File size: 11,048 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 |
// Auto-generated file. Do not edit!
// Template: src/bf16-gemm/c8-neonbf16.c.in
// Generator: tools/xngen
//
// Copyright 2022 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <assert.h>
#include <arm_neon.h>
#include <xnnpack/gemm.h>
void xnn_bf16_gemm_minmax_ukernel_3x4c8__neonbf16_bfmlal(
size_t mr,
size_t nc,
size_t kc,
const void* restrict a,
size_t a_stride,
const void* restrict w_ptr,
void* restrict c,
size_t cm_stride,
size_t cn_stride,
const union xnn_bf16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(mr != 0);
assert(mr <= 3);
assert(nc != 0);
assert(kc != 0);
assert(kc % sizeof(bfloat16_t) == 0);
assert(a != NULL);
assert(w_ptr != NULL);
assert(c != NULL);
const bfloat16_t* a0 = (const bfloat16_t*) a;
bfloat16_t* c0 = (bfloat16_t*) c;
const bfloat16_t* a1 = (const bfloat16_t*) ((uintptr_t) a0 + a_stride);
bfloat16_t* c1 = (bfloat16_t*) ((uintptr_t) c0 + cm_stride);
if XNN_UNPREDICTABLE(mr < 2) {
a1 = a0;
c1 = c0;
}
const bfloat16_t* a2 = (const bfloat16_t*) ((uintptr_t) a1 + a_stride);
bfloat16_t* c2 = (bfloat16_t*) ((uintptr_t) c1 + cm_stride);
if XNN_UNPREDICTABLE(mr <= 2) {
a2 = a1;
c2 = c1;
}
const bfloat16_t* w = (const bfloat16_t*) w_ptr;
do {
float32x4_t vacc0x0 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
float32x4_t vacc0x1 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
float32x4_t vacc0x2 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
float32x4_t vacc0x3 = vcvt_f32_bf16(vld1_lane_bf16(w, vreinterpret_bf16_u16(vdup_n_u16(0)), 0)); w += 1;
float32x4_t vacc1x0 = vacc0x0;
float32x4_t vacc1x1 = vacc0x1;
float32x4_t vacc1x2 = vacc0x2;
float32x4_t vacc1x3 = vacc0x3;
float32x4_t vacc2x0 = vacc0x0;
float32x4_t vacc2x1 = vacc0x1;
float32x4_t vacc2x2 = vacc0x2;
float32x4_t vacc2x3 = vacc0x3;
size_t k = kc;
for (; k >= 8 * sizeof(bfloat16_t); k -= 8 * sizeof(bfloat16_t)) {
const bfloat16x8_t va0 = vld1q_bf16(a0); a0 += 8;
const bfloat16x8_t va1 = vld1q_bf16(a1); a1 += 8;
const bfloat16x8_t va2 = vld1q_bf16(a2); a2 += 8;
const bfloat16x8_t vb0 = vld1q_bf16(w); w += 8;
const bfloat16x8_t vb1 = vld1q_bf16(w); w += 8;
const bfloat16x8_t vb2 = vld1q_bf16(w); w += 8;
const bfloat16x8_t vb3 = vld1q_bf16(w); w += 8;
vacc0x0 = vbfmlalbq_f32(vacc0x0, va0, vb0);
vacc1x0 = vbfmlalbq_f32(vacc1x0, va1, vb0);
vacc2x0 = vbfmlalbq_f32(vacc2x0, va2, vb0);
vacc0x1 = vbfmlalbq_f32(vacc0x1, va0, vb1);
vacc1x1 = vbfmlalbq_f32(vacc1x1, va1, vb1);
vacc2x1 = vbfmlalbq_f32(vacc2x1, va2, vb1);
vacc0x2 = vbfmlalbq_f32(vacc0x2, va0, vb2);
vacc1x2 = vbfmlalbq_f32(vacc1x2, va1, vb2);
vacc2x2 = vbfmlalbq_f32(vacc2x2, va2, vb2);
vacc0x3 = vbfmlalbq_f32(vacc0x3, va0, vb3);
vacc1x3 = vbfmlalbq_f32(vacc1x3, va1, vb3);
vacc2x3 = vbfmlalbq_f32(vacc2x3, va2, vb3);
vacc0x0 = vbfmlaltq_f32(vacc0x0, va0, vb0);
vacc1x0 = vbfmlaltq_f32(vacc1x0, va1, vb0);
vacc2x0 = vbfmlaltq_f32(vacc2x0, va2, vb0);
vacc0x1 = vbfmlaltq_f32(vacc0x1, va0, vb1);
vacc1x1 = vbfmlaltq_f32(vacc1x1, va1, vb1);
vacc2x1 = vbfmlaltq_f32(vacc2x1, va2, vb1);
vacc0x2 = vbfmlaltq_f32(vacc0x2, va0, vb2);
vacc1x2 = vbfmlaltq_f32(vacc1x2, va1, vb2);
vacc2x2 = vbfmlaltq_f32(vacc2x2, va2, vb2);
vacc0x3 = vbfmlaltq_f32(vacc0x3, va0, vb3);
