ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 04:19:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK-Tile] Enable vectorized reads on all layouts & improve perf. (#1835)

Browse Source 
* Refactor universal gemm policy.

* Adapt example to refactor changes.

* Introduce static encoding pattern

* Adding shuffled encoding patterns.

* Fix err in reverse tuple.

* Add transpose_tile2d

* Small refactoring + doc

* Enable reading on contiguous dimension in all layouts.

* Transpose A/B register tile if needed for comp v3 pipeline.

* Take contiguous dim size when calculating dram vector load size.

* A/B smem pack size taken from WarpGemm attributes

* Update B LDS layout and setup tile distribution pattern at class level.

* Fix static assert.

* Fix errors in examples.

* Formatting & fix IsTranspose

* Fix VectorSize & refactor.

* Add error loging messages.

* Fix VecLoadSize and TranspseC for mem pipeline.

* Update unit-tests & disable mem pipeline.

* Clang format

* Update include/ck_tile/core/tensor/tile_window.hpp

Co-authored-by: jakpiase <jakub.piasecki@amd.com>

* Fix compilation and reviewers comments.

* Refactor unit-test. Fallback to non-universal gemm.

Need to use GemmPipelineAGmemBGmemCRegV1 for now,
since GemmKernel is now supporting also non-K major vector reads.

---------

Co-authored-by: jakpiase <jakub.piasecki@amd.com>

[ROCm/composable_kernel commit: 39dc25a9b8]
This commit is contained in:
Adam Osewski
2025-01-27 16:37:19 +01:00
committed by GitHub
parent b1251e8b41
commit 38fd0bbc6f
31 changed files with 1393 additions and 688 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -70,9 +70,7 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
@@ -103,4 +101,26 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
#include "run_gemm_example.inc"
int run_gemm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
}
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

36
example/ck_tile/03_gemm/run_gemm_example.inc
View File

@@ -217,39 +217,3 @@ int run_gemm_example_with_layouts(int argc,
return pass;
}
int run_gemm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
}
else if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
// TODO: Fixme: with latest changes to GemmPipelineAGmemBGmemCRegV1DefaultPolicy below do not
// work.
// else if(a_layout == "C" && b_layout == "C")
// {
// return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
// }
// else if(a_layout == "C" && b_layout == "R")
// {
// return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
// }
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
}

68
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -28,8 +28,8 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr ck_tile::index_t M_Warp_Tile = 32;
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
#endif
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
// Compute friendly for Intrawave scheduler
constexpr ck_tile::index_t M_Tile = 256;
constexpr ck_tile::index_t N_Tile = 256;
@@ -48,6 +48,8 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool TransposeC = false;
constexpr int kBlockPerCu = 1;
// ===============================================
@@ -62,7 +64,8 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using GemmUniversalTraits = ck_tile::
TileGemmUniversalTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout, TransposeC>;
using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
@@ -85,14 +88,15 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
BDataType,
AccDataType,
GemmShape,
Traits,
GemmUniversalTraits,
scheduler,
has_hot_loop_v,
tail_number_v>;
using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
using GemmPipeline =
GEMM_PIPELINE<UniversalGemmProblem, ck_tile::UniversalGemmPipelineAgBgCrPolicy>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
@@ -117,6 +121,21 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
if(has_hot_loop)
{
#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
if(tail_num == ck_tile::TailNumber::Full)
{
Run(ck_tile::bool_constant<true>{},
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
}
else
{
std::ostringstream err;
err << "For compute pipeline tail number should always be Full, but have \"" << tail_num
<< "\" which is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
<< "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
// Tail pipeline One to Seven
if(tail_num == ck_tile::TailNumber::One)
{
@@ -177,6 +196,7 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Seven>{});
}
}
#endif
}
else
{
@@ -201,4 +221,38 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
#include "run_gemm_example.inc"
int run_gemm_example(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
using Row = ck_tile::tensor_layout::gemm::RowMajor;
using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
}
else if(a_layout == "R" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
else if(a_layout == "C" && b_layout == "C")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Col{}, Row{});
}
else if(a_layout == "C" && b_layout == "R")
{
return run_gemm_example_with_layouts(argc, argv, Col{}, Row{}, Row{});
}
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");
}
}
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

4
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View File

@@ -72,9 +72,7 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;

2
example/ck_tile/16_batched_gemm/batched_gemm.hpp
View File

@@ -39,7 +39,7 @@ auto create_args(int argc, char* argv[])
.insert("stride_b", "0", "Tensor B stride")
.insert("stride_c", "0", "Tensor C stride")
.insert("a_layout", "R", "A tensor data layout - Row by default")
.insert("b_layout", "R", "B tensor data layout - Row by default")
.insert("b_layout", "C", "B tensor data layout - Row by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("batch_stride_a", "32768", "Batch A stride")
.insert("batch_stride_b", "16384", "Batch B stride")

