1#version 450 2#define LOCAL_SZ_X 256 3layout(push_constant) uniform pushBlock { 4 int in_h; 5 int in_w; 6 int out_h; 7 int out_w; 8 int stride_h; 9 int stride_w; 10 int pad_h; 11 int pad_w; 12 int filter_h; 13 int filter_w; 14 int dilation_h; 15 int dilation_w; 16 int channels; 17 int batch; 18 int has_bias; 19 int M; 20 int K; 21 int N; 22 int basic_shader_batch_idx; 23 int basic_shader_partition_idx; 24 int basic_shader_partition_size; 25} p; 26 27layout(binding = 0) readonly buffer Input0{ 28 float in_buffer[]; 29}; 30layout(binding = 1) readonly buffer Input1 { 31 float bias_data[]; 32}; 33layout(binding = 2) readonly buffer Input3{ 34 float weight_data[]; 35}; 36layout(binding = 3) writeonly buffer Output{ 37 float out_buffer[]; 38}; 39layout(local_size_x = LOCAL_SZ_X, local_size_y = 1, local_size_z = 1) in; 40 41/* 42 Each work item compute one output cell 43*/ 44void main() 45{ 46 int gx = int(gl_GlobalInvocationID.x); 47 int gy = int(gl_GlobalInvocationID.y); 48 int gz = int(gl_GlobalInvocationID.z); 49 if(gx < p.out_w && gy < p.out_h && gz < p.channels) 50 { 51 float sum = 0.0f; 52 53 int org_y = gy * p.stride_h - p.pad_h; 54 int org_x = gx * p.stride_w - p.pad_w; 55 56 int weight_off = gz * p.filter_h * p.filter_w; 57 int input_off = (p.basic_shader_batch_idx * p.channels + gz) * p.in_h * p.in_w + org_y * p.in_w + org_x; 58 for(int y = 0; y < p.filter_h; y++) 59 { 60 for(int x = 0; x < p.filter_w; x++) 61 { 62 if(org_y + y * p.dilation_h >= 0 && org_y + y * p.dilation_h < p.in_h && org_x + x * p.dilation_w >= 0 && org_x + x * p.dilation_w < p.in_w) 63 { 64 sum += in_buffer[input_off + x * p.dilation_w] * weight_data[weight_off + x]; 65 } 66 } 67 weight_off += p.filter_w; 68 input_off += p.in_w * p.dilation_h; 69 } 70 71 int offset = (p.basic_shader_batch_idx * p.channels + gz) * p.out_h * p.out_w + gy * p.out_w + gx; 72 if (p.has_bias == 1) 73 out_buffer[offset] = sum + bias_data[gz]; 74 else 75 out_buffer[offset] = sum; 76 } 77} 78