| /* |
| * Copyright (c) 2017-2018 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. |
| */ |
| #pragma once |
| |
| #ifdef __arm__ |
| |
| #include <arm_neon.h> |
| |
| template<> |
| void MergeResults<8, 6, false>(float *out, const float *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const float *bias, Activation act, bool append) { |
| const float *inptr = in; |
| prefetch_6x(inptr); |
| prefetch_6x(inptr + 96); |
| |
| float nullbias[8]; |
| float minval = - std::numeric_limits<float>::infinity(); |
| float maxval = std::numeric_limits<float>::infinity(); |
| |
| switch(act.type) |
| { |
| default: |
| case Activation::Type::None: |
| break; |
| case Activation::Type::BoundedReLU: |
| maxval = static_cast<float>(act.param1); |
| /* fall through */ |
| case Activation::Type::ReLU: |
| minval = 0.0f; |
| break; |
| } |
| |
| float32x4_t minv = vdupq_n_f32(minval); |
| float32x4_t maxv = vdupq_n_f32(maxval); |
| |
| if (!append && !bias) |
| { |
| memset(nullbias, 0, (8 * sizeof(float))); |
| } |
| |
| for (int y=y0; y<ymax; y+=8) { |
| float *outptr0 = out + (y * ldout) + x0; |
| float *outptr1 = outptr0 + ldout; |
| float *outptr2 = outptr1 + ldout; |
| float *outptr3 = outptr2 + ldout; |
| float *outptr4 = outptr3 + ldout; |
| float *outptr5 = outptr4 + ldout; |
| |
| prefetch_2x(outptr0); |
| prefetch_2x(outptr1); |
| prefetch_2x(outptr2); |
| prefetch_2x(outptr3); |
| prefetch_2x(outptr4); |
| prefetch_2x(outptr5); |
| |
| for (int i=x0; i<xmax; i+=8) { |
| float dummyres[8]; |
| |
| /* Make sure we throw away results if Y isn't a multiple of 8. |
| * We do this by pointing the result pointer at a dummy buffer |
| * we later discard. */ |
| if ((y+5) >= ymax) { |
| switch ((y + 5) - ymax) { |
| case 4: |
| outptr1 = dummyres; |
| /* fall through */ |
| case 3: |
| outptr2 = dummyres; |
| /* fall through */ |
| case 2: |
| outptr3 = dummyres; |
| /* fall through */ |
| case 1: |
| outptr4 = dummyres; |
| /* fall through */ |
| case 0: |
| outptr5 = dummyres; |
| break; |
| |
| default: |
| UNREACHABLE("Impossible."); |
| } |
| } |
| |
| if (append) { |
| /* Append mode: Read, activate, write. */ |
| |
| /* For ragged X, manually copy over the valid results. */ |
| if ((i+7) >= xmax) { |
| for (int xi=0; xi<8; xi++) { |
| if ((i+xi) < xmax) { |
| *outptr0 = std::min(std::max(minval, inptr[xi] + *outptr0), maxval); |
| outptr0++; |
| *outptr1 = std::min(std::max(minval, inptr[xi + 8] + *outptr1), maxval); |
| outptr1++; |
| *outptr2 = std::min(std::max(minval, inptr[xi + 16] + *outptr2), maxval); |
| outptr2++; |
| *outptr3 = std::min(std::max(minval, inptr[xi + 24] + *outptr3), maxval); |
| outptr3++; |
| *outptr4 = std::min(std::max(minval, inptr[xi + 32] + *outptr4), maxval); |
| outptr4++; |
| *outptr5 = std::min(std::max(minval, inptr[xi + 40] + *outptr5), maxval); |
| outptr5++; |
| } |
| } |
| inptr += 48; |
| } else { |
| /* Optimized routine to copy an entire block */ |
| __asm __volatile ( |
