[Navi3x-LWPCK-545] Block-wise GEMM + Real GEMM_WMMA_FP16 (#541)

* wmma_op + unit test

* add arch limitation to wmma test

* change arch limitation

* Refactor + Add all type unit test(int4 compile failed)

* Add f32_16x16x16_bf16 unit test

* tempsave

* tempsave

* tempsave

* runtime bug, cannot find symbol

* workaround for incorrect HIP warpSize return value

* debugging

* tempsave

* Correctness OK, waiting for optimization

* Tidy up + format

* temp save

* temp save, reproduce the v_bfi_b32 issue

* add inline asm for wmmaop test

* tidy up

* clean some debug purpose code

* discard some codes

* clang format

* clang format

* compiler issue fixed + increase tile size

example/01_gemm/CMakeLists.txt

@@ -35,3 +35,8 @@ add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
 
 add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
 add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
+
+add_custom_target(example_gemm_wmma)
+add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
+add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
+

example/01_gemm/gemm_wmma_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using CDataType        = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer|MRepeat|NRepeat|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    | WMMA| WMMA|       |       |   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN|MWmmaPerWave|NWmmaPerWave|        _MBlock_MWaveMPerWmma| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     |       |       | Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|        _NBlock_NWaveNPerWmma|  _NWaveNPerWmma|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |      |      |      |      |    |     |     |       |       |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   256,     8,   8,   16,   16,      4,      4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,              S<1, 32, 1,  8>,               8, 1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/amd_wmma.hpp"
+
+namespace ck {
+
+enum struct WmmaInstr
+{
+    wmma_f32_16x16x16_f16 = 0,
+    wmma_f32_16x16x16_bf16,
+    wmma_f16_16x16x16_f16,
+    wmma_bf16_16x16x16_bf16,
+    wmma_i32_16x16x16_iu8,
+    wmma_i32_16x16x16_iu4
+};
+
+/*
+ *  WMMA Wave Tile Always MxNxK = 16x16x16
+ *  WAVE32
+        -----------------------------------
+        |RC0| | | | | | | | | | | | | | | |	   SubGroup 0
+        |RC1| | | | | | | | | | | | | | | |
+        |RC2| | | | | | | | | | | | | | | |
+        |RC3|T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|
+        |RC4|0|0|0|0|0|0|0|0|0|1|1|1|1|1|1|
+        |RC5|1|2|3|4|5|6|7|8|9|0|1|2|3|4|5|
+        |RC6| | | | | | | | | | | | | | | |
+        |RC7| | | | | | | | | | | | | | | |
+        -----------------------------------
+        |   | | | | | | | | | | | | | | | |	   SubGroup 1
+        |   | | | | | | | | | | | | | | | |
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|
+        | 1 |1|1|1|2|2|2|2|2|2|2|2|2|2|3|3|
+        | 6 |7|8|9|0|1|2|3|4|5|6|7|8|9|0|1|
+        |   | | | | | | | | | | | | | | | |
+        |   | | | | | | | | | | | | | | | |
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+
+
+ *  WAVE64
+        -----------------------------------
+        |RC0|T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|	   SubGroup 0
+        |RC1|0|0|0|0|0|0|0|0|0|1|1|1|1|1|1|
+        |RC2|1|2|3|4|5|6|7|8|9|0|1|2|3|4|5|
+        |RC3|T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|
+        -----------------------------------
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|    SubGroup 1
+        | 1 |1|1|1|2|2|2|2|2|2|2|2|2|2|3|3|
+        | 6 |7|8|9|0|1|2|3|4|5|6|7|8|9|0|1|
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|	   SubGroup 2
+        | 3 |3|3|3|3|3|3|3|4|4|4|4|4|4|4|4|
+        | 2 |3|4|5|6|7|8|9|0|1|2|3|4|5|6|7|
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|    SubGroup 3
+        | 4 |4|5|5|5|5|5|5|5|5|5|5|6|6|6|6|
+        | 8 |9|0|1|2|3|4|5|6|7|8|9|0|1|2|3|
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+
+*   RC = Register for storing accumalted result
+*	T  = Thread ID
+*/
+
+template <WmmaInstr Instr, index_t WaveSize, typename = void>
+struct wmma_type
+{
+};
+
+// A-swizzled
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    // * Data Pixel
+    static constexpr index_t m_per_wmma      = 16;
+    static constexpr index_t n_per_wmma      = 16;
+    static constexpr index_t k_per_wmma      = 16;
+    static constexpr index_t src_a_data_size = 2;
+    static constexpr index_t src_b_data_size = 2;
+    static constexpr index_t acc_data_size   = 4;
+    // * Thread mapping inside wave, num_thread_per_subgroups always alone N direction
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size = Number<WaveSize>{};
+    // * Fixed in Navi3x, Will be wave mode dependent on Navi4x
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    // * num_acc_vgprs_per_wave alone M direction
+    // * num_subgroups alone M direction
