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 = "0123456789ABCDEFGHIJKLMN" 8#include <assert.h> 9 10#include <xmmintrin.h> 11 12#include <xnnpack/igemm.h> 13 14 15void xnn_f32_igemm_ukernel_${MR}x${NR}__sse_dup( 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_output_params params[restrict static 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 while (k >= 4 * sizeof(float)) { 78 $for M in range(MR): 79 const __m128 va${M} = _mm_loadu_ps(a${M}); 80 a${M} += 4; 81 82 $for L in range(4): 83 $LLLL = str(L) * 4 84 85 $for M in range(MR): 86 const __m128 va${M}c${LLLL} = _mm_shuffle_ps(va${M}, va${M}, _MM_SHUFFLE(${L}, ${L}, ${L}, ${L})); 87 88 $for N in range(0, NR, 4): 89 const __m128 vb${ABC[N:N+4]}c${L} = _mm_load_ps(w + ${L * NR + N}); 90 91 $for N in range(0, NR, 4): 92 $for M in range(MR): 93 vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(va${M}c${LLLL}, vb${ABC[N:N+4]}c${L})); 94 95 w += ${4 * NR}; 96 k -= 4 * sizeof(float); 97 } 98 if XNN_UNLIKELY(k != 0) { 99 do { 100 const __m128 vb${ABC[0:4]} = _mm_load_ps(w); 101 $for N in range(4, NR, 4): 102 const __m128 vb${ABC[N:N+4]} = _mm_load_ps(w + ${N}); 103 w += ${NR}; 104 105 $for M in range(MR): 106 const __m128 va${M} = _mm_load1_ps(a${M}); 107 a${M} += 1; 108 109 $for M in range(MR): 110 $for N in range(0, NR, 4): 111 vacc${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]})); 112 k -= sizeof(float); 113 } while (k != 0); 114 } 115 p -= ${MR} * sizeof(void*); 116 } while (p != 0); 117 118 const __m128 vmax = _mm_load_ps(params->sse.max); 119 $for N in range(0, NR, 4): 120 $for M in range(MR): 121 vacc${M}x${ABC[N:N+4]} = _mm_min_ps(vacc${M}x${ABC[N:N+4]}, vmax); 122 123 const __m128 vmin = _mm_load_ps(params->sse.min); 124 $for N in range(0, NR, 4): 125 $for M in range(MR): 126 vacc${M}x${ABC[N:N+4]} = _mm_max_ps(vacc${M}x${ABC[N:N+4]}, vmin); 127 128 if XNN_LIKELY(nc >= ${NR}) { 129 $for M in reversed(range(MR)): 130 _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]}); 131 $for N in range(4, NR, 4): 132 _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 133 c${M} = (float*) ((uintptr_t) c${M} + cn_stride); 134 135 a = (const float**restrict) ((uintptr_t) a - ks); 136 nc -= ${NR}; 137 } else { 138 $for LOG2N in reversed(range(NR.bit_length())): 139 $if NR != 1 << LOG2N: 140 if (nc & ${1 << LOG2N}) { 141 $if LOG2N >= 2: 142 $for M in reversed(range(MR)): 143 _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]}); 144 $for N in range(4, 1 << LOG2N, 4): 145 _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 146 147 $for M in reversed(range(MR)): 148 $for N in range(0, 1 << (LOG2N - 1), 4): 149 vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]}; 150 151 $for M in reversed(range(MR)): 152 c${M} += ${1 << LOG2N}; 153 $elif LOG2N == 1: 154 $for M in reversed(range(MR)): 155 _mm_storel_pi((__m64*) c${M}, vacc${M}x${ABC[0:4]}); 156 157 $for M in reversed(range(MR)): 158 vacc${M}x${ABC[0:4]} = _mm_movehl_ps(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]}); 159 160 $for M in reversed(range(MR)): 161 c${M} += 2; 162 $elif LOG2N == 0: 163 $for M in reversed(range(MR)): 164 _mm_store_ss(c${M}, vacc${M}x${ABC[0:4]}); 165 } 166 167 nc = 0; 168 } 169 } while (nc != 0); 170} 171