File size: 17,646 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
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
// 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 DATATYPE in ["F32", "QC4", "QC8"]
$if DATATYPE == "QC8" and SSE == 2:
  $assert NR % 8 == 0
$elif DATATYPE == "QC4":
  $assert NR == 8
$else:
  $assert NR % 4 == 0
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$SSE_HEADER = "immintrin.h" if AVX else {1: "immintrin.h", 2: "emmintrin.h", 4: "smmintrin.h"}[SSE]
#include <assert.h>

#include <${SSE_HEADER}>

#include <xnnpack/gemm.h>
$if DATATYPE == "QC8" and SSE == 4:
  #include <xnnpack/unaligned.h>


$ISA = {0: "avx", 3: "fma3"}[FMA] if AVX else {1: "sse", 2: "sse2", 4: "sse41"}[SSE]
$DATATYPE_SPEC = {"F32": "f32", "QC8": "f32_qc8w", "QC4": "f32_qc4w"}[DATATYPE]
void xnn_${DATATYPE_SPEC}_gemm${"inc" if INC else ""}_minmax_ukernel_${MR}x${NR}__${ISA}_dup(
    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 __m128i vminus_kernel_zero_point = _mm_load_si128((const __m128i *) params->sse.minus_kernel_zero_point);
    const __m128i vmask = _mm_load_si128((const __m128i *) params->sse.mask);

  do {
    $if INC:
      $for M in range(MR):
        $for N in range(0, NR, 4):
          __m128 vacc${M}x${ABC[N:N+4]} = _mm_load_ps(acc + ${M*NR+N});
      acc += ${MR*NR};
    $else:
      $for N in range(0, NR, 4):
        $if DATATYPE == "F32":
          __m128 vacc0x${ABC[N:N+4]} = _mm_load_ps(w + ${N});
        $else:
          __m128 vacc0x${ABC[N:N+4]} = _mm_loadu_ps((const float*) w + ${N});
      $for M in range(1, MR):
        $for N in range(0, NR, 4):
          __m128 vacc${M}x${ABC[N:N+4]} = vacc0x${ABC[N:N+4]};
      $if DATATYPE == "F32":
        w += ${NR};
      $else:
        w = (const float*) w + ${NR};

    size_t k = kc;
    for (; k >= 4 * sizeof(float); k -= 4 * sizeof(float)) {
      $for M in range(MR):
        const __m128 va${M} = _mm_loadu_ps(a${M});
        a${M} += 4;

      $for L in range(4):
        $LLLL = str(L) * 4

        $for M in range(MR):
          $if SSE >= 2 and L < 3:
            const __m128 va${M}c${LLLL} = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va${M}), _MM_SHUFFLE(${L}, ${L}, ${L}, ${L})));
          $elif AVX >= 1:
            const __m128 va${M}c${LLLL} = _mm_permute_ps(va${M}, _MM_SHUFFLE(${L}, ${L}, ${L}, ${L}));
          $else:
            const __m128 va${M}c${LLLL} = _mm_shuffle_ps(va${M}, va${M}, _MM_SHUFFLE(${L}, ${L}, ${L}, ${L}));

