File size: 12,544 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
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
// Copyright 2019 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.

$assert NR % 4 == 0
$assert DATATYPE in ["F32", "QC4", "QC8"]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$VMULADDQ_F32 = "vfmaq_f32" if FMA else "vmlaq_f32"
$VMULADDQ_LANE_F32 = "vfmaq_lane_f32" if FMA else "vmlaq_lane_f32"

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/gemm.h>


$ISA = ("neonfma" if DUP else "aarch64_neonfma") if FMA else "neon"
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__${ISA}_${"dup" if DUP else "lane"}_ld128(
    size_t mr,
    size_t nc,
    size_t kc,
    const float* restrict a,
    size_t a_stride,
    $if DATATYPE == "F32":
      const float* restrict w,
    $else:
      const void* restrict w,
    float* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    $if INC:
      const float* restrict acc,
    $if DATATYPE == "QC4":
      const union xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
    $else:
      const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(mr != 0);
  assert(mr <= ${MR});
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);
  $if INC:
    assert(acc != NULL);

  const float* a0 = a;
  float* c0 = c;
  $for M in range(1, MR):
    const float* a${M} = (const float*) ((uintptr_t) a${M-1} + a_stride);
    float* c${M} = (float*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
  $if DATATYPE == "QC4":
    const int16x8_t vminus_kernel_zero_point = vld1q_dup_s16(&params->scalar.minus_kernel_zero_point[0]);
    const uint8x16_t vmask = vmovq_n_u8(UINT8_C(0xF));

  do {
    $if INC:
      $for M in range(MR):
        $for N in range(0, NR, 4):
          float32x4_t vacc${M}x${ABC[N:N+4]} = vld1q_f32(acc); acc += 4;
    $else:
      $for N in range(0, NR, 4):
        $if DATATYPE == "F32":
          float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w += 4;
        $else:
          float32x4_t vacc0x${ABC[N:N+4]} = vld1q_f32(w); w = (const float*) w + 4;
      $for M in range(1, MR):
        $for N in range(0, NR, 4):
          float32x4_t vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};

    size_t k = kc;
    if XNN_LIKELY(k >= 4 * sizeof(float)) {
      do {
        $for M in range(MR):
          const float32x4_t va${M} = vld1q_f32(a${M}); a${M} += 4;

        $for L in range(4):
          $VGET_PART_F32 = "vget_low_f32" if L < 2 else "vget_high_f32"

          $if DATATYPE == "F32":
            $for N in range(0, NR, 4):
              const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4;
          $elif DATATYPE == "QC4":
            $if L % 4 == 0:
              $for N in range(0, NR, 8):
                const uint8x16_t vw${ABC[N:N+8]}c0123 = vld1q_u8(w); w = (const uint8_t*) w + 16;
                const uint8x16_t vw${ABC[N:N+8]}c02 = vandq_u8(vw${ABC[N:N+8]}c0123, vmask);
                const uint8x16_t vw${ABC[N:N+8]}c13 = vshrq_n_u8(vw${ABC[N:N+8]}c0123, 4);
                const int16x8_t vbw${ABC[N:N+8]}c0 = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vget_low_u8(vw${ABC[N:N+8]}c02)));
                const int16x8_t vbw${ABC[N:N+8]}c1 = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vget_low_u8(vw${ABC[N:N+8]}c13)));
                const int16x8_t vbw${ABC[N:N+8]}c2 = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vget_high_u8(vw${ABC[N:N+8]}c02)));
                const int16x8_t vbw${ABC[N:N+8]}c3 = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vget_high_u8(vw${ABC[N:N+8]}c13)));
            $for N in range(0, NR, 8):
              const int32x4_t vbi${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vbw${ABC[N:N+8]}c${L}));
              const int32x4_t vbi${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vbw${ABC[N:N+8]}c${L}));
            $for N in range(0, NR, 4):
              const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vbi${ABC[N:N+4]}c${L});
          $else:
            $for N in range(0, NR, 8):
              const int8x8_t vw${ABC[N:N+8]}c${L} = vld1_s8(w); w = (const int8_t*) w + 8;
            $for N in range(0, NR, 8):
              const int16x8_t vbw${ABC[N:N+8]}c${L} = vmovl_s8(vw${ABC[N:N+8]}c${L});
            $for N in range(0, NR, 8):
              const int32x4_t vbi${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vbw${ABC[N:N+8]}c${L}));
              const int32x4_t vbi${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vbw${ABC[N:N+8]}c${L}));
            $for N in range(0, NR, 4):
              const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vbi${ABC[N:N+4]}c${L});

