修改ck代码适配gfx926

3rdparty/composable_kernel/client_example/README.md

@@ -9,7 +9,10 @@ cd client_example/build
 ```
 
 ```bash
-cmake -D CMAKE_CXX_COMPILER=${ROCM_PATH}/bin/hipcc  -D CMAKE_PREFIX_PATH="${ROCM_PATH};${PATH_TO_CK_INSTALL_DIRECTORY}" ..
+cmake                                                                 \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                             \
+-D CMAKE_PREFIX_PATH="/opt/rocm;${PATH_TO_CK_INSTALL_DIRECTORY}"      \
+..
 ```
 
 ### Build client example

3rdparty/composable_kernel/cmake/googletest.cmake

@@ -27,7 +27,6 @@ message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLA
 FetchContent_Declare(
   googletest
   GIT_REPOSITORY http://10.0.50.24/Mirrors/googletest.git
-# GIT_REPOSITORY /work/home/zhangshao/installer/googletest
   GIT_TAG        b85864c64758dec007208e56af933fc3f52044ee
 )
 

3rdparty/composable_kernel/compile.sh

@@ -7,7 +7,7 @@ cmake
 -D CMAKE_CXX_FLAGS="-O3"                                                                          \
 -D CMAKE_BUILD_TYPE=Release                                                                       \
 -D GPU_TARGETS="gfx906;gfx926"                                                                           \
--D CMAKE_INSTALL_PREFIX=~/composable_kernel/install_ck  \
+-D CMAKE_INSTALL_PREFIX=~/composable_kernel-develop/install_ck  \
 ..
 
 cd -

3rdparty/composable_kernel/include/ck/ck.hpp

@@ -33,7 +33,7 @@
 // buffer resource
 #ifndef __HIP_DEVICE_COMPILE__ // for host code
 #define CK_BUFFER_RESOURCE_3RD_DWORD -1
-#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)||defined(__gfx926__)  || defined(__gfx908__) || \
+#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx926__) || defined(__gfx908__) || \
     defined(__gfx90a__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
 #elif defined(__gfx1030__) // for GPU code
@@ -46,7 +46,7 @@
 #ifndef __HIP_DEVICE_COMPILE__                   // for host code, define nothing
 #elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
 #define CK_USE_AMD_V_MAC_F32
-#elif defined(__gfx906__)|| defined(__gfx926__) || defined(__gfx908__) || defined(__gfx90a__) || \
+#elif defined(__gfx906__) || defined(__gfx926__) || defined(__gfx908__) || defined(__gfx90a__) || \
     defined(__gfx1030__) // for GPU code
 #define CK_USE_AMD_V_FMAC_F32
 #define CK_USE_AMD_V_DOT2_F32_F16

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp

@@ -225,13 +225,13 @@ struct BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_
 
         auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
             a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
-        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
             b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
 
         constexpr auto threadwise_contraction =
             ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
                 FloatA,
-                FloatB,
+                FloatA,
                 FloatC,
                 decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
                 decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
@@ -394,8 +394,8 @@ struct BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_
         Sequence<0, 1, 2, 3>>;                            // SrcVectorTensorContiguousDimOrder
 
     using BThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
-        FloatB,
-        FloatB,
+        FloatB, // src
+        FloatA, // dst
         decltype(b_block_desc_bk0_bn0_bn1_bk1_),
         decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
         Sequence<BK0PerThread, 1, BN1PerThreadBN11, BK1>, // SliceLengths

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp

@@ -134,7 +134,7 @@ __global__ void
             const Block2CTileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__)||defined(__gfx926__)  || defined(__gfx1030__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx926__) || defined(__gfx1030__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -709,7 +709,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
         namespace ctc = tensor_layout::convolution;
 
         // check device
-        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030"))
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
         {
             return false;
         }

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp

@@ -106,7 +106,7 @@ __global__ void
             const Block2CTileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__)|| defined(__gfx926__) || defined(__gfx1030__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx926__) || defined(__gfx1030__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -600,7 +600,7 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
         namespace ctc = tensor_layout::convolution;
 
         // check device
-        if(!(ck::get_device_name() == "gfx906" ||ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030"))
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
         {
             return false;
         }

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp

@@ -1391,7 +1391,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Dl
     static bool IsSupportedArgument(const Argument& arg)
     {
         // check device
