Skip to content

GitLab

Commit db843529 authored by coderfeli's avatar coderfeli
Browse files

fix warnings and revert cmake and fix clang format

parent 5765ba51
Show whitespace changes
Inline Side-by-side
Showing with 188 additions and 207 deletions
CMakeLists.txt
View file @ db843529
......@@ -516,6 +516,10 @@ include_directories(BEFORE
)
SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV")
if(BUILD_DEV)
add_compile_options(-Werror)
add_compile_options(-Weverything)
endif()
message("CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
if("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang")
......
cmake/EnableCompilerWarnings.cmake
View file @ db843529
......@@ -66,6 +66,7 @@ else()
-Wunreachable-code
-Wunused
-Wno-reserved-identifier
-Werror
-Wno-option-ignored
-Wsign-compare
-Wno-extra-semi-stmt
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle.hpp
View file @ db843529
......@@ -137,7 +137,6 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
using Base::AMmaKStride;
using Base::BMmaKStride;
......@@ -271,10 +270,8 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
......@@ -285,10 +282,8 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
ABlockBuffer& a_block_buf1,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
index_t num_loop) const
......@@ -296,8 +291,6 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
__builtin_amdgcn_sched_barrier(0);
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
......@@ -349,7 +342,8 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec =
b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>, Number<0>{}>();
b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>,
Number<0>{}>();
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
......@@ -400,7 +394,8 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec =
b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>, Number<1>{}>();
b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>,
Number<1>{}>();
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
......@@ -450,7 +445,8 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec =
b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>, Number<0>{}>();
b_blockwise_copy
.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>, Number<0>{}>();
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
......@@ -458,14 +454,12 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
......@@ -492,7 +486,8 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle<BlockGemmPipelineScheduler::Intr
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec =
b_blockwise_copy.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>, Number<1>{}>();
b_blockwise_copy
.template GetSrcThreadScratchIdx<Sequence<0, k0, 0>, Number<1>{}>();
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp
View file @ db843529
......@@ -68,7 +68,9 @@ template <typename ALayout,
typename ComputeTypeB = ComputeTypeA,
typename LDSTypeA = ComputeTypeA,
typename LDSTypeB = ComputeTypeB>
struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle : public DeviceGemmMultiD_Xdl_CShuffle_V3<ALayout,
struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
: public DeviceGemmMultiD_Xdl_CShuffle_V3<
ALayout,
BLayout,
DsLayout,
CLayout,
......@@ -172,7 +174,6 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle : public DeviceGemmMultiD_Xd
LDSTypeA,
LDSTypeB>;
using Argument = typename GridwiseGemm::Argument;
// Invoker
......@@ -267,7 +268,9 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle : public DeviceGemmMultiD_Xd
constexpr index_t minimum_occupancy =
BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave ? 1 : 2;
// static_assert(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 && has_main_k_block_loop, "only impl BlockGemmPipelineVersion::v3 and has mainloop right now");
// static_assert(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 &&
// has_main_k_block_loop, "only impl BlockGemmPipelineVersion::v3 and has mainloop right
// now");
if(has_main_k_block_loop)
{
// Tail number always full
......@@ -284,8 +287,8 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle : public DeviceGemmMultiD_Xd
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<GridwiseGemm,
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
......@@ -310,8 +313,8 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle : public DeviceGemmMultiD_Xd
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<GridwiseGemm,
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp
View file @ db843529
......@@ -134,8 +134,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
static constexpr index_t NumDTensor = DsDataType::Size();
using mfma_selector = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>;
static constexpr index_t KPack = math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk);
static constexpr index_t KLane = mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
static constexpr index_t KPack =
math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk);
static constexpr index_t KLane =
mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
static constexpr index_t KRepeat = KPerBlock / KLane / KPack;
static constexpr index_t NLane = NPerXdl;
static constexpr index_t NWave = NPerBlock / NPerXdl / NXdlPerWave;
......@@ -322,10 +324,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
{
constexpr index_t NkSwizzle = BlockSize * KPack;
constexpr index_t NkSwizzleNumber = Number<NkSwizzle>{};
return make_naive_tensor_descriptor(
make_tuple(N0, K0, NkSwizzleNumber),
make_tuple(K0 * NkSwizzle, NkSwizzleNumber, I1)
);
return make_naive_tensor_descriptor(make_tuple(N0, K0, NkSwizzleNumber),
make_tuple(K0 * NkSwizzle, NkSwizzleNumber, I1));
}
__host__ __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
......@@ -423,9 +423,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
__host__ __device__ static constexpr auto
MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
{
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, NPerXdl>(BBlockDesc_BK0_N_BK1{});
return MakeGemmMmaTileDescriptor<NXdlPerWave, NWave, NPerXdl>(BBlockDesc_BK0_N_BK1{});
}
template <typename ELayout>
......@@ -943,7 +941,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
__device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
......@@ -956,7 +953,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
}
using BlockwiseGemmPipe =
remove_cvref_t<decltype(BlockwiseGemmXdlops_pipeline_bpreshuffle<BlkGemmPipeSched,
remove_cvref_t<decltype(BlockwiseGemmXdlops_pipeline_bpreshuffle<
BlkGemmPipeSched,
BlockSize,
LDSTypeA,
LDSTypeB,
......@@ -982,17 +980,12 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
// constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
// b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for C shuffle in LDS
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
......@@ -1259,8 +1252,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
{
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bpreshuffled = MakeBGridDescriptor_Preshuffled(
problem.BN0Shuffled, problem.BK0Shuffled);
const auto b_grid_desc_bpreshuffled =
MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N<CLayout>(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
......@@ -1294,10 +1287,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * (NPerBlock / NLane / NWave)) ;
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
__builtin_amdgcn_readfirstlane(block_n_id * (NPerBlock / NLane / NWave));
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
......@@ -1339,19 +1329,19 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
// using BThreadClusterLengths = Sequence<1, 1, BlockSize>;
// using BBlockTransferClusterArrangeOrder = Sequence<0, 1, 2>;
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
auto b_blockwise_copy = ThreadGroupTensorSliceTransfer_v4r1<
ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<1, KRepeat, KPack * BlockSize>,
Sequence<1, 1, BlockSize>, //BThreadClusterLengths,
Sequence<0, 1, 2>, //BBlockTransferClusterArrangeOrder,
Sequence<1, 1, BlockSize>, // BThreadClusterLengths,
Sequence<0, 1, 2>, // BBlockTransferClusterArrangeOrder,
BDataType,
LDSTypeB,
decltype(b_grid_desc_bpreshuffled),
decltype(b_block_desc_bk0_n_bk1),
Sequence<0, 1, 2>,//BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>, // BBlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
......@@ -1361,8 +1351,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
2>(
b_grid_desc_bpreshuffled,
2>(b_grid_desc_bpreshuffled,
make_multi_index(n_block_data_idx_on_grid, 0, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
......@@ -1370,20 +1359,12 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
ck::tensor_operation::element_wise::PassThrough{});
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
// Cast after lds
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<LDSTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto a_block_buf1 = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<LDSTypeA*>(p_shared1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<LDSTypeB*>(p_shared) +
a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB),
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, KRepeat, 0);
......@@ -1404,10 +1385,8 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
a_block_buf1,
a_block_slice_copy_step,
b_grid_desc_bpreshuffled,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
c_thread_buf,
num_k_block_main_loop);
......@@ -1419,7 +1398,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
......@@ -1672,7 +1650,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
});
}
}
};
} // namespace ck
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
View file @ db843529
......@@ -268,7 +268,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1
}
template <typename SeqIdx, index_t ThreadScratchId = 0>
__device__ constexpr auto GetSrcThreadScratchIdx(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
__device__ constexpr auto
GetSrcThreadScratchIdx(Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
using vector_t = typename vector_type_maker<SrcData, SrcScalarPerVector>::type::type;
return src_thread_scratch_tuple_(thread_scratch_id).template GetAsType<vector_t>(SeqIdx{});
......
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