|
| 1 | +// Copyright 2024 Google LLC |
| 2 | +// |
| 3 | +// This source code is licensed under the BSD-style license found in the |
| 4 | +// LICENSE file in the root directory of this source tree. |
| 5 | + |
| 6 | +$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" |
| 7 | +#include <assert.h> |
| 8 | + |
| 9 | +#include <immintrin.h> |
| 10 | + |
| 11 | +#include <xnnpack/common.h> |
| 12 | +#include <xnnpack/reduce.h> |
| 13 | +#include <xnnpack/math.h> |
| 14 | + |
| 15 | + |
| 16 | +$UNROLL = CHANNELS >> 3 |
| 17 | +void xnn_f32_rdsum_ukernel_${ACCUMULATORS}p${ACCUMULATORS}x__avx_c${CHANNELS}( |
| 18 | + size_t rows, |
| 19 | + size_t channels, |
| 20 | + const float* input, |
| 21 | + size_t input_stride, |
| 22 | + const float* zero, |
| 23 | + float* output, |
| 24 | + const union xnn_f32_scale_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 25 | +{ |
| 26 | + assert(rows != 0); |
| 27 | + assert(channels != 0); |
| 28 | + assert(input != NULL); |
| 29 | + assert(output != NULL); |
| 30 | + |
| 31 | + const __m256 vscale = _mm256_set1_ps(params->avx.scale); |
| 32 | + |
| 33 | + size_t input_increment = ${ACCUMULATORS} * input_stride; |
| 34 | + for (; channels >= ${CHANNELS}; channels -= ${CHANNELS}) { |
| 35 | + const float* i0 = input; |
| 36 | + $for i in range(1, ACCUMULATORS): |
| 37 | + const float* i${i} = (const float*) ((uintptr_t) input + ${i} * input_stride); |
| 38 | + |
| 39 | + $for i in range(UNROLL): |
| 40 | + __m256 vacc${i} = _mm256_setzero_ps(); |
| 41 | + |
| 42 | + for (int r = rows; r > 0; r -= ${ACCUMULATORS}) { |
| 43 | + $for N in range(1, ACCUMULATORS, 2): |
| 44 | + if XNN_UNPREDICTABLE(r < ${N+1}) { |
| 45 | + i${N} = zero; |
| 46 | + } |
| 47 | + if XNN_UNPREDICTABLE(r <= ${N+1}) { |
| 48 | + i${N+1} = zero; |
| 49 | + } |
| 50 | + $for c in range(UNROLL): |
| 51 | + __m256 vin${c}; |
| 52 | + $for j in range(ACCUMULATORS): |
| 53 | + $for c in range(UNROLL): |
| 54 | + vin${c} = _mm256_loadu_ps(&i${j}[${c*8}]); |
| 55 | + $for c in range(UNROLL): |
| 56 | + vacc${c} = _mm256_add_ps(vin${c}, vacc${c}); |
| 57 | + $for N in range(0, ACCUMULATORS): |
| 58 | + i${N} = (const float*) ((uintptr_t) i${N} + input_increment); |
| 59 | + } |
| 60 | + $for i in range(UNROLL): |
| 61 | + vacc${i} = _mm256_mul_ps(vacc${i}, vscale); |
| 62 | + |
| 63 | + const float* o = output; |
| 64 | + $for i in range(0, UNROLL): |
| 65 | + __m256 vo${i} = _mm256_loadu_ps(o); o += 8; |
| 66 | + $for i in range(0, UNROLL): |
| 67 | + vacc${i} = _mm256_add_ps(vo${i}, vacc${i}); |
| 68 | + $for i in range(0, UNROLL): |
| 69 | + _mm256_storeu_ps(output, vacc${i}); output += 8; |
| 70 | + |
| 71 | + input = (const float*) ((uintptr_t) input + ${CHANNELS} * sizeof(float)); |
| 72 | + } |
| 73 | + __m256i vmask; |
| 74 | + if (channels != 0) { |
