Merge branch 'develop' into dl_conv_multiple_d

profiler/src/profile_batchnorm_bwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_backward_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormBwdArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormBwdArgParser()  = default;
+    ~BatchnormBwdArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 4 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type           = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar = std::atoi(argv[optind++]);
+        init_method         = std::atoi(argv[optind++]);
+        time_kernel         = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int profile_batchnorm_backward(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_backward_impl;
+
+    BatchnormBwdArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F16, F32, F32, F32, F16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F32, F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<BF16, F32, F32, F32, BF16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F64, F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_bwd", "Batchnorm backward", profile_batchnorm_backward);

profiler/src/profile_batchnorm_fwd.cpp

@@ -6,7 +6,8 @@
 #include <getopt.h>
 
 #include "ck/library/utility/host_common_util.hpp"
-#include "profiler/include/profile_batchnorm_forward_impl.hpp"
+#include "profiler/profile_batchnorm_forward_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using ck::index_t;
 
@@ -214,3 +215,5 @@ int profile_batchnorm_forward(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION("bnorm_fwd", "Batchnorm forward", profile_batchnorm_forward);

profiler/src/profile_conv_bwd_data.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_bwd_data_impl.hpp"
+#include "profiler/profile_conv_bwd_data_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 namespace {
 
@@ -24,10 +25,13 @@ enum struct ConvDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "conv_bwd_data"
+#define OP_DESC "Convolution Backward Data"
+
 static void print_helper_msg()
 {
     std::cout
-        << "arg1: tensor operation (conv_bwd_data: Convolution Backward Data)\n"
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
         << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
         << "                 1: Input fp16, Weight fp16, Output fp16\n"
         << "                 2: Input bf16, Weight bf16, Output bf16\n"
@@ -182,3 +186,5 @@ int profile_conv_bwd_data(int argc, char* argv[])
 
     return 1;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_bwd_data);

profiler/src/profile_conv_fwd.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_impl.hpp"
+#include "profiler/profile_conv_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 namespace {
 
@@ -24,11 +25,14 @@ enum struct ConvDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "conv_fwd"
+#define OP_DESC "Convolution Forward"
+
 static void print_helper_msg()
 {
     std::cout
         // clang-format-off
-        << "arg1: tensor operation (conv_fwd: Convolution Forward)\n"
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
         << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
         << "                 1: Input fp16, Weight fp16, Output fp16\n"
         << "                 2: Input bf16, Weight bf16, Output bf16\n"
@@ -184,3 +188,5 @@ int profile_conv_fwd(int argc, char* argv[])
 
     return 1;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd);

profiler/src/profile_conv_fwd_bias_relu.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_bias_relu_impl.hpp"
+#include "profiler/profile_conv_fwd_bias_relu_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct ConvDataType
 {
@@ -32,11 +33,14 @@ enum struct ConvOutputLayout
     NHWK, // 1
 };
 
+#define OP_NAME "conv_fwd_bias_relu"
+#define OP_DESC "Convolution Forward+Bias+ReLU"
+
 int profile_conv_fwd_bias_relu(int argc, char* argv[])
 {
     if(argc != 25)
     {
-        printf("arg1: tensor operation (conv_fwd_bias_relu: ForwardConvolution+Bias+ReLu)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16)\n");
         printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
         printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
@@ -114,3 +118,5 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu);

profiler/src/profile_conv_fwd_bias_relu_add.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp"
+#include "profiler/profile_conv_fwd_bias_relu_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct ConvDataType
 {
@@ -32,12 +33,14 @@ enum struct ConvOutputLayout
     NHWK, // 1
 };
 
