Merge remote-tracking branch 'origin/develop' into myamlak/cgemm

library/src/tensor_operation_instance/gpu/convnd_bwd_data/CMakeLists.txt

@@ -14,7 +14,7 @@ set(DEVICE_CONVND_BWD_DATA_INSTANCE_SOURCE
    device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp;
 ) 
 
-add_library(device_convnd_bwd_data_instance SHARED ${DEVICE_CONVND_BWD_DATA_INSTANCE_SOURCE})
+add_library(device_convnd_bwd_data_instance OBJECT ${DEVICE_CONVND_BWD_DATA_INSTANCE_SOURCE})
 target_compile_features(device_convnd_bwd_data_instance PUBLIC)
 set_target_properties(device_convnd_bwd_data_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
 install(TARGETS device_convnd_bwd_data_instance LIBRARY DESTINATION lib) 

library/src/tensor_operation_instance/gpu/device_conv2d.cpp

+#include <stdlib.h>
+#include "config.hpp"
+#include "device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
+#include "element_wise_operation.hpp"
+#include "device_operation_instance.hpp"
+#include "host_interface.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_conv2d_fwd_instance {
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>& instances);
+
+} // namespace device_conv2d_fwd_instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+struct DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl
+{
+    std::unique_ptr<DeviceConvFwdPtr_t::BaseArgument>
+    MakeArgumentPointer(void* in_ptr,
+                        void* wei_ptr,
+                        void* out_ptr,
+                        size_t N,
+                        size_t K,
+                        size_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads) const
+    {
+        return el->MakeArgumentPointer(in_ptr,
+                                       wei_ptr,
+                                       out_ptr,
+                                       N,
+                                       K,
+                                       C,
+                                       input_spatial_lengths,
+                                       filter_spatial_lengths,
+                                       output_spatial_lengths,
+                                       conv_filter_strides,
+                                       conv_filter_dilations,
+                                       input_left_pads,
+                                       input_right_pads,
+                                       PassThrough{},
+                                       PassThrough{},
+                                       PassThrough{});
+    }
+    std::unique_ptr<DeviceConvFwdPtr_t::BaseInvoker> MakeInvokerPointer() const
+    {
+        return el->MakeInvokerPointer();
+    }
+
+    std::string GetTypeString() { return el->GetTypeString(); }
+    bool IsSupportedArgument(const DeviceConvFwdPtr_t::BaseArgument* arg)
+    {
+        return el->IsSupportedArgument(arg);
+    }
+
+    ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough> el;
+};
+
+DeviceConvFwdPtr_t::DeviceConvFwdPtr_t() : pImpl(nullptr) {}
+DeviceConvFwdPtr_t::~DeviceConvFwdPtr_t()                    = default;
+DeviceConvFwdPtr_t::DeviceConvFwdPtr_t(DeviceConvFwdPtr_t&&) = default;
+DeviceConvFwdPtr_t::DeviceConvFwdPtr_t(DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl& other)
+    : pImpl(std::make_unique<DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl>(std::move(other)))
+{
+}
+
+std::unique_ptr<DeviceConvFwdPtr_t::BaseArgument>
+DeviceConvFwdPtr_t::MakeArgumentPointer(void* in_ptr,
+                                        void* wei_ptr,
+                                        void* out_ptr,
+                                        size_t N,
+                                        size_t K,
+                                        size_t C,
+                                        std::vector<ck::index_t> input_spatial_lengths,
+                                        std::vector<ck::index_t> filter_spatial_lengths,
+                                        std::vector<ck::index_t> output_spatial_lengths,
+                                        std::vector<ck::index_t> conv_filter_strides,
+                                        std::vector<ck::index_t> conv_filter_dilations,
+                                        std::vector<ck::index_t> input_left_pads,
+                                        std::vector<ck::index_t> input_right_pads) const
+{
+    return pImpl->MakeArgumentPointer(in_ptr,
+                                      wei_ptr,
+                                      out_ptr,
+                                      N,
+                                      K,
+                                      C,
+                                      input_spatial_lengths,
+                                      filter_spatial_lengths,
+                                      output_spatial_lengths,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads);
+}
+
+std::unique_ptr<DeviceConvFwdPtr_t::BaseInvoker> DeviceConvFwdPtr_t::MakeInvokerPointer() const
+{
+    return pImpl->MakeInvokerPointer();
+}
+
+std::string DeviceConvFwdPtr_t::GetTypeString() { return pImpl->GetTypeString(); }
+bool DeviceConvFwdPtr_t::IsSupportedArgument(const DeviceConvFwdPtr_t::BaseArgument* arg_ptr)
+{
+    return pImpl->IsSupportedArgument(arg_ptr);
+}
+
+using namespace ck::tensor_operation::device::device_conv2d_fwd_instance;
+void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances)
+{
+    std::vector<
+        ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>
+        local_instances;
+    add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(local_instances);
+    for(auto& kinder : local_instances)
+    {
+        DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl tmp{std::move(kinder)};
+        instances.emplace_back(tmp);
+    }
+    return;
+}
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances)
+{
+    std::vector<
+        ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>
+        local_instances;
+    add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(local_instances);
+    for(auto& kinder : local_instances)
+    {
+        DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl tmp{std::move(kinder)};
+        instances.emplace_back(tmp); // Perhaps we can do better
+    }
+    return;
+}
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances)
+{
+    std::vector<
+        ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>
+        local_instances;
+    add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(local_instances);
+    for(auto& kinder : local_instances)
+    {
+        DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl tmp{std::move(kinder)};
+        instances.emplace_back(tmp); // Perhaps we can do better
+    }
+    return;
+}
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances)
+{
+    std::vector<
+        ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>
+        local_instances;
+    add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(local_instances);
+    for(auto& kinder : local_instances)
+    {
+        DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl tmp{std::move(kinder)};
+        instances.emplace_back(tmp); // Perhaps we can do better
+    }
+    return;
+}
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances)
+{
+    std::vector<
+        ck::tensor_operation::device::DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough>>
+        local_instances;
+    add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(local_instances);
+    for(auto& kinder : local_instances)
+    {
+        DeviceConvFwdPtr_t::DeviceConvFwdPtrImpl tmp{std::move(kinder)};
+        instances.emplace_back(tmp);
+    }
+    return;
+}

