Home
last modified time | relevance | path

Searched refs:vi18x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c278 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
282 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
410 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
412 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c278 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
282 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
408 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
410 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c278 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
282 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
408 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
410 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c333 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
339 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
519 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
523 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
649 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
651 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c333 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
339 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
522 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
526 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
654 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
656 vacc01234567p0 = _mm256_fmadd_ps(vi18x01234567, vk18x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c333 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
339 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
519 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
523 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
649 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
651 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c278 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
282 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
410 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
412 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c333 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
339 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
522 const __m256 vi18x01234567 = _mm256_loadu_ps(i18); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
526 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
654 const __m256 vi18x01234567 = _mm256_maskload_ps(i18, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
656 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi18x01234567, vk18x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c242 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
244 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
352 const float16x8_t vi18x01234567 = vld1q_f16(i18); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
354 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c242 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
244 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
354 const float16x8_t vi18x01234567 = vld1q_f16(i18); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
356 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c297 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
301 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
434 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
436 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
546 const float16x8_t vi18x01234567 = vld1q_f16(i18); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
548 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c297 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
301 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
431 const float16x8_t vi18x01234567 = vld1q_f16(i18); i18 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
433 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
541 const float16x8_t vi18x01234567 = vld1q_f16(i18); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
543 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi18x01234567, vk18x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()