[GEMM] Gemm universal device operation (#1154)

* Optimize GEMM on MI200/300:
1. Add new blockwise gemm pipeline
2. Add irregular splitk intances

* clang format + typo fix

* Fix a bug

* initial commit

* Add more instances to irregular splitk

* blkgemm pipeline v1~4 prototype

* Sanity Checked. Known issue:
1. Poor performance of splitk
2. Register spill on blkgemmpipeline v3

* Sanity and Performance fix:
1. fix a bug related to sanity in grouped b2c mapping
2. fix a bug related to sanity and performance in splitk offset

* Sanity and API update:
1. Remove prefetch stage
2. Fix valid check bug
3, Add first gemm_universal instance into ckProfiler

* Add NN instances for gemm universal

* 1. Add NT instances for gemm_universal
2. Fix a bug about Kpadding in gemm_universal

* Fix a bug regarding padding Odd K number

* remove kernel print

* Fix KPadding bug...

* Update safety check

* another try to fix kpadding..

* Sanity checked

* new instances..

* clang format+typo fix

* remove clang format script's change

* Add non-hotloop compile option

* 1. Add fp16xfp8 example
2. pull packed convert f8 from pr1150

* Some miscs.. opt and fix

* Add pipeline description docs

* Split universal gemm instance library to cut profiler compiling time

* uncomment cmakefile

* Fix a bug caused by blockwise_gemm_pipe_v2

* reduce default splitk to 1

* Add 224x256x64 tile size

* update, including:
1. Experiment pipeline 5~7
2. Optimization for pipeline 4
3. Organized instance library

* temp save

* temp save

* Permuted lds layout, sanity and function checked

* clang format

* Move OOB check from RunRead to RunWrite, for better software pipeline.
TODO: agpr spill when NN layout

* clangformat

* A/B splitpipe scheduler for v3

* Fix two bugs

* bug fix

* fix a bug in oob check

* Example for mixed fp16_fp8 gemm

* Clean experimental code blocks

* Add mixed precision gemm into profiler

* tempsave

* optimize m/n major lds layout

* Add RRR GEMM  mixed precision instances

* Optimize f8 matrix transpose

* Add test_gemm_universal

* A/B spilt schedule for blkpip v5

* Take ds_read2 into iglp scheduling scheme

* format

* fixed cmake

* Add llvm-option into CI cmake flag

---------
Co-authored-by: Jing Zhang <jizhan@amd.com>

profiler/src/profile_gemm_universal.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_universal_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+    F8_F16_F16,     // 4
+    F16_F8_F16,     // 5
+    F16_F16_F16_F8, // 6
+};
+
+#define OP_NAME "gemm_universal"
+#define OP_DESC "Universal GEMM"
+
+int profile_gemm_universal(int argc, char* argv[])
+{
+    if(argc != 15 && argc != 17)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16, "
+               "comp f8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg14: split k into  mulitiple batch\n");
+        printf("optional:\n");
+        printf("arg15: number of warm-up cycles (default 1)\n");
+        printf("arg16: number of iterations (default 10)\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+    const int KBatch  = std::stoi(argv[14]);
+
+    int n_warmup = 1;
+    int n_iter   = 10;
+    if(argc == 17)
+    {
+        n_warmup = std::stoi(argv[15]);
+        n_iter   = std::stoi(argv[16]);
+    }
+
+    using F32 = float;
+    using F16 = ck::half_t;
+    using F8  = ck::f8_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using CDataType   = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_universal_impl<ADataType,
+                                                              BDataType,
+                                                              AccDataType,
+                                                              CDataType,
+                                                              ALayout,
+                                                              BLayout,
+                                                              CLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideC < 0) ? DefaultStrideC : StrideC,
+            KBatch,
+            n_warmup,
+            n_iter);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_universal);

script/cmake-ck-dev.sh

@@ -8,7 +8,7 @@ MY_PROJECT_SOURCE=$1
 cmake                                                                                             \
 -D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
 -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
--D CMAKE_CXX_FLAGS="-std=c++17 -O3 -ftemplate-backtrace-limit=0  -fPIE  -Wno-gnu-line-marker"     \
+-D CMAKE_CXX_FLAGS="-Xclang -mllvm -Xclang -enable-post-misched=0 -std=c++17 -O3 -ftemplate-backtrace-limit=0  -fPIE  -Wno-gnu-line-marker"     \
 -D CMAKE_BUILD_TYPE=Release                                                                       \
 -D BUILD_DEV=ON                                                                                   \
 -D GPU_TARGETS="gfx908;gfx90a;gfx940"                                                             \

test/CMakeLists.txt

@@ -149,6 +149,7 @@ add_subdirectory(gemm)
 add_subdirectory(gemm_add)
 add_subdirectory(gemm_layernorm)
 add_subdirectory(gemm_split_k)
+add_subdirectory(gemm_universal)
 add_subdirectory(gemm_reduce)
 add_subdirectory(batched_gemm)
 add_subdirectory(batched_gemm_reduce)