library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt

@@ -35,10 +35,9 @@ set(DEVICE_GEMM_INSTANCE_SOURCE
    device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp;
 )
 
-add_library(device_gemm_instance SHARED ${DEVICE_GEMM_INSTANCE_SOURCE})
+add_library(device_gemm_instance OBJECT ${DEVICE_GEMM_INSTANCE_SOURCE})
 
 target_compile_features(device_gemm_instance PUBLIC)
 set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
-install(TARGETS device_gemm_instance LIBRARY DESTINATION lib)
 
 clang_tidy_check(device_gemm_instance)

library/src/tensor_operation_instance/gpu/gemm_bias2d/CMakeLists.txt

@@ -10,9 +10,7 @@ set(DEVICE_GEMM_BIAS2D_INSTANCE_SOURCE
    device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_nk_mn_instance.cpp;
 )
 
-add_library(device_gemm_bias2d_instance SHARED ${DEVICE_GEMM_BIAS2D_INSTANCE_SOURCE})
-target_compile_features(device_gemm_bias2d_instance PUBLIC)
+add_library(device_gemm_bias2d_instance OBJECT ${DEVICE_GEMM_BIAS2D_INSTANCE_SOURCE})
 set_target_properties(device_gemm_bias2d_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
-install(TARGETS device_gemm_bias2d_instance LIBRARY DESTINATION lib)
 
 clang_tidy_check(device_gemm_bias2d_instance)

library/src/tensor_operation_instance/gpu/gemm_bias_relu/CMakeLists.txt

@@ -6,9 +6,7 @@ set(DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE
    device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instance.cpp;
 )
 
-add_library(device_gemm_bias_relu_instance SHARED ${DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE}) 
-target_compile_features(device_gemm_bias_relu_instance PUBLIC)
+add_library(device_gemm_bias_relu_instance OBJECT ${DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE}) 
 set_target_properties(device_gemm_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
-install(TARGETS device_gemm_bias_relu_instance LIBRARY DESTINATION lib)
 
 clang_tidy_check(device_gemm_bias_relu_instance)

library/src/tensor_operation_instance/gpu/gemm_bias_relu_add/CMakeLists.txt

@@ -6,9 +6,7 @@ set(DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE
    device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instance.cpp;
 )
 