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma, index_t NPerWmma, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_f32_16x16x16_f16_w32<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_f32_16x16x16_f16_w64<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+    }
+};
+
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_f32_16x16x16_bf16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 4;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma, index_t NPerWmma, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_f32_16x16x16_bf16_w32<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_f32_16x16x16_bf16_w64<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+    }
+};
+
+#ifdef CK_UNPACKED_ACC_DESC_LOGIC
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_f16_16x16x16_f16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 2;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma,
+              index_t NPerWmma,
+              index_t Opsel,
+              class FloatA,
+              class FloatB,
+              class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_f16_16x16x16_f16_w32<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_f16_16x16x16_f16_w64<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+    }
+};
+
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_bf16_16x16x16_bf16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 2;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma,
+              index_t NPerWmma,
+              index_t Opsel,
+              class FloatA,
+              class FloatB,
+              class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_bf16_16x16x16_bf16_w32<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_bf16_16x16x16_bf16_w64<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+    }
+};
+
+#endif
+
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_i32_16x16x16_iu8,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 4;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma,
+              index_t NPerWmma,
+              bool neg_a,
+              bool neg_b,
+              bool clamp,
+              class FloatA,
+              class FloatB,
+              class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_i32_16x16x16_iu8_w32<MPerWmma, NPerWmma, neg_a, neg_b, clamp>::Run(
+                a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_i32_16x16x16_iu8_w64<MPerWmma, NPerWmma, neg_a, neg_b, clamp>::Run(
+                a, b, reg_c);
+        }
+    }
+};
+
+template <typename src_type_a,
+          typename src_type_b,
+          typename dst_type,
+          index_t MPerWmma,
+          index_t NPerWmma>
+struct WmmaSelector
+{
+    template <typename src_type_a_,
+              typename src_type_b_,
+              typename dst_type_,
+              index_t MPerWmma_,
+              index_t NPerWmma_>
+    static constexpr auto GetWmma();
+
+    template <>
+    static constexpr auto GetWmma<half_t, half_t, float, 16, 16>()
+    {
+        return WmmaInstr::wmma_f32_16x16x16_f16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<bhalf_t, bhalf_t, float, 16, 16>()
+    {
+        return WmmaInstr::wmma_f32_16x16x16_bf16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<half_t, half_t, half_t, 16, 16>()
+    {
+        return WmmaInstr::wmma_f16_16x16x16_f16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<bhalf_t, bhalf_t, bhalf_t, 16, 16>()
+    {
+        return WmmaInstr::wmma_bf16_16x16x16_bf16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<int8_t, int8_t, int, 16, 16>()
+    {
+        return WmmaInstr::wmma_i32_16x16x16_iu8;
+    }
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    template <>
+    static constexpr auto GetWmma<int4_t, int, 16, 16>()
+    {
+        return WmmaInstr::wmma_i32_16x16x16_iu4;
+    }
+#endif
+    // get_warp_size do not return the correct wavesize, hardcode to 32 as workaround
+    static constexpr auto selected_wmma =
+        wmma_type<GetWmma<src_type_a, src_type_b, dst_type, MPerWmma, NPerWmma>(), Number<32>{}>{};
+
+    __host__ __device__ constexpr WmmaSelector()
+    {
+        static_assert(selected_wmma.m_per_wmma == 16, "WRONG! WMMA_M must equal to 16");
+
+        static_assert(selected_wmma.m_per_wmma == 16, "WRONG! WMMA_M must equal to 16");
+
+        static_assert(selected_wmma.k_per_wmma == 16, "WRONG! WMMA_M must equal to 16");
+
+        static_assert(selected_wmma.wave_size * selected_wmma.num_acc_vgprs_per_wave *
+                              selected_wmma.acc_data_size ==
+                          selected_wmma.m_per_wmma * selected_wmma.n_per_wmma * 4,
+                      "WRONG! Invalid Number of Accumulator Register");
+    }
+};
+
+template <typename src_type_a,
+          typename src_type_b,
+          typename dst_type,
+          index_t MPerWmma,
+          index_t NPerWmma,
+          index_t KPack,
+          bool TransposeC = false>
+struct WmmaGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    using CIndex   = MultiIndex<2>;
+    using CIndex4D = MultiIndex<4>;
+
+    __host__ __device__ constexpr WmmaGemm()
+    {
