/* * 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_cs.h" #if defined(DATA_TYPE_FP16) precision mediump float; #endif // DATA_TYPE_FP16 /** Performs a pooling function * * @note The data type must be passed at compile time using "#define DATA_TYPE_NAME". e.g. "#define DATA_TYPE_FP32" * @note The pool size must be passed at compile time using "#define POOLING_LAYER_n". e.g. "#define POOLING_LAYER_2" * n must be one of these: 2, 3, 7, N * Pool size must be passed using POOL_SIZE if POOLING_LAYER_N is defined. e.g. POOL_SIZE=13; * @note In case of average pooling the following information must be passed at compile time: * POOL_AVG must be provided otherwise max pooling will be performed. * MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) * STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions * PAD_X and PAD_Y which are the pooling paddings in x and y dimension * * @param[in] src_ptr Pointer to the source image. Supported data types: F32/F16 * @param[in] src_attrs The attributes of the source image * @param[out] dst_ptr Pointer to the destination image. Supported data types: same as @p src_ptr * @param[in] src_attrs The attributes of the destination image */ SHADER_PARAMS_DECLARATION { Tensor3DAttributes src_attrs; Tensor3DAttributes dst_attrs; }; // Common definitions #if defined(POOL_AVG) || defined(POOL_L2) #define POOL_OP(res, a, b) ((res) = (a) + (b)) #define POOL_OP_float(res, a, b) (res = a + b) #define POOL_OP_vec2(res, a, b) ((res) = (a) + (b)) #else /* defined(POOL_AVG) || defined(POOL_L2) */ #define POOL_OP(res, a, b) \ (res) = (a); \ if(isnan(a.x) || (a.x < b.x)) \ { \ res.x = b.x; \ } \ if(isnan(a.y) || (a.y < b.y)) \ { \ res.y = b.y; \ } \ if(isnan(a.z) || (a.z < b.z)) \ { \ res.z = b.z; \ } \ if(isnan(a.w) || (a.w < b.w)) \ { \ res.w = b.w; \ } #define POOL_OP_float(res, a, b) \ (res) = (a); \ if(isnan(a) || (a < b)) \ { \ res = b; \ } #define POOL_OP_vec2(res, a, b) \ (res) = (a); \ if(isnan(a.x) || (a.x < b.x)) \ { \ res.x = b.x; \ } \ if(isnan(a.y) || (a.y < b.y)) \ { \ res.y = b.y; \ } #endif /* defined(POOL_AVG) || defined(POOL_L2) */ #if defined(POOL_L2) #define POW2_OP(x, vec_size) ((x) * (x)) #else /* defined(POOL_L2) */ #define POW2_OP(x, vec_size) (x) #endif /* defined(POOL_L2) */ #define DIV_OP(x, y) (x * (1.f / y)) #define SQRT_OP(x) sqrt((x)) #if defined(DATA_TYPE_FP32) float calculate_max(const int, Tensor3DIterator, const int, const int, const int, const int, const int, const int); float calculate_avg(const int, Tensor3DIterator, const int, const int, const int, const int, const int, const int); TENSOR_DECLARATION(1, srcBuffer, float, src_ptr, src_shift, 2, readonly); TENSOR_DECLARATION(2, dstBuffer, float, dst_ptr, dst_shift, 2, writeonly); #if defined(POOL_SIZE) // Set the initial value for the pooling operation accordingly with the data type #if defined(POOL_AVG) || defined(POOL_L2) #define INITIAL_VALUE 0.0f #else /* defined(POOL_AVG) || defined(POOL_L2) */ #define INITIAL_VALUE -3.402823466385289e+38 #endif // POOL_AVG #endif //POOL_SIZE float calculate_max(const int pool_size, Tensor3DIterator src_iter, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y) { int start_x = int(gl_GlobalInvocationID.x) * stride_x - pad_x; int start_y = int(gl_GlobalInvocationID.y) * stride_y - pad_y; int end_x = int(min(start_x + pool_size, upper_bound_w)); int end_y = int(min(start_y + pool_size, upper_bound_h)); float data_max; data_max = LOAD_CURRENT_ITEM(src_ptr, src_iter); for(int i = 0; (start_y + i) < end_y; ++i) { for(int j = 0; (start_x + j) < end_x; ++j) { float data = LOAD(src_ptr, TENSOR3D_OFFSET(src_iter, j, i, 0)); POOL_OP_float(data_max, data_max, data); } } return data_max; } float calculate_avg(const int pool_size, Tensor3DIterator src_iter, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y) { int start_x = int(gl_GlobalInvocationID.x) * stride_x - pad_x; int start_y = int(gl_GlobalInvocationID.y) * stride_y - pad_y; int end_x = int(min(start_x + pool_size, upper_bound_w)); int end_y = int(min(start_y + pool_size, upper_bound_h)); float data_total = 0.0f; for(int i = 0; (start_x + i) < end_x; i++) { for(int j = 0; (start_y + j) < end_y; ++j) { float data = LOAD(src_ptr, TENSOR3D_OFFSET(src_iter, i, j, 0)); if(isnan(data)) { data = 0.0f; } #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data = POW2_OP(data, 1); #endif /* defined(POOL_L2) */ data_total = data_total + data; } } #if defined(EXCLUDE_PADDING) start_x = max(0, start_x); start_y = max(0, start_y); #endif /* defined(EXCLUDE_PADDING) */ return data_total / float((end_y - start_y) * (end_x - start_x)); } #if defined(POOLING_LAYER_2) || defined(POOLING_LAYER_3) || defined(POOLING_LAYER_7) #if defined(POOLING_LAYER_2) #define POOL_SIZE 2 #elif defined(POOLING_LAYER_3) #define POOL_SIZE 3 #elif defined(POOLING_LAYER_7) #define POOL_SIZE 7 #else // POOLING_LAYER_n #error Please define POOLING_LAYER_N instead. #endif // POOLING_LAYER_n void main(void) { // Get pixels pointer Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); //Load and calculate data float res; #if defined(POOL_AVG) || defined(POOL_L2) res = calculate_avg(POOL_SIZE, src_iter, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y); #else /*POOL_AVG*/ res = calculate_max(POOL_SIZE, src_iter, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y); #endif /*POOL_AVG*/ #if defined(POOL_L2) // Take square root of the result in L2 pooling res = SQRT_OP(res); #endif /* defined(POOL_L2) */ // Store result STORE_CURRENT_ITEM(dst_ptr, dst_iter, res); } #elif defined(POOLING_LAYER_3_OPTIMIZED) #define POOLING3x3_STRIDE1(res, input_ptr, input_iter) \ vec4 data00 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0)); \ vec2 data01 = VLOAD2(vec2, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(4)); \ vec4 data10 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0)); \ vec2 data11 = VLOAD2(vec2, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(4)); \ vec4 data20 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0)); \ vec2 data21 = VLOAD2(vec2, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(4)); \ data00 = POW2_OP(data00, 4); \ data01 = POW2_OP(data01, 2); \ data10 = POW2_OP(data10, 4); \ data11 = POW2_OP(data11, 2); \ data20 = POW2_OP(data20, 4); \ data21 = POW2_OP(data21, 2); \ \ vec4 values000; \ vec4 values001; \ vec4 values010; \ vec4 values100; \ vec4 values101; \ vec4 values11; \ vec4 values200; \ vec4 values201; \ vec4 values21; \ values000.xyzw = data00.xyzy; \ values001.xyzw = data00.zwzw; \ values010.x = data01.x; \ values010.y = data00.w; \ values010.zw = data01.xy; \ values100.xyzw = data10.xyzy; \ values101.xyzw = data10.zwzw; \ values11.x = data11.x; \ values11.y = data10.w; \ values11.zw = data11.xy; \ values200.xyzw = data20.xyzy; \ values201.xyzw = data20.zwzw; \ values21.x = data21.x; \ values21.y = data20.w; \ values21.zw = data21.xy; \ POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw); \ POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw); \ POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \ POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw)) #define POOLING3x3_STRIDE2(res, input_ptr, input_iter) \ vec4 data000 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0)); \ vec4 data001 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(4)); \ float data010 = LOAD(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(8)); \ vec4 data100 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0)); \ vec4 data101 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(4)); \ float data11 = LOAD(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(8)); \ vec4 data200 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0)); \ vec4 data201 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(4)); \ float data21 = LOAD(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(8)); \ data000 = POW2_OP(data000, 4); \ data001 = POW2_OP(data001, 4); \ data010 = POW2_OP(data010, 1); \ data100 = POW2_OP(data100, 4); \ data101 = POW2_OP(data101, 4); \ data11 = POW2_OP(data11, 1); \ data200 = POW2_OP(data200, 4); \ data201 = POW2_OP(data201, 4); \ data21 = POW2_OP(data21, 1); \ \ vec4 values000; \ vec4 values001; \ vec4 values010; \ vec4 values100; \ vec4 values101; \ vec4 values11; \ vec4 values200; \ vec4 values201; \ vec4 values21; \ values000.xyzw = data000.xyzz; \ values001.xyzw = vec4(data000.w, data001.xxy); \ values010.xyzw = vec4(data001.zzw, data010); \ values100.xyzw = data100.xyzz; \ values101.xyzw = vec4(data100.w, data101.xxy); \ values11.xyzw = vec4(data101.zzw, data11); \ values200.xyzw = data200.xyzz; \ values201.xyzw = vec4(data200.w, data201.xxy); \ values21.xyzw = vec4(data201.zzw, data21); \ POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw); \ POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw); \ POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \ POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw)) #define POOLING3x3_STRIDE3(res, input_ptr, input_iter) \ vec4 data000 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0)); \ vec4 data001 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(4)); \ vec4 data010 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(8)); \ vec4 data100 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0)); \ vec4 data101 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(4)); \ vec4 data11 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(8)); \ vec4 data200 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0)); \ vec4 data201 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(4)); \ vec4 data21 = VLOAD4(vec4, input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(8)); \ data000 = POW2_OP(data000, 4); \ data001 = POW2_OP(data001, 4); \ data010 = POW2_OP(data010, 4); \ data100 = POW2_OP(data100, 4); \ data101 = POW2_OP(data101, 4); \ data11 = POW2_OP(data11, 4); \ data200 = POW2_OP(data200, 4); \ data201 = POW2_OP(data201, 4); \ data21 = POW2_OP(data21, 4); \ \ POOL_OP(data000.xyzw, data000.xyzw, data100.xyzw); \ POOL_OP(data001.xyzw, data001.xyzw, data101.xyzw); \ POOL_OP(data010.xyzw, data010.xyzw, data11.xyzw); \ POOL_OP(data000.xyzw, data000.xyzw, data200.xyzw); \ POOL_OP(data001.xyzw, data001.xyzw, data201.xyzw); \ POOL_OP(data010.xyzw, data010.xyzw, data21.xyzw); \ POOL_OP(res.xyzw, vec4(data000.xw, data001.z, data010.y), vec4(data000.y, data001.xw, data010.z)); \ POOL_OP(res.xyzw, res.xyzw, vec4(data000.z, data001.y, data010.xw)) void main(void) { // Get pixels pointer Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); vec4 res; // Perform pooling 3x3 for 4 output elements #if STRIDE_X == 1 POOLING3x3_STRIDE1(res, src_ptr, src_iter); #elif STRIDE_X == 2 POOLING3x3_STRIDE2(res, src_ptr, src_iter); #elif STRIDE_X == 3 POOLING3x3_STRIDE3(res, src_ptr, src_iter); #endif /*STRIDE_X == 1*/ // Divide by pool region in case of average pooling #if defined(POOL_AVG) || defined(POOL_L2) ivec4 start_x = ((ivec4(int(gl_GlobalInvocationID.x) * 4) + ivec4(0, 1, 2, 3)) * (ivec4(STRIDE_X))) - (ivec4(PAD_X)); int start_y = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y; ivec4 end_x = min((start_x + (ivec4(3))), (ivec4(MAX_WIDTH))); int end_y = min((start_y + 3), MAX_HEIGHT); #if defined(EXCLUDE_PADDING) start_x = max(ivec4(0), start_x); start_y = max(0, start_y); #endif /* defined(EXCLUDE_PADDING) */ res *= (vec4((1.f)) / vec4((ivec4(end_y - start_y)) * (end_x - start_x))); #endif /*POOL_AVG*/ #if defined(POOL_L2) // Take square root of the result in L2 pooling res = SQRT_OP(res); #endif /* defined(POOL_L2) */ VSTORE4_CURRENT_ITEM(dst_ptr, dst_iter, res); } #elif defined(POOLING_LAYER_N) void main(void) { // Get pixels pointer Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); vec4 vdata0 = vec4(INITIAL_VALUE); vec4 vdata1 = vec4(INITIAL_VALUE); float sdata = float(INITIAL_VALUE); for(int y = 0; y < int(POOL_SIZE); y++) { int