GEMM batched/splitK/cgemm/grouped int4 examples (#383)

* Grouped GEmm int4.

* Formatting + fix K dimension for int8.

* Batched Gemm int4 example.

* CGEMM int4 example.

* Include inc filese in clang-format.

* SplitK int4 example

* Refactoring of performance measurement.

* Fix #ifdef statements.

Co-authored-by: Adam Osewski <aosewski@amd.com>

example/24_batched_gemm/CMakeLists.txt

@@ -1,4 +1,17 @@
+add_custom_target(example_batched_gemm_xdl)
+
 add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
 add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
 add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
 add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
+
+add_dependencies(example_batched_gemm_xdl
+                 example_batched_gemm_xdl_fp32
+                 example_batched_gemm_xdl_fp16
+                 example_batched_gemm_xdl_bfp16
+                 example_batched_gemm_xdl_int8)
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
+endif()

example/24_batched_gemm/batched_gemm_xdl_int4.cpp

@@ -0,0 +1,99 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = ck::int4_t;
+using BDataType        = ck::int4_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = ck::int4_t;
+
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+using KernelEDataType = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+    // clang-format off
+        < ALayout,              //ALayout
+          BLayout,              //BLayout
+          DsLayout,             //DsLayout
+          ELayout,              //ELayout
+          KernelADataType,      //ADataType    
+          KernelBDataType,      //BDataType   
+          AccDataType,          //AccDataType
+          CShuffleDataType,     //CShuffleDataType
+          DsDataType,           //DsDataType
+          KernelEDataType,      //EDataType
+          AElementOp,           //AElementwiseOperation
+          BElementOp,           //BElementwiseOperation
+          CDEElementOp,         //CDEElementwiseOperation
+          GemmDefault,          //GEMMSpecialization
+          1,                    // NumGemmKPrefetchStage
+          256,                  // BlockSize
+          256,                  // MPerBlock
+          128,                  // NPerBlock
+          64,                   // KPerBlock
+          16,                   // AK1
+          16,                   // BK1
+          32,                   // MPerXdl
+          32,                   // NPerXdl
+          4,                    // MXdlPerWave
+          2,                    // NXdlPerWave
+          S<4, 64, 1>,          // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+          S<1, 0, 2>,           // ABlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // ABlockTransfer SrcAccessOrder
+          2,                    // ABlockTransfer SrcVectorDim
+          16,                   // ABlockTransfer SrcScalarPerVector
+          16,                   // ABlockTransfer DstScalarPerVector_K1
+          1,                    // ABlockLdsExtraM
+          S<4, 64, 1>,          // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+          S<1, 0, 2>,           // BBlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // BBlockTransfer SrcAccessOrder
+          2,                    // BBlockTransfer SrcVectorDim
+          16,                   // BBlockTransfer SrcScalarPerVector
+          16,                   // BBlockTransfer DstScalarPerVector_K1
+          1,                    // BBlockLdsExtraN
+          1,                    // CShuffleMXdlPerWavePerShuffle
+          1,                    // CShuffleNXdlPerWavePerShuffle
+          S<1, 64, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl
+          16>;                  // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+#define BUILD_INT4_EXAMPLE
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }

example/24_batched_gemm/run_batched_gemm_example.inc

@@ -1,3 +1,5 @@
+#include <random>
+
 #pragma once
 
 struct ProblemSize final
@@ -28,7 +30,23 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
 {
     using namespace ck::literals;
 
-    auto& [M, N, K, stride_A, stride_B, stride_C, batch_stride_A, batch_stride_B, batch_stride_C, batch_count] = problem_size;
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType));
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
+    static_assert(sizeof(EDataType) == sizeof(KernelEDataType));
+#endif
+
+    auto& [M,
+           N,
+           K,
+           stride_A,
+           stride_B,
+           stride_C,
+           batch_stride_A,
+           batch_stride_B,
+           batch_stride_C,
+           batch_count] = problem_size;
 
     // GEMM shape
     auto f_host_tensor_descriptor = [](std::size_t batch_count_,
@@ -53,9 +71,13 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
         f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
     Tensor<BDataType> b_g_k_n(
         f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
-
+#ifdef BUILD_INT4_EXAMPLE
+    Tensor<KernelEDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+#else
     Tensor<EDataType> e_g_m_n_device_result(
         f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+#endif
 
     std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
     std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
@@ -78,9 +100,16 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
     DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
     DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
 
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> a_g_m_k_converted(a_g_m_k);
+    const Tensor<KernelBDataType> b_g_k_n_converted(b_g_k_n);
+
+    a_device_buf.ToDevice(a_g_m_k_converted.mData.data());
+    b_device_buf.ToDevice(b_g_k_n_converted.mData.data());
+#else
     a_device_buf.ToDevice(a_g_m_k.mData.data());
     b_device_buf.ToDevice(b_g_k_n.mData.data());
-
+#endif
     auto a_element_op   = AElementOp{};
     auto b_element_op   = BElementOp{};
     auto cde_element_op = CDEElementOp{};
@@ -116,28 +145,21 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
             "not support this GEMM problem");
     }
 
-    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
-
-    std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
-    std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
-                            sizeof(BDataType) * batch_count * K * N +
-                            sizeof(EDataType) * batch_count * M * N;
-
-    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
-
-    float gb_per_sec = num_btype / 1.E6 / ave_time;
-
-    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
-              << gemm.GetTypeString() << std::endl;
-
+    invoker.Run(argument, StreamConfig{nullptr, false});
     bool pass = true;
 
     if(config.do_verification)
     {
         c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
 
-        using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
-        ReferenceBatchedGemm<ADataType, BDataType, EDataType, AccDataType, AElementOp, BElementOp, CDEElementOp>;
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             EDataType,
+                                                             AccDataType,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CDEElementOp>;
 
         auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
         auto ref_invoker      = ref_batched_gemm.MakeInvoker();
@@ -150,8 +172,29 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
 
         ref_invoker.Run(ref_argument);
 
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<EDataType> e_device_result_converted(e_g_m_n_device_result);
+        pass &= ck::utils::check_err(e_device_result_converted.mData, e_g_m_n_host_result.mData);
+
+#else
         pass = ck::utils::check_err(
-            e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
+            e_g_m_n_device_result.mData, e_g_m_n_host_result.mData, "Error: Incorrect results c");
+#endif
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
+        std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
+                                sizeof(BDataType) * batch_count * K * N +
+                                sizeof(EDataType) * batch_count * M * N;
+
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
     }
 
     return pass ? 0 : 1;
@@ -162,9 +205,12 @@ bool run_batched_gemm_example(int argc, char* argv[])
     ProblemSize problem_size;
     ExecutionConfig config;
 
-    problem_size.M = 256 * (rand() % 16 + 1);
-    problem_size.N = 128 * (rand() % 16 + 1);
-    problem_size.K = 64 * (rand() % 16 + 1);
+    std::mt19937 gen(11939);
+    std::uniform_int_distribution<int> dis(0, 15);
+
+    problem_size.M = 256 * (dis(gen) + 1);
+    problem_size.N = 128 * (dis(gen) + 1);
+    problem_size.K = 64 * (dis(gen) + 2);
 
     problem_size.stride_A = problem_size.K;
     problem_size.stride_B = problem_size.K;