Merge branch 'develop' into gemm_elementwise_gemm

profiler/include/profiler/profile_pool3d_fwd_impl.hpp

@@ -102,11 +102,22 @@ bool profile_pool3d_fwd_impl(PoolFwdInputParams& in_params, PoolFwdKernelParams&
     Tensor<IndexDataType> out_indices_n_c_do_ho_wo_device(
         f_host_tensor_descriptor(N, C, Do, Ho, Wo));
 
+    constexpr int inDataRangeTensor1{1};
+    constexpr int inDataRangeTensor2{5};
+    constexpr double inDataRangeTensor3{0.5};
+
     switch(in_params.init_method)
     {
-    case 0: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
-    case 1: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
-    default: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    case 0:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{inDataRangeTensor1});
+        break;
+    case 1:
+        in_n_c_di_hi_wi.GenerateTensorValue(
+            GeneratorTensor_2<InDataType>{-inDataRangeTensor2, inDataRangeTensor2});
+        break;
+    default:
+        in_n_c_di_hi_wi.GenerateTensorValue(
+            GeneratorTensor_3<InDataType>{-inDataRangeTensor3, inDataRangeTensor3});
     }
 
     DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_di_hi_wi.mDesc.GetElementSpaceSize());
@@ -229,12 +240,39 @@ bool profile_pool3d_fwd_impl(PoolFwdInputParams& in_params, PoolFwdKernelParams&
         {
             out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());
 
-            auto tolerance = 1e-3;
-            bool pass      = ck::utils::check_err(out_n_c_do_ho_wo_device.mData,
+            auto number_of_accumulations = 1;
+            static_assert(
+                ReduceOpId == ck::ReduceTensorOp::AVG || ReduceOpId == ck::ReduceTensorOp::MAX,
+                "Warning: Unhandled ReduceOpId for setting up the number of accumulations!");
+
+            if constexpr(ReduceOpId == ck::ReduceTensorOp::AVG)
+            {
+                for(size_t i = 0; i < kernel_params.window_spatial_lengths.size(); ++i)
+                {
+                    number_of_accumulations *= kernel_params.window_spatial_lengths.at(i);
+                }
+            }
+
+            auto absolute_error_threshold = 1.0;
+            switch(in_params.init_method)
+            {
+            case 0: absolute_error_threshold = static_cast<double>(inDataRangeTensor1); break;
+            case 1: absolute_error_threshold = static_cast<double>(inDataRangeTensor2); break;
+            default: absolute_error_threshold = inDataRangeTensor3;
+            }
+
+            absolute_error_threshold =
+                ck::utils::get_absolute_threshold<ComputeDataType, OutDataType>(
+                    absolute_error_threshold, number_of_accumulations);
+            auto relative_error_threshold =
+                ck::utils::get_relative_threshold<ComputeDataType, OutDataType>(
+                    number_of_accumulations);
+
+            bool pass = ck::utils::check_err(out_n_c_do_ho_wo_device.mData,
                                              out_n_c_do_ho_wo_host.mData,
                                              "Error: Incorrect results",
-                                             tolerance,
-                                             tolerance);
+                                             relative_error_threshold,
+                                             absolute_error_threshold);
 
             if constexpr(OutputIndex)
             {

profiler/src/CMakeLists.txt

@@ -59,6 +59,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
   list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
   list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
@@ -141,6 +142,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   endif()
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)

profiler/src/profile_gemm_multiply_multiply.cpp

@@ -27,6 +27,7 @@ enum struct GemmDataType
     F16_F8_F16,     // 5
     F16_F16_F16_F8, // 6
     F8_F8_BF16,     // 7
+    INT8_INT8_BF16, // 8
 };
 