-add_library(device_gemm_bias_relu_add_instance SHARED ${DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE}) 
-target_compile_features(device_gemm_bias_relu_add_instance PUBLIC)
+add_library(device_gemm_bias_relu_add_instance OBJECT ${DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE}) 
 set_target_properties(device_gemm_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
-install(TARGETS device_gemm_bias_relu_add_instance LIBRARY DESTINATION lib)
 
 clang_tidy_check(device_gemm_bias_relu_add_instance)

library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt

@@ -6,7 +6,7 @@ set(DEVICE_GROUPED_GEMM_INSTANCE_SOURCE
    device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp;
 )
 
-add_library(device_grouped_gemm_instance SHARED ${DEVICE_GROUPED_GEMM_INSTANCE_SOURCE}) 
+add_library(device_grouped_gemm_instance OBJECT ${DEVICE_GROUPED_GEMM_INSTANCE_SOURCE}) 
 
 target_compile_features(device_grouped_gemm_instance PUBLIC)
 set_target_properties(device_grouped_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)

library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt

@@ -38,9 +38,7 @@ set(DEVICE_REDUCE_INSTANCE_SOURCE
    device_reduce_instance_multiblock_partial_reduce_b16_f32_b16.cpp;
 )
 
-add_library(device_reduce_instance SHARED ${DEVICE_REDUCE_INSTANCE_SOURCE}) 
-target_compile_features(device_reduce_instance PUBLIC)
+add_library(device_reduce_instance OBJECT ${DEVICE_REDUCE_INSTANCE_SOURCE}) 
 set_target_properties(device_reduce_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
-install(TARGETS device_reduce_instance LIBRARY DESTINATION lib) 
 
 clang_tidy_check(device_reduce_instance)

library/src/utility/CMakeLists.txt

@@ -8,14 +8,14 @@ include_directories(BEFORE
     ${PROJECT_SOURCE_DIR}/library/include/ck/library/utility
 )
 
-set(CONV_FWD_UTIL_SOURCE
-    conv_fwd_util.cpp
+set(CONV_UTIL_SOURCE
+    conv_util.cpp
 )
 
-add_library(conv_fwd_util SHARED ${CONV_FWD_UTIL_SOURCE})
-target_link_libraries(conv_fwd_util PRIVATE host_tensor)
-target_compile_features(conv_fwd_util PUBLIC)
-set_target_properties(conv_fwd_util PROPERTIES POSITION_INDEPENDENT_CODE ON)
-target_include_directories(conv_fwd_util SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
+add_library(conv_util SHARED ${CONV_UTIL_SOURCE})
+target_link_libraries(conv_util PRIVATE host_tensor)
+target_compile_features(conv_util PUBLIC)
+set_target_properties(conv_util PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_include_directories(conv_util SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
 
-clang_tidy_check(conv_fwd_util)
+clang_tidy_check(conv_util)

library/src/utility/conv_fwd_util.cpp → library/src/utility/conv_util.cpp

 
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 
 namespace ck {
 namespace utils {
@@ -37,16 +37,16 @@ std::size_t get_flops(ck::index_t N,
 }
 
 ConvParams::ConvParams()
-    : num_dim_spatial(2),
-      N(128),
-      K(256),
-      C(192),
-      filter_spatial_lengths(2, 3),
-      input_spatial_lengths(2, 71),
-      conv_filter_strides(2, 2),
-      conv_filter_dilations(2, 1),
-      input_left_pads(2, 1),
-      input_right_pads(2, 1)
+    : num_dim_spatial_(2),
+      N_(128),
+      K_(256),
+      C_(192),
+      filter_spatial_lengths_(2, 3),
+      input_spatial_lengths_(2, 71),
+      conv_filter_strides_(2, 2),
+      conv_filter_dilations_(2, 1),
+      input_left_pads_(2, 1),
+      input_right_pads_(2, 1)
 {
 }
 