        $if DATATYPE == "F32":
          $for N in range(0, NR, 4):
            const __m128 vb${ABC[N:N+4]}c${L} = _mm_load_ps(w + ${L * NR + N});
        $elif DATATYPE == "QC4":
          $if L % 4 == 0:
            $if SSE >= 4:
              $for N in range(0, NR, 8):
                const __m128i vbwi${ABC[N:N+8]}c01 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 2})));
                const __m128i vbwi${ABC[N:N+8]}c23 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 2 + 8})));
              $for N in range(0, NR, 8):
                __m128i vbw${ABC[N:N+8]}c0 = _mm_and_si128(vbwi${ABC[N:N+8]}c01, vmask);
                __m128i vbw${ABC[N:N+8]}c1 = _mm_srli_epi16(vbwi${ABC[N:N+8]}c01, 4);
                __m128i vbw${ABC[N:N+8]}c2 = _mm_and_si128(vbwi${ABC[N:N+8]}c23, vmask);
                __m128i vbw${ABC[N:N+8]}c3 = _mm_srli_epi16(vbwi${ABC[N:N+8]}c23, 4);
            $else:
              $for N in range(0, NR, 8):
                __m128i vbi${ABC[N:N+8]}c0123 = _mm_loadu_si128((const __m128i *) ((const int8_t*) w + ${N * 2}));
              $for N in range(0, NR, 8):
                __m128i vbwi${ABC[N:N+8]}c01 = _mm_unpacklo_epi8(vbi${ABC[N:N+8]}c0123, vbi${ABC[N:N+8]}c0123);
                __m128i vbwi${ABC[N:N+8]}c23 = _mm_unpackhi_epi8(vbi${ABC[N:N+8]}c0123, vbi${ABC[N:N+8]}c0123);
                __m128i vbw${ABC[N:N+8]}c0 = _mm_and_si128(vbwi${ABC[N:N+8]}c01, vmask);
                __m128i vbw${ABC[N:N+8]}c1 = _mm_srli_epi16(vbwi${ABC[N:N+8]}c01, 12);
                __m128i vbw${ABC[N:N+8]}c2 = _mm_and_si128(vbwi${ABC[N:N+8]}c23, vmask);
                __m128i vbw${ABC[N:N+8]}c3 = _mm_srli_epi16(vbwi${ABC[N:N+8]}c23, 12);
          $for N in range(0, NR, 8):
            $if SSE >= 4:
              __m128i vbi${ABC[N:N+4]}c${L} = _mm_cvtepu16_epi32(vbw${ABC[N:N+8]}c${L});
            $else:
              __m128i vbi${ABC[N:N+4]}c${L} = _mm_unpacklo_epi16(vbw${ABC[N:N+8]}c${L}, _mm_setzero_si128());
            __m128i vbi${ABC[N+4:N+8]}c${L} = _mm_unpackhi_epi16(vbw${ABC[N:N+8]}c${L}, _mm_setzero_si128());
            vbi${ABC[N:N+4]}c${L} = _mm_add_epi32(vbi${ABC[N:N+4]}c${L}, vminus_kernel_zero_point);
            vbi${ABC[N+4:N+8]}c${L} = _mm_add_epi32(vbi${ABC[N+4:N+8]}c${L}, vminus_kernel_zero_point);
            const __m128 vb${ABC[N:N+4]}c${L} = _mm_cvtepi32_ps(vbi${ABC[N:N+4]}c${L});
            const __m128 vb${ABC[N+4:N+8]}c${L} = _mm_cvtepi32_ps(vbi${ABC[N+4:N+8]}c${L});
        $elif DATATYPE == "QC8":
          $if SSE >= 4:
            $for N in range(0, NR, 4):
              const __m128i vbi${ABC[N:N+4]}c${L} = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32((const int8_t*) w + ${L * NR + N})));
            $for N in range(0, NR, 4):
              const __m128 vb${ABC[N:N+4]}c${L} = _mm_cvtepi32_ps(vbi${ABC[N:N+4]}c${L});
          $else:
            $for N in range(0, NR, 8):
              const __m128i vb${ABC[N:N+8]}c${L} = _mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${L * NR + N}));
            $for N in range(0, NR, 8):
              const __m128i vbw${ABC[N:N+8]}c${L} = _mm_unpacklo_epi8(vb${ABC[N:N+8]}c${L}, vb${ABC[N:N+8]}c${L});
              const __m128 vb${ABC[N:N+4]}c${L} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(vbw${ABC[N:N+8]}c${L}, vbw${ABC[N:N+8]}c${L}), 24));
              const __m128 vb${ABC[N+4:N+8]}c${L} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(vbw${ABC[N:N+8]}c${L}, vbw${ABC[N:N+8]}c${L}), 24));

        $for N in range(0, NR, 4):
          $for M in range(MR):
            $if FMA == 3:
              vacc${M}x${ABC[N:N+4]} = _mm_fmadd_ps(va${M}c${LLLL}, vb${ABC[N:N+4]}c${L}, vacc${M}x${ABC[N:N+4]});
            $else:
              vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(va${M}c${LLLL}, vb${ABC[N:N+4]}c${L}));