 #define OP_NAME "gemm_multiply_multiply"
@@ -39,7 +40,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
         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->f8; 7: f8->bf16, "
-               "comp f8)\n");
+               "comp f8; 8: int8->bf16)\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");
@@ -89,6 +90,8 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
     using F32  = float;
     using BF16 = ck::bhalf_t;
     using F8   = ck::f8_t;
+    using I8   = int8_t;
+    using I32  = int;
 
     using Row = ck::tensor_layout::gemm::RowMajor;
     using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -162,6 +165,11 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
         return profile(
             F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
     }
+    else if(data_type == GemmDataType::INT8_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(
+            I8{}, I8{}, I8{}, I32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
+    }
     else
     {
         std::cout << "this data_type & layout is not implemented" << std::endl;

profiler/src/profile_gemm_universal.cpp

@@ -57,6 +57,25 @@ int profile_gemm_universal(int argc, char* argv[])
         exit(1);
     }
 
+    int M;
+    int N;
+    int StrideA;
+    int StrideB;
+    // Analyze the unsupported matrix shapes, switch the M and N number
+    if(std::stoi(argv[9]) % 8 != 0 && std::stoi(argv[8]) % 8 == 0)
+    {
+        M       = std::stoi(argv[9]);
+        StrideA = std::stoi(argv[12]);
+        N       = std::stoi(argv[8]);
+        StrideB = std::stoi(argv[11]);
+    }
+    else
+    {
+        M       = std::stoi(argv[8]);
+        StrideA = std::stoi(argv[11]);
+        N       = std::stoi(argv[9]);
+        StrideB = std::stoi(argv[12]);
+    }
     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]);
@@ -64,12 +83,8 @@ int profile_gemm_universal(int argc, char* argv[])
     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]);
 
@@ -86,7 +101,9 @@ int profile_gemm_universal(int argc, char* argv[])
     using F32  = float;
     using F16  = ck::half_t;
     using BF16 = ck::bhalf_t;
-    using F8   = ck::f8_t;
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
+    using F8 = ck::f8_t;
+#endif
 
     using Row = ck::tensor_layout::gemm::RowMajor;
     using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -147,6 +164,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(F16{}, F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
     }
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
     else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         return profile(F16{}, F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
@@ -163,6 +181,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(F8{}, F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
     }
+#endif
     else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         return profile(BF16{}, BF16{}, BF16{}, F32{}, BF16{}, Row{}, Row{}, Row{});
@@ -179,6 +198,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(BF16{}, BF16{}, BF16{}, F32{}, BF16{}, Col{}, Row{}, Row{});
     }
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
     else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Row{}, Row{});
@@ -187,6 +207,7 @@ int profile_gemm_universal(int argc, char* argv[])
     {
         return profile(F8{}, F8{}, F8{}, F32{}, BF16{}, Row{}, Col{}, Row{});
     }
+#endif
     else
     {
         std::cout << "this data_type & layout is not implemented" << std::endl;

