File size: 12,252 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
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
// 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_4x4c8__neonbf16_bfdot(
    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 <= 4);
  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* a3 = (const bfloat16_t*) ((uintptr_t) a2 + a_stride);
  bfloat16_t* c3 = (bfloat16_t*) ((uintptr_t) c2 + cm_stride);
  if XNN_UNPREDICTABLE(mr != 4) {
    a3 = a2;
    c3 = c2;
  }

  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;
    float32x4_t vacc3x0 = vacc0x0;
    float32x4_t vacc3x1 = vacc0x1;
    float32x4_t vacc3x2 = vacc0x2;
    float32x4_t vacc3x3 = 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 va3 = vld1q_bf16(a3); a3 += 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 = vbfdotq_f32(vacc0x0, va0, vb0);
      vacc1x0 = vbfdotq_f32(vacc1x0, va1, vb0);
      vacc2x0 = vbfdotq_f32(vacc2x0, va2, vb0);
      vacc3x0 = vbfdotq_f32(vacc3x0, va3, vb0);
      vacc0x1 = vbfdotq_f32(vacc0x1, va0, vb1);
      vacc1x1 = vbfdotq_f32(vacc1x1, va1, vb1);
      vacc2x1 = vbfdotq_f32(vacc2x1, va2, vb1);
      vacc3x1 = vbfdotq_f32(vacc3x1, va3, vb1);
      vacc0x2 = vbfdotq_f32(vacc0x2, va0, vb2);
      vacc1x2 = vbfdotq_f32(vacc1x2, va1, vb2);
      vacc2x2 = vbfdotq_f32(vacc2x2, va2, vb2);
      vacc3x2 = vbfdotq_f32(vacc3x2, va3, vb2);
      vacc0x3 = vbfdotq_f32(vacc0x3, va0, vb3);
      vacc1x3 = vbfdotq_f32(vacc1x3, va1, vb3);
      vacc2x3 = vbfdotq_f32(vacc2x3, va2, vb3);
      vacc3x3 = vbfdotq_f32(vacc3x3, va3, 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 va3 = vld1q_bf16(a3); a3 = (const bfloat16_t*) ((uintptr_t) a3 + 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 = vbfdotq_f32(vacc0x0, va0x0, vb0);
      const bfloat16x8_t va1x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm0));
      vacc1x0 = vbfdotq_f32(vacc1x0, va1x0, vb0);
      const bfloat16x8_t va2x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm0));
      vacc2x0 = vbfdotq_f32(vacc2x0, va2x0, vb0);
      const bfloat16x8_t va3x0 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va3), vm0));
      vacc3x0 = vbfdotq_f32(vacc3x0, va3x0, vb0);
      const bfloat16x8_t va0x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm1));
      vacc0x1 = vbfdotq_f32(vacc0x1, va0x1, vb1);
      const bfloat16x8_t va1x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm1));
      vacc1x1 = vbfdotq_f32(vacc1x1, va1x1, vb1);
      const bfloat16x8_t va2x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm1));
      vacc2x1 = vbfdotq_f32(vacc2x1, va2x1, vb1);
      const bfloat16x8_t va3x1 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va3), vm1));
      vacc3x1 = vbfdotq_f32(vacc3x1, va3x1, vb1);
      const bfloat16x8_t va0x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm2));
      vacc0x2 = vbfdotq_f32(vacc0x2, va0x2, vb2);
      const bfloat16x8_t va1x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm2));
      vacc1x2 = vbfdotq_f32(vacc1x2, va1x2, vb2);
      const bfloat16x8_t va2x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm2));
      vacc2x2 = vbfdotq_f32(vacc2x2, va2x2, vb2);
      const bfloat16x8_t va3x2 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va3), vm2));
      vacc3x2 = vbfdotq_f32(vacc3x2, va3x2, vb2);
      const bfloat16x8_t va0x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va0), vm3));
      vacc0x3 = vbfdotq_f32(vacc0x3, va0x3, vb3);
      const bfloat16x8_t va1x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va1), vm3));
      vacc1x3 = vbfdotq_f32(vacc1x3, va1x3, vb3);
      const bfloat16x8_t va2x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va2), vm3));
      vacc2x3 = vbfdotq_f32(vacc2x3, va2x3, vb3);
      const bfloat16x8_t va3x3 = vreinterpretq_bf16_u16(vbicq_u16(vreinterpretq_u16_bf16(va3), vm3));
      vacc3x3 = vbfdotq_f32(vacc3x3, va3x3, 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 vacc3x01 = vpaddq_f32(vacc3x0, vacc3x1);
    const float32x4_t vacc0x23 = vpaddq_f32(vacc0x2, vacc0x3);
    const float32x4_t vacc1x23 = vpaddq_f32(vacc1x2, vacc1x3);
    const float32x4_t vacc2x23 = vpaddq_f32(vacc2x2, vacc2x3);
    const float32x4_t vacc3x23 = vpaddq_f32(vacc3x2, vacc3x3);