+#define OP_NAME "conv_fwd_bias_relu_add"
+#define OP_DESC "Convolution Forward+Bias+ReLU+Add"
+
 int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
 {
     if(argc != 25)
     {
-        printf(
-            "arg1: tensor operation (conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLu+Add)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16)\n");
         printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
         printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
@@ -115,3 +118,5 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu_add);

profiler/src/profile_gemm.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_impl.hpp"
+#include "profiler/profile_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct GemmMatrixLayout
 {
@@ -24,9 +25,12 @@ enum struct GemmDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "gemm"
+#define OP_DESC "GEMM"
+
 static void print_helper_msg()
 {
-    std::cout << "arg1: tensor operation (gemm: GEMM)\n"
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
               << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
               << "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
               << "                     1: A[m, k] * B[n, k] = C[m, n];\n"
@@ -184,3 +188,5 @@ int profile_gemm(int argc, char* argv[])
         return 1;
     }
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm);

profiler/src/profile_gemm_add_add_fastgelu.cpp

@@ -6,7 +6,11 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_add_add_fastgelu_impl.hpp"
+#include "profiler/profile_gemm_add_add_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_add_fastgelu"
+#define OP_DESC "GEMM+Add+Add+FastGeLU"
 
 int profile_gemm_add_add_fastgelu(int argc, char* argv[])
 {
@@ -29,7 +33,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
     if(argc != 16)
     {
         // clang-format off
-        printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU)\n");
+        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: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
         printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
@@ -150,3 +154,5 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
         return 1;
     }
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_add_fastgelu);

profiler/src/profile_gemm_bias_add_reduce.cpp

@@ -6,7 +6,11 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_bias_add_reduce_impl.hpp"
+#include "profiler/profile_gemm_bias_add_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_bias_add_reduce"
+#define OP_DESC "GEMM+Bias+Add+Reduce"
 
 int profile_gemm_bias_add_reduce(int argc, char* argv[])
 {
@@ -26,7 +30,7 @@ int profile_gemm_bias_add_reduce(int argc, char* argv[])
 
     if(!(argc == 14 || argc == 15))
     {
-        printf("arg1: tensor operation (gemm: GEMM+bias+add+Reduce)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16)\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");
@@ -159,3 +163,5 @@ int profile_gemm_bias_add_reduce(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bias_add_reduce);

profiler/src/profile_gemm_bilinear.cpp

@@ -6,7 +6,11 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_bilinear_impl.hpp"
+#include "profiler/profile_gemm_bilinear_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_bilinear"
+#define OP_DESC "GEMM+Bilinear"
 
 int profile_gemm_bilinear(int argc, char* argv[])
 {
@@ -29,7 +33,7 @@ int profile_gemm_bilinear(int argc, char* argv[])
     if(argc != 17)
     {
         // clang-format off
-        printf("arg1: tensor operation (gemm_bilinear: GEMM+Bilinear)\n");
+        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: E[m, n] = alpha * A[m, k] * B[k, n] + beta * D[m, n];\n");
         printf("                     1: E[m, n] = alpha * A[m, k] * B[n, k] + beta * D[m, n];\n");
@@ -144,3 +148,5 @@ int profile_gemm_bilinear(int argc, char* argv[])
         return 1;
     }
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bilinear);

profiler/src/profile_gemm_reduce.cpp

@@ -6,7 +6,11 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_reduce_impl.hpp"
+#include "profiler/profile_gemm_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_reduce"
+#define OP_DESC "GEMM+Reduce"
 
 int profile_gemm_reduce(int argc, char* argv[])
 {
@@ -26,7 +30,7 @@ int profile_gemm_reduce(int argc, char* argv[])
 
     if(!(argc == 14 || argc == 15))
     {
-        printf("arg1: tensor operation (gemm: GEMM+Reduce)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16)\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");
@@ -146,3 +150,5 @@ int profile_gemm_reduce(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_reduce);