| 75 | + input_increment = ${ACCUMULATORS} * input_stride; |
| 76 | + const float* i0 = input; |
| 77 | + $for i in range(1, ACCUMULATORS): |
| 78 | + const float* i${i} = (const float*) ((uintptr_t) input + ${i} * input_stride); |
| 79 | + __m256 vacc[${UNROLL}]; |
| 80 | + $for i in range(UNROLL): |
| 81 | + vacc[${i}] = _mm256_setzero_ps(); |
| 82 | + |
| 83 | + const size_t num_full_chunks = channels >> 3; |
| 84 | + const size_t num_chunks = round_up_po2(channels, 8) >> 3; |
| 85 | + const size_t remainder = channels & 0x7; |
| 86 | + for (int r = rows; r > 0; r -= ${ACCUMULATORS}) { |
| 87 | + $for N in range(1, ACCUMULATORS, 2): |
| 88 | + if XNN_UNPREDICTABLE(r < ${N+1}) { |
| 89 | + i${N} = zero; |
| 90 | + } |
| 91 | + if XNN_UNPREDICTABLE(r <= ${N+1}) { |
| 92 | + i${N+1} = zero; |
| 93 | + } |
| 94 | + for (int i = 0; i < num_full_chunks; ++i) { |
| 95 | + $for c in range(ACCUMULATORS): |
| 96 | + vacc[i] = _mm256_add_ps(_mm256_loadu_ps(&i${c}[i*8]), vacc[i]); |
| 97 | + } |
| 98 | + |
| 99 | + if (remainder) { |
| 100 | + vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - (channels & 0x7) * sizeof(float))); |
| 101 | + $for c in range(ACCUMULATORS): |
| 102 | + vacc[num_full_chunks] = _mm256_add_ps(_mm256_maskload_ps(&i${c}[num_full_chunks*8], vmask), vacc[num_full_chunks]); |
| 103 | + } |
| 104 | + $for N in range(ACCUMULATORS): |
| 105 | + i${N} = (const float*) ((uintptr_t) i${N} + input_increment); |
| 106 | + } |
| 107 | + for (size_t i = 0; i < num_chunks; ++i) { |
| 108 | + vacc[i] = _mm256_mul_ps(vacc[i], vscale); |
| 109 | + } |
| 110 | + |
| 111 | + __m256 vo[${UNROLL}]; |
| 112 | + const float* o = output; |
| 113 | + for (int i = 0; i < channels >> 3; ++i) { |
| 114 | + vo[i] = _mm256_loadu_ps(o); o += 8; |
| 115 | + } |
| 116 | + for (int i = 0; i < channels >> 3; ++i) { |
| 117 | + vacc[i] = _mm256_add_ps(vo[i], vacc[i]); |
| 118 | + } |
| 119 | + for (int i = 0; i < channels >> 3; ++i) { |
| 120 | + _mm256_storeu_ps(output, vacc[i]); output += 8; |
| 121 | + } |
| 122 | + if (remainder) { |
| 123 | + const size_t pos = num_full_chunks; |
| 124 | + __m256 vout = vacc[pos]; |
| 125 | + const __m256 vdata = _mm256_maskload_ps(output, vmask); |
| 126 | + vout = _mm256_add_ps(vout, vdata); |
| 127 | + __m128 vout_lo = _mm256_castps256_ps128(vout); |
| 128 | + if (channels & 4) { |
| 129 | + _mm_storeu_ps(output, vout_lo); |
| 130 | + vout_lo = _mm256_extractf128_ps(vout, 1); |
| 131 | + output += 4; |
| 132 | + } |
| 133 | + if (channels & 2) { |
| 134 | + _mm_storel_pi((__m64*) output, vout_lo); |
| 135 | + vout_lo = _mm_movehl_ps(vout_lo, vout_lo); |
| 136 | + output += 2; |
| 137 | + } |
| 138 | + if (channels & 1) { |
| 139 | + _mm_store_ss(output, vout_lo); |
| 140 | + } |
| 141 | + } |
| 142 | + } |
| 143 | +} |
0 commit comments