Skip to content

GitLab

"stubs/vscode:/vscode.git/clone" did not exist on "f0a400464b4aeebc1421a128b5f48c8da358265a"
Unverified Commit 10bb8110 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Misc fixes (#24) 

* use cast_pointer_to_generic_address_space() in v6r1 kernel wrapper, DynamcBuffer and buffer_load take customized invalid-element-value, add buffer_load/store for fp64

* use remove_cvref_t
parent a7a758d8
Show whitespace changes
Inline Side-by-side
Showing with 267 additions and 252 deletions
composable_kernel/include/tensor_description/tensor_adaptor.hpp
View file @ 10bb8110
...@@ -189,8 +189,7 @@ struct TensorAdaptor ...@@ -189,8 +189,7 @@ struct TensorAdaptor
bool is_known = true; bool is_known = true;
static_for<0, Transforms::Size(), 1>{}([&](auto i) { static_for<0, Transforms::Size(), 1>{}([&](auto i) {
is_known &= is_known &= remove_cvref_t<decltype(Transforms{}[i])>::IsKnownAtCompileTime();
remove_cv_t<remove_reference_t<decltype(Transforms{}[i])>>::IsKnownAtCompileTime();
}); });
return is_known && is_known_at_compile_time<ElementSize>::value; return is_known && is_known_at_compile_time<ElementSize>::value;
......
composable_kernel/include/tensor_description/tensor_descriptor.hpp
View file @ 10bb8110
...@@ -185,8 +185,7 @@ struct TensorDescriptor ...@@ -185,8 +185,7 @@ struct TensorDescriptor
bool is_known = true; bool is_known = true;
static_for<0, Transforms::Size(), 1>{}([&](auto i) { static_for<0, Transforms::Size(), 1>{}([&](auto i) {
is_known &= is_known &= remove_cvref_t<decltype(Transforms{}[i])>::IsKnownAtCompileTime();
remove_cv_t<remove_reference_t<decltype(Transforms{}[i])>>::IsKnownAtCompileTime();
}); });
return is_known && is_known_at_compile_time<ElementSize>::value && return is_known && is_known_at_compile_time<ElementSize>::value &&
...@@ -587,11 +586,11 @@ __host__ __device__ constexpr bool coordinate_has_valid_offset(const TensorDesc& ...@@ -587,11 +586,11 @@ __host__ __device__ constexpr bool coordinate_has_valid_offset(const TensorDesc&
template <typename TensorDesc> template <typename TensorDesc>
using TensorCoordinate_t = decltype(make_tensor_coordinate( using TensorCoordinate_t = decltype(make_tensor_coordinate(
TensorDesc{}, MultiIndex<remove_cv_t<remove_reference_t<TensorDesc>>::GetNumOfDimension()>{})); TensorDesc{}, MultiIndex<remove_cvref_t<TensorDesc>::GetNumOfDimension()>{}));
template <typename TensorDesc> template <typename TensorDesc>
using TensorCoordinateStep_t = decltype(make_tensor_coordinate_step( using TensorCoordinateStep_t = decltype(make_tensor_coordinate_step(
TensorDesc{}, MultiIndex<remove_cv_t<remove_reference_t<TensorDesc>>::GetNumOfDimension()>{})); TensorDesc{}, MultiIndex<remove_cvref_t<TensorDesc>::GetNumOfDimension()>{}));
} // namespace ck } // namespace ck
#endif #endif
composable_kernel/include/tensor_operation/blockwise_gemm_dlops_v3.hpp
View file @ 10bb8110
...@@ -110,12 +110,10 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3 ...@@ -110,12 +110,10 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
const BThreadBuffer& b_thread_buf, const BThreadBuffer& b_thread_buf,
CThreadBuffer& c_thread_buf) const CThreadBuffer& c_thread_buf) const
{ {
static_assert(is_same<remove_cv_t<remove_reference_t<typename ABlockBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<FloatA>>>::value && is_same<remove_cvref_t<typename ABlockBuffer::type>, remove_cvref_t<FloatA>>::value &&
is_same<remove_cv_t<remove_reference_t<typename BThreadBuffer::type>>, is_same<remove_cvref_t<typename BThreadBuffer::type>, remove_cvref_t<FloatB>>::value &&
remove_cv_t<remove_reference_t<FloatB>>>::value && is_same<remove_cvref_t<typename CThreadBuffer::type>, remove_cvref_t<FloatC>>::value &&
is_same<remove_cv_t<remove_reference_t<typename CThreadBuffer::type>>,
remove_cv_t<remove_reference_t<FloatC>>>::value &&
"wrong! inconsistent type"); "wrong! inconsistent type");
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
......
