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$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32" 9$VMULADDQ_LANE_F32 = "vfmaq_lane_f32" if FMA else "vmlaq_lane_f32" 10 11#include <assert.h> 12 13#include <arm_neon.h> 14 15#include <xnnpack/igemm.h> 16 17 18void xnn_f32_igemm_minmax_ukernel_${MR}x${NR}__${"neonfma" if FMA else "neon"}_${"dup" if DUP else "lane"}_ld64( 19 size_t mr, 20 size_t nc, 21 size_t kc, 22 size_t ks, 23 const float**restrict a, 24 const float*restrict w, 25 float*restrict c, 26 size_t cm_stride, 27 size_t cn_stride, 28 size_t a_offset, 29 const float* zero, 30 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) 31{ 32 assert(mr != 0); 33 assert(mr <= ${MR}); 34 assert(nc != 0); 35 assert(kc != 0); 36 assert(kc % sizeof(float) == 0); 37 assert(ks != 0); 38 assert(ks % (${MR} * sizeof(void*)) == 0); 39 assert(a_offset % sizeof(float) == 0); 40 assert(a != NULL); 41 assert(w != NULL); 42 assert(c != NULL); 43 44 float* c0 = c; 45 $for M in range(1, MR): 46 float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride); 47 $if M % 2 == 0: 48 if XNN_UNPREDICTABLE(mr <= ${M}) { 49 c${M} = c${M-1}; 50 } 51 $elif M + 1 == MR: 52 if XNN_UNPREDICTABLE(mr != ${M+1}) { 53 c${M} = c${M-1}; 54 } 55 $else: 56 if XNN_UNPREDICTABLE(mr < ${M+1}) { 57 c${M} = c${M-1}; 58 } 59 60 do { 61 $for N in range(0, NR, 4): 62 float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w += 4; 63 $for M in range(1, MR): 64 $for N in range(0, NR, 4): 65 float32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]}; 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 for (; k >= 2 * sizeof(float); k -= 2 * sizeof(float)) { 79 $for M in range(MR): 80 const float32x2_t va${M} = vld1_f32(a${M}); a${M} += 2; 81 82 $for L in range(2): 83 $for N in range(0, NR, 4): 84 const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4; 85 86 $if DUP: 87 $for M in range(MR): 88 const float32x4_t va${M}c${L} = vdupq_lane_f32(va${M}, ${L}); 89 $for N in range(0, NR, 4): 90 $for M in range(MR): 91 vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, va${M}c${L}, vb${ABC[N:N+4]}c${L}); 92 $else: 93 $for N in range(0, NR, 4): 94 $for M in range(MR): 95 vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_LANE_F32}(vacc${M}x${ABC[N:N+4]}, vb${ABC[ N:N+4]}c${L}, va${M}, ${L}); 96 97 } 98 if XNN_UNLIKELY(k != 0) { 99 $for M in range(MR): 100 const float32x4_t va${M} = vld1q_dup_f32(a${M}); 101 102 $for N in range(0, NR, 4): 103 const float32x4_t vb${ABC[N:N+4]} = vld1q_f32(w); w += 4; 104 105 $for N in range(0, NR, 4): 106 $for M in range(MR): 107 vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]}); 108 } 109 p -= ${MR} * sizeof(void*); 110 } while (p != 0); 111 112 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); 113 $for N in range(0, NR, 4): 114 $for M in range(MR): 115 vacc${M}x${ABC[N:N+4]} = vminq_f32(vacc${M}x${ABC[N:N+4]}, vmax); 116 117 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); 118 $for N in range(0, NR, 4): 119 $for M in range(MR): 120 vacc${M}x${ABC[N:N+4]} = vmaxq_f32(vacc${M}x${ABC[N:N+4]}, vmin); 121 122 if XNN_LIKELY(nc >= ${NR}) { 123 $for M in reversed(range(MR)): 124 vst1q_f32(c${M}, vacc${M}x${ABC[0:4]}); 125 $for N in range(4, NR, 4): 126 vst1q_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 127 c${M} = (float*) ((uintptr_t) c${M} + cn_stride); 128 129 a = (const float**restrict) ((uintptr_t) a - ks); 130 nc -= ${NR}; 131 } else { 132 $for LOG2N in reversed(range(NR.bit_length())): 133 $if LOG2N == 1: 134 $for M in reversed(range(MR)): 135 float32x2_t vacc${M}x${ABC[0:2]} = vget_low_f32(vacc${M}x${ABC[0:4]}); 136 $if 1 << LOG2N != NR: 137 if (nc & ${1 << LOG2N}) { 138 $if LOG2N >= 2: 139 $for N in range(0, 1 << LOG2N, 4): 140 $for M in reversed(range(MR)): 141 vst1q_f32(c${M}, vacc${M}x${ABC[N:N+4]}); c${M} += 4; 142 143 $for M in reversed(range(MR)): 144 $for N in range(0, 1 << (LOG2N - 1), 4): 145 vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]}; 146 $elif LOG2N == 1: 147 $for M in reversed(range(MR)): 148 vst1_f32(c${M}, vacc${M}x${ABC[0:2]}); c${M} += 2; 149 150 $for M in reversed(range(MR)): 151 vacc${M}x${ABC[0:2]} = vget_high_f32(vacc${M}x${ABC[0:4]}); 152 $elif LOG2N == 0: 153 $for M in reversed(range(MR)): 154 vst1_lane_f32(c${M}, vacc${M}x${ABC[0:2]}, 0); 155 } 156 157 nc = 0; 158 } 159 } while (nc != 0); 160} 161