Home
last modified time | relevance | path

Searched refs:vi24x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
318 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
434 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
436 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
318 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
432 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
434 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
318 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
432 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
434 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
393 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
555 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
559 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
673 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
675 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
393 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
558 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
562 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
678 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
680 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
393 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
555 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
559 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
673 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
675 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c314 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
318 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
434 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
436 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c387 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
393 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
558 const __m256 vi24x01234567 = _mm256_loadu_ps(i24); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
562 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
678 const __m256 vi24x01234567 = _mm256_maskload_ps(i24, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
680 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c266 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
268 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
376 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
378 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c266 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
268 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
378 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
380 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c339 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
343 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
458 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
460 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
570 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
572 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c339 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
343 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
455 const float16x8_t vi24x01234567 = vld1q_f16(i24); i24 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
457 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
565 const float16x8_t vi24x01234567 = vld1q_f16(i24); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
567 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi24x01234567, vk24x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()