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