vacc1x3 = vbfmlaltq_f32(vacc1x3, va1, vb3);
vacc2x3 = vbfmlaltq_f32(vacc2x3, va2, vb3);
}
if XNN_UNLIKELY(k != 0) {
const bfloat16x8_t va0 = vld1q_bf16(a0); a0 = (const bfloat16_t*) ((uintptr_t) a0 + k);
const bfloat16x8_t va1 = vld1q_bf16(a1); a1 = (const bfloat16_t*) ((uintptr_t) a1 + k);
const bfloat16x8_t va2 = vld1q_bf16(a2); a2 = (const bfloat16_t*) ((uintptr_t) a2 + k);
const bfloat16x8_t vb0 = vld1q_bf16(w); w += 8;
const bfloat16x8_t vb1 = vld1q_bf16(w); w += 8;
const bfloat16x8_t vb2 = vld1q_bf16(w); w += 8;
const bfloat16x8_t vb3 = vld1q_bf16(w); w += 8;
const uint16x8_t vm0 = vceqq_u16(vreinterpretq_u16_bf16(vb0), vmovq_n_u16(0));
const uint16x8_t vm1 = vceqq_u16(vreinterpretq_u16_bf16(vb1), vmovq_n_u16(0));
const uint16x8_t vm2 = vceqq_u16(vreinterpretq_u16_bf16(vb2), vmovq_n_u16(0));
const uint16x8_t vm3 = vceqq_u16(vreinterpretq_u16_bf16(vb3), vmovq_n_u16(0));
const bfloat16x8_t va0x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm0));
vacc0x0 = vbfmlalbq_f32(vacc0x0, va0x0, vb0);
vacc0x0 = vbfmlaltq_f32(vacc0x0, va0x0, vb0);
const bfloat16x8_t va1x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm0));
vacc1x0 = vbfmlalbq_f32(vacc1x0, va1x0, vb0);
vacc1x0 = vbfmlaltq_f32(vacc1x0, va1x0, vb0);
const bfloat16x8_t va2x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm0));
vacc2x0 = vbfmlalbq_f32(vacc2x0, va2x0, vb0);
vacc2x0 = vbfmlaltq_f32(vacc2x0, va2x0, vb0);
const bfloat16x8_t va0x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm1));
vacc0x1 = vbfmlalbq_f32(vacc0x1, va0x1, vb1);
vacc0x1 = vbfmlaltq_f32(vacc0x1, va0x1, vb1);
const bfloat16x8_t va1x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm1));
vacc1x1 = vbfmlalbq_f32(vacc1x1, va1x1, vb1);
vacc1x1 = vbfmlaltq_f32(vacc1x1, va1x1, vb1);
const bfloat16x8_t va2x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm1));
vacc2x1 = vbfmlalbq_f32(vacc2x1, va2x1, vb1);
vacc2x1 = vbfmlaltq_f32(vacc2x1, va2x1, vb1);
const bfloat16x8_t va0x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm2));
vacc0x2 = vbfmlalbq_f32(vacc0x2, va0x2, vb2);
vacc0x2 = vbfmlaltq_f32(vacc0x2, va0x2, vb2);
const bfloat16x8_t va1x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm2));
vacc1x2 = vbfmlalbq_f32(vacc1x2, va1x2, vb2);
vacc1x2 = vbfmlaltq_f32(vacc1x2, va1x2, vb2);
const bfloat16x8_t va2x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm2));
vacc2x2 = vbfmlalbq_f32(vacc2x2, va2x2, vb2);
vacc2x2 = vbfmlaltq_f32(vacc2x2, va2x2, vb2);
const bfloat16x8_t va0x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm3));
vacc0x3 = vbfmlalbq_f32(vacc0x3, va0x3, vb3);
vacc0x3 = vbfmlaltq_f32(vacc0x3, va0x3, vb3);
const bfloat16x8_t va1x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm3));
vacc1x3 = vbfmlalbq_f32(vacc1x3, va1x3, vb3);
vacc1x3 = vbfmlaltq_f32(vacc1x3, va1x3, vb3);
const bfloat16x8_t va2x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm3));
vacc2x3 = vbfmlalbq_f32(vacc2x3, va2x3, vb3);
vacc2x3 = vbfmlaltq_f32(vacc2x3, va2x3, vb3);
}
#if XNN_ARCH_ARM64
const float32x4_t vacc0x01 = vpaddq_f32(vacc0x0, vacc0x1);
const float32x4_t vacc1x01 = vpaddq_f32(vacc1x0, vacc1x1);
const float32x4_t vacc2x01 = vpaddq_f32(vacc2x0, vacc2x1);