test/gemm_universal/CMakeLists.txt

+add_gtest_executable(test_gemm_universal test_gemm_universal_xdl.cpp)
+if(result EQUAL 0)
+   target_link_libraries(test_gemm_universal PRIVATE utility device_gemm_universal_instance)
+ endif()

test/gemm_universal/test_gemm_universal_ut_cases.inc

+#pragma once
+
+TYPED_TEST(TestGemmUniversal_MK_KN, SmallM)
+{
+    std::vector<int> Ms{1, 2, 3, 4, 5, 6};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmUniversal_MK_NK, SmallM)
+{
+    std::vector<int> Ms{1, 2, 3, 4, 5, 6};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmUniversal_MK_KN, MidLargeM)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmUniversal_MK_NK, MidLargeM)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmUniversal_MK_KN, PaddK)
+{
+    std::vector<int> Ms{127};
+    constexpr int N = 512;
+    constexpr int K = 437;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmUniversal_MK_NK, PaddK)
+{
+    std::vector<int> Ms{127};
+    constexpr int N = 512;
+    constexpr int K = 437;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmUniversal_MK_KN, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmUniversal_MK_NK, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}

test/gemm_universal/test_gemm_universal_util.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <sstream>
+#include <tuple>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "include/ck/utility/data_type.hpp"
+#include "profiler/profile_gemm_universal_impl.hpp"
+
+namespace ck {
+namespace test {
+
+template <typename Tuple>
+class TestGemmUniversal : public testing::Test
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using F32 = float;
+
+    protected:
+    using ALayout   = std::tuple_element_t<0, Tuple>;
+    using BLayout   = std::tuple_element_t<1, Tuple>;
+    using CLayout   = Row;
+    using ADataType = std::tuple_element_t<2, Tuple>;
+    using BDataType = std::tuple_element_t<3, Tuple>;
+    using CDataType = std::tuple_element_t<4, Tuple>;
+
+    public:
+    static constexpr bool verify_     = true;
+    static constexpr int init_method_ = 1; // decimal value initialization
+    static constexpr bool log_        = false;
+    static constexpr bool bench_      = false; // measure kernel performance
+    std::vector<int> k_batches_;
+
+    void SetUp() override { k_batches_ = {1, 2, 3, 5, 8}; }
+
+    void Run(const int M,
+             const int N,
+             const int K,
+             const int StrideA,
+             const int StrideB,
+             const int StrideC)
+    {
+        for(auto kb : k_batches_)
+        {
+            RunSingle(M, N, K, StrideA, StrideB, StrideC, kb);
+        }
+    }
+
+    void RunSingle(const int M,
+                   const int N,
+                   const int K,
+                   const int StrideA,
+                   const int StrideB,
+                   const int StrideC,
+                   int kbatch   = 1,
+                   int n_warmup = 1,
+                   int n_iter   = 10)
+    {
+        bool pass = ck::profiler::profile_gemm_universal_impl<ADataType,
+                                                              BDataType,
+                                                              F32,
+                                                              CDataType,
+                                                              ALayout,
+                                                              BLayout,
+                                                              CLayout>(verify_,
+                                                                       init_method_,
+                                                                       log_,
+                                                                       bench_,
+                                                                       M,
+                                                                       N,
+                                                                       K,
+                                                                       StrideA,
+                                                                       StrideB,
+                                                                       StrideC,
+                                                                       kbatch,
+                                                                       n_warmup,
+                                                                       n_iter);
+        EXPECT_TRUE(pass);
+    }
+};
+
+} // namespace test
+} // namespace ck

test/gemm_universal/test_gemm_universal_xdl.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "gtest/gtest.h"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "test_gemm_universal_util.hpp"
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+namespace {
+
+template <typename X, typename Y>
+struct tuple_concat;
+
+template <typename... Xs, typename... Ys>
+struct tuple_concat<std::tuple<Xs...>, std::tuple<Ys...>>
+{
+    using type = std::tuple<Xs..., Ys...>;
+};
+
+} // namespace
+
+template <typename Tuple>
+class TestGemmUniversal_MK_KN
+    : public ck::test::TestGemmUniversal<typename tuple_concat<std::tuple<Row, Row>, Tuple>::type>
+{
+};
+
+template <typename Tuple>
+class TestGemmUniversal_MK_NK
+    : public ck::test::TestGemmUniversal<typename tuple_concat<std::tuple<Row, Col>, Tuple>::type>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    //         ADataType, BDataType, CDataType
+    std::tuple<      F16,       F16,       F16>,
+    std::tuple<      F16,        F8,       F16>,
+    std::tuple<      F8,        F16,       F16>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestGemmUniversal_MK_KN, KernelTypes);
+TYPED_TEST_SUITE(TestGemmUniversal_MK_NK, KernelTypes);
+
+#include "test_gemm_universal_ut_cases.inc"