      $if DATATYPE == "F32":
        w += ${NR * 4};
      $elif DATATYPE == "QC4":
        w = (const int8_t*) w + ${NR * 4 // 2};
      $else:
        w = (const int8_t*) w + ${NR * 4};
    }
    if XNN_UNLIKELY(k >= 2 * sizeof(float)) {
      $for M in range(MR):
        $if SSE >= 2:
          const __m128 va${M} = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i *) a${M}));
        $else:
          const __m128 va${M} = _mm_loadl_pi(_mm_undefined_ps(), (const __m64*) a${M});
        a${M} += 2;
      $for L in range(2):
        $LLLL = str(L) * 4

        $for M in range(MR):
          $if SSE >= 2 and L < 3:
            const __m128 va${M}c${LLLL} = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(va${M}), _MM_SHUFFLE(${L}, ${L}, ${L}, ${L})));
          $elif AVX >= 1:
            const __m128 va${M}c${LLLL} = _mm_permute_ps(va${M}, _MM_SHUFFLE(${L}, ${L}, ${L}, ${L}));
          $else:
            const __m128 va${M}c${LLLL} = _mm_shuffle_ps(va${M}, va${M}, _MM_SHUFFLE(${L}, ${L}, ${L}, ${L}));

        $if DATATYPE == "F32":
          $for N in range(0, NR, 4):
            const __m128 vb${ABC[N:N+4]}c${L} = _mm_load_ps(w + ${L * NR + N});
        $elif DATATYPE == "QC4":
          $if L % 2 == 0:
            $if SSE >= 4:
              $for N in range(0, NR, 8):
                const __m128i vbwi${ABC[N:N+8]}c01 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 2})));
              $for N in range(0, NR, 8):
                __m128i vbw${ABC[N:N+8]}c0 = _mm_and_si128(vbwi${ABC[N:N+8]}c01, vmask);
                __m128i vbw${ABC[N:N+8]}c1 = _mm_srli_epi16(vbwi${ABC[N:N+8]}c01, 4);
            $else:
              $for N in range(0, NR, 8):
                __m128i vbi${ABC[N:N+8]}c01 = _mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 2}));
              $for N in range(0, NR, 8):
                __m128i vbwi${ABC[N:N+8]}c01 = _mm_unpacklo_epi8(vbi${ABC[N:N+8]}c01, vbi${ABC[N:N+8]}c01);
                __m128i vbw${ABC[N:N+8]}c0 = _mm_and_si128(vbwi${ABC[N:N+8]}c01, vmask);
                __m128i vbw${ABC[N:N+8]}c1 = _mm_srli_epi16(vbwi${ABC[N:N+8]}c01, 12);
          $for N in range(0, NR, 8):
            $if SSE >= 4:
              __m128i vbi${ABC[N:N+4]}c${L} = _mm_cvtepu16_epi32(vbw${ABC[N:N+8]}c${L});
            $else:
              __m128i vbi${ABC[N:N+4]}c${L} = _mm_unpacklo_epi16(vbw${ABC[N:N+8]}c${L}, _mm_setzero_si128());
            __m128i vbi${ABC[N+4:N+8]}c${L} = _mm_unpackhi_epi16(vbw${ABC[N:N+8]}c${L}, _mm_setzero_si128());
            vbi${ABC[N:N+4]}c${L} = _mm_add_epi32(vbi${ABC[N:N+4]}c${L}, vminus_kernel_zero_point);
            vbi${ABC[N+4:N+8]}c${L} = _mm_add_epi32(vbi${ABC[N+4:N+8]}c${L}, vminus_kernel_zero_point);
            const __m128 vb${ABC[N:N+4]}c${L} = _mm_cvtepi32_ps(vbi${ABC[N:N+4]}c${L});
            const __m128 vb${ABC[N+4:N+8]}c${L} = _mm_cvtepi32_ps(vbi${ABC[N+4:N+8]}c${L});
        $elif DATATYPE == "QC8":
          $if SSE >= 4:
            $for N in range(0, NR, 4):
              const __m128i vbi${ABC[N:N+4]}c${L} = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32((const int8_t*) w + ${L * NR + N})));
            $for N in range(0, NR, 4):
              const __m128 vb${ABC[N:N+4]}c${L} = _mm_cvtepi32_ps(vbi${ABC[N:N+4]}c${L});
          $else:
            $for N in range(0, NR, 8):
              const __m128i vb${ABC[N:N+8]}c${L} = _mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${L * NR + N}));
            $for N in range(0, NR, 8):
              const __m128i vbw${ABC[N:N+8]}c${L} = _mm_unpacklo_epi8(vb${ABC[N:N+8]}c${L}, vb${ABC[N:N+8]}c${L});
              const __m128 vb${ABC[N:N+4]}c${L} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(vbw${ABC[N:N+8]}c${L}, vbw${ABC[N:N+8]}c${L}), 24));
              const __m128 vb${ABC[N+4:N+8]}c${L} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(vbw${ABC[N:N+8]}c${L}, vbw${ABC[N:N+8]}c${L}), 24));