profiler/src/profile_gemm_splitk.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_splitk_impl.hpp"
+#include "profiler/profile_gemm_splitk_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct GemmMatrixLayout
 {
@@ -24,11 +25,14 @@ enum struct GemmDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "gemm_splitk"
+#define OP_DESC "Split-K GEMM"
+
 int profile_gemm_splitk(int argc, char* argv[])
 {
     if(argc != 15)
     {
-        printf("arg1: tensor operation (gemm_splitk: Split-K GEMM)\n");
+        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");
@@ -146,3 +150,5 @@ int profile_gemm_splitk(int argc, char* argv[])
         return 1;
     }
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_splitk);

profiler/src/profile_grouped_conv_bwd_weight.cpp

@@ -6,7 +6,8 @@
 #include <iostream>
 #include <numeric>
 
-#include "profiler/include/profile_grouped_conv_bwd_weight_impl.hpp"
+#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 namespace {
 
@@ -23,9 +24,12 @@ enum struct ConvDataType
     BF16_F32_BF16, // 2
 };
 
+#define OP_NAME "grouped_conv_bwd_weight"
+#define OP_DESC "Grouped Convolution Backward Weight"
+
 static void print_helper_msg()
 {
-    std::cout << "arg1: tensor operation (conv_bwd_weight: Convolution Backward Weight\n"
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
               << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
               << "                 1: Input fp16, Weight fp16, Output fp16\n"
               << "                 2: Input bf16, Weight fp32, Output bf16)\n"
@@ -174,3 +178,5 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
 
     return 1;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_bwd_weight);

profiler/src/profile_grouped_conv_fwd.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_grouped_conv_fwd_impl.hpp"
+#include "profiler/profile_grouped_conv_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 namespace {
 
@@ -24,11 +25,14 @@ enum struct ConvDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "grouped_conv_fwd"
+#define OP_DESC "Grouped Convolution Forward"
+
 static void print_helper_msg()
 {
     std::cout
         // clang-format off
-        << "arg1: tensor operation (grouped_conv_fwd: Grouped Convolution Forward)\n"
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
         << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
         << "                 1: Input fp16, Weight fp16, Output fp16\n"
         << "                 2: Input bf16, Weight bf16, Output bf16\n"
@@ -252,3 +256,5 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
 
     return 1;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_fwd);

profiler/src/profile_grouped_gemm.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_grouped_gemm_impl.hpp"
+#include "profiler/profile_grouped_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct GemmMatrixLayout
 {
@@ -44,11 +45,14 @@ 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))
     {
-        printf("arg1: tensor operation (grouped_gemm: Grouped GEMM)\n");
+        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");
@@ -161,3 +165,5 @@ int profile_grouped_gemm(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm);

profiler/src/profile_groupnorm.cpp

@@ -5,8 +5,9 @@
 #include <vector>
 #include <unordered_map>
 
-#include "profiler/include/data_type_enum.hpp"
-#include "profiler/include/profile_groupnorm_impl.hpp"
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_groupnorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using ck::index_t;
 
@@ -43,9 +44,12 @@ struct GroupnormArgParser
     }
 };
 
+#define OP_NAME "groupnorm"
+#define OP_DESC "Group Normalization"
+
 void print_help_groupnorm()
 {
-    std::cout << "arg1: tensor operation (groupnorm: Group normalization)\n"
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
               << "arg2: data type (0: fp16; 1: fp32)\n"
               << "arg3: verification (0: no; 1: yes)\n"
               << "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
@@ -104,3 +108,5 @@ int profile_groupnorm(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_groupnorm);

profiler/src/profile_layernorm.cpp

@@ -5,8 +5,9 @@
 #include <vector>
 #include <unordered_map>
 
-#include "profiler/include/data_type_enum.hpp"
-#include "profiler/include/profile_layernorm_impl.hpp"
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_layernorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using ck::index_t;
 
@@ -96,3 +97,5 @@ int profile_layernorm(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION("layernorm", "Layer Normalization", profile_layernorm);

profiler/src/profile_reduce.cpp

@@ -13,8 +13,9 @@
 
 #include "ck/library/utility/host_common_util.hpp"
 
