Sheri Zhang | 79144a6 | 2021-02-08 17:43:04 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/Helpers.h" |
| 25 | #include "arm_compute/core/ITensor.h" |
| 26 | #include "arm_compute/core/Types.h" |
| 27 | #include "arm_compute/core/utils/misc/Traits.h" |
| 28 | #include "src/core/NEON/wrapper/intrinsics/intrinsics.h" |
| 29 | #include "src/core/cpu/kernels/pooling/neon/list.h" |
| 30 | #include "src/core/helpers/WindowHelpers.h" |
| 31 | |
| 32 | #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) |
| 33 | |
| 34 | namespace arm_compute |
| 35 | { |
| 36 | namespace cpu |
| 37 | { |
| 38 | namespace |
| 39 | { |
| 40 | void pooling2_f16_maxpool_indices(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window) |
| 41 | { |
| 42 | const int window_start_x = window.x().start(); |
| 43 | const int window_end_x = window.x().end(); |
| 44 | const int window_step_x = 8; |
| 45 | |
| 46 | Window window_out = window; |
| 47 | window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 48 | |
| 49 | Iterator in(src, window_src); |
| 50 | Iterator out(dst0, window_out); |
| 51 | Iterator indices(dst1, window_out); |
| 52 | |
| 53 | const int pool_pad_top = pool_info.pad_stride_info.pad_top(); |
| 54 | const int pool_pad_left = pool_info.pad_stride_info.pad_left(); |
| 55 | |
| 56 | int pool_stride_x = 0; |
| 57 | int pool_stride_y = 0; |
| 58 | std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride(); |
| 59 | |
| 60 | const int pad_right = src->info()->padding().right; |
| 61 | const int in_stride_y = static_cast<int>(src->info()->strides_in_bytes().y()); |
| 62 | const int in_stride_z = static_cast<int>(src->info()->strides_in_bytes().z()); |
| 63 | |
| 64 | execute_window_loop(window_out, [&](const Coordinates & id) |
| 65 | { |
| 66 | const int idx_width = id.y() * pool_stride_x; |
| 67 | const int idx_height = id.z() * pool_stride_y; |
| 68 | const int pool_limit_y = pool_pad_top - idx_height; |
| 69 | const int pool_limit_x = pool_pad_left - idx_width; |
| 70 | |
| 71 | const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); |
| 72 | const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); |
| 73 | const int in_x0_offset = (pool_start_x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast<int>(src->info()->strides_in_bytes().z()); |
| 74 | const int in_x1_offset = (pool_start_x + 1 - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (pool_start_y - pool_pad_top) * static_cast<int> |
| 75 | (src->info()->strides_in_bytes().z()); |
| 76 | const int in_x2_offset = (pool_start_x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast<int> |
| 77 | (src->info()->strides_in_bytes().z()); |
| 78 | const int in_x3_offset = (pool_start_x + 1 - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (pool_start_y + 1 - pool_pad_top) * static_cast<int> |
| 79 | (src->info()->strides_in_bytes().z()); |
| 80 | |
| 81 | int x_off = window_start_x; |
| 82 | for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) |
| 83 | { |
| 84 | const auto in_x0_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x0_offset) + x_off; |
| 85 | const auto in_x1_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x1_offset) + x_off; |
| 86 | const auto in_x2_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x2_offset) + x_off; |
| 87 | const auto in_x3_ptr = reinterpret_cast<const float16_t *>(in.ptr() + in_x3_offset) + x_off; |
| 88 | const auto v_x0 = vld1q_f16(in_x0_ptr); |
| 89 | const auto v_x1 = vld1q_f16(in_x1_ptr); |
| 90 | const auto v_x2 = vld1q_f16(in_x2_ptr); |
| 91 | const auto v_x3 = vld1q_f16(in_x3_ptr); |
| 92 | float16x8_t vres = vmaxq_f16(vmaxq_f16(v_x2, v_x3), vmaxq_f16(v_x0, v_x1)); |
| 93 | // Store result |
