Home
last modified time | relevance | path

Searched refs:vaccGHIJKLMN (Results 1 – 21 of 21) sorted by relevance

/external/XNNPACK/src/qs8-vadd/gen/
Dminmax-avx2-mul32-ld64-x24.c47 …__m256i vaccGHIJKLMN = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vxGHIJKLMN, vx_mul… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24() local
51 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vyGHIJKLMN, vy_multiplier)); in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24()
55 …GHIJKLMN = _mm256_add_epi32(_mm256_and_si256(vaccGHIJKLMN, vremainder_mask), _mm256_srai_epi32(vac… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24()
59 …vaccGHIJKLMN = _mm256_sub_epi32(_mm256_sra_epi32(vaccGHIJKLMN, vshift), _mm256_cmpgt_epi32(vremGHI… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24()
62 …adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extracti128_si256(vaccGHIJ… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24()
Dminmax-avx2-mul32-ld64-x32.c49 …__m256i vaccGHIJKLMN = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vxGHIJKLMN, vx_mul… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32() local
54 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vyGHIJKLMN, vy_multiplier)); in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32()
59 …GHIJKLMN = _mm256_add_epi32(_mm256_and_si256(vaccGHIJKLMN, vremainder_mask), _mm256_srai_epi32(vac… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32()
64 …vaccGHIJKLMN = _mm256_sub_epi32(_mm256_sra_epi32(vaccGHIJKLMN, vshift), _mm256_cmpgt_epi32(vremGHI… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32()
68 …__m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), v… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32()
Dminmax-neon-ld64-x24.c79 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24() local
82 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24()
Dminmax-neon-ld64-x32.c91 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() local
95 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup24x9-minmax-avx2-mul32.c85 …__m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32() local
98 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi0xGHIJKLMN, vk0xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
110 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi1xGHIJKLMN, vk1xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
122 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi2xGHIJKLMN, vk2xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
134 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
146 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
158 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi5xGHIJKLMN, vk5xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
170 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi6xGHIJKLMN, vk6xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
182 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
194 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi8xGHIJKLMN, vk8xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
[all …]
Dup32x9-minmax-avx2-mul32.c85 …__m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t))); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32() local
101 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi0xGHIJKLMN, vk0xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
116 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi1xGHIJKLMN, vk1xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
131 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi2xGHIJKLMN, vk2xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
146 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi3xGHIJKLMN, vk3xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
161 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi4xGHIJKLMN, vk4xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
176 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi5xGHIJKLMN, vk5xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
191 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi6xGHIJKLMN, vk6xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
206 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi7xGHIJKLMN, vk7xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
221 vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_mullo_epi32(vi8xGHIJKLMN, vk8xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
[all …]
Dup32x9-minmax-avx2-mul16.c85 …__m256i vaccGHIJKLMN = _mm256_loadu_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int32_t))); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() local
101 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod0x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
116 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod1x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
131 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod2x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
146 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod3x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
161 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod4x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
176 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod5x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
191 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod6x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
206 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod7x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
221 …vaccGHIJKLMN = _mm256_add_epi32(vaccGHIJKLMN, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(vprod8x… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
[all …]
Dup24x9-minmax-neon-mul16.c247 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN), voutput… in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
250 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
