Home
last modified time | relevance | path

Searched refs:vk16x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c269 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
270 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
403 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
404 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c269 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
270 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
401 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
402 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c269 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
270 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
401 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
402 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c319 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
321 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
510 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
511 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
642 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
643 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c319 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
321 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
513 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
514 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
647 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
648 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c319 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
321 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
510 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
511 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
642 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
643 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c269 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
270 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
403 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
404 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c319 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
321 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
513 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
514 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
647 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
648 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c235 const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
236 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
345 const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
346 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c235 const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
236 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
347 const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
348 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c285 const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
287 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
427 const float16x8_t vk16x01234567 = vld1q_f16(w + 264); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
428 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
539 const float16x8_t vk16x01234567 = vld1q_f16(w + 272); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
540 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c285 const float16x8_t vk16x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
287 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
424 const float16x8_t vk16x01234567 = vld1q_f16(w + 264); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
425 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
534 const float16x8_t vk16x01234567 = vld1q_f16(w + 272); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
535 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()