composable_kernel/include/tensor_operation/threadwise_contraction_dlops.hpp
View file @ 10bb8110
...@@ -55,18 +55,15 @@ struct ThreadwiseGemmDlops_km0m1_kn0n1_m0m1n0n1 ...@@ -55,18 +55,15 @@ struct ThreadwiseGemmDlops_km0m1_kn0n1_m0m1n0n1
CBuffer& c_buf, CBuffer& c_buf,
COriginIdx) COriginIdx)
{ {
static_assert( static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<AOriginIdx>>>::value && is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<BOriginIdx>>>::value && is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
is_known_at_compile_time<remove_cv_t<remove_reference_t<COriginIdx>>>::value,
"wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time"); "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
static_assert(is_same<remove_cv_t<remove_reference_t<typename ABuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<FloatA>>>::value && is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
is_same<remove_cv_t<remove_reference_t<typename BBuffer::type>>, is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
remove_cv_t<remove_reference_t<FloatB>>>::value && is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
is_same<remove_cv_t<remove_reference_t<typename CBuffer::type>>,
remove_cv_t<remove_reference_t<FloatC>>>::value &&
"wrong! inconsistent type"); "wrong! inconsistent type");
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
...@@ -157,18 +154,15 @@ struct ThreadwiseContractionDlops_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_ ...@@ -157,18 +154,15 @@ struct ThreadwiseContractionDlops_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_
CBuffer& c_buf, CBuffer& c_buf,
COriginIdx) COriginIdx)
{ {
static_assert( static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<AOriginIdx>>>::value && is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<BOriginIdx>>>::value && is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
is_known_at_compile_time<remove_cv_t<remove_reference_t<COriginIdx>>>::value,
"wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time"); "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
static_assert(is_same<remove_cv_t<remove_reference_t<typename ABuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<FloatA>>>::value && is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
is_same<remove_cv_t<remove_reference_t<typename BBuffer::type>>, is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
remove_cv_t<remove_reference_t<FloatB>>>::value && is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
is_same<remove_cv_t<remove_reference_t<typename CBuffer::type>>,
remove_cv_t<remove_reference_t<FloatC>>>::value &&
"wrong! inconsistent type"); "wrong! inconsistent type");
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
......
composable_kernel/include/tensor_operation/threadwise_gemm_dlops_v3.hpp
View file @ 10bb8110
...@@ -41,18 +41,15 @@ struct ThreadwiseGemmDlops_km_kn_mn_v3 ...@@ -41,18 +41,15 @@ struct ThreadwiseGemmDlops_km_kn_mn_v3
CDesc::IsKnownAtCompileTime(), CDesc::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time"); "wrong! Desc should be known at compile-time");
static_assert( static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<AOriginIdx>>>::value && is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<BOriginIdx>>>::value && is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
is_known_at_compile_time<remove_cv_t<remove_reference_t<COriginIdx>>>::value,
"wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time"); "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
static_assert(is_same<remove_cv_t<remove_reference_t<typename ABuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<FloatA>>>::value && is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
is_same<remove_cv_t<remove_reference_t<typename BBuffer::type>>, is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
remove_cv_t<remove_reference_t<FloatB>>>::value && is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
is_same<remove_cv_t<remove_reference_t<typename CBuffer::type>>,
remove_cv_t<remove_reference_t<FloatC>>>::value &&
"wrong! inconsistent type"); "wrong! inconsistent type");
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_set.hpp
View file @ 10bb8110
...@@ -30,11 +30,11 @@ struct ThreadwiseTensorSliceSet_v1 ...@@ -30,11 +30,11 @@ struct ThreadwiseTensorSliceSet_v1
static_assert(Buffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer"); static_assert(Buffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
static_assert(is_known_at_compile_time<remove_cv_t<remove_reference_t<OriginIdx>>>::value, static_assert(is_known_at_compile_time<remove_cvref_t<OriginIdx>>::value,
"wrong! OriginIdx need to be known at compile-time"); "wrong! OriginIdx need to be known at compile-time");
// Desc is known at compile-time // Desc is known at compile-time
constexpr auto desc = remove_cv_t<remove_reference_t<Desc>>{}; constexpr auto desc = remove_cvref_t<Desc>{};
// OriginIdx is known at compile-time // OriginIdx is known at compile-time
constexpr auto origin_idx = to_multi_index(OriginIdx{}); constexpr auto origin_idx = to_multi_index(OriginIdx{});
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer.hpp
View file @ 10bb8110
...@@ -95,18 +95,13 @@ struct ThreadwiseTensorSliceTransfer_v1r3 ...@@ -95,18 +95,13 @@ struct ThreadwiseTensorSliceTransfer_v1r3
static_assert(SrcDesc::IsKnownAtCompileTime(), static_assert(SrcDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc need to known at compile-time"); "wrong! SrcDesc need to known at compile-time");
static_assert( static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value,
is_known_at_compile_time<remove_cv_t<remove_reference_t<SrcSliceOriginIdx>>>::value,
"wrong! SrcSliceOrigin need to known at compile-time"); "wrong! SrcSliceOrigin need to known at compile-time");
static_assert(SrcBuffer::IsStaticBuffer(), "wrong! SrcBuffer need to be StaticBuffer"); static_assert(SrcBuffer::IsStaticBuffer(), "wrong! SrcBuffer need to be StaticBuffer");
// static_assert(is_same<remove_cv_t<remove_reference_t<typename SrcBuffer::type>>,
// remove_cv_t<remove_reference_t<SrcData>>>::value,
//"wrong! SrcBuffer data type is wrong");
// SrcDesc and src_slice_origin_idx are known at compile-time // SrcDesc and src_slice_origin_idx are known at compile-time
constexpr auto src_desc = remove_cv_t<remove_reference_t<SrcDesc>>{}; constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{}); constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
...@@ -421,16 +416,15 @@ struct ThreadwiseTensorSliceTransfer_v2 ...@@ -421,16 +416,15 @@ struct ThreadwiseTensorSliceTransfer_v2
static_assert(DstDesc::IsKnownAtCompileTime(), static_assert(DstDesc::IsKnownAtCompileTime(),
"wrong! DstDesc need to known at compile-time"); "wrong! DstDesc need to known at compile-time");
static_assert( static_assert(is_known_at_compile_time<remove_cvref_t<DstSliceOriginIdx>>::value,
is_known_at_compile_time<remove_cv_t<remove_reference_t<DstSliceOriginIdx>>>::value,
"wrong! DstSliceOrigin need to known at compile-time"); "wrong! DstSliceOrigin need to known at compile-time");