@@ -60,23 +60,23 @@ ConvParams::ConvParams(ck::index_t n_dim,
                        const std::vector<ck::index_t>& dilations,
                        const std::vector<ck::index_t>& left_pads,
                        const std::vector<ck::index_t>& right_pads)
-    : num_dim_spatial(n_dim),
-      N(n_batch),
-      K(n_out_channels),
-      C(n_in_channels),
-      filter_spatial_lengths(filters_len),
-      input_spatial_lengths(input_len),
-      conv_filter_strides(strides),
-      conv_filter_dilations(dilations),
-      input_left_pads(left_pads),
-      input_right_pads(right_pads)
+    : num_dim_spatial_(n_dim),
+      N_(n_batch),
+      K_(n_out_channels),
+      C_(n_in_channels),
+      filter_spatial_lengths_(filters_len),
+      input_spatial_lengths_(input_len),
+      conv_filter_strides_(strides),
+      conv_filter_dilations_(dilations),
+      input_left_pads_(left_pads),
+      input_right_pads_(right_pads)
 {
-    if(ck::type_convert<ck::index_t>(filter_spatial_lengths.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(input_spatial_lengths.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(conv_filter_strides.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(conv_filter_dilations.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(input_left_pads.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(input_right_pads.size()) != num_dim_spatial)
+    if(ck::type_convert<ck::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(input_right_pads_.size()) != num_dim_spatial_)
     {
         throw(
             std::runtime_error("ConvParams::GetOutputSpatialLengths: "
@@ -86,27 +86,28 @@ ConvParams::ConvParams(ck::index_t n_dim,
 
 std::vector<ck::index_t> ConvParams::GetOutputSpatialLengths() const
 {
-    if(ck::type_convert<ck::index_t>(filter_spatial_lengths.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(input_spatial_lengths.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(conv_filter_strides.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(conv_filter_dilations.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(input_left_pads.size()) != num_dim_spatial ||
-       ck::type_convert<ck::index_t>(input_right_pads.size()) != num_dim_spatial)
+    if(ck::type_convert<ck::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
+       ck::type_convert<ck::index_t>(input_right_pads_.size()) != num_dim_spatial_)
     {
         throw(
             std::runtime_error("ConvParams::GetOutputSpatialLengths: "
                                "parameter size is different from number of declared dimensions!"));
     }
 
-    std::vector<ck::index_t> out_spatial_len(num_dim_spatial, 0);
-    for(ck::index_t i = 0; i < num_dim_spatial; ++i)
+    std::vector<ck::index_t> out_spatial_len(num_dim_spatial_, 0);
+    for(ck::index_t i = 0; i < num_dim_spatial_; ++i)
     {
         // XEff = (X - 1) * conv_dilation_w + 1;
         // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
-        const ck::index_t idx_eff = (filter_spatial_lengths[i] - 1) * conv_filter_dilations[i] + 1;
+        const ck::index_t idx_eff =
+            (filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
         out_spatial_len[i] =
-            (input_spatial_lengths[i] + input_left_pads[i] + input_right_pads[i] - idx_eff) /
-                conv_filter_strides[i] +
+            (input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - idx_eff) /
+                conv_filter_strides_[i] +
             1;
     }
     return out_spatial_len;
@@ -116,40 +117,40 @@ ConvParams parse_conv_params(int num_dim_spatial, int arg_idx, char* const argv[
 {
     ck::utils::conv::ConvParams params;
 
-    params.num_dim_spatial = num_dim_spatial;
-    params.N               = std::stoi(argv[arg_idx++]);
-    params.K               = std::stoi(argv[arg_idx++]);
-    params.C               = std::stoi(argv[arg_idx++]);
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
 
-    params.filter_spatial_lengths.resize(num_dim_spatial);
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_spatial_lengths.resize(num_dim_spatial);
+    params.input_spatial_lengths_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_strides.resize(num_dim_spatial);
+    params.conv_filter_strides_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.conv_filter_dilations.resize(num_dim_spatial);
+    params.conv_filter_dilations_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_left_pads.resize(num_dim_spatial);
+    params.input_left_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
     }
-    params.input_right_pads.resize(num_dim_spatial);
+    params.input_right_pads_.resize(num_dim_spatial);
     for(int i = 0; i < num_dim_spatial; ++i)
     {
-        params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
     }
 