profiler/src/profile_gemm_universal_batched.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdint>
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_universal_batched_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_universal_batched.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    BF16_BF16_BF16, // 0
+    F8_F8_BF16,     // 1
+};
+
+#define OP_NAME "gemm_universal_batched"
+#define OP_DESC "Batched GEMM Universal"
+
+int profile_batched_gemm_universal(int argc, char* argv[])
+{
+    if(argc != 18 && argc != 21)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: bf16, 1: fp8->bf16)\n");
+        printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
+        printf("                     2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     3: A[g, k, m] * B[g, n, k] = C[g, 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=n0, 1=yes)\n");
+        printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
+        printf("optional:\n");
+        printf("arg18: number of warm-up cycles (default 1)\n");
+        printf("arg19: number of iterations (default 10)\n");
+        printf("arg20: memory for rotating buffer (default 0, size in MB)\n");
+        // clang-format on
+        exit(1);
+    }
+
+    int n_warmup      = 1;
+    int n_iter        = 10;
+    uint64_t rotating = 0;
+    if(argc == 21)
+    {
+        n_warmup = std::stoi(argv[18]);
+        n_iter   = std::stoi(argv[19]);
+        rotating = std::stoull(argv[20]) * 1024 * 1024;
+    }
+
+    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 BatchStrideA = std::stoi(argv[14]);
+    const int BatchStrideB = std::stoi(argv[15]);
+    const int BatchStrideC = std::stoi(argv[16]);
+
+    const int BatchCount = std::stoi(argv[17]);
+
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
+    using F8 = ck::f8_t;
+#endif
+    using BF16 = ck::bhalf_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile =
+        [&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) {
+            using ADataType  = decltype(a_type);
+            using BDataType  = decltype(b_type);
+            using DsDataType = ck::Tuple<>;
+            using CDataType  = decltype(c_type);
+
+            using ALayout  = decltype(a_layout);
+            using BLayout  = decltype(b_layout);
+            using DsLayout = ck::Tuple<>;
+            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;
+
+            const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
+            const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
+            const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+            const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
+            const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
+            const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
+
+            const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
+            const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
+            const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
+
+            using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+            using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+            using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+            using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmV2MultiD<ALayout,
+                                                                                     BLayout,
+                                                                                     DsLayout,
+                                                                                     CLayout,
+                                                                                     ADataType,
+                                                                                     BDataType,
+                                                                                     DsDataType,
+                                                                                     CDataType,
+                                                                                     AElementOp,
+                                                                                     BElementOp,
+                                                                                     CElementOp>;
+
+            bool pass = ck::profiler::profile_gemm_universal_batched_impl<ADataType,
+                                                                          BDataType,
+                                                                          CDataType,
+                                                                          ALayout,
+                                                                          BLayout,
+                                                                          CLayout,
+                                                                          AElementOp,
+                                                                          BElementOp,
+                                                                          CElementOp,
+                                                                          DeviceOp>(do_verification,
+                                                                                    init_method,
+                                                                                    do_log,
+                                                                                    time_kernel,
+                                                                                    M,
+                                                                                    N,
+                                                                                    K,
+                                                                                    BatchStrideA_,
+                                                                                    BatchStrideB_,
+                                                                                    BatchStrideC_,
+                                                                                    StrideA_,
+                                                                                    StrideB_,
+                                                                                    StrideC_,
+                                                                                    BatchCount,
+                                                                                    n_warmup,
+                                                                                    n_iter,
+                                                                                    rotating);
+
+            return pass ? 0 : 1;
+        };
+
+    if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
+    }
+#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
+    else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F8{}, F8{}, BF16{}, Row{}, Col{}, Row{});
+    }
+#endif
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_universal);

profiler/src/profile_grouped_conv_bwd_weight.cpp

@@ -25,7 +25,8 @@ enum struct ConvDataType
     F16_F16_F16,        // 1
     BF16_F32_BF16,      // 2
     F16_F16_F16_BF8_F8, // 3
-    I8_I8_I8            // 4
+    I8_I8_I8,           // 4
+    BF16_BF16_BF16,     // 5
 };
 
 #define OP_NAME "grouped_conv_bwd_weight"
@@ -38,7 +39,8 @@ static void print_helper_msg()
               << "                 1: Input fp16, Weight fp16, Output fp16\n"
               << "                 2: Input bf16, Weight fp32, Output bf16\n"
               << "                 3: Input fp16, Weight fp16, Output fp16, Gemm bf8@fp8\n"
-              << "                 4: Input int8, Weight int8, Output int8)\n"
+              << "                 4: Input int8, Weight int8, Output int8\n"
+              << "                 5: Input bf16, Weight bf16, Output bf16)\n"
               << "arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, "
                  "N, K, Ho, Wo]\n"
               << "                     1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, "
@@ -180,6 +182,10 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
             // fp32 atomic add is used for weight tensor in bf16 kernel
             return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
         }
+        if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
+        }
     }
     else if(num_dim_spatial == 2 && layout == ConvLayout::NGCHW_GKYXC_NGKHW)
     {
@@ -187,6 +193,11 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
         {
             return profile(I2, NGCHW{}, GKYXC{}, NGKHW{}, F16{}, F16{}, F16{}, F16{}, F16{});
         }
+        if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I2, NGCHW{}, GKYXC{}, NGKHW{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
+        }
     }
     if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
     {
@@ -224,6 +235,11 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
             // fp32 atomic add is used for weight tensor in bf16 kernel
             return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
         }
+        if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(
+                I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
+        }
         if(data_type == ConvDataType::F16_F16_F16_BF8_F8)
         {
             return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{}, BF8{}, F8{});
@@ -240,6 +256,11 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
         {
             return profile(I3, NGCDHW{}, GKZYXC{}, NGKDHW{}, F16{}, F16{}, F16{}, F16{}, F16{});
         }
+        if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(
+                I3, NGCDHW{}, GKZYXC{}, NGKDHW{}, BF16{}, BF16{}, BF16{}, BF16{}, BF16{});
+        }
     }
 