    float32x4_t vacc0x0123 = vpaddq_f32(vacc0x01, vacc0x23);
    float32x4_t vacc1x0123 = vpaddq_f32(vacc1x01, vacc1x23);
    float32x4_t vacc2x0123 = vpaddq_f32(vacc2x01, vacc2x23);
    float32x4_t vacc3x0123 = vpaddq_f32(vacc3x01, vacc3x23);
#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 vsum3x0 = vadd_f32(vget_low_f32(vacc3x0), vget_high_f32(vacc3x0));
    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 vsum3x1 = vadd_f32(vget_low_f32(vacc3x1), vget_high_f32(vacc3x1));
    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 vsum3x2 = vadd_f32(vget_low_f32(vacc3x2), vget_high_f32(vacc3x2));
    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));
    const float32x2_t vsum3x3 = vadd_f32(vget_low_f32(vacc3x3), vget_high_f32(vacc3x3));

    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));
    float32x4_t vacc3x0123 = vcombine_f32(vpadd_f32(vsum3x0, vsum3x1), vpadd_f32(vsum3x2, vsum3x3));
#endif

    const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
    vacc0x0123 = vminq_f32(vacc0x0123, vmax);
    vacc1x0123 = vminq_f32(vacc1x0123, vmax);
    vacc2x0123 = vminq_f32(vacc2x0123, vmax);
    vacc3x0123 = vminq_f32(vacc3x0123, vmax);

    const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
    vacc0x0123 = vmaxq_f32(vacc0x0123, vmin);
    vacc1x0123 = vmaxq_f32(vacc1x0123, vmin);
    vacc2x0123 = vmaxq_f32(vacc2x0123, vmin);
    vacc3x0123 = vmaxq_f32(vacc3x0123, vmin);

    bfloat16x4_t vout0x0123 = vcvt_bf16_f32(vacc0x0123);
    bfloat16x4_t vout1x0123 = vcvt_bf16_f32(vacc1x0123);
    bfloat16x4_t vout2x0123 = vcvt_bf16_f32(vacc2x0123);
    bfloat16x4_t vout3x0123 = vcvt_bf16_f32(vacc3x0123);

    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);
      vst1_bf16(c3, vout3x0123);
      c3 = (bfloat16_t*) ((uintptr_t) c3 + 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);
      a3 = (const bfloat16_t*) ((uintptr_t) a3 - 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;
        vst1_lane_u32((void*) c3, vreinterpret_u32_bf16(vout3x0123), 0); c3 += 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));
        vout3x0123 = vreinterpret_bf16_u16(vext_u16(vreinterpret_u16_bf16(vout3x0123), vreinterpret_u16_bf16(vout3x0123), 2));
      }
      if (nc & 1) {
        vst1_lane_bf16(c0, vout0x0123, 0);
        vst1_lane_bf16(c1, vout1x0123, 0);
        vst1_lane_bf16(c2, vout2x0123, 0);
        vst1_lane_bf16(c3, vout3x0123, 0);
      }

      nc = 0;
    }
  } while (nc != 0);
}