MNN/source/backend/opencl/execution/cl/interp_mnn_cl.cpp

54 lines
3.2 KiB
C++

#include "opencl_source_map.hpp"
namespace MNN {
const char* interp =
"#ifdef MNN_SUPPORT_FP16\n"
"#pragma OPENCL EXTENSION cl_khr_fp16 : enable\n"
"#endif\n"
"#define GLOBAL_SIZE_3_DIMS "" __private const int global_size_dim0,__private const int global_size_dim1,__private const int global_size_dim2,\n"
"#define DEAL_NON_UNIFORM_DIM3(input1, input2, input3) "" if (input1 >= global_size_dim0 || input2 >= global_size_dim1 || input3 >= global_size_dim2) { "" return; "" }\n"
"__constant sampler_t SAMPLER=CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;\n"
"__kernel void interp(GLOBAL_SIZE_3_DIMS __read_only image2d_t input,__write_only image2d_t output,\n"
" __private const float height_scale,__private const float width_scale,\n"
" __private const float height_offset,__private const float width_offset,\n"
" __private const int input_height,__private const int input_width,\n"
" __private const int out_height) {\n"
" const int output_channel_block_idx=get_global_id(0);\n"
" const int output_width_block_idx=get_global_id(1);\n"
" const int output_batch_height_block_idx=get_global_id(2);\n"
" DEAL_NON_UNIFORM_DIM3(output_channel_block_idx,output_width_block_idx,output_batch_height_block_idx);\n"
" const int output_channel_block_idxs=global_size_dim0;\n"
" const int output_width=global_size_dim1;\n"
" const int output_batch_idx=output_batch_height_block_idx/out_height;\n"
" const int output_height_idx=output_batch_height_block_idx % out_height;\n"
" const float scale_height=output_height_idx*height_scale+height_offset;\n"
" const float scale_width=output_width_block_idx*width_scale+width_offset;\n"
"#define CLAMP(val,min_val,max_val) max(min(val,max_val),min_val)\n"
" const int height_floor=(int)floor(scale_height);\n"
" const int height_lf=CLAMP(height_floor,0,input_height-1);\n"
" const int height_uf=CLAMP(height_floor+1,0,input_height-1);\n"
" \n"
" const int width_floor=(int)floor(scale_width);\n"
" const int width_lf=CLAMP(width_floor,0,input_width-1);\n"
" const int width_uf=CLAMP(width_floor+1,0,input_width-1);\n"
" const float height_gap=scale_height-height_floor;\n"
" const float width_gap=scale_width-width_floor;\n"
" const int input_width_offset=mul24(output_channel_block_idx,input_width);\n"
" const int input_height_offset=mul24(output_batch_idx,input_height);\n"
" float4 top_left =\n"
" read_imagef(input,SAMPLER,(int2)(input_width_offset+width_lf,input_height_offset+height_lf));\n"
" float4 top_right =\n"
" read_imagef(input,SAMPLER,(int2)(input_width_offset+width_uf,input_height_offset+height_lf));\n"
" float4 bottom_left =\n"
" read_imagef(input,SAMPLER,(int2)(input_width_offset+width_lf,input_height_offset+height_uf));\n"
" float4 bottom_right =\n"
" read_imagef(input,SAMPLER,(int2)(input_width_offset+width_uf,input_height_offset+height_uf));\n"
" float4 top=mad((top_right-top_left),width_gap,top_left);\n"
" float4 bottom=mad((bottom_right-bottom_left),width_gap,bottom_left);\n"
" float4 out=mad((bottom-top),height_gap,top);\n"
" const int out_image_w=mad24(output_channel_block_idx,output_width,output_width_block_idx);\n"
" const int out_image_h=mad24(output_batch_idx,out_height,output_height_idx);\n"
" write_imagef(output,(int2)(out_image_w,out_image_h),out);\n"
"}\n"
;
}