     std::cout << "this data_type & layout is not implemented" << std::endl;

profiler/src/profile_layernorm_fwd.cpp

@@ -85,7 +85,7 @@ int profile_layernorm(int argc, char* argv[])
 
         if(data_type == ck::DataTypeEnum::Half)
         {
-            ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
+            ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F16, false, rank>(
                 do_verification, init_method, do_log, time_kernel, length);
         }
         else if(data_type == ck::DataTypeEnum::Float)

python/ck4inductor/grouped_conv_fwd/gen_instances.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+import logging
+import os
+import subprocess
+from dataclasses import replace
+from functools import lru_cache
+from typing import List
+
+from ..util import library_path
+
+from .op import CKGroupedConvFwdOp
+
+log = logging.getLogger(__name__)
+
+
+def _ck_conv_instances_path():
+    conv_instances_path = os.path.join(  # noqa: F821
+        library_path(),
+        "include",
+        "ck",
+        "library",
+        "tensor_operation_instance",
+        "gpu",
+        "grouped_conv_fwd",
+    )
+    if not os.path.exists(conv_instances_path):
+        log.error(
+            "CK library conv instances path %s does not exist", conv_instances_path
+        )
+        return None
+    return conv_instances_path
+
+
+def parse_instances(str_instances: List[str]) -> List[CKGroupedConvFwdOp]:
+    """
+    Parse the lines containing Grouped Convolution Forward template instances
+    into `CKGroupedConvFwdOp` instances
+    """
+
+    def maybe_int(s):
+        try:
+            return int(s)
+        except ValueError:
+            return s
+
+    op_instances = []
+    # TODO: maybe use libclang for parsing C++ code in the future
+    # to avoid this hacky parsing logic below ? :) - copilot
+    for line in str_instances:
+        s_template_args = line.split("DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3")[
+            -1
+        ].strip("<>, ")
+        template_args = []
+        i_current = 0
+        while i_current < len(s_template_args):
+            if s_template_args[i_current] == " ":
+                # skip whitespace
+                i_current += 1
+                continue
+            elif s_template_args[i_current : i_current + 2] == "S<":
+                # parse template S<Index...>
+                i_next = s_template_args.find(">", i_current)
+                template_args.append(
+                    tuple(map(int, s_template_args[i_current + 2 : i_next].split(",")))
+                )
+                i_current = i_next + 2
+            else:
+                # all string attributes must be either type aliases or global constants in C++
+                i_next = s_template_args.find(",", i_current)
+                template_args.append(
+                    maybe_int(
+                        s_template_args[i_current : i_next if i_next != -1 else None]
+                    )
+                )
+                if i_next != -1:
+                    i_current = i_next + 1
+            if i_next == -1:
+                break
+
+        template_args[0] = -1  # n_dim_spatial
+        template_args[3] = tuple()  # ds_layout
+        template_args[9] = tuple()  # ds_element_dtype
+
+        new_instance = CKGroupedConvFwdOp(
+            *template_args,  # type: ignore[arg-type]
+        )
+
+        op_instances.append(new_instance)
+    return op_instances
+
+
+@lru_cache(None)
+def gen_conv_ops_library() -> List[CKGroupedConvFwdOp]:
+    """
+    Parse the Grouped Convolution Forward instances
+    defined in the Composable Kernel library folder.
+    """
+    ck_library_dir = _ck_conv_instances_path()
+    if not ck_library_dir:
+        return []
+
+    grep_result = subprocess.run(
+        [
+            "grep",
+            "-inR",
+            "DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3",
+            ck_library_dir,
+        ],
+        capture_output=True,
+        text=True,
+    )
+
+    op_instances = parse_instances(grep_result.stdout.strip().split("\n"))
+
+    log.debug("ck instances from library: %d", len(op_instances))
+
+    schedulers = [
+        "BlockGemmPipelineScheduler::Intrawave",
+        "BlockGemmPipelineScheduler::Interwave",
+    ]
+    conv_specs = [
+        "ConvolutionForwardSpecialization::Default",
+        "ConvolutionForwardSpecialization::Filter1x1Pad0",
+        "ConvolutionForwardSpecialization::Filter1x1Stride1Pad0",
+        "ConvolutionForwardSpecialization::OddC",
+    ]
+
+    # substitute templated args by looping through their domains
+    substitute_instances = []
+    for instance in op_instances:
+        sub_scheduler = (
+            instance.block_gemm_pipeline_scheduler == "BlkGemmPipeSched"
+        )
+        sub_spec = instance.conv_forward_specialization == "ConvSpec"
+        schedulers_range = (
+            schedulers if sub_scheduler else [instance.block_gemm_pipeline_scheduler]
+        )
+        spec_range = conv_specs if sub_spec else [instance.conv_forward_specialization]
+        for scheduler in schedulers_range:
+            for spec in spec_range:
+                for channels_last in [True, False]:
+                    if channels_last:
+                        a_layout = "NHWGC"
+                        e_layout = "NHWGK"
+                    else:
+                        a_layout = "NGCHW"
+                        e_layout = "NGKHW"
+                    substitute_instances.append(
+                        replace(
+                            instance,
+                            block_gemm_pipeline_scheduler=scheduler,
+                            conv_forward_specialization=spec,
+                            gemm_specialization="GemmSpecialization::MNKPadding",
+                            n_dim_spatial=2,
+                            a_layout=a_layout,
+                            b_layout="GKYXC",
+                            e_layout=e_layout,
+                        )
+                    )
+
+    return substitute_instances
+
+
+if __name__ == "__main__":
+    print(gen_conv_ops_library())

