ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-18 09:38:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

fixed

Browse Source
This commit is contained in:
mtgu0705
2025-01-02 19:41:04 +08:00
parent 3eee7eda7c
commit 2d50ecbeb5
2 changed files with 44 additions and 107 deletions
Download Patch File Download Diff File Expand all files Collapse all files

64
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
View File

@@ -1146,36 +1146,7 @@ struct ThreadwiseTensorSliceTransfer_v4
});
}
if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value &&
is_same<remove_cvref_t<DstData>, half_t>::value)
{
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
// DstData)
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
constexpr index_t pack_size = 8;
static_assert(SrcScalarPerVector % pack_size == 0, "");
using src_v_t = typename vector_type_maker_t<SrcData, pack_size / PackedSize>::type;
using dst_v_t = typename vector_type_maker_t<DstData, pack_size>::type;
static_for<0, SrcScalarPerVector / pack_size, 1>{}([&](auto i) {
ck::tensor_operation::element_wise::PassThroughPack8{}(
dst_tmp_vector.template AsType<dst_v_t>()(i),
src_tmp_vector.template AsType<src_v_t>()[i]);
});
// copy data from dst_tmp_vector into dst_buf
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t dst_offset = dst_desc.CalculateOffset(
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
});
}
else if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value &&
is_same<remove_cvref_t<DstData>, f8_t>::value)
if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value)
{
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
// DstData)
@@ -1361,7 +1332,6 @@ struct ThreadwiseTensorSliceTransfer_v4
}
else if constexpr(SrcBuffer::IsStaticBuffer())
{
static_assert(false, "");
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t src_offset = src_desc.CalculateOffset(
src_ref_to_origin_disp_idx + data_to_origin_disp_idx +
@@ -1371,10 +1341,8 @@ struct ThreadwiseTensorSliceTransfer_v4
});
}
if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value &&
is_same<remove_cvref_t<DstData>, half_t>::value)
if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value)
{
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
// DstData)
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
@@ -1405,34 +1373,6 @@ struct ThreadwiseTensorSliceTransfer_v4
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
});
}
else if constexpr(is_same<remove_cvref_t<SrcData>, pk_i4_t>::value &&
is_same<remove_cvref_t<DstData>, f8_t>::value)
{
// copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
// DstData)
vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
constexpr index_t pack_size = 8;
static_assert(SrcScalarPerVector % pack_size == 0, "");
using src_v_t = typename vector_type_maker_t<SrcData, pack_size / PackedSize>::type;
using dst_v_t = typename vector_type_maker_t<DstData, pack_size>::type;
static_for<0, SrcScalarPerVector / pack_size, 1>{}([&](auto i) {
ck::tensor_operation::element_wise::PassThroughPack8{}(
dst_tmp_vector.template AsType<dst_v_t>()(i),
src_tmp_vector.template AsType<src_v_t>()[i]);
});
// copy data from dst_tmp_vector into dst_buf
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t dst_offset = dst_desc.CalculateOffset(
dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
});
}
else if constexpr(is_same<remove_cvref_t<SrcData>, f8_t>::value &&
is_same<remove_cvref_t<DstData>, half_t>::value &&
SrcScalarPerVector % 2 == 0)