Home
last modified time | relevance | path

Searched refs:vi23x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c308 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
312 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
430 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
432 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c308 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
312 vacc01234567p0 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
428 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
430 vacc01234567p0 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c308 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
312 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
428 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
430 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c378 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
384 vacc01234567p0 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
549 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
553 vacc01234567p0 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
669 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
671 vacc01234567p0 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c378 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
384 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
552 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
556 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
674 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
676 vacc01234567p1 = _mm256_fmadd_ps(vi23x01234567, vk23x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c378 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
384 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
549 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
553 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
669 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
671 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c308 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
312 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
430 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
432 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c378 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
384 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
552 const __m256 vi23x01234567 = _mm256_loadu_ps(i23); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
556 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
674 const __m256 vi23x01234567 = _mm256_maskload_ps(i23, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
676 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi23x01234567, vk23x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c262 const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
264 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
372 const float16x8_t vi23x01234567 = vld1q_f16(i23); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
374 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c262 const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
264 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
374 const float16x8_t vi23x01234567 = vld1q_f16(i23); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
376 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c332 const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
336 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
454 const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
456 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
566 const float16x8_t vi23x01234567 = vld1q_f16(i23); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
568 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c332 const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
336 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
451 const float16x8_t vi23x01234567 = vld1q_f16(i23); i23 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
453 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
561 const float16x8_t vi23x01234567 = vld1q_f16(i23); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
563 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi23x01234567, vk23x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()