static_assert(is_same<remove_cv_t<remove_reference_t<typename DstBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<DstData>>>::value && is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value &&
"wrong! inconsistent type"); "wrong! inconsistent type");
// DstDesc and dst_slice_origin_idx are known at compile-time // DstDesc and dst_slice_origin_idx are known at compile-time
constexpr auto dst_desc = remove_cv_t<remove_reference_t<DstDesc>>{}; constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
constexpr auto dst_slice_origin_idx = DstSliceOriginIdx{}; constexpr auto dst_slice_origin_idx = DstSliceOriginIdx{};
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
...@@ -742,8 +736,8 @@ struct ThreadwiseTensorSliceTransfer_v3 ...@@ -742,8 +736,8 @@ struct ThreadwiseTensorSliceTransfer_v3
SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds, SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
"wrong!"); "wrong!");
static_assert(is_same<remove_cv_t<remove_reference_t<typename SrcBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<SrcData>>>::value, is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
"wrong! SrcBuffer and SrcData data type are inconsistent"); "wrong! SrcBuffer and SrcData data type are inconsistent");
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
...@@ -899,8 +893,8 @@ struct ThreadwiseTensorSliceTransfer_v3 ...@@ -899,8 +893,8 @@ struct ThreadwiseTensorSliceTransfer_v3
DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds, DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
"wrong!"); "wrong!");
static_assert(is_same<remove_cv_t<remove_reference_t<typename DstBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<DstData>>>::value, is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
"wrong! SrcBuffer or DstBuffer data type is wrong"); "wrong! SrcBuffer or DstBuffer data type is wrong");
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
...@@ -1315,24 +1309,21 @@ struct ThreadwiseTensorSliceTransfer_v4 ...@@ -1315,24 +1309,21 @@ struct ThreadwiseTensorSliceTransfer_v4
static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(), static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc and DstDesc need to known at compile-time"); "wrong! SrcDesc and DstDesc need to known at compile-time");
static_assert(is_same<remove_cv_t<remove_reference_t<typename SrcBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<SrcData>>>::value && is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
is_same<remove_cv_t<remove_reference_t<typename DstBuffer::type>>, is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
remove_cv_t<remove_reference_t<DstData>>>::value,
"wrong! SrcBuffer or DstBuffer data type is wrong"); "wrong! SrcBuffer or DstBuffer data type is wrong");
static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer"); static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
static_assert( static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
is_known_at_compile_time< is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
remove_cv_t<remove_reference_t<SrcRefToOriginDisplacement>>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<DstOriginIdx>>>::value,
"wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known " "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
"at compile-time"); "at compile-time");
// SrcDesc and DstDesc are known at compile-time // SrcDesc and DstDesc are known at compile-time
constexpr auto src_desc = remove_cv_t<remove_reference_t<SrcDesc>>{}; constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
constexpr auto dst_desc = remove_cv_t<remove_reference_t<DstDesc>>{}; constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
// SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{}); constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v2.hpp
View file @ 10bb8110
...@@ -80,8 +80,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -80,8 +80,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1
SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds, SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
"wrong!"); "wrong!");
static_assert(is_same<remove_cv_t<remove_reference_t<typename SrcBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<SrcData>>>::value, is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
"wrong! SrcBuffer and SrcData data type are inconsistent"); "wrong! SrcBuffer and SrcData data type are inconsistent");
// tensor descriptor for src_vector // tensor descriptor for src_vector
...@@ -248,8 +248,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -248,8 +248,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1
DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds, DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
"wrong!"); "wrong!");
static_assert(is_same<remove_cv_t<remove_reference_t<typename DstBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<DstData>>>::value, is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
"wrong! SrcBuffer or DstBuffer data type is wrong"); "wrong! SrcBuffer or DstBuffer data type is wrong");
// tensor descriptor for dst_vector // tensor descriptor for dst_vector
...@@ -669,24 +669,21 @@ struct ThreadwiseTensorSliceTransfer_v4r1 ...@@ -669,24 +669,21 @@ struct ThreadwiseTensorSliceTransfer_v4r1
static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(), static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc and DstDesc need to known at compile-time"); "wrong! SrcDesc and DstDesc need to known at compile-time");
static_assert(is_same<remove_cv_t<remove_reference_t<typename SrcBuffer::type>>, static_assert(
remove_cv_t<remove_reference_t<SrcData>>>::value && is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
is_same<remove_cv_t<remove_reference_t<typename DstBuffer::type>>, is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
remove_cv_t<remove_reference_t<DstData>>>::value,
"wrong! SrcBuffer or DstBuffer data type is wrong"); "wrong! SrcBuffer or DstBuffer data type is wrong");
static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer"); static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
static_assert( static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
is_known_at_compile_time< is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
remove_cv_t<remove_reference_t<SrcRefToOriginDisplacement>>>::value &&
is_known_at_compile_time<remove_cv_t<remove_reference_t<DstOriginIdx>>>::value,
"wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known " "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
"at compile-time"); "at compile-time");
// SrcDesc and DstDesc are known at compile-time // SrcDesc and DstDesc are known at compile-time
constexpr auto src_desc = remove_cv_t<remove_reference_t<SrcDesc>>{}; constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
constexpr auto dst_desc = remove_cv_t<remove_reference_t<DstDesc>>{}; constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
// SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{}); constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
......
