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$assert NR % 4 == 0 7$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 8#include <assert.h> 9 10#include <xmmintrin.h> 11 12#include <xnnpack/igemm.h> 13 14 15void xnn_f32_igemm_minmax_ukernel_${MR}x${NR}s4__sse( 16 size_t mr, 17 size_t nc, 18 size_t kc, 19 size_t ks, 20 const float**restrict a, 21 const float*restrict w, 22 float*restrict c, 23 size_t cm_stride, 24 size_t cn_stride, 25 size_t a_offset, 26 const float* zero, 27 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 28{ 29 assert(mr != 0); 30 assert(mr <= ${MR}); 31 assert(nc != 0); 32 assert(kc != 0); 33 assert(kc % sizeof(float) == 0); 34 assert(ks != 0); 35 assert(ks % (${MR} * sizeof(void*)) == 0); 36 assert(a_offset % sizeof(float) == 0); 37 assert(a != NULL); 38 assert(w != NULL); 39 assert(c != NULL); 40 41 float* c0 = c; 42 $for M in range(1, MR): 43 float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride); 44 $if M % 2 == 0: 45 if XNN_UNPREDICTABLE(mr <= ${M}) { 46 c${M} = c${M-1}; 47 } 48 $elif M + 1 == MR: 49 if XNN_UNPREDICTABLE(mr != ${M+1}) { 50 c${M} = c${M-1}; 51 } 52 $else: 53 if XNN_UNPREDICTABLE(mr < ${M+1}) { 54 c${M} = c${M-1}; 55 } 56 57 do { 58 __m128 vacc0x${ABC[0:4]} = _mm_load_ps(w); 59 $for N in range(4, NR, 4): 60 __m128 vacc0x${ABC[N:N+4]} = _mm_load_ps(w + ${N}); 61 $for M in range(1, MR): 62 $for N in range(0, NR, 4): 63 __m128 vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]}; 64 w += ${NR}; 65 66 size_t p = ks; 67 do { 68 $for M in range(MR): 69 const float* restrict a${M} = a[${M}]; 70 assert(a${M} != NULL); 71 if XNN_UNPREDICTABLE(a${M} != zero) { 72 a${M} = (const float*) ((uintptr_t) a${M} + a_offset); 73 } 74 a += ${MR}; 75 76 size_t k = kc; 77 while (k >= 4 * sizeof(float)) { 78 $for M in range(MR): 79 __m128 va${M} = _mm_loadu_ps(a${M}); 80 a${M} += 4; 81 82 $for L in range(4): 83 84 $for N in range(0, NR, 4): 85 const __m128 vb${ABC[N:N+4]}c${L} = _mm_load_ps(w + ${L * NR + N}); 86 87 $for N in range(0, NR, 4): 88 $for M in range(MR): 89 vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(va${M}, vb${ABC[N:N+4]}c${L})); 90 91 $if L + 1 != 4: 92 $for M in range(MR): 93 va${M} = _mm_shuffle_ps(va${M}, va${M}, _MM_SHUFFLE(0, 3, 2, 1)); 94 95 w += ${4 * NR}; 96 k -= 4 * sizeof(float); 97 } 98 if XNN_UNLIKELY(k != 0) { 99 $for M in range(MR): 100 __m128 va${M} = _mm_loadu_ps(a${M}); 101 a${M} = (const float*) ((uintptr_t) a${M} + k); 102 103 $for L in range(4): 104 105 $for N in range(0, NR, 4): 106 const __m128 vb${ABC[N:N+4]}c${L} = _mm_load_ps(w + ${L * NR + N}); 107 108 $for N in range(0, NR, 4): 109 $for M in range(MR): 110 vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(_mm_andnot_ps(_mm_cmpeq_ps(_mm_setzero_ps(), vb${ABC[N:N+4]}c${L}), va${M}), vb${ABC[N:N+4]}c${L})); 111 112 $if L + 1 != 4: 113 $for M in range(MR): 114 va${M} = _mm_shuffle_ps(va${M}, va${M}, _MM_SHUFFLE(0, 3, 2, 1)); 115 116 w += ${4 * NR}; 117 } 118 p -= ${MR} * sizeof(void*); 119 } while (p != 0); 120 121 const __m128 vmax = _mm_load_ps(params->sse.max); 122 $for N in range(0, NR, 4): 123 $for M in range(MR): 124 vacc${M}x${ABC[N:N+4]} = _mm_min_ps(vacc${M}x${ABC[N:N+4]}, vmax); 125 126 const __m128 vmin = _mm_load_ps(params->sse.min); 127 $for N in range(0, NR, 4): 128 $for M in range(MR): 129 vacc${M}x${ABC[N:N+4]} = _mm_max_ps(vacc${M}x${ABC[N:N+4]}, vmin); 130 131 if XNN_LIKELY(nc >= ${NR}) { 132 $for M in reversed(range(MR)): 133 _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]}); 134 $for N in range(4, NR, 4): 135 _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 136 c${M} = (float*) ((uintptr_t) c${M} + cn_stride); 137 138 a = (const float**restrict) ((uintptr_t) a - ks); 139 nc -= ${NR}; 140 } else { 141 $for LOG2N in reversed(range(NR.bit_length())): 142 $if NR != 1 << LOG2N: 143 if (nc & ${1 << LOG2N}) { 144 $if LOG2N >= 2: 145 $for M in reversed(range(MR)): 146 _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]}); 147 $for N in range(4, 1 << LOG2N, 4): 148 _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 149 150 $for M in reversed(range(MR)): 151 $for N in range(0, 1 << (LOG2N - 1), 4): 152 vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]}; 153 154 $for M in reversed(range(MR)): 155 c${M} += ${1 << LOG2N}; 156 $elif LOG2N == 1: 157 $for M in reversed(range(MR)): 158 _mm_storel_pi((__m64*) c${M}, vacc${M}x${ABC[0:4]}); 159 160 $for M in reversed(range(MR)): 161 vacc${M}x${ABC[0:4]} = _mm_movehl_ps(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}); 162 163 $for M in reversed(range(MR)): 164 c${M} += 2; 165 $elif LOG2N == 0: 166 $for M in reversed(range(MR)): 167 _mm_store_ss(c${M}, vacc${M}x${ABC[0:4]}); 168 } 169 170 nc = 0; 171 } 172 } while (nc != 0); 173} 174