python/ck4inductor/grouped_conv_fwd/op.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+from dataclasses import asdict, dataclass
+from typing import Optional, Tuple
+
+
+@dataclass
+class CKGroupedConvFwdOp:
+    n_dim_spatial: int
+    a_layout: str
+    b_layout: str
+    ds_layout: Tuple[str]
+    e_layout: str
+    a_element_dtype: str
+    b_element_dtype: str
+    acc_dtype: str
+    c_shuffle_dtype: str
+    ds_element_dtype: Tuple[str]
+    e_element_dtype: str
+    a_elementwise_op: str
+    b_elementwise_op: str
+    cde_elementwise_op: str
+    conv_forward_specialization: str
+    gemm_specialization: str
+
+    block_size: int
+    m_per_block: int
+    n_per_block: int
+    k_per_block: int
+    ak1: int
+    bk1: int
+    m_per_xdl: int
+    n_per_xdl: int
+    m_xdl_per_wave: int
+    n_xdl_per_wave: int
+    a_block_transfer_thread_cluster_lengths_ak0_m_ak1: Tuple[int, int, int]
+    a_block_transfer_thread_cluster_arrange_order: Tuple[int, int, int]
+    a_block_transfer_src_access_order: Tuple[int, int, int]
+    a_block_transfer_src_vector_dim: int
+    a_block_transfer_src_scalar_per_vector: int
+    a_block_transfer_dst_scalar_per_vector_ak1: int
+    a_block_lds_extra_m: bool
+
+    b_block_transfer_thread_cluster_lengths_bk0_n_bk1: Tuple[int, int, int]
+    b_block_transfer_thread_cluster_arrange_order: Tuple[int, int, int]
+    b_block_transfer_src_access_order: Tuple[int, int, int]
+
+    b_block_transfer_src_vector_dim: int
+    b_block_transfer_src_scalar_per_vector: int
+    b_block_transfer_dst_scalar_per_vector_bk1: int
+    b_block_lds_extra_n: bool
+
+    c_shuffle_m_xdl_per_wave_per_shuffle: int
+    c_shuffle_n_xdl_per_wave_per_shuffle: int
+    cde_block_transfer_cluster_lengths_m_block_m_per_block_n_block_n_per_block: Tuple[  # noqa
+        int,
+        int,
+        int,
+        int,
+    ]
+    cde_block_transfer_scalar_per_vector_n_per_block: int
+    block_gemm_pipeline_scheduler: str
+    block_gemm_pipeline_version: str
+
+    a_compute_dtype: Optional[str] = None
+    b_compute_dtype: Optional[str] = None
+
+    def name(self):
+        # cpp alias for template instance
+        return (
+            f"ck_device_grouped_convolution_fwd_multiple_abd_xdl_c_shuffle_v3_"
+            f"{self.key_name()}"
+        )
+
+    def key_name(self):
+        # TBD; must be unique per instance. Intended to use as dict key
+        return "_".join(
+            [
+                "K"
+                + field_name.replace("_", "").lower()
+                + "V"
+                + (
+                    "x".join(map(str, iter(field_value)))
+                    if isinstance(field_value, tuple)
+                    else str(field_value).replace(":", "")
+                )
+                for field_name, field_value in self.dict_items()
+            ]
+        )
+
+    def dict_items(self):
+        return asdict(self).items()

