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}__sse_load1( 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)]) 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 += ${MR}; 75 76 size_t k = kc; 77 do { 78 const __m128 vb${ABC[0:4]} = _mm_load_ps(w); 79 $for N in range(4, NR, 4): 80 const __m128 vb${ABC[N:N+4]} = _mm_load_ps(w + ${N}); 81 w += ${NR}; 82 83 $for M in range(MR): 84 const __m128 va${M} = _mm_load1_ps(a${M}); 85 a${M} += 1; 86 87 $for M in range(MR): 88 $for N in range(0, NR, 4): 89 vacc${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]})); 90 k -= sizeof(float); 91 } while (k != 0); 92 p -= ${MR} * sizeof(void*); 93 } while (p != 0); 94 95 const __m128 vmax = _mm_load_ps(params->sse.max); 96 $for N in range(0, NR, 4): 97 $for M in range(MR): 98 vacc${M}x${ABC[N:N+4]} = _mm_min_ps(vacc${M}x${ABC[N:N+4]}, vmax); 99 100 const __m128 vmin = _mm_load_ps(params->sse.min); 101 $for N in range(0, NR, 4): 102 $for M in range(MR): 103 vacc${M}x${ABC[N:N+4]} = _mm_max_ps(vacc${M}x${ABC[N:N+4]}, vmin); 104 105 if XNN_LIKELY(nc >= ${NR}) { 106 $for M in reversed(range(MR)): 107 _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]}); 108 $for N in range(4, NR, 4): 109 _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 110 c${M} = (float*) ((uintptr_t) c${M} + cn_stride); 111 112 a = (const float**restrict) ((uintptr_t) a - ks); 113 nc -= ${NR}; 114 } else { 115 $for LOG2N in reversed(range(NR.bit_length())): 116 $if NR != 1 << LOG2N: 117 if (nc & ${1 << LOG2N}) { 118 $if LOG2N >= 2: 119 $for M in reversed(range(MR)): 120 _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]}); 121 $for N in range(4, 1 << LOG2N, 4): 122 _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 123 124 $for M in reversed(range(MR)): 125 $for N in range(0, 1 << (LOG2N - 1), 4): 126 vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]}; 127 128 $for M in reversed(range(MR)): 129 c${M} += ${1 << LOG2N}; 130 $elif LOG2N == 1: 131 $for M in reversed(range(MR)): 132 _mm_storel_pi((__m64*) c${M}, vacc${M}x${ABC[0:4]}); 133 134 $for M in reversed(range(MR)): 135 vacc${M}x${ABC[0:4]} = _mm_movehl_ps(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}); 136 137 $for M in reversed(range(MR)): 138 c${M} += 2; 139 $elif LOG2N == 0: 140 $for M in reversed(range(MR)): 141 _mm_store_ss(c${M}, vacc${M}x${ABC[0:4]}); 142 } 143 144 nc = 0; 145 } 146 } while (nc != 0); 147} 148