| 94 | vst1q_f16(reinterpret_cast<float16_t *>(out.ptr()) + x_off, vres); |
| 95 | |
Manuel Bottini | ca62c6f | 2021-03-23 11:50:34 +0000 | [diff] [blame^] | 96 | const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NHWC); |
Sheri Zhang | 79144a6 | 2021-02-08 17:43:04 +0000 | [diff] [blame] | 97 | const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; |
| 98 | const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; |
| 99 | const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * src->info()->tensor_shape()[1]; |
| 100 | const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; |
| 101 | const uint32x4_t voffset_x0_0 = { offset_x0, offset_x0 + 1, offset_x0 + 2, offset_x0 + 3 }; |
| 102 | const uint32x4_t voffset_x0_1 = { offset_x0 + 4, offset_x0 + 5, offset_x0 + 6, offset_x0 + 7 }; |
| 103 | const uint16x8_t voffset_x0 = vcombine_u16(vmovn_u32(voffset_x0_0), vmovn_u32(voffset_x0_1)); |
| 104 | const uint32x4_t voffset_x1_0 = { offset_x1, offset_x1 + 1, offset_x1 + 2, offset_x1 + 3 }; |
| 105 | const uint32x4_t voffset_x1_1 = { offset_x1 + 4, offset_x1 + 5, offset_x1 + 6, offset_x1 + 7 }; |
| 106 | const uint16x8_t voffset_x1 = vcombine_u16(vmovn_u32(voffset_x1_0), vmovn_u32(voffset_x1_1)); |
| 107 | const uint32x4_t voffset_x2_0 = { offset_x2, offset_x2 + 1, offset_x2 + 2, offset_x2 + 3 }; |
| 108 | const uint32x4_t voffset_x2_1 = { offset_x2 + 4, offset_x2 + 5, offset_x2 + 6, offset_x2 + 7 }; |
| 109 | const uint16x8_t voffset_x2 = vcombine_u16(vmovn_u32(voffset_x2_0), vmovn_u32(voffset_x2_1)); |
| 110 | const uint32x4_t voffset_x3_0 = { offset_x3, offset_x3 + 1, offset_x3 + 2, offset_x3 + 3 }; |
| 111 | const uint32x4_t voffset_x3_1 = { offset_x3 + 4, offset_x3 + 5, offset_x3 + 6, offset_x3 + 7 }; |
| 112 | const uint16x8_t voffset_x3 = vcombine_u16(vmovn_u32(voffset_x3_0), vmovn_u32(voffset_x3_1)); |
| 113 | const uint16x8_t tmp_indices0 = vbslq_u16(vcgeq_f16(v_x0, v_x1), voffset_x0, voffset_x1); |
| 114 | const uint16x8_t tmp_indices1 = vbslq_u16(vcgeq_f16(v_x2, v_x3), voffset_x2, voffset_x3); |
| 115 | const uint16x8_t tmp_indices2 = vbslq_u16(vcgeq_f16(vmaxq_f16(v_x0, v_x1), vmaxq_f16(v_x2, v_x3)), tmp_indices0, tmp_indices1); |
| 116 | const uint32x4_t tmp_indeces3_0 = vmovl_u16(vget_low_u16(tmp_indices2)); |
| 117 | const uint32x4_t tmp_indeces3_1 = vmovl_u16(vget_high_u16(tmp_indices2)); |
| 118 | // Store indicies |
| 119 | vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off, tmp_indeces3_0); |
| 120 | vst1q_u32(reinterpret_cast<uint32_t *>(indices.ptr() + 16) + x_off, tmp_indeces3_1); |
| 121 | } |
| 122 | |
| 123 | // Left-overs loop |
| 124 | for(; x_off < window_end_x; ++x_off) |
| 125 | { |
| 126 | const auto x0 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x0_offset) + x_off); |
| 127 | const auto x1 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x1_offset) + x_off); |
| 128 | const auto x2 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x2_offset) + x_off); |
| 129 | const auto x3 = *(reinterpret_cast<const float16_t *>(in.ptr() + in_x3_offset) + x_off); |
| 130 | float16_t res = std::max(std::max(x2, x3), std::max(x0, x1)); |
| 131 | |
| 132 | // Store result |
| 133 | *(reinterpret_cast<float16_t *>(out.ptr()) + x_off) = res; |
| 134 | |
Manuel Bottini | ca62c6f | 2021-03-23 11:50:34 +0000 | [diff] [blame^] | 135 | const uint32_t offset_base = offset_no_padding<float16_t>(in.offset(), id, *src->info(), pool_stride_x, pool_stride_y, DataLayout::NHWC); |
Sheri Zhang | 79144a6 | 2021-02-08 17:43:04 +0000 | [diff] [blame] | 136 | const uint32_t offset_x0 = (uint32_t)offset_base / sizeof(float16_t) + x_off; |
| 137 | const uint32_t offset_x1 = (uint32_t)offset_x0 + in_stride_y / sizeof(float16_t) - pad_right; |
