merge down from public repo

profiler/src/CMakeLists.txt

@@ -29,6 +29,7 @@ set(PROFILER_SOURCES
     profile_batchnorm_fwd.cpp
     profile_batchnorm_bwd.cpp
     profile_batchnorm_infer.cpp
+    profile_grouped_gemm_fastgelu.cpp
 )
 
 set(PROFILER_EXECUTABLE ckProfiler)
@@ -68,4 +69,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instan
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
 rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)

profiler/src/profile_grouped_gemm.cpp

@@ -29,6 +29,11 @@ enum struct GemmDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "grouped_gemm"
+#define OP_DESC "Grouped GEMM"
+
+namespace {
+
 std::vector<int> argToIntArray(char* input)
 {
     std::vector<int> out;
@@ -45,9 +50,6 @@ std::vector<int> argToIntArray(char* input)
     return out;
 }
 
-#define OP_NAME "grouped_gemm"
-#define OP_DESC "Grouped GEMM"
-
 int profile_grouped_gemm(int argc, char* argv[])
 {
     if(!(argc == 14))
@@ -166,4 +168,6 @@ int profile_grouped_gemm(int argc, char* argv[])
     return 0;
 }
 
+} // anonymous namespace
+
 REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm);

profiler/src/profile_grouped_gemm_fastgelu.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_gemm_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+    MK_KN_NM, // 4
+    MK_NK_NM, // 5
+    KM_KN_NM, // 6
+    KM_NK_NM, // 7
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "grouped_gemm_fastgelu"
+#define OP_DESC "Grouped GEMM+FastGelu"
+
+namespace {
+
+std::vector<int> argToIntArray(char* input)
+{
+    std::vector<int> out;
+
+    std::istringstream in(input);
+
+    std::string item;
+
+    while(std::getline(in, item, ','))
+    {
+        out.push_back(std::stoi(item));
+    }
+
+    return out;
+}
+
+int profile_grouped_gemm_fastgelu(int argc, char* argv[])
+{
+    if(!(argc == 14))
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
+               "64,64 64,64 128,128)\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const auto Ms = argToIntArray(argv[8]);
+    const auto Ns = argToIntArray(argv[9]);
+    const auto Ks = argToIntArray(argv[10]);
+
+    const auto StrideAs = argToIntArray(argv[11]);
+    const auto StrideBs = argToIntArray(argv[12]);
+    const auto StrideCs = argToIntArray(argv[13]);
+
+    if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+} // anonymous namespace
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm_fastgelu);

script/cmake-ck-dev.sh

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

test/grouped_convnd_bwd_weight/grouped_convnd_bwd_weight.cpp

@@ -43,7 +43,7 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
                 DataType,
                 DataType,
                 DataType>(true,  // do_verification
-                          1,     // init_method integer value
+                          1,     // init_method: integer value
                           false, // do_log
                           false, // time_kernel
                           param,
@@ -60,9 +60,9 @@ TYPED_TEST_SUITE(TestGroupedConvndBwdWeight, KernelTypes);
 TYPED_TEST(TestGroupedConvndBwdWeight, Test1D)
 {
     this->conv_params.clear();
-    this->conv_params.push_back({1, 4, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
-    this->conv_params.push_back({1, 4, 64, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
-    this->conv_params.push_back({1, 4, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 2, 32, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
     this->template Run<1>();
 }
 
@@ -70,11 +70,11 @@ TYPED_TEST(TestGroupedConvndBwdWeight, Test2D)
 {
     this->conv_params.clear();
     this->conv_params.push_back(
-        {2, 4, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+        {2, 2, 64, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
     this->conv_params.push_back(
-        {2, 4, 8, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+        {2, 2, 4, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
     this->conv_params.push_back(
-        {2, 4, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+        {2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
     this->template Run<2>();
 }
 
@@ -82,10 +82,10 @@ TYPED_TEST(TestGroupedConvndBwdWeight, Test3D)
 {
     this->conv_params.clear();
     this->conv_params.push_back(
-        {3, 4, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+        {3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
     this->conv_params.push_back(
-        {3, 4, 8, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+        {3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
     this->conv_params.push_back(
-        {3, 4, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+        {3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
     this->template Run<3>();
 }