composable_kernel/include/utility/amd_buffer_addressing.hpp
View file @ 10bb8110
...@@ -225,13 +225,49 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w ...@@ -225,13 +225,49 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w
index_t src_wave_addr_offset) index_t src_wave_addr_offset)
{ {
static_assert( static_assert(
(is_same<T, double>::value && (N == 1 || N == 2 || N == 4)) ||
(is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
(is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) || (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)), (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
"wrong! not implemented"); "wrong! not implemented");
if constexpr(is_same<T, float>::value) if constexpr(is_same<T, double>::value)
{
// use fp32 load to mimic fp64 load
if constexpr(N == 1)
{
const float2_t tmp = llvm_amdgcn_raw_buffer_load_fp32x2(
src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
return as_type<double>(tmp);
}
else if constexpr(N == 2)
{
const float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4(
src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
return as_type<double2_t>(tmp);
}
else if constexpr(N == 4)
{
const float4_t f32_0 = llvm_amdgcn_raw_buffer_load_fp32x4(
src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
const float4_t f32_1 =
llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
src_thread_addr_offset,
src_wave_addr_offset + 4 * sizeof(float),
0);
vector_type<double, 4> tmp;
tmp.AsType<double2_t>()(Number<0>{}) = as_type<double2_t>(f32_0);
tmp.AsType<double2_t>()(Number<1>{}) = as_type<double2_t>(f32_1);
return tmp.AsType<double4_t>()(Number<0>{});
}
}
else if constexpr(is_same<T, float>::value)
{ {
if constexpr(N == 1) if constexpr(N == 1)
{ {
...@@ -283,25 +319,11 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w ...@@ -283,25 +319,11 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w
} }
else if constexpr(N == 8) else if constexpr(N == 8)
{ {
#if 0 // use fp32 load to mimic fp16 load
vector_type<half_t, 8> tmp;
tmp.AsType<half4_t>()(Number<0>{}) = llvm_amdgcn_raw_buffer_load_fp16x4(
src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
tmp.AsType<half4_t>()(Number<1>{}) =
llvm_amdgcn_raw_buffer_load_fp16x4(src_wave_buffer_resource,
src_thread_addr_offset,
src_wave_addr_offset + 4 * sizeof(half_t),
0);
return tmp.AsType<half8_t>()(Number<0>{});
#else
float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4( float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4(
src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0); src_wave_buffer_resource, src_thread_addr_offset, src_wave_addr_offset, 0);
return as_type<half8_t>(tmp); return as_type<half8_t>(tmp);
#endif
} }
} }
else if constexpr(is_same<T, int32_t>::value) else if constexpr(is_same<T, int32_t>::value)
...@@ -433,17 +455,19 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -433,17 +455,19 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
index_t dst_wave_addr_offset) index_t dst_wave_addr_offset)
{ {
static_assert( static_assert(
(is_same<T, double>::value && (N == 1 || N == 2)) ||
(is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) || (is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) ||
(is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
(is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)) || (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)) ||
(is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) || (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
(is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)),
"wrong! not implemented"); "wrong! not implemented");
if constexpr(is_same<T, float>::value) if constexpr(is_same<T, double>::value)
{ {
// use fp32 store to mimic fp64 store
if constexpr(N == 1) if constexpr(N == 1)
{ {
llvm_amdgcn_raw_buffer_store_fp32(src_thread_data, llvm_amdgcn_raw_buffer_store_fp32x2(as_type<float2_t>(src_thread_data),
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
...@@ -451,26 +475,18 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -451,26 +475,18 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
} }
else if constexpr(N == 2) else if constexpr(N == 2)
{ {
llvm_amdgcn_raw_buffer_store_fp32x2(src_thread_data, llvm_amdgcn_raw_buffer_store_fp32x4(as_type<float4_t>(src_thread_data),
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset,
0);
}
else if constexpr(N == 4)
{
llvm_amdgcn_raw_buffer_store_fp32x4(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
} }
} }
else if constexpr(is_same<T, int32_t>::value) else if constexpr(is_same<T, float>::value)
{ {
if constexpr(N == 1) if constexpr(N == 1)
{ {
llvm_amdgcn_raw_buffer_store_i32(src_thread_data, llvm_amdgcn_raw_buffer_store_fp32(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
...@@ -478,7 +494,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -478,7 +494,7 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
} }
else if constexpr(N == 2) else if constexpr(N == 2)
{ {
llvm_amdgcn_raw_buffer_store_i32x2(src_thread_data, llvm_amdgcn_raw_buffer_store_fp32x2(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
...@@ -486,18 +502,18 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -486,18 +502,18 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
} }
else if constexpr(N == 4) else if constexpr(N == 4)
{ {
llvm_amdgcn_raw_buffer_store_i32x4(src_thread_data, llvm_amdgcn_raw_buffer_store_fp32x4(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
} }
} }
else if constexpr(is_same<T, int8_t>::value) else if constexpr(is_same<T, half_t>::value)
{ {
if constexpr(N == 1) if constexpr(N == 1)
{ {
llvm_amdgcn_raw_buffer_store_i8(src_thread_data, llvm_amdgcn_raw_buffer_store_fp16(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
...@@ -505,58 +521,69 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -505,58 +521,69 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
} }
else if constexpr(N == 2) else if constexpr(N == 2)
{ {
#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE llvm_amdgcn_raw_buffer_store_fp16x2(src_thread_data,
llvm_amdgcn_raw_buffer_store_i8x2(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
#else }
llvm_amdgcn_raw_buffer_store_i16(as_type<int16_t>(src_thread_data), else if constexpr(N == 4)
{
