!7765 transpose fp16 slice calc moved to ReSize and mark run into concat int32

Merge pull request !7765 from zhaozhenlong/lite/issue/I1X1X3_face_detection_fp16_error

mindspore/lite/src/runtime/kernel/arm/base/strided_slice.cc

@@ -136,6 +136,7 @@ kernel::LiteKernel *CpuStridedSliceKernelCreator(const std::vector<lite::Tensor
   return kernel;
 }
 
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_StridedSlice, CpuStridedSliceKernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_StridedSlice, CpuStridedSliceKernelCreator)
 REG_KERNEL(kCPU, kNumberTypeInt8, PrimitiveType_StridedSlice, CpuStridedSliceKernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/transpose_fp16.cc

@@ -31,10 +31,6 @@ using mindspore::schema::PrimitiveType_Transpose;
 
 namespace mindspore::kernel {
 int TransposeFp16CPUKernel::Init() {
-  TransposeParameter *param = reinterpret_cast<TransposeParameter *>(this->op_parameter_);
-  num_unit_ = static_cast<int>(in_tensors_[kInputIndex]->shape().at(param->perm_[kNHWC_H]));
-  thread_h_num_ = MSMIN(thread_num_, num_unit_);
-  thread_h_stride_ = UP_DIV(num_unit_, thread_h_num_);
   if (!InferShapeDone()) {
     return RET_OK;
   }
@@ -42,9 +38,12 @@ int TransposeFp16CPUKernel::Init() {
 }
 
 int TransposeFp16CPUKernel::ReSize() {
+  TransposeParameter *param = reinterpret_cast<TransposeParameter *>(this->op_parameter_);
+  num_unit_ = static_cast<int>(in_tensors_[kInputIndex]->shape().at(param->perm_[kNHWC_H]));
+  thread_h_num_ = MSMIN(thread_num_, num_unit_);
+  thread_h_stride_ = UP_DIV(num_unit_, thread_h_num_);
   auto &in_tensor = in_tensors_.front();
   auto &out_tensor = out_tensors_.front();
-  auto param = reinterpret_cast<TransposeParameter *>(op_parameter_);
   auto in_shape = in_tensor->shape();
   auto out_shape = out_tensor->shape();
   param->strides_[param->num_axes_ - 1] = 1;

mindspore/lite/src/runtime/kernel/arm/fp32/arithmetic.cc

@@ -325,8 +325,10 @@ kernel::LiteKernel *CpuArithmeticFp32KernelCreator(const std::vector<lite::Tenso
 
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Mul, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeInt, PrimitiveType_Mul, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_Mul, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Add, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeInt, PrimitiveType_Add, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_Add, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Sub, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Div, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_RealDiv, CpuArithmeticFp32KernelCreator)
@@ -338,6 +340,8 @@ REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_FloorDiv, CpuArithmeticFp32Ke
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_FloorMod, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeInt, PrimitiveType_FloorDiv, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeInt, PrimitiveType_FloorMod, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_FloorDiv, CpuArithmeticFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_FloorMod, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SquaredDifference, CpuArithmeticFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Eltwise, CpuArithmeticFp32KernelCreator)
 

mindspore/lite/src/runtime/kernel/arm/fp32/slice.cc

@@ -109,5 +109,6 @@ kernel::LiteKernel *CpuSliceFp32KernelCreator(const std::vector<lite::Tensor *>
   return kernel;
 }
 
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_Slice, CpuSliceFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Slice, CpuSliceFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/scheduler.cc

@@ -296,7 +296,8 @@ kernel::LiteKernel *Scheduler::ScheduleNode(const std::vector<Tensor *> &in_tens
 TypeId Scheduler::GetFirstFp32Fp16OrInt8Type(const std::vector<Tensor *> &in_tensors) {
   for (const auto &tensor : in_tensors) {
     auto dtype = tensor->data_type();
-    if (dtype == kNumberTypeFloat32 || dtype == kNumberTypeFloat16 || dtype == kNumberTypeInt8) {
+    if (dtype == kNumberTypeFloat32 || dtype == kNumberTypeFloat16 || dtype == kNumberTypeInt8 ||
+        dtype == kNumberTypeInt32) {
       return dtype;
     }
   }
@@ -341,7 +342,8 @@ kernel::SubGraphType Scheduler::GetKernelSubGraphType(kernel::LiteKernel *kernel
   } else if (desc.arch == kernel::KERNEL_ARCH::kCPU) {
     if (desc.data_type == kNumberTypeFloat16) {
       return kernel::kCpuFP16SubGraph;
-    } else if (desc.data_type == kNumberTypeFloat32 || desc.data_type == kNumberTypeInt8) {
+    } else if (desc.data_type == kNumberTypeFloat32 || desc.data_type == kNumberTypeInt8 ||
+               desc.data_type == kNumberTypeInt32) {
       return kernel::kCpuFP32SubGraph;
     }
   }