x = 0; for(; x <= (int(POOL_SIZE) - 8); x += 8) { vec4 data2 = VLOAD4(vec4, src_ptr, TENSOR3D_OFFSET(src_iter, x, y, 0)); vec4 data3 = VLOAD4(vec4, src_ptr, TENSOR3D_OFFSET(src_iter, x, y, 0) + uint(4)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data2 *= data2; data3 *= data3; #endif /* defined(POOL_L2) */ POOL_OP(vdata0, vdata0, data2); POOL_OP(vdata1, vdata1, data3); } // Leftover for(; x < int(POOL_SIZE); ++x) { float data4 = LOAD(src_ptr, TENSOR3D_OFFSET(src_iter, x, y, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data4 *= data4; #endif /* defined(POOL_L2) */ POOL_OP_float(sdata, sdata, data4); } } //Reduce result vec4 reduce4; POOL_OP(reduce4, vdata0.xyzw, vdata1.xyzw); vec2 reduce2; POOL_OP_vec2(reduce2, reduce4.xy, reduce4.zw); float res; POOL_OP_float(res, reduce2.x, reduce2.y); POOL_OP_float(res, res, sdata); #if defined(POOL_AVG) || defined(POOL_L2) { // Divide by pool region in case of average pooling int start_x = int(gl_GlobalInvocationID.x) * STRIDE_X - PAD_X; int start_y = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y; int end_x = int(min(start_x + POOL_SIZE, MAX_WIDTH)); int end_y = int(min(start_y + POOL_SIZE, MAX_HEIGHT)); #if defined(EXCLUDE_PADDING) start_x = max(0, start_x); start_y = max(0, start_y); #endif /* defined(EXCLUDE_PADDING) */ float res1 = float((end_y - start_y) * (end_x - start_x)); res = DIV_OP(res, res1); } #endif /* defined(POOL_AVG) || defined(POOL_L2) */ #if defined(POOL_L2) // Take square root of the result in L2 pooling res = SQRT_OP(res); #endif /* defined(POOL_L2) */ // Store result STORE_CURRENT_ITEM(dst_ptr, dst_iter, res); } #endif // POOLING_LAYER_N #elif defined(DATA_TYPE_FP16) vec2 calculate_max(const int, Tensor3DIterator, const int, const int, const int, const int, const int, const int); vec2 calculate_avg(const int, Tensor3DIterator, const int, const int, const int, const int, const int, const int); TENSOR_DECLARATION(1, srcBuffer, uint, src_ptr, src_shift, 2, readonly); TENSOR_DECLARATION(2, dstBuffer, uint, dst_ptr, dst_shift, 2, writeonly); #if defined(POOL_SIZE) // Set the initial value for the pooling operation accordingly with the data type #if defined(POOL_AVG) || defined(POOL_L2) #define INITIAL_VALUE 0.0f #else /* defined(POOL_AVG) || defined(POOL_L2) */ #define INITIAL_VALUE -65504.0f #endif //POOL_AVG #endif //POOL_SIZE vec2 calculate_max(const int pool_size, Tensor3DIterator src_iter, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y) { int start_x1 = int(gl_GlobalInvocationID.x) * stride_x - pad_x; int start_y1 = int(gl_GlobalInvocationID.y) * stride_y - pad_y; int end_x1 = int(min(start_x1 + pool_size, upper_bound_w)); int end_y1 = int(min(start_y1 + pool_size, upper_bound_h)); int start_x2 = start_x1 + stride_x; int start_y2 = start_y1; int end_x2 = int(min(start_x2 + pool_size, upper_bound_w)); int end_y2 = int(min(start_y2 + pool_size, upper_bound_h)); //Initialize maximum vec2 data_max = vec2(0); //Load and Set initial maximum1 vec2 data_init1 = LOAD_UNPACK2_CURRENT_ITEM_HALF(src_ptr, src_iter); data_max.x = data_init1.x; //Load and Set initial maximum2 if(end_x1 < upper_bound_w) { if((stride_x % 2) == 0) { vec2 data_init2 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, stride_x, 0, 0)); data_max.y = data_init2.x; } else { vec2 data_init2 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, stride_x - 1, 0, 0)); data_max.y = data_init2.y; } } for(int i = 0; (start_y1 + i) < end_y1; i++) for(int j = 0; (start_x1 + j) < end_x1; j = j + 2) { //Calculate maximum1 if((start_x1 + j + 1) < end_x1) { vec2 data1 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, j, i, 0)); float data_mr1; POOL_OP_float(data_mr1, data1.x, data1.y); POOL_OP_float(data_max.x, data_max.x, data_mr1); } else { vec2 data1 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, j, i, 0)); POOL_OP_float(data_max.x, data_max.x, data1.x); } //Calculate maximum2 if((start_x2 + j) < end_x2 && end_x1 < upper_bound_w) { if((stride_x % 2) == 0) { vec2 data2 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, (j + stride_x), i, 0)); if((start_x2 + j + 1) < end_x2) { float data_mr2; POOL_OP_float(data_mr2, data2.x, data2.y); POOL_OP_float(data_max.y, data_max.y, data_mr2); } else { POOL_OP_float(data_max.y, data_max.y, data2.x); } } else { vec2 data2 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, (j + stride_x - 1), i, 0)); vec2 data3 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, (j + stride_x + 1), i, 0)); if((start_x2 + j + 1) < end_x2) { float data_mr2; POOL_OP_float(data_mr2, data3.x, data2.y); POOL_OP_float(data_max.y, data_max.y, data_mr2); } else { POOL_OP_float(data_max.y, data_max.y, data2.y); } } } } return data_max; } vec2 calculate_avg(const int pool_size, Tensor3DIterator src_iter, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y) { int start_x1 = (2 * int(gl_GlobalInvocationID.x)) * stride_x - pad_x; int start_y1 = int(gl_GlobalInvocationID.y) * stride_y - pad_y; int end_x1 = int(min(start_x1 + pool_size, upper_bound_w)); int end_y1 = int(min(start_y1 + pool_size, upper_bound_h)); int start_x2 = start_x1 + stride_x; int start_y2 = start_y1; int end_x2 = int(min(start_x2 + pool_size, upper_bound_w)); int end_y2 = int(min(start_y2 + pool_size, upper_bound_h)); //Initialize sum float data_total1 = float(0); float data_total2 = float(0); for(int i = 0; (start_y1 + i) < end_y1; i++) for(int j = 0; (start_x1 + j) < end_x1; j = j + 2) { vec2 data1 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, j, i, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data1 = POW2_OP(data1, 2); #endif /* defined(POOL_L2) */ //Calculate sum1 if((start_x1 + j + 1) < end_x1) { data_total1 = data_total1 + data1.x + data1.y; } else { data_total1 = data_total1 + data1.x; } //Calculate sum2 if((start_x2 + j) < end_x2 && end_x1 <= upper_bound_w) { if((stride_x % 2) == 0) { vec2 data2 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, (j + stride_x), i, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data2 = POW2_OP(data2, 2); #endif /* defined(POOL_L2) */ if((start_x2 + j + 1) < end_x2) { data_total2 = data_total2 + data2.x + data2.y; } else { data_total2 = data_total2 + data2.x; } } else { vec2 data2 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, (j + stride_x - 1), i, 0)); vec2 data3 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, (j + stride_x + 1), i, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data2 = POW2_OP(data2, 2); data3 = POW2_OP(data3, 2); #endif /* defined(POOL_L2) */ if((start_x2 + j + 1) < end_x2) { data_total2 = data_total2 + data3.x + data2.y; } else { data_total2 = data_total2 + data2.y; } } } } #if defined(EXCLUDE_PADDING) start_x1 = max(0, start_x1); start_y1 = max(0, start_y1); start_x2 = max(0, start_x2); start_y2 = max(0, start_y2); #endif /* defined(EXCLUDE_PADDING) */ //Calculate average vec2 data_avg; data_avg.x = data_total1 / float((end_y1 - start_y1) * (end_x1 - start_x1)); data_avg.y = data_total2 / float((end_y2 - start_y2) * (end_x2 - start_x2)); return data_avg; } #if defined(POOLING_LAYER_2) || defined(POOLING_LAYER_3) || defined(POOLING_LAYER_7) #if defined(POOLING_LAYER_2) #define POOL_SIZE 2 #elif defined(POOLING_LAYER_3) #define POOL_SIZE 3 #elif defined(POOLING_LAYER_7) #define POOL_SIZE 7 #else // POOLING_LAYER_n #error Please define POOLING_LAYER_N instead. #endif // POOLING_LAYER_n void main(void) { // Get pixels pointer Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); //Load and calculate data vec2 data; #if defined(POOL_AVG) || defined(POOL_L2) data = calculate_avg(POOL_SIZE, src_iter, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y); #else /*POOL_AVG*/ data = calculate_max(POOL_SIZE, src_iter, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y); #endif /*POOL_AVG*/ #if defined(POOL_L2) // Take square root of the result in L2 pooling data = SQRT_OP(data); #endif /* defined(POOL_L2) */ // Store result STORE_PACK2_CURRENT_ITEM_HALF(dst_ptr, dst_iter, data); } #elif defined(POOLING_LAYER_3_OPTIMIZED) #define POOLING3x3_STRIDE1_fp16(res, input_ptr, input_iter) \ vec4 data00 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0)); \ vec2 data01 = LOAD_UNPACK2_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(2)); \ vec4 data10 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0)); \ vec2 data11 = LOAD_UNPACK2_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(2)); \ vec4 data20 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0)); \ vec2 data21 = LOAD_UNPACK2_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(2)); \ data00 = POW2_OP(data00, 4); \ data01 = POW2_OP(data01, 2); \ data10 = POW2_OP(data10, 4); \ data11 = POW2_OP(data11, 2); \ data20 = POW2_OP(data20, 4); \ data21 = POW2_OP(data21, 2); \ \ vec4 values000; \ vec4 values001; \ vec4 values010; \ vec4 values100; \ vec4 values101; \ vec4 values11; \ vec4 values200; \ vec4 values201; \ vec4 values21; \ values000.xyzw = data00.xyzy; \ values001.xyzw = data00.zwzw; \ values010.x = data01.x; \ values010.y = data00.w; \ values010.zw = data01.xy; \ values100.xyzw = data10.xyzy; \ values101.xyzw = data10.zwzw; \ values11.x = data11.x; \ values11.y = data10.w; \ values11.zw = data11.xy; \ values200.xyzw = data20.xyzy; \ values201.xyzw = data20.zwzw; \ values21.x = data21.x; \ values21.y = data20.w; \ values21.zw = data21.xy; \ POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw); \ POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw); \ POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \ POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw)) #define POOLING3x3_STRIDE2_fp16(res, input_ptr, input_iter) \ vec4 data000; \ vec4 data001; \ float data010; \ vec4 data100; \ vec4 data101; \ float data11; \ vec4 data200; \ vec4 data201; \ float data21; \ vec2 datamiddle0; \ vec2 datamiddle1; \ vec2 datamiddle2; \ data000 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0)); \ data001 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(2)); \ datamiddle0 = LOAD_UNPACK2_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(4)); \ data010 = datamiddle0.x; \ data100 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0)); \ data101 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(2)); \ datamiddle1 = LOAD_UNPACK2_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(4)); \ data11 = datamiddle1.x; \ data200 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0)); \ data201 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(2)); \ datamiddle2 = LOAD_UNPACK2_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(4)); \ data21 = datamiddle2.x; \ data000 = POW2_OP(data000, 4); \ data001 = POW2_OP(data001, 4); \ data010 = POW2_OP(data010, 1); \ data100 = POW2_OP(data100, 4); \ data101 = POW2_OP(data101, 4); \ data11 = POW2_OP(data11, 1); \ data200 = POW2_OP(data200, 4); \ data201 = POW2_OP(data201, 4); \ data21 = POW2_OP(data21, 1); \ \ vec4 values000; \ vec4 values001; \ vec4 values010; \ vec4 values100; \ vec4 values101; \ vec4 values11; \ vec4 values200; \ vec4 values201; \ vec4 values21; \ values000.xyzw = data000.xyzz; \ values001.xyzw = vec4(data000.w, data001.xxy); \ values010.xyzw = vec4(data001.zzw, data010); \ values100.xyzw = data100.xyzz; \ values101.xyzw = vec4(data100.w, data101.xxy); \ values11.xyzw = vec4(data101.zzw, data11); \ values200.xyzw = data200.xyzz; \ values201.xyzw = vec4(data200.w, data201.xxy); \ values21.xyzw = vec4(data201.zzw, data21); \ POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw); \ POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw); \ POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw); \ POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw); \ POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \ POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw)) #define POOLING3x3_STRIDE3_fp16(res, input_ptr, input_iter) \ vec4 data000 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0)); \ vec4 data001 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(2)); \ vec4 data010 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 0, 0) + uint(4)); \ vec4 data100 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0)); \ vec4 data101 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(2)); \ vec4 data11 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 1, 0) + uint(4)); \ vec4 data200 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0)); \ vec4 data201 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(2)); \ vec4 data21 = VLOAD2_UNPACK4_HALF(input_ptr, TENSOR3D_OFFSET(input_iter, 0, 2, 0) + uint(4)); \ data000 = POW2_OP(data000, 4); \ data001 = POW2_OP(data001, 4); \ data010 = POW2_OP(data010, 4); \ data100 = POW2_OP(data100, 4); \ data101 = POW2_OP(data101, 4); \ data11 = POW2_OP(data11, 4); \ data200 = POW2_OP(data200, 4); \ data201 = POW2_OP(data201, 4); \ data21 = POW2_OP(data21, 4); \ \ POOL_OP(data000.xyzw, data000.xyzw, data100.xyzw); \ POOL_OP(data001.xyzw, data001.xyzw, data101.xyzw); \ POOL_OP(data010.xyzw, data010.xyzw, data11.xyzw); \ POOL_OP(data000.xyzw, data000.xyzw, data200.xyzw); \ POOL_OP(data001.xyzw, data001.xyzw, data201.xyzw); \ POOL_OP(data010.xyzw, data010.xyzw, data21.xyzw); \ POOL_OP(res.xyzw, vec4(data000.xw, data001.z, data010.y), vec4(data000.y, data001.xw, data010.z)); \ POOL_OP(res.xyzw, res.xyzw, vec4(data000.z, data001.y, data010.xw)) void main(void) { // Get pixels pointer Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); vec4 res; // Perform pooling 3x3 for 4 output elements #if STRIDE_X == 1 POOLING3x3_STRIDE1_fp16(res, src_ptr, src_iter); #elif STRIDE_X == 2 POOLING3x3_STRIDE2_fp16(res, src_ptr, src_iter); #elif STRIDE_X == 3 POOLING3x3_STRIDE3_fp16(res, src_ptr, src_iter); #endif /*STRIDE_X == 1*/ // Divide by pool region in case of average pooling #if defined(POOL_AVG) || defined(POOL_L2) ivec4 start_x = ((ivec4(int(gl_GlobalInvocationID.x) * 4) + ivec4(0, 1, 2, 3)) * (ivec4(STRIDE_X))) - (ivec4(PAD_X)); int start_y = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y; ivec4 end_x = min((start_x + (ivec4(3))), (ivec4(MAX_WIDTH))); int end_y = min((start_y + 3), MAX_HEIGHT); #if defined(EXCLUDE_PADDING) start_x = max(ivec4(0), start_x); start_y = max(0, start_y); #endif /* defined(EXCLUDE_PADDING) */ res *= (vec4((1.f)) / vec4((ivec4(end_y - start_y)) * (end_x - start_x))); #endif /*POOL_AVG*/ #if defined(POOL_L2) // Take square root of the result in L2 pooling res = SQRT_OP(res); #endif /* defined(POOL_L2) */ VSTORE2_PACK4_CURRENT_ITEM_HALF(dst_ptr, dst_iter, res); } #elif defined(POOLING_LAYER_N) void main(void) { // Get pixels pointer Tensor3DIterator src_iter = CONVERT_TO_TENSOR3D_ITERATOR(src_attrs, src_shift); Tensor3DIterator dst_iter = CONVERT_TO_TENSOR3D_ITERATOR(dst_attrs, dst_shift); vec4 vdata00 = vec4(INITIAL_VALUE); vec4 vdata01 = vec4(INITIAL_VALUE); vec4 vdata10 = vec4(INITIAL_VALUE); vec4 vdata11 = vec4(INITIAL_VALUE); vec2 sdata = vec2(INITIAL_VALUE); for(int y = 0; y < int(POOL_SIZE); y++) { int x = 0; for(; x <= (int(POOL_SIZE) - 8); x += 8) { vec4 data2 = VLOAD2_UNPACK4_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x, y, 0)); vec4 data3 = VLOAD2_UNPACK4_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x, y, 0) + uint(2)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data2 *= data2; data3 *= data3; #endif /* defined(POOL_L2) */ POOL_OP(vdata00, vdata00, data2); POOL_OP(vdata10, vdata10, data3); } // Leftover for(; x < int(POOL_SIZE); x = x + 2) { vec2 data4middle = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x, y, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data4middle *= data4middle; #endif /* defined(POOL_L2) */ if((x + 1) >= int(POOL_SIZE)) { POOL_OP_float(sdata.x, sdata.x, data4middle.x); } else { float data4; POOL_OP_float(data4, data4middle.x, data4middle.y); POOL_OP_float(sdata.x, sdata.x, data4); } } } for(int