python/ck4inductor/universal_gemm/gen_instances.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
 import logging
 import os
 import subprocess
-from dataclasses import fields, replace
+from dataclasses import replace
 from functools import lru_cache, partial
 from typing import List
 

python/ck4inductor/universal_gemm/op.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
 from dataclasses import asdict, dataclass
 from typing import Optional, Tuple
 

python/ck4inductor/util.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
 import functools
 import os
 
 
 @functools.lru_cache(None)
 def library_path():
-    return os.path.join(os.path.dirname(__file__), 'library')
+    return os.path.join(os.path.dirname(__file__), "library")

script/convert_miopen_driver_to_profiler.py

@@ -65,8 +65,9 @@ def parse_data_type(args):
         if args.ck_profier_op == "grouped_conv_fwd":
             args.data_type = 3
     if args.data_type == "bfp16":
-        if args.ck_profier_op == "grouped_conv_bwd_weight" or \
-           args.ck_profier_op == "grouped_conv_bwd_data" or \
+        if args.ck_profier_op == "grouped_conv_bwd_weight":
+            args.data_type = 5
+        if args.ck_profier_op == "grouped_conv_bwd_data" or \
            args.ck_profier_op == "grouped_conv_fwd":
             args.data_type = 2
 

script/process_perf_data.py

@@ -133,12 +133,12 @@ def parse_logfile(logfile):
             if 'Best Perf' in line:
                 lst=line.split()
                 res.append(lst[4])
-    elif 'onnx_gemm' in logfile or 'mixed_gemm' in logfile:
+    elif 'onnx_gemm' in logfile:
         for line in open(logfile):
             if 'Best Perf' in line:
                 lst=line.split()
                 res.append(lst[33])
-    elif 'splitK_gemm' in logfile:
+    elif 'splitK_gemm' in logfile or 'mixed_gemm' in logfile:
         for line in open(logfile):
             if 'Best Perf' in line:
                 lst=line.split()

script/process_qa_data.sh

@@ -22,6 +22,7 @@ python3 process_perf_data.py perf_gemm_bilinear.log
 python3 process_perf_data.py perf_reduction.log
 python3 process_perf_data.py perf_splitK_gemm.log
 python3 process_perf_data.py perf_onnx_gemm.log
+python3 process_perf_data.py perf_mixed_gemm.log
 