+        static_assert(NPerWmma == 16 && MPerWmma == 16,
+                      "Only support GemmNPerWmma == 16 and GemmMPerWmma == 16 for wmma");
+
+        static_assert(KPack == wmma_instr.k_per_wmma, "KPack should be k_per_wmma");
+    }
+
+    // WMMA output supporting C = A * B
+    // Vector Write
+    // MPerWMMA_NPerWMMA -> MSubGroup_..._NPerWMMA_MAccVgprPerWave
+    template <typename CDesc_MBlockxRepeat_MWave_MPerWMMA_NBlockxRepeat_NWave_NPerWMMA>
+    __host__ __device__ static constexpr auto
+    MakeCDesc_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
+        const CDesc_MBlockxRepeat_MWave_MPerWMMA_NBlockxRepeat_NWave_NPerWMMA&
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma)
+    {
+        const auto MBlockxRepeat =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I0);
+        const auto NBlockxRepeat =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I3);
+        const auto MWave =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I1);
+        const auto NWave =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I4);
+
+        return transform_tensor_descriptor(
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma,
+            make_tuple(
+                make_pass_through_transform(MBlockxRepeat),
+                make_pass_through_transform(MWave),
+                make_unmerge_transform(make_tuple(Number<wmma_instr.num_subgroups>{},
+                                                  Number<wmma_instr.num_acc_vgprs_per_wave>{})),
+                make_pass_through_transform(NBlockxRepeat),
+                make_pass_through_transform(NWave),
+                make_pass_through_transform(Number<wmma_instr.num_thread_per_subgroups>{})),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2, 6>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}));
+    }
+
+    __device__ static constexpr index_t GetRegSizePerWmma()
+    {
+        return wmma_instr.num_acc_vgprs_per_wave;
+    }
+
+    __device__ static constexpr index_t GetWaveSize() { return wmma_instr.wave_size; }
+
+    template <class FloatA, class FloatB, class FloatC>
+    __device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
+    {
+        static_assert(
+            (is_same<src_type_a, half_t>::value && is_same<src_type_b, half_t>::value &&
+             is_same<dst_type, float>::value) ||
+                (is_same<src_type_a, bhalf_t>::value && is_same<src_type_b, bhalf_t>::value &&
+                 is_same<dst_type, float>::value) ||
+                (is_same<src_type_a, half_t>::value && is_same<src_type_b, half_t>::value &&
+                 is_same<dst_type, half_t>::value) ||
+                (is_same<src_type_a, bhalf_t>::value && is_same<src_type_b, bhalf_t>::value &&
+                 is_same<dst_type, bhalf_t>::value) ||
+                (is_same<src_type_a, int8_t>::value && is_same<src_type_b, int8_t>::value &&
+                 is_same<dst_type, int32_t>::value)
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+                || (is_same<src_type_a, int4_t>::value && is_same<src_type_b, int4_t>::value &&
+                    is_same<dst_type, int32_t>::value)
+#endif
+                ,
+            "base type couple must be (half, float), (bhalf, float), (half, half), (bhalf, bhalf), "
+            "(int8, int32) or (int4, int32)!");
+        if constexpr(!TransposeC)
+        {
+            wmma_instr.template run<MPerWmma, NPerWmma>(p_a_wave, p_b_wave, p_c_thread);
+        }
+        else
+        {
+            wmma_instr.template run<MPerWmma, NPerWmma>(p_b_wave, p_a_wave, p_c_thread);
+        }
+    }
+
+    __device__ static auto GetLaneId() { return get_thread_local_1d_id() % wmma_instr.wave_size; }
+
+    __device__ static auto GetSubGroupId()
+    {
+        return (GetLaneId() / wmma_instr.num_thread_per_subgroups) % wmma_instr.num_subgroups;
+    }
+
+    __device__ static auto GetLaneIdUnderSubGroup()
+    {
+        return GetLaneId() % wmma_instr.num_thread_per_subgroups;
+    }
+    __device__ static auto GetSwizzledLaneIdLow()
+    {
+        return ((GetLaneIdUnderSubGroup() & 1) << 3) | (GetLaneIdUnderSubGroup() >> 1);
+    }
+
+    __host__ __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        return GetSwizzledLaneIdLow();
+    }
+
+    __host__ __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        return GetLaneIdUnderSubGroup();
+    }
+
+    __device__ static CIndex GetBeginOfThreadBlk()
+    {
+        index_t n_offset = GetLaneIdUnderSubGroup();
+        index_t m_offset = GetSubGroupId() * wmma_instr.num_acc_vgprs_per_wave;
+
+        return TransposeC ? CIndex{n_offset, m_offset} : CIndex{m_offset, n_offset};
+    }
+
+    static constexpr auto wmma =
+        WmmaSelector<src_type_a, src_type_b, dst_type, MPerWmma, NPerWmma>{};
+    static constexpr auto wmma_instr = wmma.selected_wmma;
+
+    __host__ __device__ static constexpr auto
+    GetCMSubGroupNThreadPerSubGroupMAccVgprsThreadBlkLengths()
+    {
+        return make_tuple(I1, I1, Number<wmma_instr.num_acc_vgprs_per_wave>{});
+    }
+};
+
+} // namespace ck

