Home
last modified time | relevance | path

Searched refs:vk10x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c233 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
234 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
379 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
380 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c233 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
234 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
377 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
378 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c233 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
234 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
377 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
378 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c265 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
267 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
474 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
475 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
618 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
619 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c265 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
267 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
477 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
478 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
623 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
624 vacc01234567p0 = _mm256_fmadd_ps(vi10x01234567, vk10x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c265 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
267 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
474 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
475 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
618 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
619 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c233 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
234 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
379 const __m256 vk10x01234567 = _mm256_load_ps(w + 88); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
380 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c265 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
267 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
477 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
478 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
623 const __m256 vk10x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
624 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi10x01234567, vk10x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c211 const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
212 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
321 const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
322 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c211 const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
212 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
323 const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
324 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c243 const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
245 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
403 const float16x8_t vk10x01234567 = vld1q_f16(w + 168); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
404 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
515 const float16x8_t vk10x01234567 = vld1q_f16(w + 176); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
516 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c243 const float16x8_t vk10x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
245 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
400 const float16x8_t vk10x01234567 = vld1q_f16(w + 168); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
401 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
510 const float16x8_t vk10x01234567 = vld1q_f16(w + 176); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
511 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi10x01234567, vk10x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()