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/gemm.h> 16 17 18void xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__${"neonfma" if FMA else "neon"}_${"dup" if DUP else "lane"}_ld128( 19 size_t mr, 20 size_t nc, 21 size_t kc, 22 const float* restrict a, 23 size_t a_stride, 24 const float* restrict w, 25 float* restrict c, 26 size_t cm_stride, 27 size_t cn_stride, 28 $if INC: 29 const float*restrict acc, 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(a != NULL); 38 assert(w != NULL); 39 assert(c != NULL); 40 $if INC: 41 assert(acc != NULL); 42 43 const float* a0 = a; 44 float* c0 = c; 45 $for M in range(1, MR): 46 const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride); 47 float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride); 48 $if M % 2 == 0: 49 if XNN_UNPREDICTABLE(mr <= ${M}) { 50 a${M} = a${M-1}; 51 c${M} = c${M-1}; 52 } 53 $elif M + 1 == MR: 54 if XNN_UNPREDICTABLE(mr != ${M+1}) { 55 a${M} = a${M-1}; 56 c${M} = c${M-1}; 57 } 58 $else: 59 if XNN_UNPREDICTABLE(mr < ${M+1}) { 60 a${M} = a${M-1}; 61 c${M} = c${M-1}; 62 } 63 64 do { 65 $if INC: 66 $for M in range(MR): 67 $for N in range(0, NR, 4): 68 float32x4_t vacc${M}x${ABC[N:N+4]} = vld1q_f32(acc); acc += 4; 69 $else: 70 $for N in range(0, NR, 4): 71 float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w += 4; 72 $for M in range(1, MR): 73 $for N in range(0, NR, 4): 74 float32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]}; 75 76 size_t k = kc; 77 for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) { 78 $for M in range(MR): 79 const float32x4_t va${M} = vld1q_f32(a${M}); a${M} += 4; 80 81 $for L in range(4): 82 $VGET_PART_F32 = "vget_low_f32" if L < 2 else "vget_high_f32" 83 84 $for N in range(0, NR, 4): 85 const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4; 86 87 $if DUP: 88 $for M in range(MR): 89 const float32x4_t va${M}c${L} = vdupq_lane_f32(${VGET_PART_F32}(va${M}), ${L % 2}); 90 $for N in range(0, NR, 4): 91 $for M in range(MR): 92 vacc${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}); 93 $else: 94 $for N in range(0, NR, 4): 95 $for M in range(MR): 96 vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_LANE_F32}(vacc${M}x${ABC[N:N+4]}, vb${ABC[N:N+4]}c${L}, ${VGET_PART_F32}(va${M}), ${L % 2}); 97 } 98 if XNN_UNLIKELY(k != 0) { 99 do { 100 $for M in range(MR): 101 const float32x4_t va${M} = vld1q_dup_f32(a${M}); a${M} += 1; 102 103 $for N in range(0, NR, 4): 104 const float32x4_t vb${ABC[N:N+4]} = vld1q_f32(w); w += 4; 105 106 $for N in range(0, NR, 4): 107 $for M in range(MR): 108 vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]}); 109 110 k -= sizeof(float); 111 } while (k != 0); 112 } 113 const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max); 114 $for N in range(0, NR, 4): 115 $for M in range(MR): 116 vacc${M}x${ABC[N:N+4]} = vminq_f32(vacc${M}x${ABC[N:N+4]}, vmax); 117 118 const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min); 119 $for N in range(0, NR, 4): 120 $for M in range(MR): 121 vacc${M}x${ABC[N:N+4]} = vmaxq_f32(vacc${M}x${ABC[N:N+4]}, vmin); 122 123 if XNN_LIKELY(nc >= ${NR}) { 124 $for M in reversed(range(MR)): 125 vst1q_f32(c${M}, vacc${M}x${ABC[0:4]}); 126 $for N in range(4, NR, 4): 127 vst1q_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]}); 128 c${M} = (float*) ((uintptr_t) c${M} + cn_stride); 129 130 $for M in reversed(range(MR)): 131 a${M} = (const float*) ((uintptr_t) a${M} - kc); 132 133 nc -= ${NR}; 134 135 } else { 136 $for LOG2N in reversed(range(NR.bit_length())): 137 $if NR != 1 << LOG2N: 138 if (nc & ${1 << LOG2N}) { 139 $if LOG2N >= 2: 140 $for N in range(0, 1 << LOG2N, 4): 141 $for M in reversed(range(MR)): 142 vst1q_f32(c${M}, vacc${M}x${ABC[N:N+4]}); c${M} += 4; 143 144 $for M in reversed(range(MR)): 145 $for N in range(0, 1 << (LOG2N - 1), 4): 146 vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 << LOG2N)+4]}; 147 $elif LOG2N == 1: 148 $for M in reversed(range(MR)): 149 vst1_f32(c${M}, vacc${M}x${ABC[0:2]}); c${M} += 2; 150 151 $for M in reversed(range(MR)): 152 vacc${M}x${ABC[0:2]} = vget_high_f32(vacc${M}x${ABC[0:4]}); 153 $elif LOG2N == 0: 154 $for M in reversed(range(MR)): 155 vst1_lane_f32(c${M}, vacc${M}x${ABC[0:2]}, 0); 156 } 157 $if LOG2N == 2: 158 $for M in reversed(range(MR)): 159 float32x2_t vacc${M}x${ABC[0:2]} = vget_low_f32(vacc${M}x${ABC[0:4]}); 160 161 nc = 0; 162 } 163 } while (nc != 0); 164}