          $if DUP:
            $for M in range(MR):
              const float32x4_t va${M}c${L} = vdupq_lane_f32(${VGET_PART_F32}(va${M}), ${L % 2});
            $for N in range(0, NR, 4):
              $for M in range(MR):
                vacc${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});
          $else:
            $for N in range(0, NR, 4):
              $for M in range(MR):
                vacc${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});
        k -= 4 * sizeof(float);
      } while (k >= 4 * sizeof(float));
    }

    if XNN_UNLIKELY(k != 0) {
      if XNN_UNLIKELY(k & (2 * sizeof(float))) {
        $for M in range(MR):
          const float32x2_t va${M} = vld1_f32(a${M}); a${M} += 2;

        $for L in range(2):
          $VGET_PART_F32 = "vget_low_f32"

          $if DATATYPE == "F32":
            $for N in range(0, NR, 4):
              const float32x4_t vb${ABC[N:N+4]}c${L} = vld1q_f32(w); w += 4;
          $elif DATATYPE == "QC4":
            $if L % 2 == 0:
              $for N in range(0, NR, 8):
                const uint8x8_t vw${ABC[N:N+8]}c01 = vld1_u8(w); w = (const uint8_t*) w + 8;
                const uint8x8_t vw${ABC[N:N+8]}c0 = vand_u8(vw${ABC[N:N+8]}c01, vget_low_u8(vmask));
                const uint8x8_t vw${ABC[N:N+8]}c1 = vshr_n_u8(vw${ABC[N:N+8]}c01, 4);
                const int16x8_t vbw${ABC[N:N+8]}c0 = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vw${ABC[N:N+8]}c0));
                const int16x8_t vbw${ABC[N:N+8]}c1 = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vw${ABC[N:N+8]}c1));
            $for N in range(0, NR, 8):
              const int32x4_t vbi${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vbw${ABC[N:N+8]}c${L}));
              const int32x4_t vbi${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vbw${ABC[N:N+8]}c${L}));
            $for N in range(0, NR, 4):
              const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vbi${ABC[N:N+4]}c${L});
          $else:
            $for N in range(0, NR, 8):
              const int8x8_t vw${ABC[N:N+8]}c${L} = vld1_s8(w); w = (const int8_t*) w + 8;
            $for N in range(0, NR, 8):
              const int16x8_t vbw${ABC[N:N+8]}c${L} = vmovl_s8(vw${ABC[N:N+8]}c${L});
            $for N in range(0, NR, 8):
              const int32x4_t vbi${ABC[N:N+4]}c${L} = vmovl_s16(vget_low_s16(vbw${ABC[N:N+8]}c${L}));
              const int32x4_t vbi${ABC[N+4:N+8]}c${L} = vmovl_s16(vget_high_s16(vbw${ABC[N:N+8]}c${L}));
            $for N in range(0, NR, 4):
              const float32x4_t vb${ABC[N:N+4]}c${L} = vcvtq_f32_s32(vbi${ABC[N:N+4]}c${L});

          $if DUP:
            $for M in range(MR):
              const float32x4_t va${M}c${L} = vdupq_lane_f32(va${M}, ${L});
            $for N in range(0, NR, 4):
              $for M in range(MR):
                vacc${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});
          $else:
            $for N in range(0, NR, 4):
              $for M in range(MR):
                vacc${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});
      }
      if XNN_UNLIKELY(k & (1 * sizeof(float))) {
        $for M in range(MR):
          const float32x4_t va${M} = vld1q_dup_f32(a${M}); a${M} += 1;