llvm_amdgcn_raw_buffer_store_fp16x4(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
#endif
} }
else if constexpr(N == 4) else if constexpr(N == 8)
{ {
#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE vector_type<half_t, 8> tmp{src_thread_data};
llvm_amdgcn_raw_buffer_store_i8x4(src_thread_data,
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<0>{}],
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
#else
llvm_amdgcn_raw_buffer_store_i32(as_type<int32_t>(src_thread_data), llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<1>{}],
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset + 4 * sizeof(half_t),
0);
}
}
else if constexpr(is_same<T, int32_t>::value)
{
if constexpr(N == 1)
{
llvm_amdgcn_raw_buffer_store_i32(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
#endif
} }
else if constexpr(N == 8) else if constexpr(N == 2)
{ {
llvm_amdgcn_raw_buffer_store_i32x2(as_type<int32x2_t>(src_thread_data), llvm_amdgcn_raw_buffer_store_i32x2(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
} }
else if constexpr(N == 16) else if constexpr(N == 4)
{ {
llvm_amdgcn_raw_buffer_store_i32x4(as_type<int32x4_t>(src_thread_data), llvm_amdgcn_raw_buffer_store_i32x4(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
} }
} }
else if constexpr(is_same<T, half_t>::value) else if constexpr(is_same<T, int8_t>::value)
{ {
if constexpr(N == 1) if constexpr(N == 1)
{ {
llvm_amdgcn_raw_buffer_store_fp16(src_thread_data, llvm_amdgcn_raw_buffer_store_i8(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
...@@ -564,34 +591,50 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src ...@@ -564,34 +591,50 @@ __device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src
} }
else if constexpr(N == 2) else if constexpr(N == 2)
{ {
llvm_amdgcn_raw_buffer_store_fp16x2(src_thread_data, #if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
llvm_amdgcn_raw_buffer_store_i8x2(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
#else
llvm_amdgcn_raw_buffer_store_i16(as_type<int16_t>(src_thread_data),
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset,
0);
#endif
} }
else if constexpr(N == 4) else if constexpr(N == 4)
{ {
llvm_amdgcn_raw_buffer_store_fp16x4(src_thread_data, #if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
llvm_amdgcn_raw_buffer_store_i8x4(src_thread_data,
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
#else
llvm_amdgcn_raw_buffer_store_i32(as_type<int32_t>(src_thread_data),
dst_wave_buffer_resource,
dst_thread_addr_offset,
dst_wave_addr_offset,
0);
#endif
} }
else if constexpr(N == 8) else if constexpr(N == 8)
{ {
vector_type<half_t, 8> tmp{src_thread_data}; llvm_amdgcn_raw_buffer_store_i32x2(as_type<int32x2_t>(src_thread_data),
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<0>{}],
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset, dst_wave_addr_offset,
0); 0);
}
llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<1>{}], else if constexpr(N == 16)
{
llvm_amdgcn_raw_buffer_store_i32x4(as_type<int32x4_t>(src_thread_data),
dst_wave_buffer_resource, dst_wave_buffer_resource,
dst_thread_addr_offset, dst_thread_addr_offset,
dst_wave_addr_offset + 4 * sizeof(half_t), dst_wave_addr_offset,
0); 0);
} }
} }
...@@ -720,7 +763,7 @@ __device__ void amd_buffer_atomic_add_impl(const typename vector_type<T, N>::typ ...@@ -720,7 +763,7 @@ __device__ void amd_buffer_atomic_add_impl(const typename vector_type<T, N>::typ
} }
// buffer_load requires: // buffer_load requires:
// 1) p_src_wave must be in global memory space // 1) p_src_wave must point to global memory space
// 2) p_src_wave must be a wavewise pointer. // 2) p_src_wave must be a wavewise pointer.
// It is user's responsibility to make sure that is true. // It is user's responsibility to make sure that is true.
template <typename T, index_t N> template <typename T, index_t N>
...@@ -754,7 +797,7 @@ amd_buffer_load_invalid_element_return_return_zero(const T* p_src_wave, ...@@ -754,7 +797,7 @@ amd_buffer_load_invalid_element_return_return_zero(const T* p_src_wave,
} }
// buffer_load requires: // buffer_load requires:
// 1) p_src_wave must be in global memory space // 1) p_src_wave must point to global memory space
// 2) p_src_wave must be a wavewise pointer. // 2) p_src_wave must be a wavewise pointer.
// It is user's responsibility to make sure that is true. // It is user's responsibility to make sure that is true.
template <typename T, index_t N> template <typename T, index_t N>
...@@ -782,7 +825,7 @@ amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave, ...@@ -782,7 +825,7 @@ amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave,
} }
// buffer_store requires: // buffer_store requires:
// 1) p_dst_wave must be global memory // 1) p_dst_wave must point to global memory
// 2) p_dst_wave must be a wavewise pointer. // 2) p_dst_wave must be a wavewise pointer.
// It is user's responsibility to make sure that is true. // It is user's responsibility to make sure that is true.
template <typename T, index_t N> template <typename T, index_t N>
...@@ -816,7 +859,7 @@ __device__ void amd_buffer_store(const typename vector_type_maker<T, N>::type::t ...@@ -816,7 +859,7 @@ __device__ void amd_buffer_store(const typename vector_type_maker<T, N>::type::t
} }
// buffer_atomic_add requires: // buffer_atomic_add requires:
// 1) p_dst_wave must be global memory // 1) p_dst_wave must point to global memory
// 2) p_dst_wave must be a wavewise pointer. // 2) p_dst_wave must be a wavewise pointer.
// It is user's responsibility to make sure that is true. // It is user's responsibility to make sure that is true.
template <typename T, index_t N> template <typename T, index_t N>
......
composable_kernel/include/utility/array.hpp
View file @ 10bb8110
...@@ -48,7 +48,7 @@ struct Array<TData, 0> ...@@ -48,7 +48,7 @@ struct Array<TData, 0>
template <typename X, typename... Xs> template <typename X, typename... Xs>
__host__ __device__ constexpr auto make_array(X&& x, Xs&&... xs) __host__ __device__ constexpr auto make_array(X&& x, Xs&&... xs)
{ {
using data_type = remove_cv_t<remove_reference_t<X>>; using data_type = remove_cvref_t<X>;
return Array<data_type, sizeof...(Xs) + 1>{{std::forward<X>(x), std::forward<Xs>(xs)...}}; return Array<data_type, sizeof...(Xs) + 1>{{std::forward<X>(x), std::forward<Xs>(xs)...}};
} }
......