-#include "profiler/include/profile_reduce_impl.hpp"
-#include "profiler/include/data_type_enum.hpp"
+#include "profiler/profile_reduce_impl.hpp"
+#include "profiler/data_type_enum.hpp"
+#include "profiler_operation_registry.hpp"
 
 using namespace std;
 
@@ -429,3 +430,5 @@ int profile_reduce(int argc, char* argv[])
 
     return (0);
 };
+
+REGISTER_PROFILER_OPERATION("reduce", "Reduce", profile_reduce);

profiler/src/profile_softmax.cpp

@@ -5,7 +5,8 @@
 #include <vector>
 #include <unordered_map>
 
-#include "profiler/include/profile_softmax_impl.hpp"
+#include "profiler/profile_softmax_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using ck::index_t;
 using ck::profiler::SoftmaxDataType;
@@ -164,3 +165,5 @@ int profile_softmax(int argc, char* argv[])
 //     profile_normalization(argc, argv);
 //     return 0;
 // }
+
+REGISTER_PROFILER_OPERATION("softmax", "Softmax", profile_softmax);

profiler/src/profiler.cpp

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
-#include <cstring>
+#include <cstdlib>
+#include <iostream>
 
-int profile_gemm(int, char*[]);
-int profile_gemm_splitk(int, char*[]);
-int profile_gemm_bilinear(int, char*[]);
-int profile_gemm_add_add_fastgelu(int, char*[]);
-int profile_gemm_reduce(int, char*[]);
-int profile_gemm_bias_add_reduce(int, char*[]);
-int profile_batched_gemm(int, char*[]);
-int profile_batched_gemm_gemm(int, char*[]);
-int profile_batched_gemm_add_relu_gemm_add(int, char*[]);
-int profile_batched_gemm_reduce(int, char*[]);
-int profile_grouped_gemm(int, char*[]);
-int profile_conv_fwd(int, char*[]);
-int profile_conv_fwd_bias_relu(int, char*[]);
-int profile_conv_fwd_bias_relu_add(int, char*[]);
-int profile_conv_bwd_data(int, char*[]);
-int profile_grouped_conv_fwd(int, char*[]);
-int profile_grouped_conv_bwd_weight(int, char*[]);
-int profile_softmax(int, char*[]);
-int profile_layernorm(int, char*[]);
-int profile_groupnorm(int, char*[]);
-int profile_reduce(int, char*[]);
-int profile_batchnorm_forward(int, char*[]);
+#include "profiler_operation_registry.hpp"
 
 static void print_helper_message()
 {
-    // clang-format off
-    printf("arg1: tensor operation (gemm: GEMM\n"
-           "                        gemm_splitk: Split-K GEMM\n"
-           "                        gemm_bilinear: GEMM+Bilinear\n"
-           "                        gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"
-           "                        gemm_reduce: GEMM+Reduce\n"
-           "                        gemm_bias_add_reduce: GEMM+Bias+Add+Reduce\n"
-           "                        batched_gemm: Batched GEMM\n"
-           "                        batched_gemm_gemm: Batched+GEMM+GEMM\n"
-           "                        batched_gemm_add_relu_gemm_add: Batched+GEMM+bias+gelu+GEMM+bias\n"
-           "                        batched_gemm_reduce: Batched GEMM+Reduce\n"
-           "                        grouped_gemm: Grouped GEMM\n"
-           "                        conv_fwd: Convolution Forward\n"
-           "                        conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
-           "                        conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
-           "                        conv_bwd_data: Convolution Backward Data\n"
-           "                        grouped_conv_fwd: Grouped Convolution Forward\n"
-           "                        grouped_conv_bwd_weight: Grouped Convolution Backward Weight\n"
-           "                        softmax: Softmax\n"
-           "                        reduce: Reduce\n"
-	   "                        bnorm_fwd: Batchnorm forward\n");
-    // clang-format on
+    std::cout << "arg1: tensor operation " << ProfilerOperationRegistry::GetInstance() << std::endl;
 }
 