| // Rows 0-1 |
| "VLD1.32 {d0-d3}, [%[inptr]]!\n" |
| "VLD1.32 {d8-d11}, [%[outptr0]]\n" |
| "VLD1.32 {d4-d7}, [%[inptr]]!\n" |
| "VLD1.32 {d12-d15}, [%[outptr1]]\n" |
| |
| "VADD.f32 q4, q4, q0\n" |
| ASM_PREFETCH("[%[inptr], #352]") |
| "VADD.f32 q5, q5, q1\n" |
| "VADD.f32 q6, q6, q2\n" |
| "VADD.f32 q7, q7, q3\n" |
| ASM_PREFETCH("[%[inptr], #416]") |
| "VMAX.f32 q4, q4, %q[minv]\n" |
| "VMAX.f32 q5, q5, %q[minv]\n" |
| "VMAX.f32 q6, q6, %q[minv]\n" |
| ASM_PREFETCH("[%[inptr], #480]") |
| "VMAX.f32 q7, q7, %q[minv]\n" |
| "VMIN.f32 q4, q4, %q[maxv]\n" |
| "VMIN.f32 q5, q5, %q[maxv]\n" |
| "VST1.32 {d8-d11}, [%[outptr0]]!\n" |
| "VMIN.f32 q6, q6, %q[maxv]\n" |
| "VMIN.f32 q7, q7, %q[maxv]\n" |
| "VST1.32 {d12-d15}, [%[outptr1]]!\n" |
| |
| // Rows 2-3 |
| "VLD1.32 {d0-d3}, [%[inptr]]!\n" |
| "VLD1.32 {d8-d11}, [%[outptr2]]\n" |
| "VLD1.32 {d4-d7}, [%[inptr]]!\n" |
| "VLD1.32 {d12-d15}, [%[outptr3]]\n" |
| |
| "VADD.f32 q4, q4, q0\n" |
| ASM_PREFETCH("[%[outptr0], #96]") |
| "VADD.f32 q5, q5, q1\n" |
| "VADD.f32 q6, q6, q2\n" |
| "VADD.f32 q7, q7, q3\n" |
| ASM_PREFETCH("[%[outptr1], #96]") |
| "VMAX.f32 q4, q4, %q[minv]\n" |
| "VMAX.f32 q5, q5, %q[minv]\n" |
| "VMAX.f32 q6, q6, %q[minv]\n" |
| ASM_PREFETCH("[%[outptr2], #128]") |
| "VMAX.f32 q7, q7, %q[minv]\n" |
| "VMIN.f32 q4, q4, %q[maxv]\n" |
| "VMIN.f32 q5, q5, %q[maxv]\n" |
| "VST1.32 {d8-d11}, [%[outptr2]]!\n" |
| "VMIN.f32 q6, q6, %q[maxv]\n" |
| "VMIN.f32 q7, q7, %q[maxv]\n" |
| "VST1.32 {d12-d15}, [%[outptr3]]!\n" |
| |
| // Rows 4-5 |
| "VLD1.32 {d0-d3}, [%[inptr]]!\n" |
| "VLD1.32 {d8-d11}, [%[outptr4]]\n" |
| "VLD1.32 {d4-d7}, [%[inptr]]!\n" |
| "VLD1.32 {d12-d15}, [%[outptr5]]\n" |
| |
| "VADD.f32 q4, q4, q0\n" |
| ASM_PREFETCH("[%[outptr3], #96]") |
| "VADD.f32 q5, q5, q1\n" |
| "VADD.f32 q6, q6, q2\n" |
| "VADD.f32 q7, q7, q3\n" |
| ASM_PREFETCH("[%[outptr4], #128]") |
| "VMAX.f32 q4, q4, %q[minv]\n" |
| "VMAX.f32 q5, q5, %q[minv]\n" |
| "VMAX.f32 q6, q6, %q[minv]\n" |
| ASM_PREFETCH("[%[outptr5], #128]") |
| "VMAX.f32 q7, q7, %q[minv]\n" |
| "VMIN.f32 q4, q4, %q[maxv]\n" |
| "VMIN.f32 q5, q5, %q[maxv]\n" |
| "VST1.32 {d8-d11}, [%[outptr4]]!\n" |
| "VMIN.f32 q6, q6, %q[maxv]\n" |
| "VMIN.f32 q7, q7, %q[maxv]\n" |
| "VST1.32 {d12-d15}, [%[outptr5]]!\n" |
| : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), |
| [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [inptr] "+r" (inptr) |
| : [minv] "w" (minv), [maxv] "w" (maxv) |
| : "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "memory" |
| ); |
| } |
| } else { |
| /* Bias mode: Add bias to everything, then min/max/write as before. */ |
| const float *biasptr = bias ? bias + i : nullbias; |
| |
| /* For ragged X, manually copy over the valid results. */ |
| if ((i+7) >= xmax) { |
| for (int xi=0; xi<7; xi++) { |
| if ((i+xi) < xmax) { |
| *outptr0 = std::min(std::max(minval, inptr[xi] + biasptr[xi]), maxval); |
| outptr0++; |
| *outptr1 = std::min(std::max(minval, inptr[xi + 8] + biasptr[xi]), maxval); |