composable_kernel/include/utility/data_type.hpp
View file @ 10bb8110
...@@ -73,6 +73,13 @@ struct scalar_type<vector_type<T, N>> ...@@ -73,6 +73,13 @@ struct scalar_type<vector_type<T, N>>
}; };
// //
template <>
struct scalar_type<double>
{
using type = double;
static constexpr index_t vector_size = 1;
};
template <> template <>
struct scalar_type<float> struct scalar_type<float>
{ {
...@@ -864,6 +871,10 @@ struct vector_type<T, 256> ...@@ -864,6 +871,10 @@ struct vector_type<T, 256>
} }
}; };
// fp64
using double2_t = typename vector_type<double, 2>::type;
using double4_t = typename vector_type<double, 4>::type;
// fp32 // fp32
using float2_t = typename vector_type<float, 2>::type; using float2_t = typename vector_type<float, 2>::type;
using float4_t = typename vector_type<float, 4>::type; using float4_t = typename vector_type<float, 4>::type;
......
composable_kernel/include/utility/dynamic_buffer.hpp
View file @ 10bb8110
...@@ -39,18 +39,15 @@ struct DynamicBuffer ...@@ -39,18 +39,15 @@ struct DynamicBuffer
} }
template <typename X, template <typename X,
typename enable_if< typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
is_same<typename scalar_type<remove_cv_t<remove_reference_t<X>>>::type, typename scalar_type<remove_cvref_t<T>>::type>::value,
typename scalar_type<remove_cv_t<remove_reference_t<T>>>::type>::value,
bool>::type = false> bool>::type = false>
__host__ __device__ constexpr auto Get(index_t i, bool is_valid_element) const __host__ __device__ constexpr auto Get(index_t i, bool is_valid_element) const
{ {
// X contains multiple T // X contains multiple T
constexpr index_t scalar_per_t_vector = constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
scalar_type<remove_cv_t<remove_reference_t<T>>>::vector_size;
constexpr index_t scalar_per_x_vector = constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
scalar_type<remove_cv_t<remove_reference_t<X>>>::vector_size;
static_assert(scalar_per_x_vector % scalar_per_t_vector == 0, static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
"wrong! X need to be multiple T"); "wrong! X need to be multiple T");
...@@ -67,14 +64,13 @@ struct DynamicBuffer ...@@ -67,14 +64,13 @@ struct DynamicBuffer
if constexpr(InvalidElementUseNumericalZeroValue) if constexpr(InvalidElementUseNumericalZeroValue)
{ {
return amd_buffer_load_invalid_element_return_return_zero< return amd_buffer_load_invalid_element_return_return_zero<remove_cvref_t<T>,
remove_cv_t<remove_reference_t<T>>, t_per_x>(
t_per_x>(p_data_, i, is_valid_element, element_space_size_); p_data_, i, is_valid_element, element_space_size_);
} }
else else
{ {
return amd_buffer_load_invalid_element_return_customized_value< return amd_buffer_load_invalid_element_return_customized_value<remove_cvref_t<T>,
remove_cv_t<remove_reference_t<T>>,
t_per_x>( t_per_x>(
p_data_, i, is_valid_element, element_space_size_, invalid_element_value_); p_data_, i, is_valid_element, element_space_size_, invalid_element_value_);
} }
...@@ -94,18 +90,15 @@ struct DynamicBuffer ...@@ -94,18 +90,15 @@ struct DynamicBuffer
} }
template <typename X, template <typename X,
typename enable_if< typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
is_same<typename scalar_type<remove_cv_t<remove_reference_t<X>>>::type, typename scalar_type<remove_cvref_t<T>>::type>::value,
typename scalar_type<remove_cv_t<remove_reference_t<T>>>::type>::value,
bool>::type = false> bool>::type = false>
__host__ __device__ void Set(index_t i, bool is_valid_element, const X& x) __host__ __device__ void Set(index_t i, bool is_valid_element, const X& x)
{ {
// X contains multiple T // X contains multiple T
constexpr index_t scalar_per_t_vector = constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
scalar_type<remove_cv_t<remove_reference_t<T>>>::vector_size;
constexpr index_t scalar_per_x_vector = constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
scalar_type<remove_cv_t<remove_reference_t<X>>>::vector_size;
static_assert(scalar_per_x_vector % scalar_per_t_vector == 0, static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
"wrong! X need to be multiple T"); "wrong! X need to be multiple T");
...@@ -115,7 +108,7 @@ struct DynamicBuffer ...@@ -115,7 +108,7 @@ struct DynamicBuffer
#if CK_USE_AMD_BUFFER_ADDRESSING #if CK_USE_AMD_BUFFER_ADDRESSING
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector; constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
amd_buffer_store<remove_cv_t<remove_reference_t<T>>, t_per_x>( amd_buffer_store<remove_cvref_t<T>, t_per_x>(
x, p_data_, i, is_valid_element, element_space_size_); x, p_data_, i, is_valid_element, element_space_size_);
#else #else
if(is_valid_element) if(is_valid_element)
...@@ -136,70 +129,65 @@ struct DynamicBuffer ...@@ -136,70 +129,65 @@ struct DynamicBuffer
// ISA, so I try to let compiler emit IR "store<i32, 4>" which would be lower to // ISA, so I try to let compiler emit IR "store<i32, 4>" which would be lower to
// ds_write_b128 // ds_write_b128
// TODO: remove this after compiler fix // TODO: remove this after compiler fix
if constexpr(is_same<typename scalar_type<remove_cv_t<remove_reference_t<T>>>::type, if constexpr(is_same<typename scalar_type<remove_cvref_t<T>>::type, int8_t>::value)
int8_t>::value)
{ {
