/* * Copyright (c) 2017 ARM Limited. * * SPDX-License-Identifier: MIT * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to * deal in the Software without restriction, including without limitation the * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in all * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in; #include "helpers.h" layout(std140) uniform shader_params { TENSOR3D_PARAM_DECLARATION(src); TENSOR3D_PARAM_DECLARATION(dst); TENSOR3D_PARAM_DECLARATION(weights); #ifdef BIAS VECTOR_PARAM_DECLARATION(biases); #endif /* BIAS */ uint weights_stride_w; uint weights_depth; }; #define LOAD12(r, name, offset) \ r.x = LOAD4(name, offset); \ r.y = LOAD4(name, offset + uint(1)); \ r.z = LOAD4(name, offset + uint(2)) #define LOAD3X3(r, name) \ r[0] = LOAD4(name, tensor3D_offset(name, 0, 0, 0)); \ r[1] = LOAD4(name, tensor3D_offset(name, 1, 0, 0)); \ r[2] = LOAD4(name, tensor3D_offset(name, 2, 0, 0)); \ r[3] = LOAD4(name, tensor3D_offset(name, 0, 1, 0)); \ r[4] = LOAD4(name, tensor3D_offset(name, 1, 1, 0)); \ r[5] = LOAD4(name, tensor3D_offset(name, 2, 1, 0)); \ r[6] = LOAD4(name, tensor3D_offset(name, 0, 2, 0)); \ r[7] = LOAD4(name, tensor3D_offset(name, 1, 2, 0)); \ r[8] = LOAD4(name, tensor3D_offset(name, 2, 2, 0)) #if defined(PROCESS_1_ELEMENT) BUFFER_DECLARATION(src, 1, float, readonly); BUFFER_DECLARATION(dst, 2, float, writeonly); BUFFER_DECLARATION(weights, 3, float, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, float, readonly); #endif /* BIAS */ /** This kernel performs a direct convolution to convolve the low three dimensions. * * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); #endif /* BIAS */ float pixels = CONVERT(0, float); uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w >> 2; for(int d = 0; d < int(weights_depth); ++d) { vec3 temp; vec3 w; LOAD12(temp, src, offset(src, 0, 0)); LOAD12(w, weights, tensor3D_offset(weights, 0, 0, 0)); pixels += temp.x * w[0] + temp.y * w[1] + temp.z * w[2]; LOAD12(temp, src, offset(src, 0, 1)); LOAD12(w, weights, tensor3D_offset(weights, 0, 1, 0)); pixels += temp.x * w[0] + temp.y * w[1] + temp.z * w[2]; LOAD12(temp, src, offset(src, 0, 2)); LOAD12(w, weights, tensor3D_offset(weights, 0, 2, 0)); pixels += temp.x * w[0] + temp.y * w[1] + temp.z * w[2]; src.current_offset += src_stride_z >> 2; weights.current_offset += weights_stride_z >> 2; } #ifdef BIAS pixels += LOAD4(biases, vector_offset(biases, int(z_index))); #endif /* BIAS */ STORE4(dst, CURRENT_OFFSET(dst), pixels); } #elif defined(PROCESS_8_ELEMENT) BUFFER_DECLARATION(src, 1, vec4, readonly); BUFFER_DECLARATION(dst, 2, vec4, writeonly); BUFFER_DECLARATION(weights, 3, float, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, float, readonly); #endif /* BIAS */ #if STRIDE_X == 2 #define CONVOLVE1x3(offset, w) convolve1x3_stride2(offset, w) #elif STRIDE_X == 1 /* STRIDE_X == 1 */ #define CONVOLVE1x3(offset, w) convolve1x3_stride1(offset, w) #else /* STRIDE_X not equals 1 or 2 */ #error STRIDE_X larger than 2 is not supported #endif /* STRIDE_X == 2 */ vec4[2] convolve1x3_stride1(uint offset, vec3 w) { vec4 middle; vec4 right; vec4 tmp[3]; vec4 r[2]; LOAD3(tmp, src, offset); middle = vec4(tmp[0].yzw, tmp[1].x); right = vec4(tmp[0].zw, tmp[1].xy); r[0] = tmp[0] * w[0] + middle * w[1] + right * w[2]; middle = vec4(tmp[1].yzw, tmp[2].x); right = vec4(tmp[1].zw, tmp[2].xy); r[1] = tmp[1] * w[0] + middle * w[1] + right * w[2]; return r; } vec4[2] convolve1x3_stride2(uint offset, vec3 w) { vec4 left; vec4 middle; vec4 right; vec4 tmp[3]; vec4 r[2]; LOAD3(tmp, src, offset); left = vec4(tmp[0].xz, tmp[1].xz); middle = vec4(tmp[0].yw, tmp[1].yw); right = vec4(tmp[0].z, tmp[1].xz, tmp[2].x); r[0] = left * w[0] + middle * w[1] + right * w[2]; LOAD2(tmp, src, offset + ((uint(3) * src_stride_x) >> 2)); left = vec4(tmp[2].xz, tmp[0].xz); middle = vec4(tmp[2].yw, tmp[0].yw); right = vec4(tmp[2].z, tmp[0].xz, tmp[1].x); r[1] = left * w[0] + middle * w[1] + right * w[2]; return r; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 8 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); #endif /* BIAS */ vec4 pixels[2]; pixels[0] = vec4(0); pixels[1] = vec4(0); uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w >> 2; for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once vec3 w; vec4 r[2]; // first line LOAD3(w, weights, tensor3D_offset(weights, 0, 0, 0)); r = CONVOLVE1x3(src.current_offset >> uint(2), w); pixels[0] += r[0]; pixels[1] += r[1]; // second line LOAD3(w, weights, tensor3D_offset(weights, 0, 1, 0)); r = CONVOLVE1x3((src.current_offset + (src_stride_y >> 2)) >> uint(2), w); pixels[0] += r[0]; pixels[1] += r[1]; // third line LOAD3(w, weights, tensor3D_offset(weights, 0, 2, 0)); r = CONVOLVE1x3((src.current_offset + (src_stride_y >> 1)) >> uint(2), w); pixels[0] += r[0]; pixels[1] += r[1]; src.current_offset += src_stride_z >> 2; weights.current_offset += weights_stride_z >> 2; } #ifdef BIAS float b; LOAD1(b, biases, vector_offset(biases, int(z_index))); pixels[0] += vec4(b); pixels[1] += vec4(b); #endif /* BIAS */ STORE2(dst, dst.current_offset >> uint(2), pixels); } #elif defined(PROCESS_4_ELEMENT) BUFFER_DECLARATION(src, 1, vec4, readonly); BUFFER_DECLARATION(dst, 2, vec4, writeonly); BUFFER_DECLARATION(weights, 3, float, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, float, readonly); #endif /* BIAS */ #if STRIDE_X == 2 #define CONVOLVE1x3(offset, w) convolve1x3_stride2(offset, w) #elif STRIDE_X == 1 /* STRIDE_X == 1 */ #define CONVOLVE1x3(offset, w) convolve1x3_stride1(offset, w) #else /* STRIDE_X not equals 1 or 2 */ #error STRIDE_X larger than 2 is not supported #endif /* STRIDE_X == 2 */ vec4 convolve1x3_stride1(uint offset, vec3 w) { vec4 tmp[2]; vec4 middle; vec4 right; LOAD2(tmp, src, offset); middle = vec4(tmp[0].yzw, tmp[1].x); right = vec4(tmp[0].zw, tmp[1].xy); tmp[1] = tmp[0] * w[0] + middle * w[1] + right * w[2]; return tmp[1]; } vec4 convolve1x3_stride2(uint offset, vec3 w) { vec4 left; vec4 middle; vec4 right; vec4 tmp[3]; LOAD3(tmp, src, offset); left = vec4(tmp[0].xz, tmp[1].xz); middle = vec4(tmp[0].yw, tmp[1].yw); right = vec4(tmp[0].z, tmp[1].xz, tmp[2].x); tmp[0] = left * w[0] + middle * w[1] + right * w[2]; return tmp[0]; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); #endif /* BIAS */ vec4 pixels; pixels = vec4(0); uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w >> 2; for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once vec3 w; // first line LOAD3(w, weights, tensor3D_offset(weights, 0, 0, 0)); pixels += CONVOLVE1x3(src.current_offset >> uint(2), w); // second line LOAD3(w, weights, tensor3D_offset(weights, 0, 1, 0)); pixels += CONVOLVE1x3((src.current_offset + (src_stride_y >> 2)) >> uint(2), w); // third line LOAD3(w, weights, tensor3D_offset(weights, 0, 2, 0)); pixels += CONVOLVE1x3((src.current_offset + (src_stride_y >> 1)) >> uint(2), w); src.current_offset += src_stride_z >> 2; weights.current_offset += weights_stride_z >> 2; } #ifdef BIAS float b; LOAD1(b, biases, vector_offset(biases, int(z_index))); pixels += vec4(b); #endif /* BIAS */ STORE1(dst, dst.current_offset >> uint(2), pixels); } #elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS) BUFFER_DECLARATION(src, 1, vec4, readonly); BUFFER_DECLARATION(dst, 2, vec4, writeonly); BUFFER_DECLARATION(weights, 3, float, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, float, readonly); #endif /* BIAS */ #define CONVOLVE1x3(left, middle, right, w) convolve1x3_stride1(left, middle, right, w) vec4 convolve1x3_stride1(vec4 left, vec4 middle, vec4 right, vec3 w) { vec4 r; r = left * w[0] + middle * w[1] + right * w[2]; return r; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x3 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases); #endif /* BIAS */ vec4 pixels[3]; pixels[0] = vec4(0); pixels[1] = vec4(0); pixels[2] = vec4(0); uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w >> 2; for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once vec3 w[3]; LOAD3(w[0], weights, tensor3D_offset(weights, 0, 0, 0)); LOAD3(w[1], weights, tensor3D_offset(weights, 0, 1, 0)); LOAD3(w[2], weights, tensor3D_offset(weights, 0, 2, 0)); vec4 s[2]; vec4 middle; vec4 right; // first line LOAD2(s, src, src.current_offset >> uint(2)); middle = vec4(s[0].yzw, s[1].x); right = vec4(s[0].zw, s[1].xy); pixels[0] += CONVOLVE1x3(s[0], middle, right, w[0]); // second line LOAD2(s, src, (src.current_offset + (src_stride_y >> 2)) >> uint(2)); middle = vec4(s[0].yzw, s[1].x); right = vec4(s[0].zw, s[1].xy); pixels[0] += CONVOLVE1x3(s[0], middle, right, w[1]); pixels[1] += CONVOLVE1x3(s[0], middle, right, w[0]); // third line LOAD2(s, src, (src.current_offset + (src_stride_y >> 1)) >> uint(2)); middle = vec4(s[0].yzw, s[1].x); right = vec4(s[0].zw, s[1].xy); pixels[0] += CONVOLVE1x3(s[0], middle, right, w[2]); pixels[1] += CONVOLVE1x3(s[0], middle, right, w[1]); pixels[2] += CONVOLVE1x3(s[0], middle, right, w[0]); // forth line LOAD2(s, src, (src.current_offset + (uint(3) * (src_stride_y >> 2))) >> uint(2)); middle = vec4(s[0].yzw, s[1].x); right = vec4(s[0].zw, s[1].xy); pixels[1] += CONVOLVE1x3(s[0], middle, right, w[2]); pixels[2] += CONVOLVE1x3(s[0], middle, right, w[1]); // fifth line LOAD2(s, src, (src.current_offset + (src_stride_y)) >> uint(2)); middle = vec4(s[0].yzw, s[1].x); right = vec4(s[0].zw, s[1].xy); pixels[2] += CONVOLVE1x3(s[0], middle, right, w[2]); src.current_offset += src_stride_z >> 2; weights.current_offset += weights_stride_z >> 2; } #ifdef BIAS float b; LOAD1(b, biases, vector_offset(biases, int(z_index))); pixels[0] += vec4(b); pixels[1] += vec4(b); pixels[2] += vec4(b); #endif /* BIAS */ STORE1(dst, dst.current_offset >> uint(2), pixels[0]); STORE1(dst, (dst.current_offset + (dst_stride_y >> 2)) >> uint(2), pixels[1]); STORE1(dst, (dst.current_offset + (dst_stride_y >> 1)) >> uint(2), pixels[2]); } #elif defined(PROCESS_X_8ELEMENTS_Y_3ELEMENTS_FP16) precision mediump float; BUFFER_DECLARATION(src, 1, uvec4, readonly); BUFFER_DECLARATION(dst, 2, uvec4, writeonly); BUFFER_DECLARATION(weights, 3, uint, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, uint, readonly); #endif /* BIAS */ #define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w) vec4[2] convolve1x3_stride1(vec4 tmp[3], vec3 w) { vec4 middle; vec4 right; vec4 r[2]; middle = vec4(tmp[0].yzw, tmp[1].x); right = vec4(tmp[0].zw, tmp[1].xy); r[0] = tmp[0] * w[0] + middle * w[1] + right * w[2]; middle = vec4(tmp[1].yzw, tmp[2].x); right = vec4(tmp[1].zw, tmp[2].xy); r[1] = tmp[1] * w[0] + middle * w[1] + right * w[2]; return r; } vec4[3] load_and_unpack(uint offset) { uvec4 packed_s[2]; vec4 s[3]; LOAD1(packed_s[0], src, offset); LOAD1(packed_s[1], src, offset + uint(1)); ; s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); s[1] = vec4(unpackHalf2x16(packed_s[0].z), unpackHalf2x16(packed_s[0].w)); s[2] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); return s; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 8x3 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT_FP16(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases); #endif /* BIAS */ uvec2 packed_d[2]; uvec4 vd; vec4 pixels[3][2]; int i, j; for(i = 0; i < 3; i++) { for(j = 0; j < 2; j++) { pixels[i][j] = vec4(0); } } uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w; for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once uvec2 packed_w[3]; LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2); LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2); LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2); vec3 w[3]; w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x); w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x); w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x); uvec4 packed_s[2]; vec4 s[3]; vec4 r[2]; uint offset; // first line offset = src.current_offset >> uint(4); s = load_and_unpack(offset); r = CONVOLVE1x3(s, w[0]); pixels[0][0] += r[0]; pixels[0][1] += r[1]; // second line offset = (src.current_offset + src_stride_y) >> uint(4); s = load_and_unpack(offset); r = CONVOLVE1x3(s, w[1]); pixels[0][0] += r[0]; pixels[0][1] += r[1]; r = CONVOLVE1x3(s, w[0]); pixels[1][0] += r[0]; pixels[1][1] += r[1]; // third line offset = (src.current_offset + (src_stride_y << 1)) >> uint(4); s = load_and_unpack(offset); r = CONVOLVE1x3(s, w[2]); pixels[0][0] += r[0]; pixels[0][1] += r[1]; r = CONVOLVE1x3(s, w[1]); pixels[1][0] += r[0]; pixels[1][1] += r[1]; r = CONVOLVE1x3(s, w[0]); pixels[2][0] += r[0]; pixels[2][1] += r[1]; // forth line offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(4); s = load_and_unpack(offset); r = CONVOLVE1x3(s, w[2]); pixels[1][0] += r[0]; pixels[1][1] += r[1]; r = CONVOLVE1x3(s, w[1]); pixels[2][0] += r[0]; pixels[2][1] += r[1]; // fifth line offset = (src.current_offset + (src_stride_y << 2)) >> uint(4); s = load_and_unpack(offset); r = CONVOLVE1x3(s, w[2]); pixels[2][0] += r[0]; pixels[2][1] += r[1]; src.current_offset += src_stride_z; weights.current_offset += weights_stride_z; } #ifdef BIAS uint packed_b; float b; LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2); if(z_index % uint(2) == uint(0)) { b = unpackHalf2x16(packed_b).x; } else { b = unpackHalf2x16(packed_b).y; } for(i = 0; i < 3; i++) { for(j = 0; j < 2; j++) { pixels[i][j] += vec4(b); } } #endif /* BIAS */ packed_d[0] = uvec2(packHalf2x16(pixels[0][0].xy), packHalf2x16(pixels[0][0].zw)); packed_d[1] = uvec2(packHalf2x16(pixels[0][1].xy), packHalf2x16(pixels[0][1].zw)); vd = uvec4(packed_d[0], packed_d[1]); STORE1(dst, dst.current_offset >> uint(4), vd); packed_d[0] = uvec2(packHalf2x16(pixels[1][0].xy), packHalf2x16(pixels[1][0].zw)); packed_d[1] = uvec2(packHalf2x16(pixels[1][1].xy), packHalf2x16(pixels[1][1].zw)); vd = uvec4(packed_d[0], packed_d[1]); STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(4), vd); packed_d[0] = uvec2(packHalf2x16(pixels[2][0].xy), packHalf2x16(pixels[2][0].zw)); packed_d[1] = uvec2(packHalf2x16(pixels[2][1].xy), packHalf2x16(pixels[2][1].zw)); vd = uvec4(packed_d[0], packed_d[1]); STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(4), vd); } #elif defined(PROCESS_X_4ELEMENTS_FP16) precision mediump float; BUFFER_DECLARATION(src, 1, uvec2, readonly); BUFFER_DECLARATION(dst, 2, uvec2, writeonly); BUFFER_DECLARATION(weights, 3, uint, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, uint, readonly); #endif /* BIAS */ #if STRIDE_X == 2 #define CONVOLVE1x3(s, w) convolve1x3_stride2(s, w) #define LOAD_AND_UNPACK(offset) load_and_unpack_stride2(offset) #elif STRIDE_X == 1 /* STRIDE_X == 1 */ #define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w) #define LOAD_AND_UNPACK(offset) load_and_unpack_stride1(offset) #else /* STRIDE_X not equals 1 or 2 */ #error STRIDE_X larger than 2 is not supported #endif /* STRIDE_X == 2 */ vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w) { vec4 middle; vec4 right; vec4 r; middle = vec4(tmp[0].yzw, tmp[1].x); right = vec4(tmp[0].zw, tmp[1].xy); r = tmp[0] * w[0] + middle * w[1] + right * w[2]; return r; } vec4 convolve1x3_stride2(vec4 tmp[3], vec3 w) { vec4 left; vec4 middle; vec4 right; vec4 r; left = vec4(tmp[0].xz, tmp[1].xz); middle = vec4(tmp[0].yw, tmp[1].yw); right = vec4(tmp[0].z, tmp[1].xz, tmp[2].x); r = left * w[0] + middle * w[1] + right * w[2]; return r; } vec4[2] load_and_unpack_stride1(uint offset) { uvec2 packed_s[2]; vec4 s[2]; LOAD1(packed_s[0], src, offset); LOAD1(packed_s[1], src, offset + uint(1)); s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); return s; } vec4[3] load_and_unpack_stride2(uint offset) { uvec2 packed_s[3]; vec4 s[3]; LOAD1(packed_s[0], src, offset); LOAD1(packed_s[1], src, offset + uint(1)); LOAD1(packed_s[2], src, offset + uint(2)); s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); s[2] = vec4(unpackHalf2x16(packed_s[2].x), unpackHalf2x16(packed_s[2].y)); return s; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT_FP16(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases); #endif /* BIAS */ uvec2 packed_d; vec4 pixels = vec4(0); uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w; for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once uvec2 packed_w[3]; LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2); LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2); LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2); vec3 w[3]; w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x); w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x); w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x); #if STRIDE_X == 2 vec4 s[3]; #elif STRIDE_X == 1 /* STRIDE_X == 1 */ vec4 s[2]; #else /* STRIDE_X not equals 1 or 2 */ #error STRIDE_X larger than 2 is not supported #endif /* STRIDE_X == 2 */ vec4 r; uint offset; // first line offset = src.current_offset >> uint(3); s = LOAD_AND_UNPACK(offset); pixels += CONVOLVE1x3(s, w[0]); // second line offset = (src.current_offset + src_stride_y) >> uint(3); s = LOAD_AND_UNPACK(offset); pixels += CONVOLVE1x3(s, w[1]); // third line offset = (src.current_offset + (src_stride_y << 1)) >> uint(3); s = LOAD_AND_UNPACK(offset); pixels += CONVOLVE1x3(s, w[2]); src.current_offset += src_stride_z; weights.current_offset += weights_stride_z; } #ifdef BIAS uint packed_b; float b; LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2); if(z_index % uint(2) == uint(0)) { b = unpackHalf2x16(packed_b).x; } else { b = unpackHalf2x16(packed_b).y; } pixels += vec4(b); #endif /* BIAS */ packed_d = uvec2(packHalf2x16(pixels.xy), packHalf2x16(pixels.zw)); STORE1(dst, dst.current_offset >> uint(3), packed_d); } #elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS_FP16) precision mediump float; BUFFER_DECLARATION(src, 1, uvec2, readonly); BUFFER_DECLARATION(dst, 2, uvec2, writeonly); BUFFER_DECLARATION(weights, 3, uint, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, uint, readonly); #endif /* BIAS */ #define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w) vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w) { vec4 middle; vec4 right; vec4 r; middle = vec4(tmp[0].yzw, tmp[1].x); right = vec4(tmp[0].zw, tmp[1].xy); r = tmp[0] * w[0] + middle * w[1] + right * w[2]; return r; } vec4[2] load_and_unpack(uint offset) { uvec2 packed_s[2]; vec4 s[2]; LOAD1(packed_s[0], src, offset); LOAD1(packed_s[1], src, offset + uint(1)); s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); return s; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x3 