 int main(int argc, char* argv[])
@@ -57,101 +16,15 @@ int main(int argc, char* argv[])
     if(argc == 1)
     {
         print_helper_message();
-
-        return 0;
-    }
-    else if(strcmp(argv[1], "gemm") == 0)
-    {
-        return profile_gemm(argc, argv);
-    }
-    else if(strcmp(argv[1], "gemm_splitk") == 0)
-    {
-        return profile_gemm_splitk(argc, argv);
-    }
-    else if(strcmp(argv[1], "gemm_bilinear") == 0)
-    {
-        return profile_gemm_bilinear(argc, argv);
-    }
-    else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
-    {
-        return profile_gemm_add_add_fastgelu(argc, argv);
-    }
-    else if(strcmp(argv[1], "gemm_reduce") == 0)
-    {
-        return profile_gemm_reduce(argc, argv);
-    }
-    else if(strcmp(argv[1], "gemm_bias_add_reduce") == 0)
-    {
-        return profile_gemm_bias_add_reduce(argc, argv);
-    }
-    else if(strcmp(argv[1], "batched_gemm") == 0)
-    {
-        return profile_batched_gemm(argc, argv);
-    }
-    else if(strcmp(argv[1], "batched_gemm_gemm") == 0)
-    {
-        return profile_batched_gemm_gemm(argc, argv);
-    }
-    else if(strcmp(argv[1], "batched_gemm_add_relu_gemm_add") == 0)
-    {
-        return profile_batched_gemm_add_relu_gemm_add(argc, argv);
-    }
-    else if(strcmp(argv[1], "batched_gemm_reduce") == 0)
-    {
-        return profile_batched_gemm_reduce(argc, argv);
-    }
-    else if(strcmp(argv[1], "grouped_gemm") == 0)
-    {
-        return profile_grouped_gemm(argc, argv);
     }
-    else if(strcmp(argv[1], "conv_fwd") == 0)
+    else if(const auto operation = ProfilerOperationRegistry::GetInstance().Get(argv[1]);
+            operation.has_value())
     {
-        return profile_conv_fwd(argc, argv);
-    }
-    else if(strcmp(argv[1], "conv_fwd_bias_relu") == 0)
-    {
-        return profile_conv_fwd_bias_relu(argc, argv);
-    }
-    else if(strcmp(argv[1], "conv_fwd_bias_relu_add") == 0)
-    {
-        return profile_conv_fwd_bias_relu_add(argc, argv);
-    }
-    else if(strcmp(argv[1], "conv_bwd_data") == 0)
-    {
-        return profile_conv_bwd_data(argc, argv);
-    }
-    else if(strcmp(argv[1], "grouped_conv_fwd") == 0)
-    {
-        return profile_grouped_conv_fwd(argc, argv);
-    }
-    else if(strcmp(argv[1], "conv_bwd_weight") == 0)
-    {
-        return profile_grouped_conv_bwd_weight(argc, argv);
-    }
-    else if(strcmp(argv[1], "reduce") == 0)
-    {
-        return profile_reduce(argc, argv);
-    }
-    else if(strcmp(argv[1], "softmax") == 0)
-    {
-        return profile_softmax(argc, argv);
-    }
-    else if(strcmp(argv[1], "layernorm") == 0)
-    {
-        return profile_layernorm(argc, argv);
-    }
-    else if(strcmp(argv[1], "groupnorm") == 0)
-    {
-        return profile_groupnorm(argc, argv);
-    }
-    else if(strcmp(argv[1], "bnorm_fwd") == 0)
-    {
-        return profile_batchnorm_forward(argc, argv);
+        return (*operation)(argc, argv);
     }
     else
     {
-        print_helper_message();
-
-        return 0;
+        std::cerr << "cannot find operation: " << argv[1] << std::endl;
+        return EXIT_FAILURE;
     }
 }