include/ck/utility/amd_inline_asm.hpp

@@ -355,5 +355,11 @@ __device__ void amd_assembly_outer_product_1x4(int8x16_t a,
                                    c3);
 }
 
+// Ranged input operand
+__device__ void amd_assembly_wmma_f32_16x16x16_f16_w32(half16_t a, half16_t b, float8_t& c)
+{
+    asm volatile("v_wmma_f32_16x16x16_f16 %0, %1, %2, %0" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+
 } // namespace ck
 #endif

include/ck/utility/amd_wmma.hpp

@@ -4,11 +4,13 @@
 #ifndef CK_AMD_WMMA_HPP
 #define CK_AMD_WMMA_HPP
 
+#include "ck/utility/amd_inline_asm.hpp"
 #include "data_type.hpp"
 // TODO: Add arch limitation
 namespace ck {
 
-// wave32 only
+/********************************WAVE32 MODE***********************************************/
+
 // src: fp16, dst: fp32
 template <index_t MPerWave, index_t NPerWave>
 struct intrin_wmma_f32_16x16x16_f16_w32;
@@ -19,8 +21,13 @@ struct intrin_wmma_f32_16x16x16_f16_w32<16, 16>
     template <class FloatC>
     __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
     {
-        reg_c.template AsType<float8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(
-            reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
+        // * Inline assembly need to elimate the duplicated data load, compiler won't help you
+        // delete them.
+        amd_assembly_wmma_f32_16x16x16_f16_w32(
+            reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
+        // reg_c.template AsType<float8_t>()(Number<0>{}) =
+        // __builtin_amdgcn_wmma_f32_16x16x16_f16_w32( reg_a, reg_b, reg_c.template
+        // AsType<float8_t>()[Number<0>{}]);
     }
 };
 
@@ -98,5 +105,95 @@ struct intrin_wmma_i32_16x16x16_iu8_w32<16, 16, neg_a, neg_b, clamp>
     }
 };
 
