Review remarks

example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp

@@ -131,7 +131,7 @@ int main(int argc, char* argv[])
 
     std::size_t flop = 0, num_btype = 0;
 
-    for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+    for(std::size_t i = 0; i < gemm_shapes.size(); i++)
     {
         a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
             gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
@@ -168,7 +168,7 @@ int main(int argc, char* argv[])
         }
     }
 
-    for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+    for(std::size_t i = 0; i < gemm_shapes.size(); i++)
     {
         a_tensors_device.emplace_back(
             std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpace()));
@@ -213,7 +213,7 @@ int main(int argc, char* argv[])
 
     if(do_verification)
     {
-        for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+        for(std::size_t i = 0; i < gemm_shapes.size(); i++)
         {
             c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
             auto ref_gemm    = ReferenceGemmInstance{};

include/ck/tensor_operation/gpu/device/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp

@@ -698,7 +698,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
         }
 
         // Gridwise GEMM size
-        for(int i = 0; i < ck::type_convert<int>(arg.a_grid_desc_k0_m_k1_container_.size()); i++)
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
         {
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                             arg.b_grid_desc_k0_n_k1_container_[i],

include/ck/tensor_operation/gpu/device/device_convnd_bwd_data_xdl_ndhwc_kzyxc_ndhwk.hpp

@@ -1413,7 +1413,7 @@ struct DeviceConvndBwdDataXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho
         }
 
         // Gridwise GEMM size
-        for(int i = 0; i < ck::type_convert<int>(arg.a_grid_desc_k0_m_k1_container_.size()); i++)
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
         {
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                             arg.b_grid_desc_k0_n_k1_container_[i],

include/ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp

@@ -301,7 +301,7 @@ struct DeviceGroupedGemmXdl
 
             gemm_desc_kernel_arg_.reserve(group_count_);
 
-            for(index_t i = 0; i < ck::type_convert<index_t>(gemm_shapes.size()); i++)
+            for(std::size_t i = 0; i < gemm_shapes.size(); i++)
             {
                 const index_t M = gemm_shapes[i].M;
                 const index_t N = gemm_shapes[i].N;

library/include/ck/library/host_tensor/host_tensor.hpp

@@ -316,7 +316,7 @@ float check_error(const Tensor<T>& ref, const Tensor<T>& result)
 
     constexpr float eps = 1e-10;
 
-    for(int i = 0; i < ck::type_convert<int>(ref.mData.size()); ++i)
+    for(std::size_t i = 0; i < ref.mData.size(); ++i)
     {
         float ref_v    = ck::type_convert<float>(ref.mData[i]);
         float result_v = ck::type_convert<float>(result.mData[i]);

library/src/host_tensor/host_tensor.cpp

@@ -25,7 +25,7 @@ std::size_t HostTensorDescriptor::GetElementSize() const
 std::size_t HostTensorDescriptor::GetElementSpace() const
 {
     std::size_t space = 1;
-    for(int i = 0; i < ck::type_convert<int>(mLens.size()); ++i)
+    for(std::size_t i = 0; i < mLens.size(); ++i)
     {
         space += (mLens[i] - 1) * mStrides[i];
     }
@@ -68,7 +68,7 @@ void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream
 // FIXME: remove
 void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst)
 {
-    for(int i = 0; i < ck::type_convert<int>(src.mData.size()); ++i)
+    for(std::size_t i = 0; i < src.mData.size(); ++i)
         dst.mData[i] = ck::type_convert<float>(src.mData[i]);
 }
 #endif

profiler/include/profile_convnd_bwd_data_impl.hpp

@@ -222,7 +222,7 @@ static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
 {
     float max_diff = 1e-6;
 
-    for(int i = 0; i < ck::type_convert<int>(ref.mData.size()); ++i)
+    for(std::size_t i = 0; i < ref.mData.size(); ++i)
     {
         float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
         if(max_diff < diff)

profiler/include/profile_grouped_gemm_impl.hpp

@@ -50,12 +50,12 @@ void profile_grouped_gemm_impl(int do_verification,
                                int init_method,
                                bool do_log,
                                int nrepeat,
-                               std::vector<int> Ms,
-                               std::vector<int> Ns,
-                               std::vector<int> Ks,
-                               std::vector<int> StrideAs,
-                               std::vector<int> StrideBs,
-                               std::vector<int> StrideCs)
+                               const std::vector<int>& Ms,
+                               const std::vector<int>& Ns,
+                               const std::vector<int>& Ks,
+                               const std::vector<int>& StrideAs,
+                               const std::vector<int>& StrideBs,
+                               const std::vector<int>& StrideCs)
 {
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
@@ -83,7 +83,7 @@ void profile_grouped_gemm_impl(int do_verification,
     std::vector<Tensor<BDataType>> b_k_n;
     std::vector<Tensor<CDataType>> c_m_n_device_results;
 
-    for(int i = 0; i < ck::type_convert<int>(Ms.size()); i++)
+    for(std::size_t i = 0; i < Ms.size(); i++)
     {
         a_m_k.push_back(
             Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
@@ -144,7 +144,7 @@ void profile_grouped_gemm_impl(int do_verification,
 
     gemm_shapes.reserve(group_count);
 
-    for(int i = 0; i < ck::type_convert<int>(group_count); i++)
+    for(std::size_t i = 0; i < group_count; i++)
     {
         a_device_buf.emplace_back(
             std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpace()));
@@ -234,7 +234,7 @@ void profile_grouped_gemm_impl(int do_verification,
             float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
 
             std::size_t flop = 0, num_btype = 0;
-            for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+            for(std::size_t i = 0; i < gemm_shapes.size(); i++)
             {
                 flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
 
@@ -258,7 +258,7 @@ void profile_grouped_gemm_impl(int do_verification,
 
             if(do_verification)
             {
-                for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+                for(std::size_t i = 0; i < gemm_shapes.size(); i++)
                 {
 
                     c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());

test/gemm_split_k/gemm_split_k.cpp

@@ -45,7 +45,7 @@ static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
 {
     float max_diff = 1e-6;
 
-    for(int i = 0; i < ck::type_convert<int>(ref.mData.size()); ++i)
+    for(std::size_t i = 0; i < ref.mData.size(); ++i)
     {
         float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
         if(max_diff < diff)

test/grouped_gemm/grouped_gemm_fp16.cpp

@@ -104,7 +104,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
     b_tensors_device.reserve(group_count);
     c_tensors_device.reserve(group_count);
 
-    for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+    for(std::size_t i = 0; i < gemm_shapes.size(); i++)
     {
         a_tensors.emplace_back(Tensor<ADataType>(f_host_tensor_descriptor(
             gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
@@ -119,7 +119,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
         b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
     }
 
-    for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+    for(std::size_t i = 0; i < gemm_shapes.size(); i++)
     {
         a_tensors_device.emplace_back(
             std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize()));
@@ -147,7 +147,7 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
 
     invoker_ptr->Run(argument_ptr.get());
 
-    for(int i = 0; i < ck::type_convert<int>(gemm_shapes.size()); i++)
+    for(std::size_t i = 0; i < gemm_shapes.size(); i++)
     {
         c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
 

test/reduce/reduce_util.hpp

@@ -9,7 +9,7 @@ namespace reduce_util {
 template <typename T>
 void to_f32_vector(const Tensor<T>& src, Tensor<float>& dst)
 {
-    for(int i = 0; i < ck::type_convert<int>(src.mData.size()); ++i)
+    for(std::size_t i = 0; i < src.mData.size(); ++i)
         dst.mData[i] = type_convert<float>(src.mData[i]);
 }