        $for N in range(0, NR, 4):
          $for M in range(MR):
            $if FMA == 3:
              vacc${M}x${ABC[N:N+4]} = _mm_fmadd_ps(va${M}c${LLLL}, vb${ABC[N:N+4]}c${L}, vacc${M}x${ABC[N:N+4]});
            $else:
              vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(va${M}c${LLLL}, vb${ABC[N:N+4]}c${L}));

      $if DATATYPE == "F32":
        w += ${NR * 2};
      $elif DATATYPE == "QC4":
        w = (const int8_t*) w + ${NR * 2 // 2};
      $else:
        w = (const int8_t*) w + ${NR * 2};
      k -= 2 * sizeof(float);
    }
    if XNN_UNLIKELY(k != 0) {
      $for M in range(MR):
        const __m128 va${M} = _mm_load1_ps(a${M});
        a${M} += 1;

      $if DATATYPE == "F32":
        const __m128 vb${ABC[0:4]} = _mm_load_ps(w);
        $for N in range(4, NR, 4):
          const __m128 vb${ABC[N:N+4]} = _mm_load_ps(w + ${N});
        w += ${NR};
      $elif DATATYPE == "QC4":
        $if SSE >= 4:
          $for N in range(0, NR, 8):
            const __m128i vbwi${ABC[N:N+8]}c0 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 2})));
          $for N in range(0, NR, 8):
            __m128i vbw${ABC[N:N+8]}c0 = _mm_and_si128(vbwi${ABC[N:N+8]}c0, vmask);
        $else:
          $for N in range(0, NR, 8):
            __m128i vbi${ABC[N:N+8]}c0 = _mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N * 2}));
          $for N in range(0, NR, 8):
            __m128i vbwi${ABC[N:N+8]}c0 = _mm_unpacklo_epi8(vbi${ABC[N:N+8]}c0, vbi${ABC[N:N+8]}c0);
            __m128i vbw${ABC[N:N+8]}c0 = _mm_and_si128(vbwi${ABC[N:N+8]}c0, vmask);
        $for N in range(0, NR, 8):
          $if SSE >= 4:
            __m128i vbi${ABC[N:N+4]}c0 = _mm_cvtepu16_epi32(vbw${ABC[N:N+8]}c0);
          $else:
            __m128i vbi${ABC[N:N+4]}c0 = _mm_unpacklo_epi16(vbw${ABC[N:N+8]}c0, _mm_setzero_si128());
          __m128i vbi${ABC[N+4:N+8]}c0 = _mm_unpackhi_epi16(vbw${ABC[N:N+8]}c0, _mm_setzero_si128());
          vbi${ABC[N:N+4]}c0 = _mm_add_epi32(vbi${ABC[N:N+4]}c0, vminus_kernel_zero_point);
          vbi${ABC[N+4:N+8]}c0 = _mm_add_epi32(vbi${ABC[N+4:N+8]}c0, vminus_kernel_zero_point);
          const __m128 vb${ABC[N:N+4]} = _mm_cvtepi32_ps(vbi${ABC[N:N+4]}c0);
          const __m128 vb${ABC[N+4:N+8]} = _mm_cvtepi32_ps(vbi${ABC[N+4:N+8]}c0);
        w = (const int8_t*) w + ${NR};
      $elif DATATYPE == "QC8":
        $if SSE >= 4:
          const __m128i vbi${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32((const void*) w)));
          $for N in range(4, NR, 4):
            const __m128i vbi${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32((const int8_t*) w + ${N})));
          $for N in range(0, NR, 4):
            const __m128 vb${ABC[N:N+4]} = _mm_cvtepi32_ps(vbi${ABC[N:N+4]});
        $else:
          $for N in range(0, NR, 8):
            const __m128i vb${ABC[N:N+8]} = _mm_loadl_epi64((const __m128i *) ((const int8_t*) w + ${N}));
          $for N in range(0, NR, 8):
            const __m128i vbw${ABC[N:N+8]} = _mm_unpacklo_epi8(vb${ABC[N:N+8]}, vb${ABC[N:N+8]});
            const __m128 vb${ABC[N:N+4]} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpacklo_epi16(vbw${ABC[N:N+8]}, vbw${ABC[N:N+8]}), 24));
            const __m128 vb${ABC[N+4:N+8]} = _mm_cvtepi32_ps(_mm_srai_epi32(_mm_unpackhi_epi16(vbw${ABC[N:N+8]}, vbw${ABC[N:N+8]}), 24));
        w = (const int8_t*) w + ${NR};