     return params;
@@ -228,12 +229,12 @@ HostTensorDescriptor get_input_host_tensor_descriptor(const std::vector<std::siz
 std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParams& p)
 {
     os << "ConvParams {"
-       << "\nnum_dim_spatial: " << p.num_dim_spatial << "\nN: " << p.N << "\nK: " << p.K
-       << "\nC: " << p.C << "\nfilter_spatial_lengths: " << p.filter_spatial_lengths
-       << "\ninput_spatial_lengths: " << p.input_spatial_lengths
-       << "\nconv_filter_strides: " << p.conv_filter_strides
-       << "\nconv_filter_dilations: " << p.conv_filter_dilations
-       << "\ninput_left_pads: " << p.input_left_pads
-       << "\ninput_right_pads: " << p.input_right_pads;
+       << "\nnum_dim_spatial: " << p.num_dim_spatial_ << "\nN: " << p.N_ << "\nK: " << p.K_
+       << "\nC: " << p.C_ << "\nfilter_spatial_lengths: " << p.filter_spatial_lengths_
+       << "\ninput_spatial_lengths: " << p.input_spatial_lengths_
+       << "\nconv_filter_strides: " << p.conv_filter_strides_
+       << "\nconv_filter_dilations: " << p.conv_filter_dilations_
+       << "\ninput_left_pads: " << p.input_left_pads_
+       << "\ninput_right_pads: " << p.input_right_pads_;
     return os;
 }

profiler/CMakeLists.txt

@@ -43,7 +43,7 @@ set(PROFILER_SOURCE
 add_executable(ckProfiler ${PROFILER_SOURCE})
 
 target_link_libraries(ckProfiler PRIVATE host_tensor)
-target_link_libraries(ckProfiler PRIVATE conv_fwd_util)
+target_link_libraries(ckProfiler PRIVATE conv_util)
 target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)

profiler/include/profile_batched_gemm_impl.hpp

@@ -63,7 +63,7 @@ template <typename ADataType,
 bool profile_batched_gemm_impl(int do_verification,
                                int init_method,
                                bool do_log,
-                               int nrepeat,
+                               bool time_kernel,
                                int M,
                                int N,
                                int K,
@@ -356,11 +356,12 @@ bool profile_batched_gemm_impl(int do_verification,
         {
             std::string gemm_name = gemm_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
 
-            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
                                      sizeof(CDataType) * M * N) *
                                     BatchCount;
 

profiler/include/profile_batched_gemm_reduce_impl.hpp

@@ -53,7 +53,7 @@ template <typename ADataType,
 bool profile_batched_gemm_reduce_impl(int do_verification,
                                       int init_method,
                                       bool do_log,
-                                      int nrepeat,
+                                      bool time_kernel,
                                       int M,
                                       int N,
                                       int K,
@@ -259,30 +259,12 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
 
         if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
         {
-            // warm up
-            invoker_ptr->Run(argument_ptr.get());
+            // init DO, D1 to 0
+            d0_device_buf.SetZero();
+            d1_device_buf.SetZero();
 
-            // timing
-            float total_time = 0;
-
-            for(int i = 0; i < nrepeat; ++i)
-            {
-                // init DO, D1 to 0
-                d0_device_buf.SetZero();
-                d1_device_buf.SetZero();
-
-                KernelTimer timer;
-
-                timer.Start();
-
-                invoker_ptr->Run(argument_ptr.get());
-
-                timer.End();
-
-                total_time += timer.GetElapsedTime();
-            }
-
-            float ave_time = total_time / nrepeat;
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::string gemm_name = gemm_ptr->GetTypeString();
 

profiler/include/profile_conv_bwd_data_impl.hpp

@@ -51,7 +51,7 @@ template <int NDimSpatial,
 void profile_conv_bwd_data_impl(int do_verification,
                                 int init_method,
                                 bool do_log,
-                                int nrepeat,
+                                bool time_kernel,
                                 ck::index_t N,
                                 ck::index_t K,
                                 ck::index_t C,
@@ -228,7 +228,8 @@ void profile_conv_bwd_data_impl(int do_verification,
         {
             std::string conv_name = conv_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamControl{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 

profiler/include/profile_conv_bwd_weight_impl.hpp

 #pragma once
+
+#include "stream_config.hpp"
 #include "config.hpp"
 #include "device.hpp"
 #include "host_tensor.hpp"
@@ -43,7 +45,7 @@ template <int NDimSpatial,
 bool profile_conv_bwd_weight_impl(int do_verification,
                                   int init_method,
                                   bool do_log,
-                                  int nrepeat,
+                                  bool time_kernel,
                                   ck::index_t N,
                                   ck::index_t K,
                                   ck::index_t C,
@@ -182,6 +184,7 @@ bool profile_conv_bwd_weight_impl(int do_verification,
 
