1// Copyright 2019 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$assert NR % 8 == 0 7$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 8#include <assert.h> 9 10#include <immintrin.h> 11 12#include <xnnpack/igemm.h> 13 14 15$ISA = {0: "avx", 3: "fma3"}[FMA] 16void xnn_f32_igemm_ukernel_${MR}x${NR}__${ISA}_broadcast( 17 size_t mr, 18 size_t nc, 19 size_t kc, 20 size_t ks, 21 const float**restrict a, 22 const float*restrict w, 23 float*restrict c, 24 size_t cm_stride, 25 size_t cn_stride, 26 size_t a_offset, 27 const float* zero, 28 const union xnn_f32_output_params params[restrict static 1]) 29{ 30 assert(mr != 0); 31 assert(mr <= ${MR}); 32 assert(nc != 0); 33 assert(kc != 0); 34 assert(kc % sizeof(float) == 0); 35 assert(ks != 0); 36 assert(ks % (${MR} * sizeof(void*)) == 0); 37 assert(a_offset % sizeof(float) == 0); 38 assert(a != NULL); 39 assert(w != NULL); 40 assert(c != NULL); 41 42 float* c0 = c; 43 $for M in range(1, MR): 44 float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride); 45 $if M % 2 == 0: 46 if XNN_UNPREDICTABLE(mr <= ${M}) { 47 c${M} = c${M-1}; 48 } 49 $elif M + 1 == MR: 50 if XNN_UNPREDICTABLE(mr != ${M+1}) { 51 c${M} = c${M-1}; 52 } 53 $else: 54 if XNN_UNPREDICTABLE(mr < ${M+1}) { 55 c${M} = c${M-1}; 56 } 57 58 do { 59 __m256 vacc0x${ABC[0:8]} = _mm256_load_ps(w); 60 $for N in range(8, NR, 8): 61 __m256 vacc0x${ABC[N:N+8]} = _mm256_load_ps(w + ${N}); 62 $for M in range(1, MR): 63 $for N in range(0, NR, 8): 64 __m256 vacc${M}x${ABC[N:N+8]} = vacc0x${ABC[N:N+8]}; 65 w += ${NR}; 66 67 size_t p = ks; 68 do { 69 $for M in range(MR): 70 const float* restrict a${M} = a[${M}]; 71 assert(a${M} != NULL); 72 if XNN_UNPREDICTABLE(a${M} != zero) { 73 a${M} = (const float*) ((uintptr_t) a${M} + a_offset); 74 } 75 a += ${MR}; 76 77 size_t k = kc; 78 do { 79 const __m256 vb${ABC[0:8]} = _mm256_load_ps(w); 80 $for N in range(8, NR, 8): 81 const __m256 vb${ABC[N:N+8]} = _mm256_load_ps(w + ${N}); 82 w += ${NR}; 83 84 $for M in range(MR): 85 const __m256 va${M} = _mm256_broadcast_ss(a${M}); 86 a${M} += 1; 87 88 $for M in range(MR): 89 $for N in range(0, NR, 8): 90 $if FMA == 3: 91 vacc${M}x${ABC[N:N+8]} = _mm256_fmadd_ps(va${M}, vb${ABC[N:N+8]}, vacc${M}x${ABC[N:N+8]}); 92 $else: 93 vacc${M}x${ABC[N:N+8]} = _mm256_add_ps(vacc${M}x${ABC[N:N+8]}, _mm256_mul_ps(va${M}, vb${ABC[N:N+8]})); 94 k -= sizeof(float); 95 } while (k != 0); 96 p -= ${MR} * sizeof(void*); 97 } while (p != 0); 98 99 const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max); 100 $for N in range(0, NR, 8): 101 $for M in range(MR): 102 vacc${M}x${ABC[N:N+8]} = _mm256_min_ps(vacc${M}x${ABC[N:N+8]}, vmax); 103 104 const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min); 105 $for N in range(0, NR, 8): 106 $for M in range(MR): 107 vacc${M}x${ABC[N:N+8]} = _mm256_max_ps(vacc${M}x${ABC[N:N+8]}, vmin); 108 109 if XNN_LIKELY(nc >= ${NR}) { 110 $for M in reversed(range(MR)): 111 _mm256_storeu_ps(c${M}, vacc${M}x${ABC[0:8]}); 112 $for N in range(8, NR, 8): 113 _mm256_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+8]}); 114 c${M} = (float*) ((uintptr_t) c${M} + cn_stride); 115 116 a = (const float**restrict) ((uintptr_t) a - ks); 117 nc -= ${NR}; 118 } else { 119 $for LOG2N in reversed(range(NR.bit_length())): 120 $if NR != 1 << LOG2N: 121 if (nc & ${1 << LOG2N}) { 122 $if LOG2N >= 3: 123 $for M in reversed(range(MR)): 124 _mm256_storeu_ps(c${M}, vacc${M}x${ABC[0:8]}); 125 $for N in range(8, 1 << LOG2N, 8): 126 _mm256_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+8]}); 127 128 $for M in reversed(range(MR)): 129 $for N in range(0, 1 << (LOG2N - 1), 8): 130 vacc${M}x${ABC[N:N+8]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+8]}; 131 132 $for M in reversed(range(MR)): 133 c${M} += ${1 << LOG2N}; 134 $elif LOG2N == 2: 135 $for M in reversed(range(MR)): 136 _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]}); 137 138 $for M in reversed(range(MR)): 139 vacc${M}x${ABC[0:4]} = _mm256_extractf128_ps(vacc${M}x${ABC[0:8]}, 1); 140 141 $for M in reversed(range(MR)): 142 c${M} += 4; 143 $elif LOG2N == 1: 144 $for M in reversed(range(MR)): 145 _mm_storel_pi((__m64*) c${M}, vacc${M}x${ABC[0:4]}); 146 147 $for M in reversed(range(MR)): 148 vacc${M}x${ABC[0:4]} = _mm_movehl_ps(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}); 149 150 $for M in reversed(range(MR)): 151 c${M} += 2; 152 $elif LOG2N == 0: 153 $for M in reversed(range(MR)): 154 _mm_store_ss(c${M}, vacc${M}x${ABC[0:4]}); 155 } 156 $if LOG2N == 3: 157 $for M in reversed(range(MR)): 158 __m128 vacc${M}x${ABC[0:4]} = _mm256_castps256_ps128(vacc${M}x${ABC[0:8]}); 159 160 nc = 0; 161 } 162 } while (nc != 0); 163} 164