static_assert( static_assert((is_same<remove_cvref_t<T>, int8_t>::value &&
(is_same<remove_cv_t<remove_reference_t<T>>, int8_t>::value && is_same<remove_cvref_t<X>, int8_t>::value) ||
is_same<remove_cv_t<remove_reference_t<X>>, int8_t>::value) || (is_same<remove_cvref_t<T>, int8_t>::value &&
(is_same<remove_cv_t<remove_reference_t<T>>, int8_t>::value && is_same<remove_cvref_t<X>, int8x2_t>::value) ||
is_same<remove_cv_t<remove_reference_t<X>>, int8x2_t>::value) || (is_same<remove_cvref_t<T>, int8_t>::value &&
(is_same<remove_cv_t<remove_reference_t<T>>, int8_t>::value && is_same<remove_cvref_t<X>, int8x4_t>::value) ||
is_same<remove_cv_t<remove_reference_t<X>>, int8x4_t>::value) || (is_same<remove_cvref_t<T>, int8x4_t>::value &&
(is_same<remove_cv_t<remove_reference_t<T>>, int8x4_t>::value && is_same<remove_cvref_t<X>, int8x4_t>::value) ||
is_same<remove_cv_t<remove_reference_t<X>>, int8x4_t>::value) || (is_same<remove_cvref_t<T>, int8x8_t>::value &&
(is_same<remove_cv_t<remove_reference_t<T>>, int8x8_t>::value && is_same<remove_cvref_t<X>, int8x8_t>::value) ||
is_same<remove_cv_t<remove_reference_t<X>>, int8x8_t>::value) || (is_same<remove_cvref_t<T>, int8x16_t>::value &&
(is_same<remove_cv_t<remove_reference_t<T>>, int8x16_t>::value && is_same<remove_cvref_t<X>, int8x16_t>::value),
is_same<remove_cv_t<remove_reference_t<X>>, int8x16_t>::value),
"wrong! not implemented for this combination, please add " "wrong! not implemented for this combination, please add "
"implementation"); "implementation");
if constexpr(is_same<remove_cv_t<remove_reference_t<T>>, int8_t>::value && if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cv_t<remove_reference_t<X>>, int8_t>::value) is_same<remove_cvref_t<X>, int8_t>::value)
{ {
// HACK: cast pointer of x is bad // HACK: cast pointer of x is bad
// TODO: remove this after compiler fix // TODO: remove this after compiler fix
*c_style_pointer_cast<int8_t*>(&p_data_[i]) = *c_style_pointer_cast<int8_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int8_t*>(&x); *c_style_pointer_cast<const int8_t*>(&x);
} }
else if constexpr(is_same<remove_cv_t<remove_reference_t<T>>, int8_t>::value && else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cv_t<remove_reference_t<X>>, int8x2_t>::value) is_same<remove_cvref_t<X>, int8x2_t>::value)
{ {
// HACK: cast pointer of x is bad // HACK: cast pointer of x is bad
// TODO: remove this after compiler fix // TODO: remove this after compiler fix
*c_style_pointer_cast<int16_t*>(&p_data_[i]) = *c_style_pointer_cast<int16_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int16_t*>(&x); *c_style_pointer_cast<const int16_t*>(&x);
} }
else if constexpr(is_same<remove_cv_t<remove_reference_t<T>>, int8_t>::value && else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
is_same<remove_cv_t<remove_reference_t<X>>, int8x4_t>::value) is_same<remove_cvref_t<X>, int8x4_t>::value)
{ {
// HACK: cast pointer of x is bad // HACK: cast pointer of x is bad
// TODO: remove this after compiler fix // TODO: remove this after compiler fix
*c_style_pointer_cast<int32_t*>(&p_data_[i]) = *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32_t*>(&x); *c_style_pointer_cast<const int32_t*>(&x);
} }
else if constexpr(is_same<remove_cv_t<remove_reference_t<T>>, else if constexpr(is_same<remove_cvref_t<T>, int8x4_t>::value &&
int8x4_t>::value && is_same<remove_cvref_t<X>, int8x4_t>::value)
is_same<remove_cv_t<remove_reference_t<X>>, int8x4_t>::value)
{ {
// HACK: cast pointer of x is bad // HACK: cast pointer of x is bad
// TODO: remove this after compiler fix // TODO: remove this after compiler fix
*c_style_pointer_cast<int32_t*>(&p_data_[i]) = *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32_t*>(&x); *c_style_pointer_cast<const int32_t*>(&x);
} }
else if constexpr(is_same<remove_cv_t<remove_reference_t<T>>, else if constexpr(is_same<remove_cvref_t<T>, int8x8_t>::value &&
int8x8_t>::value && is_same<remove_cvref_t<X>, int8x8_t>::value)
is_same<remove_cv_t<remove_reference_t<X>>, int8x8_t>::value)
{ {
// HACK: cast pointer of x is bad // HACK: cast pointer of x is bad
// TODO: remove this after compiler fix // TODO: remove this after compiler fix
*c_style_pointer_cast<int32x2_t*>(&p_data_[i]) = *c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
*c_style_pointer_cast<const int32x2_t*>(&x); *c_style_pointer_cast<const int32x2_t*>(&x);
} }
else if constexpr(is_same<remove_cv_t<remove_reference_t<T>>, else if constexpr(is_same<remove_cvref_t<T>, int8x16_t>::value &&
int8x16_t>::value && is_same<remove_cvref_t<X>, int8x16_t>::value)
is_same<remove_cv_t<remove_reference_t<X>>, int8x16_t>::value)
{ {
// HACK: cast pointer of x is bad // HACK: cast pointer of x is bad
// TODO: remove this after compiler fix // TODO: remove this after compiler fix
...@@ -224,18 +212,15 @@ struct DynamicBuffer ...@@ -224,18 +212,15 @@ struct DynamicBuffer
} }
template <typename X, template <typename X,
typename enable_if< typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
is_same<typename scalar_type<remove_cv_t<remove_reference_t<X>>>::type, typename scalar_type<remove_cvref_t<T>>::type>::value,
typename scalar_type<remove_cv_t<remove_reference_t<T>>>::type>::value,
bool>::type = false> bool>::type = false>