        $if DATATYPE == "F32":
          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]} = vld1q_f32(w); w += 4;
        $elif DATATYPE == "QC4":
          $for N in range(0, NR, 8):
            const uint8x8_t vw${ABC[N:N+8]}c01 = vld1_u8(w); w = (const uint8_t*) w + 8;
            const uint8x8_t vw${ABC[N:N+8]}c0 = vand_u8(vw${ABC[N:N+8]}c01, vget_low_u8(vmask));
            const int16x8_t vbw${ABC[N:N+8]}c0 = vaddw_s8(vminus_kernel_zero_point, vreinterpret_s8_u8(vw${ABC[N:N+8]}c0));
          $for N in range(0, NR, 8):
            const int32x4_t vbi${ABC[N:N+4]}c0 = vmovl_s16(vget_low_s16(vbw${ABC[N:N+8]}c0));
            const int32x4_t vbi${ABC[N+4:N+8]}c0 = vmovl_s16(vget_high_s16(vbw${ABC[N:N+8]}c0));
          $for N in range(0, NR, 4):
            const float32x4_t vb${ABC[N:N+4]} = vcvtq_f32_s32(vbi${ABC[N:N+4]}c0);
        $else:
          $for N in range(0, NR, 8):
            const int8x8_t vw${ABC[N:N+8]} = vld1_s8(w); w = (const int8_t*) w + 8;
          $for N in range(0, NR, 8):
            const int16x8_t vbi${ABC[N:N+8]} = vmovl_s8(vw${ABC[N:N+8]});
          $for N in range(0, NR, 8):
            const int32x4_t vbi${ABC[N:N+4]} = vmovl_s16(vget_low_s16(vbi${ABC[N:N+8]}));
          $for N in range(0, NR, 8):
            const int32x4_t vbi${ABC[N+4:N+8]} = vmovl_s16(vget_high_s16(vbi${ABC[N:N+8]}));
          $for N in range(0, NR, 8):
            const float32x4_t vb${ABC[N:N+4]} = vcvtq_f32_s32(vbi${ABC[N:N+4]});
            const float32x4_t vb${ABC[N+4:N+8]} = vcvtq_f32_s32(vbi${ABC[N+4:N+8]});

        $for N in range(0, NR, 4):
          $for M in range(MR):
            vacc${M}x${ABC[N:N+4]} = ${VMULADDQ_F32}(vacc${M}x${ABC[N:N+4]}, va${M}, vb${ABC[N:N+4]});
      }
    }
    $if DATATYPE in ["QC8", "QC4"]:
      $for N in range(0, NR, 4):
        const float32x4_t vscale${ABC[N:N+4]} = vld1q_f32(w); w = ((const float*) w + 4);
      $for N in range(0, NR, 4):
        $for M in range(MR):
          vacc${M}x${ABC[N:N+4]} = vmulq_f32(vacc${M}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
    const float32x4_t vmax = vld1q_dup_f32(&params->scalar.max);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = vminq_f32(vacc${M}x${ABC[N:N+4]}, vmax);

    const float32x4_t vmin = vld1q_dup_f32(&params->scalar.min);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = vmaxq_f32(vacc${M}x${ABC[N:N+4]}, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        vst1q_f32(c${M}, vacc${M}x${ABC[0:4]});
        $for N in range(4, NR, 4):
          vst1q_f32(c${M} + ${N}, vacc${M}x${ABC[N:N+4]});
        c${M} = (float*) ((uintptr_t) c${M} + cn_stride);

      $for M in reversed(range(MR)):
        a${M} = (const float*) ((uintptr_t) a${M} - kc);

      nc -= ${NR};

    } else {
      $for LOG2N in reversed(range(NR.bit_length())):
        $if NR != 1 << LOG2N:
          if (nc & ${1 << LOG2N}) {
            $if LOG2N >= 2:
              $for N in range(0, 1 << LOG2N, 4):
                $for M in reversed(range(MR)):
                  vst1q_f32(c${M}, vacc${M}x${ABC[N:N+4]}); c${M} += 4;

              $for M in reversed(range(MR)):
                $for N in range(0, NR - (1 << LOG2N), 4):
                  vacc${M}x${ABC[N:N+4]} = vacc${M}x${ABC[N + (1 << LOG2N):N + (1 <<   LOG2N)+4]};
            $elif LOG2N == 1:
              $for M in reversed(range(MR)):
                vst1_f32(c${M}, vacc${M}x${ABC[0:2]}); c${M} += 2;

              $for M in reversed(range(MR)):
                vacc${M}x${ABC[0:2]} = vget_high_f32(vacc${M}x${ABC[0:4]});
            $elif LOG2N == 0:
              $for M in reversed(range(MR)):
                vst1_lane_f32(c${M}, vacc${M}x${ABC[0:2]}, 0);
          }
          $if LOG2N == 2:
            $for M in reversed(range(MR)):
              float32x2_t vacc${M}x${ABC[0:2]} = vget_low_f32(vacc${M}x${ABC[0:4]});

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