Home
last modified time | relevance | path

Searched refs:vi16x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c266 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
402 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c266 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
400 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c266 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
400 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c315 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
507 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
641 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c315 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
510 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
646 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c315 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
507 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
641 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c266 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
402 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c315 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
510 const __m256 vi16x01234567 = _mm256_loadu_ps(i16); 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()
646 const __m256 vi16x01234567 = _mm256_maskload_ps(i16, vmask); 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.c234 const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 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()
344 const float16x8_t vi16x01234567 = vld1q_f16(i16); 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.c234 const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 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()
346 const float16x8_t vi16x01234567 = vld1q_f16(i16); 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.c283 const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 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()
426 const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 8; 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()
538 const float16x8_t vi16x01234567 = vld1q_f16(i16); 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.c283 const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 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()
423 const float16x8_t vi16x01234567 = vld1q_f16(i16); i16 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
425 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
533 const float16x8_t vi16x01234567 = vld1q_f16(i16); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
535 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi16x01234567, vk16x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()