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT_FP16(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases); #endif /* BIAS */ uvec2 packed_d; vec4 pixels[3]; int i; for(i = 0; i < 3; i++) { pixels[i] = vec4(0); } uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w; for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once uvec2 packed_w[3]; LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2); LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2); LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2); vec3 w[3]; w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x); w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x); w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x); vec4 s[2]; vec4 r; uint offset; // first line offset = src.current_offset >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[0]); // second line offset = (src.current_offset + src_stride_y) >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[1]); pixels[1] += CONVOLVE1x3(s, w[0]); // third line offset = (src.current_offset + (src_stride_y << 1)) >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[2]); pixels[1] += CONVOLVE1x3(s, w[1]); pixels[2] += CONVOLVE1x3(s, w[0]); // forth line offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(3); s = load_and_unpack(offset); pixels[1] += CONVOLVE1x3(s, w[2]); pixels[2] += CONVOLVE1x3(s, w[1]); // fifth line offset = (src.current_offset + (src_stride_y << 2)) >> uint(3); s = load_and_unpack(offset); pixels[2] += CONVOLVE1x3(s, w[2]); src.current_offset += src_stride_z; weights.current_offset += weights_stride_z; } #ifdef BIAS uint packed_b; float b; LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2); if(z_index % uint(2) == uint(0)) { b = unpackHalf2x16(packed_b).x; } else { b = unpackHalf2x16(packed_b).y; } for(i = 0; i < 3; i++) { pixels[i] += vec4(b); } #endif /* BIAS */ packed_d = uvec2(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw)); STORE1(dst, dst.current_offset >> uint(3), packed_d); packed_d = uvec2(packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw)); STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(3), packed_d); packed_d = uvec2(packHalf2x16(pixels[2].xy), packHalf2x16(pixels[2].zw)); STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(3), packed_d); } #elif defined(PROCESS_X_4ELEMENTS_Y_4ELEMENTS_FP16) precision mediump float; BUFFER_DECLARATION(src, 1, uvec2, readonly); BUFFER_DECLARATION(dst, 2, uvec2, writeonly); BUFFER_DECLARATION(weights, 3, uint, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, uint, readonly); #endif /* BIAS */ #define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w) vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w) { vec4 middle; vec4 right; vec4 r; middle = vec4(tmp[0].yzw, tmp[1].x); right = vec4(tmp[0].zw, tmp[1].xy); r = tmp[0] * w[0] + middle * w[1] + right * w[2]; return r; } vec4[2] load_and_unpack(uint offset) { uvec2 packed_s[2]; vec4 s[2]; LOAD1(packed_s[0], src, offset); LOAD1(packed_s[1], src, offset + uint(1)); s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); return s; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x4 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT_FP16(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases); #endif /* BIAS */ uvec2 packed_d; vec4 pixels[4]; int i; for(i = 0; i < 4; i++) { pixels[i] = vec4(0); } uint z_index = gl_GlobalInvocationID.z; weights.current_offset += z_index * weights_stride_w; for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once uvec2 packed_w[3]; LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2); LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2); LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2); vec3 w[3]; w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x); w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x); w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x); vec4 s[2]; vec4 r; uint offset; // first line offset = src.current_offset >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[0]); // second line offset = (src.current_offset + src_stride_y) >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[1]); pixels[1] += CONVOLVE1x3(s, w[0]); // third line offset = (src.current_offset + (src_stride_y << 1)) >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[2]); pixels[1] += CONVOLVE1x3(s, w[1]); pixels[2] += CONVOLVE1x3(s, w[0]); // forth line offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(3); s = load_and_unpack(offset); pixels[1] += CONVOLVE1x3(s, w[2]); pixels[2] += CONVOLVE1x3(s, w[1]); pixels[3] += CONVOLVE1x3(s, w[0]); // fifth line offset = (src.current_offset + (src_stride_y << 2)) >> uint(3); s = load_and_unpack(offset); pixels[2] += CONVOLVE1x3(s, w[2]); pixels[3] += CONVOLVE1x3(s, w[1]); // sixth line offset = (src.current_offset + uint(5) * (src_stride_y)) >> uint(3); s = load_and_unpack(offset); pixels[3] += CONVOLVE1x3(s, w[2]); src.current_offset += src_stride_z; weights.current_offset += weights_stride_z; } #ifdef BIAS uint packed_b; float b; LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2); if(z_index % uint(2) == uint(0)) { b = unpackHalf2x16(packed_b).x; } else { b = unpackHalf2x16(packed_b).y; } for(i = 0; i < 4; i++) { pixels[i] += vec4(b); } #endif /* BIAS */ packed_d = uvec2(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw)); STORE1(dst, dst.current_offset >> uint(3), packed_d); packed_d = uvec2(packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw)); STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(3), packed_d); packed_d = uvec2(packHalf2x16(pixels[2].xy), packHalf2x16(pixels[2].zw)); STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(3), packed_d); packed_d = uvec2(packHalf2x16(pixels[3].xy), packHalf2x16(pixels[3].zw)); STORE1(dst, (dst.current_offset + uint(3) * (dst_stride_y)) >> uint(3), packed_d); } #elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS_Z_2ELEMENTS_FP16) precision mediump float; BUFFER_DECLARATION(src, 1, uvec2, readonly); BUFFER_DECLARATION(dst, 2, uvec2, writeonly); BUFFER_DECLARATION(weights, 3, uint, readonly); #ifdef BIAS BUFFER_DECLARATION(biases, 4, uint, readonly); #endif /* BIAS */ #define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w) vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w) { vec4 middle; vec4 right; vec4 r; middle = vec4(tmp[0].yzw, tmp[1].x); right = vec4(tmp[0].zw, tmp[1].xy); r = tmp[0] * w[0] + middle * w[1] + right * w[2]; return r; } vec4[2] load_and_unpack(uint offset) { uvec2 packed_s[2]; vec4 s[2]; LOAD1(packed_s[0], src, offset); LOAD1(packed_s[1], src, offset + uint(1)); s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y)); s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y)); return s; } /** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x3x2 elements at once * * @note This OpenGL ES shader works with stride_x = 1 and 2 * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16" * @note If biases are used then "define HAS_BIAS" has to be passed at compile time * * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) * @param[in] dst_step_y dst_stride_y * number of elements along Z processed per workitem(in bytes) * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor * @param[out] weights_ptr Pointer to the weights tensor. Supported data types: same as @p src_ptr * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes) * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes) * @param[in] weights_step_y weights_stride_y * number of elements along y processed per workitem(in bytes) * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes) * @param[in] weights_step_z weights_stride_z * number of elements along Z processed per workitem(in bytes) * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor * @param[in] biases_ptr Pointer to the biases tensor. Same as @p src_ptr * @param[in] biases_stride_x Stride of the biases tensor in X dimension (in bytes) * @param[in] biases_step_x biases_stride_x * number of elements along X processed per workitem(in bytes) * @param[in] biases_offset_first_element_in_bytes The offset of the first element in the biases tensor * @param[in] weights_stride_w Stride of the weights tensor in the 4th dimension * @param[in] weights_depth The third dimensions of the weights tensors */ void main() { Image src = CONVERT_TO_IMAGE_STRUCT_FP16(src); Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights); Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst); #ifdef BIAS Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases); #endif /* BIAS */ uvec2 packed_d; vec4 pixels[3]; int i; uint z_base_index = gl_GlobalInvocationID.z << 1; // store orginal src current offset uint s_offset = src.current_offset; weights.current_offset += z_base_index * weights_stride_w; for(int z = 0; z < 2; ++z) { uint z_index = z_base_index + uint(z); src.current_offset = s_offset; //weights.current_offset = z_index * weights_stride_w; for(i = 0; i < 3; i++) { pixels[i] = vec4(0); } for(int d = 0; d < int(weights_depth); ++d) { // load 3 weights once uvec2 packed_w[3]; LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2); LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2); LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2); vec3 w[3]; w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x); w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x); w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x); vec4 s[2]; vec4 r; uint offset; // first line offset = src.current_offset >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[0]); // second line offset = (src.current_offset + src_stride_y) >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[1]); pixels[1] += CONVOLVE1x3(s, w[0]); // third line offset = (src.current_offset + (src_stride_y << 1)) >> uint(3); s = load_and_unpack(offset); pixels[0] += CONVOLVE1x3(s, w[2]); pixels[1] += CONVOLVE1x3(s, w[1]); pixels[2] += CONVOLVE1x3(s, w[0]); // forth line offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(3); s = load_and_unpack(offset); pixels[1] += CONVOLVE1x3(s, w[2]); pixels[2] += CONVOLVE1x3(s, w[1]); // fifth line offset = (src.current_offset + (src_stride_y << 2)) >> uint(3); s = load_and_unpack(offset); pixels[2] += CONVOLVE1x3(s, w[2]); src.current_offset += src_stride_z; weights.current_offset += weights_stride_z; } #ifdef BIAS uint packed_b; float b; LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2); if(z_index % uint(2) == uint(0)) { b = unpackHalf2x16(packed_b).x; } else { b = unpackHalf2x16(packed_b).y; } for(i = 0; i < 3; i++) { pixels[i] += vec4(b); } #endif /* BIAS */ packed_d = uvec2(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw)); STORE1(dst, dst.current_offset >> uint(3), packed_d); packed_d = uvec2(packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw)); STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(3), packed_d); packed_d = uvec2(packHalf2x16(pixels[2].xy), packHalf2x16(pixels[2].zw)); STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(3), packed_d); dst.current_offset += dst_stride_z; } } #endif /* PROCESS_1_ELEMENT */