      $for N in range(0, NR, 4):
        $for M in range(MR):
          $if FMA == 3:
            vacc${M}x${ABC[N:N+4]} = _mm_fmadd_ps(va${M}, vb${ABC[N:N+4]}, vacc${M}x${ABC[N:N+4]});
          $else:
            vacc${M}x${ABC[N:N+4]} = _mm_add_ps(vacc${M}x${ABC[N:N+4]}, _mm_mul_ps(va${M}, vb${ABC[N:N+4]}));

      k -= sizeof(float);
    }

    $if DATATYPE in ["QC8", "QC4"]:
      $for N in range(0, NR, 4):
        const __m128 vscale${ABC[N:N+4]} = _mm_loadu_ps((const float*) w + ${N});
        $for M in range(MR):
          vacc${M}x${ABC[N:N+4]} = _mm_mul_ps(vacc${M}x${ABC[N:N+4]}, vscale${ABC[N:N+4]});
      w = (const float*) w + ${NR};
    const __m128 vmax = _mm_load_ps(params->sse.max);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = _mm_min_ps(vacc${M}x${ABC[N:N+4]}, vmax);

    const __m128 vmin = _mm_load_ps(params->sse.min);
    $for N in range(0, NR, 4):
      $for M in range(MR):
        vacc${M}x${ABC[N:N+4]} = _mm_max_ps(vacc${M}x${ABC[N:N+4]}, vmin);

    if XNN_LIKELY(nc >= ${NR}) {
      $for M in reversed(range(MR)):
        _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]});
        $for N in range(4, NR, 4):
          _mm_storeu_ps(c${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:
              $for M in reversed(range(MR)):
                _mm_storeu_ps(c${M}, vacc${M}x${ABC[0:4]});
                $for N in range(4, 1 << LOG2N, 4):
                  _mm_storeu_ps(c${M} + ${N}, vacc${M}x${ABC[N:N+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]};

              $for M in reversed(range(MR)):
                c${M} += ${1 << LOG2N};
            $elif LOG2N == 1:
              $for M in reversed(range(MR)):
                _mm_storel_pi((__m64*) c${M}, vacc${M}x${ABC[0:4]});

              $for M in reversed(range(MR)):
                vacc${M}x${ABC[0:4]} = _mm_movehl_ps(vacc${M}x${ABC[0:4]}, vacc${M}x${ABC[0:4]});

              $for M in reversed(range(MR)):
                c${M} += 2;
            $elif LOG2N == 0:
              $for M in reversed(range(MR)):
                _mm_store_ss(c${M}, vacc${M}x${ABC[0:4]});
          }

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