+/********************************WAVE64 MODE***********************************************/
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_f16_w64;
+
+template <>
+struct intrin_wmma_f32_16x16x16_f16_w64<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w64(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}]);
+    }
+};
+
+// src: bf16, dst: fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_bf16_w64;
+
+template <>
+struct intrin_wmma_f32_16x16x16_bf16_w64<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_f32_16x16x16_bf16_w64(
+                reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}]);
+    }
+};
+
+// src: fp16, dst: fp16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w64;
+
+template <index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w64<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+        reg_c.template AsType<half8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f16_16x16x16_f16_w64(
+            reg_a, reg_b, reg_c.template AsType<half8_t>()[Number<0>{}], Opsel);
+    }
+};
+
+// src: bf16, dst: bf16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w64;
+
+template <index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w64<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+        reg_c.template AsType<bhalf8_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64(
+                reg_a, reg_b, reg_c.template AsType<bhalf8_t>()[Number<0>{}], Opsel);
+    }
+};
+
+// src: iu8, dst: i32
+template <index_t MPerWave, index_t NPerWave, bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w64;
+
+template <bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w64<16, 16, neg_a, neg_b, clamp>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x16_t& reg_a, const int8x16_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x4_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_i32_16x16x16_iu8_w64(
+                neg_a,
+                bit_cast<int32x4_t>(reg_a),
+                neg_b,
+                bit_cast<int32x4_t>(reg_b),
+                reg_c.template AsType<int32x4_t>()[Number<0>{}],
+                clamp);
+    }
+};
+
 } // namespace ck
 #endif

test/wmma_op/wmma_op_util.hpp

@@ -97,6 +97,7 @@ builtin_wmma_naive_selector<int4x16_t,
 template <typename src_t, typename dst_t, typename acc_t, index_t acc_num>
 __global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
 {
+    __shared__ src_t p_shared[16 * 16 * 2];
     const int lIdx = threadIdx.x;
     // a and b fragments are stored in 8 VGPRs each, in packed format, so 16 elements each for a and
     // b a_frag will store one column of the 16x16 matrix tile b_frag will store one row of the
@@ -104,6 +105,9 @@ __global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
     using src_vec  = typename vector_type<src_t, 16>::type;
     src_vec a_frag = {};
     src_vec b_frag = {};
+
+    src_vec a_temp = {};
+    src_vec b_temp = {};
     // initialize c fragment to 0
     using acc_vec = StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, acc_t, 1, acc_num, true>;
     acc_vec c_thread_buf_;
@@ -111,21 +115,57 @@ __global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
     // lane is (0-31) mod 16 instead of 0-31 due to matrix replication in gfx11
     // see https://atlvsp3.amd.com/sp3_gfx11_5_instructions.pdf page 482
     // TODO: remove this dependency in gfx12 https://ontrack-internal.amd.com/browse/DEGFXSP3-101
-    const int lane = lIdx % 16;
+    const int lane    = lIdx % 16;
+    const int lane_lo = lIdx / 2;
+    const int lane_hi = lIdx % 2;
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        a_temp[ele] = a[8 * lane_hi + 16 * lane_lo + ele];
+    }
+
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        b_temp[ele] = b[8 * lane_hi + 16 * lane_lo + ele];
+    }
+
+    __syncthreads();
+
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        p_shared[8 * 16 * lane_hi + 8 * lane_lo + ele] = a_temp[ele];
+    }
+
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        p_shared[8 * 16 * lane_hi + 8 * lane_lo + ele + 16 * 16] = b_temp[ele];
+    }
+
+    asm volatile("\
+    s_waitcnt lgkmcnt(0) \n \
+    s_barrier \
+    " ::);
 
     for(int ele = 0; ele < 16; ++ele)
     {
-        b_frag[ele] = b[16 * lane + ele];
+        b_frag[ele] = p_shared[(ele / 8) * 16 * 8 + 8 * lane + ele % 8 + 16 * 16];
     }
     // follow origin design
     for(int ele = 0; ele < 16; ++ele)
     {
-        a_frag[ele] = a[16 * lane + ele];
+        a_frag[ele] = p_shared[(ele / 8) * 16 * 8 + 8 * lane + ele % 8];
     }
 
+    asm volatile("\
+    s_waitcnt lgkmcnt(0) \n \
+    s_barrier \
+    " ::);
+
     // sync threads, similar to mma_sync
-    __syncthreads();
+    // __syncthreads();
     builtin_wmma_naive_selector<src_vec, acc_vec>(a_frag, b_frag, c_thread_buf_);
+    // since only fp16_fp32 asm wmma implemented for experiment purpose, restrict test case to fp16
+    // when enable this ck::amd_assembly_wmma_f32_16x16x16_f16_w32(a_frag, b_frag,
+    // c_thread_buf_.GetVectorTypeReference(Number<0>{}).template AsType<float8_t>()(Number<0>{}));
     __syncthreads();
     // wait for results, similar to mma_sync
     static_for<0, 8, 1>{}([&](auto ele) {