254 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16() local
257 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
Dup32x9-minmax-neon-mul16.c291 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN), voutput… in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
295 int8x16_t voutGHIJKLMNOPQRSTUV = vqmovn_high_s16(vqmovn_s16(vaccGHIJKLMN), vaccOPQRSTUV); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
299 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16() local
303 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
/external/XNNPACK/src/qs8-vaddc/gen/
Dminmax-avx2-mul32-ld64-x24.c44 …__m256i vaccGHIJKLMN = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vxGHIJKLMN, vx_mul… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24() local
48 …GHIJKLMN = _mm256_add_epi32(_mm256_and_si256(vaccGHIJKLMN, vremainder_mask), _mm256_srai_epi32(vac… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24()
52 …vaccGHIJKLMN = _mm256_sub_epi32(_mm256_sra_epi32(vaccGHIJKLMN, vshift), _mm256_cmpgt_epi32(vremGHI… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24()
55 …adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extracti128_si256(vaccGHIJ… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24()
Dminmax-avx2-mul32-ld64-x32.c45 …__m256i vaccGHIJKLMN = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vxGHIJKLMN, vx_mul… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x32() local
50 …GHIJKLMN = _mm256_add_epi32(_mm256_and_si256(vaccGHIJKLMN, vremainder_mask), _mm256_srai_epi32(vac… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x32()
55 …vaccGHIJKLMN = _mm256_sub_epi32(_mm256_sra_epi32(vaccGHIJKLMN, vshift), _mm256_cmpgt_epi32(vremGHI… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x32()
59 …__m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), v… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x32()
Dminmax-neon-ld64-x24.c71 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24() local
74 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x24()
Dminmax-neon-ld64-x32.c79 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32() local
83 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32()
/external/XNNPACK/src/f16-spmm/gen/
D24x1-minmax-neonfp16arith-x2.c76 float16x8_t vaccGHIJKLMN = vaccGHIJKLMNx0; in xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith_x2() local
79 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vaccGHIJKLMNx1); in xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith_x2()
90 vaccGHIJKLMN = vfmaq_f16(vaccGHIJKLMN, vaGHIJKLMN, vb); in xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith_x2()
95 float16x8_t voutGHIJKLMN = vmulq_f16(vaccGHIJKLMN, vscale); in xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith_x2()
D32x1-minmax-neonfp16arith-x2.c82 float16x8_t vaccGHIJKLMN = vaccGHIJKLMNx0; in xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith_x2() local
86 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vaccGHIJKLMNx1); in xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith_x2()
99 vaccGHIJKLMN = vfmaq_f16(vaccGHIJKLMN, vaGHIJKLMN, vb); in xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith_x2()
105 float16x8_t voutGHIJKLMN = vmulq_f16(vaccGHIJKLMN, vscale); in xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith_x2()
D32x1-minmax-neonfp16arith.c49 float16x8_t vaccGHIJKLMN = vacc01234567; in xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith() local
62 vaccGHIJKLMN = vfmaq_f16(vaccGHIJKLMN, vaGHIJKLMN, vb); in xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith()
68 float16x8_t voutGHIJKLMN = vmulq_f16(vaccGHIJKLMN, vscale); in xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith()
D24x1-minmax-neonfp16arith.c49 float16x8_t vaccGHIJKLMN = vacc01234567; in xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith() local
60 vaccGHIJKLMN = vfmaq_f16(vaccGHIJKLMN, vaGHIJKLMN, vb); in xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith()
65 float16x8_t voutGHIJKLMN = vmulq_f16(vaccGHIJKLMN, vscale); in xnn_f16_spmm_minmax_ukernel_24x1__neonfp16arith()
/external/XNNPACK/src/qs8-gavgpool/gen/
D7x-minmax-neon-c24-acc2.c202 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN), voutput… in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() local
205 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2()
216 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() local
219 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2()
D7x-minmax-neon-c32-acc2.c241 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN), voutput… in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() local
245 int8x16_t voutGHIJKLMNOPQRSTUV = vqmovn_high_s16(vqmovn_s16(vaccGHIJKLMN), vaccOPQRSTUV); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2()
258 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() local
262 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2()
D7p7x-minmax-neon-c24-acc2.c421 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN), voutput… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
424 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
435 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local
438 int8x8_t voutGHIJKLMN = vqmovn_s16(vaccGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
D7p7x-minmax-neon-c32-acc2.c498 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN), voutput… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
502 int8x16_t voutGHIJKLMNOPQRSTUV = vqmovn_high_s16(vqmovn_s16(vaccGHIJKLMN), vaccOPQRSTUV); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
515 …const int16x8_t vaccGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN))… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local
519 … int8x16_t voutGHIJKLMNOPQRSTUV = vcombine_s8(vqmovn_s16(vaccGHIJKLMN), vqmovn_s16(vaccOPQRSTUV)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()