__host__ __device__ void AtomicAdd(index_t i, bool is_valid_element, const X& x) __host__ __device__ void AtomicAdd(index_t i, bool is_valid_element, const X& x)
{ {
// X contains multiple T // X contains multiple T
constexpr index_t scalar_per_t_vector = constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
scalar_type<remove_cv_t<remove_reference_t<T>>>::vector_size;
constexpr index_t scalar_per_x_vector = constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
scalar_type<remove_cv_t<remove_reference_t<X>>>::vector_size;
static_assert(scalar_per_x_vector % scalar_per_t_vector == 0, static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
"wrong! X need to be multiple T"); "wrong! X need to be multiple T");
...@@ -245,7 +230,7 @@ struct DynamicBuffer ...@@ -245,7 +230,7 @@ struct DynamicBuffer
#if CK_USE_AMD_BUFFER_ADDRESSING #if CK_USE_AMD_BUFFER_ADDRESSING
constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector; constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
amd_buffer_atomic_add<remove_cv_t<remove_reference_t<T>>, t_per_x>( amd_buffer_atomic_add<remove_cvref_t<T>, t_per_x>(
x, p_data_, i, is_valid_element, element_space_size_); x, p_data_, i, is_valid_element, element_space_size_);
#else #else
if(is_valid_element) if(is_valid_element)
...@@ -266,9 +251,14 @@ __host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize el ...@@ -266,9 +251,14 @@ __host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize el
return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true>{p, element_space_size}; return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, true>{p, element_space_size};
} }
template <AddressSpaceEnum_t BufferAddressSpace, typename T, typename ElementSpaceSize> template <
AddressSpaceEnum_t BufferAddressSpace,
typename T,
typename ElementSpaceSize,
typename X,
typename enable_if<is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value, bool>::type = false>
__host__ __device__ constexpr auto __host__ __device__ constexpr auto
make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, T invalid_element_value) make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element_value)
{ {
return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false>{ return DynamicBuffer<BufferAddressSpace, T, ElementSpaceSize, false>{
p, element_space_size, invalid_element_value}; p, element_space_size, invalid_element_value};
......
composable_kernel/include/utility/tuple.hpp
View file @ 10bb8110
...@@ -159,7 +159,7 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X ...@@ -159,7 +159,7 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
template <typename... Xs> template <typename... Xs>
__host__ __device__ constexpr auto make_tuple(Xs&&... xs) __host__ __device__ constexpr auto make_tuple(Xs&&... xs)
{ {
return Tuple<remove_cv_t<remove_reference_t<Xs>>...>(std::forward<Xs>(xs)...); return Tuple<remove_cvref_t<Xs>...>(std::forward<Xs>(xs)...);
} }
} // namespace ck } // namespace ck
......
composable_kernel/include/utility/tuple_helper.hpp
View file @ 10bb8110
...@@ -14,9 +14,7 @@ struct is_known_at_compile_time<Tuple<Ts...>> ...@@ -14,9 +14,7 @@ struct is_known_at_compile_time<Tuple<Ts...>>
return container_reduce( return container_reduce(
Tuple<Ts...>{}, Tuple<Ts...>{},
[](auto x, bool r) { [](auto x, bool r) {
return is_known_at_compile_time< return is_known_at_compile_time<remove_cvref_t<decltype(x)>>::value & r;
remove_cv_t<remove_reference_t<decltype(x)>>>::value &
r;
}, },
true); true);
} }
......
composable_kernel/include/utility/type.hpp
View file @ 10bb8110
...@@ -22,6 +22,9 @@ using remove_reference_t = typename std::remove_reference<T>::type; ...@@ -22,6 +22,9 @@ using remove_reference_t = typename std::remove_reference<T>::type;
template <typename T> template <typename T>
using remove_cv_t = typename std::remove_cv<T>::type; using remove_cv_t = typename std::remove_cv<T>::type;
template <typename T>
using remove_cvref_t = remove_cv_t<std::remove_reference_t<T>>;
template <typename T> template <typename T>
inline constexpr bool is_pointer_v = std::is_pointer<T>::value; inline constexpr bool is_pointer_v = std::is_pointer<T>::value;
......
composable_kernel/src/kernel_wrapper/convolution_forward_implicit_gemm_v6r1_dlops_nchw_kcyx_nkhw.cpp
View file @ 10bb8110
...@@ -374,13 +374,8 @@ extern "C" __global__ void ...@@ -374,13 +374,8 @@ extern "C" __global__ void
CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1{}, CGridDesc_GM10_BM0_BM1_GN10_BN0_BN1{},
CGridBlockCluster_BlockId_To_GM10_GN10{})); CGridBlockCluster_BlockId_To_GM10_GN10{}));
const auto desc_tuple = *reinterpret_cast<const DescTuple*>( const auto desc_tuple =
#pragma clang diagnostic push *reinterpret_cast<const DescTuple*>(cast_pointer_to_generic_address_space(p_desc_tuple));
#pragma clang diagnostic ignored "-Wold-style-cast"
// TODO: how to cast?
(const void*)p_desc_tuple
#pragma clang diagnostic pop
);
const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = desc_tuple[I0]; const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = desc_tuple[I0];
const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = desc_tuple[I1]; const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = desc_tuple[I1];
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment