original_multi_id[1],
new_src_partial_original_multi_id[2]
);
}
}
#endif
// update "mThreadSrcOriginalMultiId"
static_for<0, src_partial_original_dims.GetSize(), 1>{}([&](auto I_) {
constexpr auto I = decltype(I_){};

20
src/include/gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hip.hpp
View File

@@ -255,28 +255,14 @@ struct GridwiseConvolutionImplicitGemm_v4_nchw_kcyx_nkhw
for(index_t e = 0; e < E; e += EPerBlock)
{
#if 0
if(e == 1 * EPerBlock && get_block_1d_id() == 0)
if(e == 0 * EPerBlock && get_block_1d_id() == 0)
{
printf("id %5u %5u: "
"mThreadSrcOriginalMultiId %u %u %u %u %u %u %u %u, "
"mThreadSrcPartialOffsets %u %u %u %u, "
"mThreadSrcOffset %u, mThreadDstOffset %u \n",
get_block_1d_id(),
get_thread_local_1d_id(),
blockwise_in_copy.mThreadSrcOriginalMultiId[0],
blockwise_in_copy.mThreadSrcOriginalMultiId[1],
blockwise_in_copy.mThreadSrcOriginalMultiId[2],
blockwise_in_copy.mThreadSrcOriginalMultiId[3],
blockwise_in_copy.mThreadSrcOriginalMultiId[4],
blockwise_in_copy.mThreadSrcOriginalMultiId[5],
blockwise_in_copy.mThreadSrcOriginalMultiId[6],
blockwise_in_copy.mThreadSrcOriginalMultiId[7],
blockwise_in_copy.mThreadSrcPartialOffsets[0],
blockwise_in_copy.mThreadSrcPartialOffsets[1],
blockwise_in_copy.mThreadSrcPartialOffsets[2],
blockwise_in_copy.mThreadSrcPartialOffsets[3],
blockwise_in_copy.mThreadSrcOffset,
blockwise_in_copy.mThreadDstOffset);
blockwise_wei_copy.mThreadSrcOffset,
blockwise_wei_copy.mThreadDstOffset);
}
#endif
// marching slicing window