     // profile device Conv instances
     bool pass = true;
+
     for(auto& conv_ptr : conv_ptrs)
     {
         // using atomic, so need to reset input
@@ -189,6 +192,7 @@ bool profile_conv_bwd_weight_impl(int do_verification,
         {
             wei_device_buf.SetZero();
         }
+
         auto argument_ptr = conv_ptr->MakeArgumentPointer(
             static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
             static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
@@ -214,7 +218,8 @@ bool profile_conv_bwd_weight_impl(int do_verification,
         {
             std::string conv_name = conv_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 
@@ -242,6 +247,7 @@ bool profile_conv_bwd_weight_impl(int do_verification,
                 wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
 
                 float max_error = check_error(wei_k_c_y_x_host_result, wei_k_c_y_x_device_result);
+
                 if(max_error > 8)
                 {
                     pass = false;

profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp

@@ -42,7 +42,7 @@ template <int NDimSpatial,
 void profile_conv_fwd_bias_relu_add_impl(int do_verification,
                                          int init_method,
                                          bool do_log,
-                                         int nrepeat,
+                                         bool time_kernel,
                                          ck::index_t N,
                                          ck::index_t K,
                                          ck::index_t C,
@@ -219,7 +219,8 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
         {
             std::string conv_name = op_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 

profiler/include/profile_conv_fwd_bias_relu_atomic_add_impl.hpp

@@ -119,7 +119,7 @@ template <int NDimSpatial,
 void profile_conv_fwd_bias_relu_atomic_add_impl(int do_verification,
                                                 int init_method,
                                                 bool do_log,
-                                                int nrepeat,
+                                                bool time_kernel,
                                                 ck::index_t N,
                                                 ck::index_t K,
                                                 ck::index_t C,
@@ -275,7 +275,8 @@ void profile_conv_fwd_bias_relu_atomic_add_impl(int do_verification,
         {
             std::string conv_name = op_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 

profiler/include/profile_conv_fwd_bias_relu_impl.hpp

@@ -41,7 +41,7 @@ template <int NDimSpatial,
 void profile_conv_fwd_bias_relu_impl(int do_verification,
                                      int init_method,
                                      bool do_log,
-                                     int nrepeat,
+                                     bool time_kernel,
                                      ck::index_t N,
                                      ck::index_t K,
                                      ck::index_t C,
@@ -207,7 +207,8 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
         {
             std::string conv_name = op_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 

profiler/include/profile_convnd_bwd_data_impl.hpp

 #pragma once
 #include "config.hpp"
 #include "device.hpp"
-#include "conv_fwd_util.hpp"
+#include "conv_util.hpp"
 #include "host_tensor.hpp"
 #include "host_tensor_generator.hpp"
 #include "tensor_layout.hpp"
@@ -269,7 +269,7 @@ template <int NDimSpatial,
 bool profile_convnd_bwd_data_impl(int do_verification,
                                   int init_method,
                                   bool do_log,
-                                  int nrepeat,
+                                  bool time_kernel,
                                   ck::index_t N,
                                   ck::index_t K,
                                   ck::index_t C,
@@ -410,7 +410,8 @@ bool profile_convnd_bwd_data_impl(int do_verification,
         {
             std::string conv_name = conv_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop =
                 ck::utils::conv::get_flops(N, C, K, filter_spatial_lengths, output_spatial_lengths);

profiler/include/profile_gemm_bias_2d_impl.hpp

@@ -65,7 +65,7 @@ template <typename ADataType,
 void profile_gemm_bias_2d_impl(int do_verification,
                                int init_method,
                                bool do_log,
-                               int nrepeat,
+                               bool time_kernel,
                                int M,
                                int N,
                                int K,
@@ -259,7 +259,8 @@ void profile_gemm_bias_2d_impl(int do_verification,
         {
             std::string gemm_name = gemm_ptr->GetTypeString();
 
-            float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
             std::size_t flop = std::size_t(2) * M * N * K;