| 138 | const uint32_t offset_x2 = (uint32_t)offset_x0 + in_stride_z / sizeof(float16_t) - pad_right * src->info()->tensor_shape()[1]; |
| 139 | const uint32_t offset_x3 = (uint32_t)offset_x2 + in_stride_y / sizeof(float16_t) - pad_right; |
| 140 | const uint32_t tmp_idx0 = (x0 >= x1) ? offset_x0 : offset_x1; |
| 141 | const uint32_t tmp_idx1 = (x2 >= x3) ? offset_x2 : offset_x3; |
| 142 | const uint32_t tmp_idx2 = (std::max(x0, x1) >= std::max(x2, x3)) ? tmp_idx0 : tmp_idx1; |
| 143 | |
| 144 | // Store indices |
| 145 | *(reinterpret_cast<uint32_t *>(indices.ptr()) + x_off) = tmp_idx2; |
| 146 | } |
| 147 | }, |
| 148 | in, out, indices); |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | void poolingMxN_fp16_neon_nhwc(const ITensor *src, ITensor *dst0, ITensor *dst1, PoolingLayerInfo &pool_info, const Window &window_src, const Window &window) |
| 153 | { |
| 154 | if(pool_info.pool_size == Size2D(2, 2) && pool_info.pool_type == PoolingType::MAX && dst1) |
| 155 | { |
| 156 | pooling2_f16_maxpool_indices(src, dst0, dst1, pool_info, window_src, window); |
| 157 | } |
| 158 | const int window_start_x = window.x().start(); |
| 159 | const int window_end_x = window.x().end(); |
| 160 | const int window_step_x = 8; |
| 161 | |
| 162 | Window window_out = window; |
| 163 | window_out.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 164 | |
| 165 | Iterator in(src, window_src); |
| 166 | Iterator out(dst0, window_out); |
| 167 | |
| 168 | const int pool_size_x = pool_info.is_global_pooling ? src->info()->tensor_shape().y() : pool_info.pool_size.width; |
| 169 | const int pool_size_y = pool_info.is_global_pooling ? src->info()->tensor_shape().z() : pool_info.pool_size.height; |
| 170 | const int pool_pad_right = pool_info.pad_stride_info.pad_right(); |
| 171 | const int pool_pad_top = pool_info.pad_stride_info.pad_top(); |
| 172 | const int pool_pad_left = pool_info.pad_stride_info.pad_left(); |
| 173 | const int pool_pad_bottom = pool_info.pad_stride_info.pad_bottom(); |
| 174 | int pool_stride_x = 0; |
| 175 | int pool_stride_y = 0; |
| 176 | std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info.stride(); |
| 177 | const int upper_bound_w = src->info()->dimension(1) + (pool_info.exclude_padding ? 0 : pool_pad_right); |
| 178 | const int upper_bound_h = src->info()->dimension(2) + (pool_info.exclude_padding ? 0 : pool_pad_bottom); |
| 179 | |
| 180 | float16x8_t vres; |
| 181 | |
| 182 | execute_window_loop(window_out, [&](const Coordinates & id) |
| 183 | { |
| 184 | const int idx_width = id.y() * pool_stride_x; |
| 185 | const int idx_height = id.z() * pool_stride_y; |
| 186 | const int pool_limit_y = pool_pad_top - idx_height; |
| 187 | const int pool_limit_x = pool_pad_left - idx_width; |
| 188 | |
| 189 | const int pool_start_y = std::max(0, window_src.z().start() + pool_limit_y); |
| 190 | const int pool_end_y = std::min(pool_size_y, window_src.z().end() + pool_limit_y); |
| 191 | const int pool_start_x = std::max(0, window_src.y().start() + pool_limit_x); |
| 192 | const int pool_end_x = std::min(pool_size_x, window_src.y().end() + pool_limit_x); |
| 193 | |
| 194 | int x_off = window_start_x; |
| 195 | for(; x_off <= (window_end_x - window_step_x); x_off += window_step_x) |
| 196 | { |
| 197 | if(pool_info.pool_type != PoolingType::MAX) |
| 198 | { |
| 199 | // Calculate scale |
| 200 | const float scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, |
| 201 | pool_stride_y); |
| 202 | const float16x8_t scale_v = vdupq_n_f16(scale); |
| 203 | |
| 204 | // Perform pooling |
| 205 | vres = vdupq_n_f16(0.0f); |
| 206 | for(int y = pool_start_y; y < pool_end_y; ++y) |