-        if(!(ck::get_device_name() == "gfx906"||ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030"))
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030"))
         {
             return false;
         }

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp

@@ -205,6 +205,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
     using GridwiseGemm =
         GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
                                      ADataType,
+                                     BDataType,
                                      AccDataType,
                                      CDataType,
                                      InMemoryDataOperationEnum::Set,
@@ -364,6 +365,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
                 const auto kernel =
                     kernel_gemm_dl_v1r3<GridwiseGemm,
                                         ADataType,
+                                        BDataType,
                                         CDataType,
                                         remove_reference_t<AGridDesc_K0_M0_M1_K1>,
                                         remove_reference_t<BGridDesc_K0_N0_N1_K1>,
@@ -390,6 +392,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
                 const auto kernel =
                     kernel_gemm_dl_v1r3<GridwiseGemm,
                                         ADataType,
+                                        BDataType,
                                         CDataType,
                                         remove_reference_t<AGridDesc_K0_M0_M1_K1>,
                                         remove_reference_t<BGridDesc_K0_N0_N1_K1>,
@@ -416,6 +419,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
                 const auto kernel =
                     kernel_gemm_dl_v1r3<GridwiseGemm,
                                         ADataType,
+                                        BDataType,
                                         CDataType,
                                         remove_reference_t<AGridDesc_K0_M0_M1_K1>,
                                         remove_reference_t<BGridDesc_K0_N0_N1_K1>,
@@ -442,6 +446,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
                 const auto kernel =
                     kernel_gemm_dl_v1r3<GridwiseGemm,
                                         ADataType,
+                                        BDataType,
                                         CDataType,
                                         remove_reference_t<AGridDesc_K0_M0_M1_K1>,
                                         remove_reference_t<BGridDesc_K0_N0_N1_K1>,
@@ -483,7 +488,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::get_device_name() == "gfx906" ||ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030")
+        if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx926" || ck::get_device_name() == "gfx1030")
         {
             return GridwiseGemm::CheckValidity(
                 arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp

@@ -117,18 +117,11 @@ struct GridwiseGemmDlMultipleD_km_kn_mn
 
         // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
 
-        bool ret=(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+        return (M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
                 K0 == b_grid_desc_k0_n_k1.GetLength(I0) &&
                 K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
                 K1 == b_grid_desc_k0_n_k1.GetLength(I2)) &&
                (M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
-        if(!ret){
-            std::cout<<"M="<<M<<" N="<<N<<" K0="<<K0<<" c_grid_desc_m_n[0]="<<c_grid_desc_m_n.GetLength(I0)<<" c_grid_desc_m_n[1]="<<c_grid_desc_m_n.GetLength(I1)
-                <<" b_grid_desc_k0_n_k1[0]="<<b_grid_desc_k0_n_k1.GetLength(I0)<<" b_grid_desc_k0_n_k1[2]="<<b_grid_desc_k0_n_k1.GetLength(I2)
-                <<" a_grid_desc_k0_m_k1[2]="<<a_grid_desc_k0_m_k1.GetLength(I2)
-                <<" K1="<<K1<<" MPerBlock="<<MPerBlock<<" NPerBlock="<<NPerBlock<<" K0PerBlock="<<K0PerBlock<<std::endl;
-        }
-        return ret;
     }
 
     __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp

@@ -18,7 +18,8 @@
 namespace ck {
 
 template <typename GridwiseGemm,
-          typename FloatAB,
+          typename FloatA,
+          typename FloatB,
           typename FloatC,
           typename AGridDesc_K0_M0_M1_K1,
           typename BGridDesc_K0_N0_N1_K1,
@@ -30,23 +31,27 @@ __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_dl_v1r3(const FloatAB* __restrict__ p_a_grid,
-                            const FloatAB* __restrict__ p_b_grid,
+        kernel_gemm_dl_v1r3(const FloatA* __restrict__ p_a_grid,
+                            const FloatB* __restrict__ p_b_grid,
                             FloatC* __restrict__ p_c_grid,
                             const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
                             const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
                             const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
                             const Block2CTileMap block_2_ctile_map)
 {
-    constexpr index_t shared_block_size =
-        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+    constexpr index_t shared_block_size_of_a =
+        GridwiseGemm::GetSharedMemoryNumberOfByteToA() / sizeof(FloatA);
+    constexpr index_t shared_block_size_of_b =
+        GridwiseGemm::GetSharedMemoryNumberOfByteToB() / sizeof(FloatB);
 
-    __shared__ FloatAB p_shared_block[shared_block_size];
+    __shared__ FloatA p_shared_block_a[shared_block_size_of_a];
+    __shared__ FloatB p_shared_block_b[shared_block_size_of_b];
 
     GridwiseGemm::Run(p_a_grid,
                       p_b_grid,
                       p_c_grid,
-                      p_shared_block,
+                      p_shared_block_a,
+                      p_shared_block_b,
                       a_grid_desc_k0_m0_m1_k1,
                       b_grid_desc_k0_n0_n1_k1,
                       c_grid_desc_m0_m10_m11_n0_n10_n11,
@@ -56,7 +61,8 @@ __global__ void
 }
 
 template <index_t BlockSize,
-          typename FloatAB,
+          typename FloatA,
+          typename FloatB,
           typename FloatAcc,
           typename FloatC,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
@@ -99,7 +105,7 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
     // K1 should be Number<...>
     static constexpr auto K1 = Number<K1Value>{};
 
-    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByteToA()
     {
         // TODO: change this. I think it needs multi-dimensional alignment
         constexpr auto max_lds_align = K1;
@@ -122,7 +128,33 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
         constexpr auto b_block_aligned_space_size =
             math::integer_least_multiple(b_block_desc_k_n.GetElementSpaceSize(), max_lds_align);
 
-        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+        return 2 * a_block_aligned_space_size * sizeof(FloatA);
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByteToB()
+    {
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k_m = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k_n = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size =
+            math::integer_least_multiple(a_block_desc_k_m.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size =
+            math::integer_least_multiple(b_block_desc_k_n.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * b_block_aligned_space_size * sizeof(FloatB);
     }
 
     __host__ __device__ static constexpr bool
@@ -145,14 +177,14 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
 
     __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
     {
-        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock); // M0 * N0
 
         return grid_size;
     }
 
     __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
     {
-        const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
+        const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;  // K0 > K0PerBlock ???
 
         return has_main_k_block_loop;
     }
@@ -170,7 +202,7 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
         const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
         const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
 
-        const auto M1 = Number<MPerBlock>{};
+        const auto M1 = Number<MPerBlock>{}; // 128
         const auto M0 = M / M1;
 
         const auto a_grid_desc_k0_m0_m1_k1 =
@@ -178,8 +210,8 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
                                         make_tuple(make_pass_through_transform(K0),
                                                    make_unmerge_transform(make_tuple(M0, M1)),
                                                    make_pass_through_transform(K1)),
-                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
-                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),  // K0, M, K1
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{})); // K0, M0, M1, K1
 
         return a_grid_desc_k0_m0_m1_k1;
     }
@@ -190,7 +222,7 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
         const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
         const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
 
-        const auto N1 = Number<NPerBlock>{};
+        const auto N1 = Number<NPerBlock>{}; // 128
         const auto N0 = N / N1;
 
         const auto b_grid_desc_k0_n0_n1_k1 =
@@ -198,8 +230,8 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
                                         make_tuple(make_pass_through_transform(K0),
                                                    make_unmerge_transform(make_tuple(N0, N1)),
                                                    make_pass_through_transform(K1)),
-                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
-                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), // K0, N, K1
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{})); // K0, N0, N1, K1
 
         return b_grid_desc_k0_n0_n1_k1;
     }
@@ -210,33 +242,33 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
         const auto M = c_grid_desc_m_n.GetLength(I0);
         const auto N = c_grid_desc_m_n.GetLength(I1);
 
-        constexpr auto M1 = Number<MPerBlock>{};
-        constexpr auto N1 = Number<NPerBlock>{};
+        constexpr auto M1 = Number<MPerBlock>{}; // 128
+        constexpr auto N1 = Number<NPerBlock>{}; // 128
 
         const auto M0 = M / M1;
         const auto N0 = N / N1;
 
-        constexpr auto M11 =
-            Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
-                   M1PerThreadM111>{};
-        constexpr auto N11 =
-            Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
-                   N1PerThreadN111>{};
+        constexpr auto M11 =  // 64
+            Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) * // S<8, 2>  ==>  8*2=16
+                   M1PerThreadM111>{}; // M1PerThread 4
+        constexpr auto N11 = // 64
+            Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) * // 16
+                   N1PerThreadN111>{}; // N1PerThread 4
 
-        constexpr auto M10 = M1 / M11;
-        constexpr auto N10 = N1 / N11;
+        constexpr auto M10 = M1 / M11; // 2
+        constexpr auto N10 = N1 / N11; // 2
 
         const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
             c_grid_desc_m_n,
             make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
                        make_unmerge_transform(make_tuple(N0, N10, N11))),
-            make_tuple(Sequence<0>{}, Sequence<1>{}),
-            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+            make_tuple(Sequence<0>{}, Sequence<1>{}), // M, N
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{})); // M0, M10, M11, N0, N10, N11
 
         return c_grid_desc_m0_m10_m11_n0_n10_n11;
     }
 
-    // return block_id to C matrix tile idx (m0, n0) mapping
+    // return block_id to C matrix tile idx (m0, n0) mapping       // what a fuck ???????????? 到底生成了啥
     __host__ __device__ static constexpr auto
     MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
     {
@@ -252,10 +284,11 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
 
     template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
     __device__ static void
-    Run(const FloatAB* __restrict__ p_a_grid,
-        const FloatAB* __restrict__ p_b_grid,
+    Run(const FloatA* __restrict__ p_a_grid,
+        const FloatB* __restrict__ p_b_grid,
         FloatC* __restrict__ p_c_grid,
-        FloatAB* __restrict__ p_shared_block,
+        FloatA* __restrict__ p_shared_block_a,
+        FloatB* __restrict__ p_shared_block_b,
         const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1,
         const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1,
         const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
@@ -304,12 +337,12 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
         // TODO: check alignment
         // A matrix in LDS memory, for blockwise GEMM
         constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
-            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align); // (16, 128, 2), 2
 
         // TODO: check alignment
         // B matrix in LDS memory, for blockwise GEMM
         constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
-            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align); // (16, 128, 2), 2
 
         static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
                           a_k0_m_k1_block_desc.GetElementSpaceSize() &&
@@ -325,8 +358,8 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
             ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
             ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
             ABlockTransferThreadClusterArrangeOrder,
-            FloatAB,
-            FloatAB,
+            FloatA,
+            FloatA,
             remove_reference_t<decltype(a_grid_desc_k0_m0_m1_k1)>,
             decltype(a_block_desc_k0_m0_m1_k1),
             ABlockTransferSrcAccessOrder,
@@ -349,8 +382,8 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
             BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
             BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
             BBlockTransferThreadClusterArrangeOrder,
-            FloatAB,
-            FloatAB,
+            FloatB,
+            FloatB,
             remove_reference_t<decltype(b_grid_desc_k0_n0_n1_k1)>,
             decltype(b_block_desc_k0_n0_n1_k1),
             BBlockTransferSrcAccessOrder,
@@ -368,14 +401,14 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
         // GEMM definition
         //   c_mtx += transpose(a_mtx) * b_mtx
         //     a_mtx[K0PerBlock, MPerBlock] is in LDS
-        //     b_mtx[KPerBlocl, NPerBlock] is in LDS
+        //     b_mtx[K0PerBlocl, NPerBlock] is in LDS
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         const auto blockwise_gemm =
             BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
                 BlockSize,
-                FloatAB,
-                FloatAB,
+                FloatA,  // todo split a/b
+                FloatB,
                 FloatAcc,
                 decltype(a_k0_m_k1_block_desc),
                 decltype(b_k0_n_k1_block_desc),
@@ -400,8 +433,8 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
         constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
             b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
 
-        FloatAB* p_a_block_double = p_shared_block;
-        FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
+        FloatA* p_a_block_double = p_shared_block_a;
+        FloatB* p_b_block_double = p_shared_block_b;
 
         // register allocation for output
         auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
@@ -436,7 +469,7 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
 
         if constexpr(HasMainKBlockLoop)
         {
-            const auto K0 = a_grid_desc_k0_m0_m1_k1.GetLength(I0);
+            const auto K0 = a_grid_desc_k0_m0_m1_k1.GetLength(I0);  // K / K1(=2)
 
             index_t k_block_data_begin = 0;
 
@@ -487,7 +520,7 @@ struct GridwiseGemmDl_km_kn_mn_v1r3
                 b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
 
                 k_block_data_begin += 2 * K0PerBlock;
-            } while(k_block_data_begin < K0 - 2 * K0PerBlock);
+            } while(k_block_data_begin < K0 - 2 * K0PerBlock);  // K0PerBlock = 16
         }
 
         // LDS double buffer: tail

3rdparty/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp

@@ -151,7 +151,7 @@ struct ThreadwiseTensorSliceTransfer_v4r1
                     dst_origin_idx + data_to_origin_disp_idx + src_vector_idx);
 
                 dst_buf(Number<dst_offset>{}) = type_convert<DstData>(
-                    src_vector.template AsType<DstData>()[Number<src_vector_offset>{}]);
+                    src_vector.template AsType<SrcData>()[Number<src_vector_offset>{}]);
             });
         });
     }

3rdparty/composable_kernel/include/ck/utility/data_type.hpp

@@ -942,6 +942,10 @@ using int8x16_t = typename vector_type<int8_t, 16>::type;
 using int8x32_t = typename vector_type<int8_t, 32>::type;
 using int8x64_t = typename vector_type<int8_t, 64>::type;
 
+// u8
+using uint8x2_t  = typename vector_type<uint8_t, 2>::type;
+using uint8x4_t  = typename vector_type<uint8_t, 4>::type;
+
 // Convert X to Y
 template <typename Y, typename X>
 __host__ __device__ constexpr Y type_convert(X x)
@@ -951,6 +955,13 @@ __host__ __device__ constexpr Y type_convert(X x)
     return static_cast<Y>(x);
 }
 
+// Convert X to Y
+template <>
+__host__ __device__ constexpr half_t type_convert<half_t, uint8_t>(uint8_t x)
+{
+    return static_cast<half_t>(x);
+}
+
 // convert bfp16 to fp32
 template <>
 inline __host__ __device__ constexpr float type_convert<float, bhalf_t>(bhalf_t x)

3rdparty/composable_kernel/include/ck/utility/inner_product.hpp

@@ -9,6 +9,31 @@ namespace ck {
 template <typename TA, typename TB, typename TC>
 __device__ void inner_product(const TA& a, const TB& b, TC& c);
 
+template <>
+__device__ void inner_product<half2_t, uint8x2_t, float>(const half2_t& a, const uint8x2_t& b, float& c)
+{
+    const vector_type<half_t, 2> a_vector{a};
+    const vector_type<uint8_t, 2> b_vector{b};
+    const vector_type<half_t, 2> b_fp16_vector;
+
+    static constexpr uint32_t mask_for_elt_01     = 0x05020500;
+    static constexpr uint32_t mask_for_elt_23     = 0x05030501;
+    static constexpr uint32_t start_byte_for_fp16 = 0x64646464;
+
+    asm volatile("v_perm_b32 %0,%1,%2,%3;\n" : "=v"(((uint32_t*)&b_fp16_vector.data_)[0]) : "v"(start_byte_for_fp16), "v"(((uint32_t*)&b_vector.data_)[0]), "v"(mask_for_elt_01));
+    // asm volatile("v_perm_b32 %0,%1,%2,%3;\n" : "=v"(((uint32_t*)&b_fp16_vector.data_)[1]) : "v"(start_byte_for_fp16), "v"(((uint32_t*)&b_vector.data_)[0]), "v"(mask_for_elt_23));
+
+    static constexpr uint32_t I8s_TO_F16s_MAGIC_NUM = 0x6480;
+    asm volatile("v_sub_f16x2 %0, %1, %2;\n" : "=v"(((uint32_t*)&b_fp16_vector.data_)[0]) : "v"(((uint32_t*)&b_fp16_vector.data_)[0]), "v"(I8s_TO_F16s_MAGIC_NUM));
+    // asm volatile("v_sub_f16x2 %0, %1, %2;\n" : "=v"(((uint32_t*)&b_fp16_vector.data_)[1]) : "v"(((uint32_t*)&b_fp16_vector.data_)[1]), "v"(I8s_TO_F16s_MAGIC_NUM));
+
+    static_for<0, 2, 1>{}([&](auto i) {
+        c += type_convert<int32_t>(a_vector.AsType<half_t>()[i]) *
+             type_convert<int32_t>(b_fp16_vector.AsType<half_t>()[i]);
+    });
+    
+}
+
 template <>
 __device__ void inner_product<float, float, float>(const float& a, const float& b, float& c)
 {
@@ -71,12 +96,6 @@ inner_product<float4_t, float4_t, float>(const float4_t& a, const float4_t& b, f
                   c);
 }
 
-template <>
-__device__ void inner_product<half_t, half_t, float>(const half_t& a, const half_t& b, float& c)
-{
-  c+=static_cast<float>(a*b);
-}
-
 template <>
 __device__ void inner_product<half2_t, half2_t, float>(const half2_t& a, const half2_t& b, float& c)
 {

3rdparty/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp

@@ -15,6 +15,7 @@ namespace device {
 namespace instance {
 
 using F16 = ck::half_t;
+using I8 = int8_t;
 using F32 = float;
 
 using Row = ck::tensor_layout::gemm::RowMajor;
@@ -34,13 +35,13 @@ using device_gemm_dl_f16_f16_f16_km_kn_mn_instances = std::tuple<
         // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
         // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
         // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
-        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        DeviceGemmDl<   F16,   I8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
     // clang-format on
     >;
 
 void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
-        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        DeviceGemm<Col, Row, Row, F16, I8, F16, PassThrough, PassThrough, PassThrough>>>&
         instances)
 {
     add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_km_kn_mn_instances{});