 file=./perf_fmha_fwd_gfx942.log
 if [ -e "$file" ]; then

test/CMakeLists.txt

@@ -64,11 +64,11 @@ function(add_test_executable TEST_NAME)
     #only continue if there are some source files left on the list
     if(ARGN)
         if(ARGN MATCHES "_xdl")
-             list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
+             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
         elseif(ARGN MATCHES "_wmma")
-             list(REMOVE_ITEM TEST_TARGETS gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
         elseif(ARGN MATCHES "_smfmac")
-             list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201)
+             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
         endif()
         set_source_files_properties(${ARGN} PROPERTIES LANGUAGE HIP)
         add_executable(${TEST_NAME} ${ARGN})
@@ -141,11 +141,11 @@ function(add_gtest_executable TEST_NAME)
     #only continue if there are some source files left on the list
     if(ARGN)
         if(ARGN MATCHES "_xdl")
-             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
+             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
         elseif(ARGN MATCHES "_wmma")
-             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
         elseif(ARGN MATCHES "_smfmac")
-             list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201)
+             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
         endif()
         set_source_files_properties(${ARGN} PROPERTIES LANGUAGE HIP)
         add_executable(${TEST_NAME} ${ARGN})
@@ -210,3 +210,4 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx942" AND CK_HIP_VERSION_MAJOR GREATER_EQUAL
     add_subdirectory(smfmac_op)
 endif()
 add_subdirectory(position_embedding)
+add_subdirectory(scatter_gather)

test/ck_tile/CMakeLists.txt

 add_subdirectory(image_to_column)
+add_subdirectory(gemm)

test/ck_tile/gemm/CMakeLists.txt

+# Currently ck_tile is only built on gfx9
+if(GPU_TARGETS MATCHES "gfx9")
+    add_gtest_executable(test_ck_tile_gemm_mem_pipeline test_gemm_mem_pipeline.cpp)
+endif()

test/ck_tile/gemm/test_gemm_mem_pipeline.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "gtest/gtest.h"
+
+#include "ck_tile/host.hpp"
+#include "test_gemm_mem_pipeline_util.hpp"
+
+using F16 = ck_tile::half_t;
+using F32 = float;
+
+using Row = ck_tile::tensor_layout::gemm::RowMajor;
+using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    //         ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType, CDataType
+    std::tuple<    Row,     Col,     Row,       F16,       F16,         F32,      F16>,
+    std::tuple<    Col,     Row,     Row,       F16,       F16,         F32,      F16>,
+    std::tuple<    Row,     Row,     Row,       F16,       F16,         F32,      F16>,
+    std::tuple<    Col,     Col,     Row,       F16,       F16,         F32,      F16>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestCkTileGemmMemPipeline, KernelTypes);
+
+#include "test_gemm_mem_pipeline_ut_cases.inc"

test/ck_tile/gemm/test_gemm_mem_pipeline_ut_cases.inc

+#pragma once
+
+TYPED_TEST(TestCkTileGemmMemPipeline, SmallM)
+{
+    std::vector<int> Ms{1, 2, 3, 4, 5, 6};
+    constexpr int N = 1024;
+    constexpr int K = 320;
+
+    for(int M : Ms)
+        this->Run(M, N, K);
+}
+
+TYPED_TEST(TestCkTileGemmMemPipeline, MidLargeM)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    constexpr int N = 1024;
+    constexpr int K = 320;
+
+    for(int M : Ms)
+        this->Run(M, N, K);
+}
+
+TYPED_TEST(TestCkTileGemmMemPipeline, PaddK)
+{
+    std::vector<int> Ms{127};
+    constexpr int N = 1024;
+    constexpr int K = 432;
+
+    for(int M : Ms)
+        this->Run(M, N, K);
+}
+
+TYPED_TEST(TestCkTileGemmMemPipeline, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 1024;
+    constexpr int K = 512;
+
+    for(int M : Ms)
+        this->Run(M, N, K);
+}