| outptr1++; |
| *outptr2 = std::min(std::max(minval, inptr[xi + 16] + biasptr[xi]), maxval); |
| outptr2++; |
| *outptr3 = std::min(std::max(minval, inptr[xi + 24] + biasptr[xi]), maxval); |
| outptr3++; |
| *outptr4 = std::min(std::max(minval, inptr[xi + 32] + biasptr[xi]), maxval); |
| outptr4++; |
| *outptr5 = std::min(std::max(minval, inptr[xi + 40] + biasptr[xi]), maxval); |
| outptr5++; |
| } |
| } |
| inptr += 48; |
| } else { |
| /* Optimized routine to copy an entire block */ |
| __asm __volatile ( |
| // Rows 0-1 |
| "VLD1.32 {d8-d11}, [%[inptr]]!\n" |
| "VLD1.32 {d0-d3}, [%[biasptr]]\n" |
| "VLD1.32 {d12-d15}, [%[inptr]]!\n" |
| |
| "VADD.f32 q4, q4, q0\n" |
| ASM_PREFETCH("[%[inptr], #352]") |
| "VADD.f32 q5, q5, q1\n" |
| "VADD.f32 q6, q6, q0\n" |
| "VADD.f32 q7, q7, q1\n" |
| ASM_PREFETCH("[%[inptr], #416]") |
| "VMAX.f32 q4, q4, %q[minv]\n" |
| "VMAX.f32 q5, q5, %q[minv]\n" |
| "VMAX.f32 q6, q6, %q[minv]\n" |
| ASM_PREFETCH("[%[inptr], #480]") |
| "VMAX.f32 q7, q7, %q[minv]\n" |
| "VMIN.f32 q4, q4, %q[maxv]\n" |
| "VMIN.f32 q5, q5, %q[maxv]\n" |
| "VST1.32 {d8-d11}, [%[outptr0]]!\n" |
| "VMIN.f32 q6, q6, %q[maxv]\n" |
| "VMIN.f32 q7, q7, %q[maxv]\n" |
| "VST1.32 {d12-d15}, [%[outptr1]]!\n" |
| |
| // Rows 2-3 |
| "VLD1.32 {d8-d11}, [%[inptr]]!\n" |
| "VLD1.32 {d12-d15}, [%[inptr]]!\n" |
| |
| "VADD.f32 q4, q4, q0\n" |
| ASM_PREFETCH("[%[outptr0], #96]") |
| "VADD.f32 q5, q5, q1\n" |
| "VADD.f32 q6, q6, q0\n" |
| "VADD.f32 q7, q7, q1\n" |
| ASM_PREFETCH("[%[outptr1], #96]") |
| "VMAX.f32 q4, q4, %q[minv]\n" |
| "VMAX.f32 q5, q5, %q[minv]\n" |
| "VMAX.f32 q6, q6, %q[minv]\n" |
| ASM_PREFETCH("[%[outptr2], #128]") |
| "VMAX.f32 q7, q7, %q[minv]\n" |
| "VMIN.f32 q4, q4, %q[maxv]\n" |
| "VMIN.f32 q5, q5, %q[maxv]\n" |
| "VST1.32 {d8-d11}, [%[outptr2]]!\n" |
| "VMIN.f32 q6, q6, %q[maxv]\n" |
| "VMIN.f32 q7, q7, %q[maxv]\n" |
| "VST1.32 {d12-d15}, [%[outptr3]]!\n" |
| |
| // Rows 4-5 |
| "VLD1.32 {d8-d11}, [%[inptr]]!\n" |
| "VLD1.32 {d12-d15}, [%[inptr]]!\n" |
| |
| "VADD.f32 q4, q4, q0\n" |
| ASM_PREFETCH("[%[outptr3], #96]") |
| "VADD.f32 q5, q5, q1\n" |
| "VADD.f32 q6, q6, q0\n" |
| "VADD.f32 q7, q7, q1\n" |
| ASM_PREFETCH("[%[outptr4], #128]") |
| "VMAX.f32 q4, q4, %q[minv]\n" |
| "VMAX.f32 q5, q5, %q[minv]\n" |
| "VMAX.f32 q6, q6, %q[minv]\n" |
| ASM_PREFETCH("[%[outptr5], #128]") |
| "VMAX.f32 q7, q7, %q[minv]\n" |
| "VMIN.f32 q4, q4, %q[maxv]\n" |
| "VMIN.f32 q5, q5, %q[maxv]\n" |
| "VST1.32 {d8-d11}, [%[outptr4]]!\n" |
| "VMIN.f32 q6, q6, %q[maxv]\n" |
| "VMIN.f32 q7, q7, %q[maxv]\n" |
| "VST1.32 {d12-d15}, [%[outptr5]]!\n" |
| : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), |
| [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [inptr] "+r" (inptr) |
| : [minv] "w" (minv), [maxv] "w" (maxv), [biasptr] "r" (biasptr) |
| : "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "memory" |
| ); |
| } |
| } |
| } |
| } |
| } |
| |
| #endif // __arm__ |