83
example/ck_tile/16_batched_gemm/run_batched_gemm_example.inc
View File

@@ -3,13 +3,6 @@
#pragma once
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
{
return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
ck_tile::tensor_layout::gemm::RowMajor>>{};
}
auto calculate_rtol_atol(const ck_tile::index_t K,
const ck_tile::index_t kbatch,
const float max_accumulated_value)
@@ -113,16 +106,56 @@ int run_batched_gemm_example_with_layouts(int argc,
int n_warmup = arg_parser.get_int("warmup");
int n_repeat = arg_parser.get_int("repeat");
stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(c_layout));
using namespace ck_tile::literals;
ck_tile::HostTensor<ADataType> a_m_k(ck_tile::host_tensor_descriptor(
batch_count, M, K, stride_A, batch_stride_A, is_row_major(a_layout)));
ck_tile::HostTensor<BDataType> b_k_n(ck_tile::host_tensor_descriptor(
batch_count, K, N, stride_B, batch_stride_B, is_row_major(b_layout)));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(ck_tile::host_tensor_descriptor(
batch_count, M, N, stride_C, batch_stride_C, is_row_major(c_layout)));
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return ck_tile::HostTensorDescriptor({batch_count_, row, col},
{batch_stride, stride, 1_uz});
}
else
{
return ck_tile::HostTensorDescriptor({batch_count_, row, col},
{batch_stride, 1_uz, stride});
}
};
auto f_get_default_stride = [](std::size_t row,
std::size_t col,
std::size_t stride,
auto layout) {
if(stride == 0)
{
// give a chance if stride is zero, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck_tile::tensor_layout::gemm::RowMajor>)
{
return col;
}
else
{
return row;
}
}
else
return stride;
};
stride_A = f_get_default_stride(M, K, stride_A, a_layout);
stride_B = f_get_default_stride(K, N, stride_B, b_layout);
stride_C = f_get_default_stride(M, N, stride_C, c_layout);
ck_tile::HostTensor<ADataType> a_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, a_layout));
ck_tile::HostTensor<BDataType> b_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, b_layout));
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, c_layout));
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n);
@@ -158,8 +191,8 @@ int run_batched_gemm_example_with_layouts(int argc,
if(arg_parser.get_int("v") == 1)
{
ck_tile::HostTensor<CDataType> c_m_n_host_ref(ck_tile::host_tensor_descriptor(
batch_count, M, N, stride_C, batch_stride_C, is_row_major(CLayout){}));
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, CLayout{}));
c_m_n_host_ref.SetZero();
const auto b_n_k = b_k_n.transpose({0, 2, 1});
@@ -183,8 +216,8 @@ int run_batched_gemm_example_with_layouts(int argc,
}
else if(arg_parser.get_int("v") == 2)
{
ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(ck_tile::host_tensor_descriptor(
batch_count, M, N, stride_C, batch_stride_C, is_row_major(CLayout){}));
ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, CLayout{}));
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes());
c_m_n_gpu_ref.SetZero();
c_m_n_gpu_buf_ref.SetZero();
@@ -268,11 +301,11 @@ int run_batched_gemm_example(int argc, char* argv[])
std::string a_layout = arg_parser.get_str("a_layout");
std::string b_layout = arg_parser.get_str("b_layout");
if(a_layout == "R" && b_layout == "R")
{
return run_batched_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
}
else if(a_layout == "R" && b_layout == "C")
// if(a_layout == "R" && b_layout == "R")
// {
// return run_batched_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
// }
if(a_layout == "R" && b_layout == "C")
{
return run_batched_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}

5
example/ck_tile/17_grouped_gemm/grouped_gemm.cpp
View File

@@ -88,12 +88,9 @@ using CodegenPipelineProblem =
CodegenGemmShape,
CodegenGemmTraits<ALayout, BLayout, CLayout>>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>,
CodegenGemmPolicy>;
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>>;
template <typename ALayout, typename BLayout, typename CLayout>
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner,

2
example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
View File

@@ -41,7 +41,7 @@ auto create_args(int argc, char* argv[])
.insert("stride_Bs", "", "Tensor B strides - it is empty by default.")
.insert("stride_Cs", "", "Tensor C strides - it is empty by default.")
.insert("a_layout", "R", "A tensor data layout - Row by default.")
.insert("b_layout", "R", "B tensor data layout - Row by default.")
.insert("b_layout", "C", "B tensor data layout - Row by default.")
.insert("c_layout", "R", "C tensor data layout - Row by default.")
.insert("validate", "1", "0. No validation, 1. Validation on CPU.")
.insert("warmup", "10", "number of iterations before benchmark the kernel.")

17
example/ck_tile/17_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -135,12 +135,9 @@ int run_grouped_gemm_example_with_layouts(int argc,
const ck_tile::index_t N = Ns[i];
const ck_tile::index_t K = Ks[i];
stride_As[i] =
ck_tile::get_default_stride(M, N, stride_As[i], is_row_major(a_layout));
stride_Bs[i] =
ck_tile::get_default_stride(K, N, stride_Bs[i], is_row_major(b_layout));
stride_Cs[i] =
ck_tile::get_default_stride(M, N, stride_Cs[i], is_row_major(CLayout{}));
stride_As[i] = ck_tile::get_default_stride(M, N, stride_As[i], is_row_major(a_layout));
stride_Bs[i] = ck_tile::get_default_stride(K, N, stride_Bs[i], is_row_major(b_layout));
stride_Cs[i] = ck_tile::get_default_stride(M, N, stride_Cs[i], is_row_major(CLayout{}));
a_m_k_tensors.push_back(ck_tile::HostTensor<ADataType>(
ck_tile::host_tensor_descriptor(M, K, stride_As[i], is_row_major(a_layout))));
@@ -229,10 +226,10 @@ int run_grouped_gemm_example(int argc, char* argv[])
{
return run_grouped_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
else if(a_layout == "R" && b_layout == "R")
{
return run_grouped_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
}
// else if(a_layout == "R" && b_layout == "R")
// {
// return run_grouped_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
// }
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");