Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into wmma_op

library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwdDefault>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwd1x1S1P0>{});
+}
+
+void add_device_conv2d_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              GNHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Relu_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwdDefault>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwd1x1P0>{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwd1x1S1P0>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/CMakeLists.txt

@@ -27,6 +27,7 @@ set(PROFILER_SOURCE
     src/profile_layernorm.cpp
     src/profile_softmax.cpp
     src/profile_batchnorm_fwd.cpp
+    src/profile_batchnorm_bwd.cpp
 )
 
 add_executable(ckProfiler ${PROFILER_SOURCE})

profiler/src/profiler.cpp

@@ -25,6 +25,7 @@ int profile_layernorm(int, char*[]);
 int profile_groupnorm(int, char*[]);
 int profile_reduce(int, char*[]);
 int profile_batchnorm_forward(int, char*[]);
+int profile_batchnorm_backward(int, char*[]);
 
 static void print_helper_message()
 {
@@ -148,6 +149,10 @@ int main(int argc, char* argv[])
     {
         return profile_batchnorm_forward(argc, argv);
     }
+    else if(strcmp(argv[1], "bnorm_bwd") == 0)
+    {
+        return profile_batchnorm_backward(argc, argv);
+    }
     else
     {
         print_helper_message();

test/CMakeLists.txt

@@ -54,6 +54,7 @@ add_subdirectory(normalization)
 add_subdirectory(data_type)
 add_subdirectory(elementwise_normalization)
 add_subdirectory(batchnorm_fwd)
+add_subdirectory(batchnorm)
 if(GPU_TARGETS MATCHES "gfx1100")
     add_subdirectory(wmma_op)
 endif()

test/batchnorm_fwd/CMakeLists.txt → test/batchnorm/CMakeLists.txt

 add_gtest_executable(test_batchnorm_fwd_rank_4 batchnorm_fwd_rank_4.cpp)
+add_gtest_executable(test_batchnorm_bwd_rank_4 batchnorm_bwd_rank_4.cpp)
 target_link_libraries(test_batchnorm_fwd_rank_4 PRIVATE utility device_batchnorm_instance)
+target_link_libraries(test_batchnorm_bwd_rank_4 PRIVATE utility device_batchnorm_instance)