const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3);
const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3);
const float32x4_t vacc2x23 = vpaddq_f32(vacc2x2, vacc2x3);
float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23);
float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23);
float32x4_t vacc2x0123 = vpaddq_f32(vacc2x01, vacc2x23);
#else
const float32x2_t vsum0x0 = vadd_f32(vget_low_f32(vacc0x0), vget_high_f32(vacc0x0));
const float32x2_t vsum1x0 = vadd_f32(vget_low_f32(vacc1x0), vget_high_f32(vacc1x0));
const float32x2_t vsum2x0 = vadd_f32(vget_low_f32(vacc2x0), vget_high_f32(vacc2x0));
const float32x2_t vsum0x1 = vadd_f32(vget_low_f32(vacc0x1), vget_high_f32(vacc0x1));
const float32x2_t vsum1x1 = vadd_f32(vget_low_f32(vacc1x1), vget_high_f32(vacc1x1));
const float32x2_t vsum2x1 = vadd_f32(vget_low_f32(vacc2x1), vget_high_f32(vacc2x1));
const float32x2_t vsum0x2 = vadd_f32(vget_low_f32(vacc0x2), vget_high_f32(vacc0x2));
const float32x2_t vsum1x2 = vadd_f32(vget_low_f32(vacc1x2), vget_high_f32(vacc1x2));
const float32x2_t vsum2x2 = vadd_f32(vget_low_f32(vacc2x2), vget_high_f32(vacc2x2));
const float32x2_t vsum0x3 = vadd_f32(vget_low_f32(vacc0x3), vget_high_f32(vacc0x3));
const float32x2_t vsum1x3 = vadd_f32(vget_low_f32(vacc1x3), vget_high_f32(vacc1x3));
const float32x2_t vsum2x3 = vadd_f32(vget_low_f32(vacc2x3), vget_high_f32(vacc2x3));
float32x4_t vacc0x0123 = vcombine_f32(vpadd_f32(vsum0x0, vsum0x1), vpadd_f32(vsum0x2, vsum0x3));
float32x4_t vacc1x0123 = vcombine_f32(vpadd_f32(vsum1x0, vsum1x1), vpadd_f32(vsum1x2, vsum1x3));
float32x4_t vacc2x0123 = vcombine_f32(vpadd_f32(vsum2x0, vsum2x1), vpadd_f32(vsum2x2, vsum2x3));
#endif
const float32x4_t vmax = vld1q_dup_f32(¶ms->scalar.max);
vacc0x0123 = vminq_f32(vacc0x0123, vmax);
vacc1x0123 = vminq_f32(vacc1x0123, vmax);
vacc2x0123 = vminq_f32(vacc2x0123, vmax);
const float32x4_t vmin = vld1q_dup_f32(¶ms->scalar.min);
vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123);
bfloat16x4_t vout1x0123 = vcvt_bf16_f32(vacc1x0123);
bfloat16x4_t vout2x0123 = vcvt_bf16_f32(vacc2x0123);
if XNN_LIKELY(nc >= 4) {
vst1_bf16(c0, vout0x0123);
c0 = (bfloat16_t*) ((uintptr_t) c0 + cn_stride);
vst1_bf16(c1, vout1x0123);
c1 = (bfloat16_t*) ((uintptr_t) c1 + cn_stride);
vst1_bf16(c2, vout2x0123);
c2 = (bfloat16_t*) ((uintptr_t) c2 + cn_stride);
a0 = (const bfloat16_t*) ((uintptr_t) a0 - kc);
a1 = (const bfloat16_t*) ((uintptr_t) a1 - kc);
a2 = (const bfloat16_t*) ((uintptr_t) a2 - kc);
nc -= 4;
} else {
if (nc & 2) {
vst1_lane_u32((void*) c0, vreinterpret_u32_bf16(vout0x0123), 0); c0 += 2;
vst1_lane_u32((void*) c1, vreinterpret_u32_bf16(vout1x0123), 0); c1 += 2;
vst1_lane_u32((void*) c2, vreinterpret_u32_bf16(vout2x0123), 0); c2 += 2;
vout0x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout0x0123), vreinterpret_u16_bf16(vout0x0123), 2));
vout1x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout1x0123), vreinterpret_u16_bf16(vout1x0123), 2));
vout2x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout2x0123), vreinterpret_u16_bf16(vout2x0123), 2));
}
if (nc & 1) {
vst1_lane_bf16(c0, vout0x0123, 0);
vst1_lane_bf16(c1, vout1x0123, 0);
vst1_lane_bf16(c2, vout2x0123, 0);
}
nc = 0;
}
} while (nc != 0);
}
|