| 207 | { |
| 208 | for(int x = pool_start_x; x < pool_end_x; ++x) |
| 209 | { |
| 210 | const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> |
| 211 | (src->info()->strides_in_bytes().z())) + x_off); |
| 212 | |
| 213 | // Get power of 2 in case of l2 pooling and accumulate |
| 214 | if(pool_info.pool_type == PoolingType::L2) |
| 215 | { |
| 216 | vres = vaddq_f16(vres, vmulq_f16(data, data)); |
| 217 | } |
| 218 | else |
| 219 | { |
| 220 | vres = vaddq_f16(vres, data); |
| 221 | } |
| 222 | } |
| 223 | } |
| 224 | // Divide by scale |
| 225 | vres = vmulq_f16(vres, scale_v); |
| 226 | } |
| 227 | else |
| 228 | { |
| 229 | vres = vdupq_n_f16(std::numeric_limits<float>::lowest()); |
| 230 | |
| 231 | for(int y = pool_start_y; y < pool_end_y; ++y) |
| 232 | { |
| 233 | for(int x = pool_start_x; x < pool_end_x; ++x) |
| 234 | { |
| 235 | const float16x8_t data = vld1q_f16(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> |
| 236 | (src->info()->strides_in_bytes().z())) + x_off); |
| 237 | vres = vmaxq_f16(vres, data); |
| 238 | } |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | // Calculate square-root in case of l2 pooling |
| 243 | if(pool_info.pool_type == PoolingType::L2) |
| 244 | { |
| 245 | float16x8_t sqrt_reciprocal = vrsqrteq_f16(vres); |
| 246 | vres = vmulq_f16(vres, vmulq_f16(vrsqrtsq_f16(vmulq_f16(vres, sqrt_reciprocal), sqrt_reciprocal), sqrt_reciprocal)); |
| 247 | } |
| 248 | |
| 249 | // Store result |
| 250 | vst1q_f16(reinterpret_cast<float16_t *>(out.ptr()) + x_off, vres); |
| 251 | } |
| 252 | |
| 253 | // Left-overs loop |
| 254 | for(; x_off < window_end_x; ++x_off) |
| 255 | { |
| 256 | float16_t res = 0.0f; |
| 257 | |
| 258 | if(pool_info.pool_type != PoolingType::MAX) |
| 259 | { |
| 260 | // Calculate scale |
| 261 | const float16_t scale = calculate_avg_scale(pool_info.exclude_padding, DataLayout::NHWC, id, pool_size_x, pool_size_y, upper_bound_w, upper_bound_h, pool_pad_left, pool_pad_top, pool_stride_x, |
| 262 | pool_stride_y); |
| 263 | |
| 264 | for(int y = pool_start_y; y < pool_end_y; ++y) |
| 265 | { |
| 266 | for(int x = pool_start_x; x < pool_end_x; ++x) |
| 267 | { |
| 268 | const float data = *(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> |
| 269 | (src->info()->strides_in_bytes().z())) + x_off); |
| 270 | |
| 271 | // Get power of 2 in case of l2 pooling and accumulate |
| 272 | if(pool_info.pool_type == PoolingType::L2) |
| 273 | { |
| 274 | res += data * data; |
| 275 | } |
| 276 | else |
| 277 | { |
| 278 | res += data; |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | |
| 283 | // Divide by scale |
| 284 | res *= scale; |
| 285 | } |
| 286 | else |
| 287 | { |
| 288 | res = std::numeric_limits<float>::lowest(); |
| 289 | for(int y = pool_start_y; y < pool_end_y; ++y) |
| 290 | { |
| 291 | for(int x = pool_start_x; x < pool_end_x; ++x) |
| 292 | { |
| 293 | const float16_t data = *(reinterpret_cast<const float16_t *>(in.ptr() + (x - pool_pad_left) * static_cast<int>(src->info()->strides_in_bytes().y()) + (y - pool_pad_top) * static_cast<int> |
| 294 | (src->info()->strides_in_bytes().z())) + x_off); |
| 295 | res = std::max(res, data); |
| 296 | } |
| 297 | } |
| 298 | } |
| 299 | |
| 300 | // Calculate square-root in case of l2 pooling |
| 301 | if(pool_info.pool_type == PoolingType::L2) |
| 302 | { |
| 303 | res = std::sqrt(res); |
| 304 | } |
| 305 | |
| 306 | // Store result |
| 307 | *(reinterpret_cast<float16_t *>(out.ptr()) + x_off) = res; |
| 308 | } |
| 309 | }, |
| 310 | in, out); |
| 311 | } |
| 312 | } // namespace cpu |
| 313 | } // namespace arm_compute |
| 314 | |
| 315 | #endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */ |