1 #ifndef RASTER_CU_H 2 #define RASTER_CU_H 3 #include "core/TensorUtils.hpp" 4 #include "backend/cuda/core/runtime/CUDARuntime.hpp" 5 namespace MNN { 6 namespace CUDA { 7 void RasterBlit(uint8_t* output, const uint8_t* input, const int32_t* size, const int32_t* srcStride, const int32_t* dstStride, int bytes, CUDARuntime* runtime); 8 void PackC4(uint8_t* dest, const uint8_t* src, int inside, int axis, int outside, int bytes, CUDARuntime* runtime); 9 void UnpackC4(uint8_t* dest, const uint8_t* src, int inside, int axis, int outside, int bytes, CUDARuntime* runtime); 10 void BlitWithIndice(uint8_t* dest, const uint8_t* src, const int32_t* dstIndices, const int32_t* srcIndices, int dstUseIndice, int srcUseIndice, int loopCount, int dstStep, int srcStep, int srcLimit, const Tensor::InsideDescribe::Region& reg, int bytes, CUDARuntime* runtime); 11 void UnaryBlit(uint8_t* output, const uint8_t* input, const int32_t* size, const int32_t* srcStride, const int32_t* dstStride, int bytes, CUDARuntime* runtime, int opType); 12 void BinaryBlit(uint8_t* output, const uint8_t* input, const uint8_t* input1, const int32_t* size, const int32_t* srcStride, const int32_t* srcStride1, const int32_t* dstStride, halide_type_t type, CUDARuntime* runtime, int opType); 13 } 14 } 15 16 #endif