y = 0; y < int(POOL_SIZE); y++) { if((STRIDE_X % 2) == 0) { int x1 = STRIDE_X; for(; x1 <= (int(POOL_SIZE + STRIDE_X) - 8); x1 += 8) { vec4 data2 = VLOAD2_UNPACK4_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x1, y, 0)); vec4 data3 = VLOAD2_UNPACK4_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x1, y, 0) + uint(2)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data2 *= data2; data3 *= data3; #endif /* defined(POOL_L2) */ POOL_OP(vdata01, vdata01, data2); POOL_OP(vdata11, vdata11, data3); } // Leftover for(; x1 < int(POOL_SIZE + STRIDE_X); x1 = x1 + 2) { vec2 data4middle; data4middle = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x1, y, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data4middle *= data4middle; #endif /* defined(POOL_L2) */ if((x1 + 1) >= int(POOL_SIZE + STRIDE_X)) { POOL_OP_float(sdata.y, sdata.y, data4middle.x); } else { float data4; POOL_OP_float(data4, data4middle.x, data4middle.y); POOL_OP_float(sdata.y, sdata.y, data4); } } } else { vec2 dataorigin2; dataorigin2 = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, (STRIDE_X - 1), y, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling dataorigin2.y *= dataorigin2.y; #endif /* defined(POOL_L2) */ POOL_OP_float(sdata.y, sdata.y, dataorigin2.y); int x1 = STRIDE_X + 1; for(; x1 <= (int(POOL_SIZE + STRIDE_X) - 8); x1 += 8) { vec4 data2 = VLOAD2_UNPACK4_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x1, y, 0)); vec4 data3 = VLOAD2_UNPACK4_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x1, y, 0) + uint(2)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data2 *= data2; data3 *= data3; #endif /* defined(POOL_L2) */ POOL_OP(vdata01, vdata01, data2); POOL_OP(vdata11, vdata11, data3); } // Leftover for(; x1 < int(POOL_SIZE + STRIDE_X); x1 = x1 + 2) { vec2 data4middle = LOAD_UNPACK2_HALF(src_ptr, TENSOR3D_OFFSET(src_iter, x1, y, 0)); #if defined(POOL_L2) // Raise to power of 2 for L2 Pooling data4middle *= data4middle; #endif /* defined(POOL_L2) */ if((x1 + 1) >= int(POOL_SIZE + STRIDE_X)) { POOL_OP_float(sdata.y, sdata.y, data4middle.x); } else { float data4; POOL_OP_float(data4, data4middle.x, data4middle.y); POOL_OP_float(sdata.y, sdata.y, data4); } } } } //Reduce result vec4 reduce40; POOL_OP(reduce40, vdata00.xyzw, vdata10.xyzw); vec2 reduce20; POOL_OP_vec2(reduce20, reduce40.xy, reduce40.zw); vec4 reduce41; POOL_OP(reduce41, vdata01.xyzw, vdata11.xyzw); vec2 reduce21; POOL_OP_vec2(reduce21, reduce41.xy, reduce41.zw); vec2 data; POOL_OP_float(data.x, reduce20.x, reduce20.y); POOL_OP_float(data.x, data.x, sdata.x); POOL_OP_float(data.y, reduce21.x, reduce21.y); POOL_OP_float(data.y, data.y, sdata.y); #if defined(POOL_AVG) || defined(POOL_L2) { // Divide by pool region in case of average pooling int start_x1 = (2 * int(gl_GlobalInvocationID.x)) * STRIDE_X - PAD_X; int start_y1 = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y; int end_x1 = int(min(start_x1 + POOL_SIZE, MAX_WIDTH)); int end_y1 = int(min(start_y1 + POOL_SIZE, MAX_HEIGHT)); int start_x2 = start_x1 + STRIDE_X; int start_y2 = start_y1; int end_x2 = int(min(start_x2 + POOL_SIZE, MAX_WIDTH)); int end_y2 = int(min(start_y2 + POOL_SIZE, MAX_HEIGHT)); #if defined(EXCLUDE_PADDING) start_x1 = max(0, start_x1); start_y1 = max(0, start_y1); start_x2 = max(0, start_x2); start_y2 = max(0, start_y2); #endif /* defined(EXCLUDE_PADDING) */ vec2 res1; res1.x = float((end_y1 - start_y1) * (end_x1 - start_x1)); res1.y = float((end_y2 - start_y2) * (end_x2 - start_x2)); data.x = DIV_OP(data.x, res1.x); data.y = DIV_OP(data.y, res1.y); } #endif /* defined(POOL_AVG) || defined(POOL_L2) */ #if defined(POOL_L2) // Take square root of the result in L2 pooling data = SQRT_OP(data); #endif /* defined(POOL_L2) */ // Store result STORE_PACK2_CURRENT_ITEM_HALF(dst_ptr, dst_iter, data); } #endif // POOLING_LAYER_N #else // DATA_TYPE_FP32 #error Data type not supported #endif // DATA_TYPE_FP32