| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x4-psimd.c | 73 const psimd_f32 vk2x4567 = psimd_load_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
75 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x4__psimd()
|
| D | up8x4-psimd-acc2.c | 73 const psimd_f32 vk2x4567 = psimd_load_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
75 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
|
| D | up8x4-sse.c | 73 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_ukernel_up8x4__sse() local
75 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x4__sse()
|
| D | up8x4-sse-acc2.c | 73 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
75 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
|
| D | up8x9-psimd.c | 83 const psimd_f32 vk2x4567 = psimd_load_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
85 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd()
|
| D | up8x9-neon-acc2.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
79 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
|
| D | up8x9-neonfma-acc2.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
79 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
|
| D | up8x9-sse-acc2.c | 83 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
85 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
|
| D | up8x9-neon.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local
79 vacc4567p0 = vmlaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x9__neon()
|
| D | up8x9-sse.c | 83 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_ukernel_up8x9__sse() local
85 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x9__sse()
|
| D | up8x9-psimd-acc2.c | 83 const psimd_f32 vk2x4567 = psimd_load_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
85 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
|
| D | up8x9-neonfma.c | 77 const float32x4_t vk2x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
79 vacc4567p0 = vfmaq_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
|
| D | up8x25-psimd.c | 115 const psimd_f32 vk2x4567 = psimd_load_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x25__psimd() local
117 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd()
|
| D | up8x25-psimd-acc2.c | 115 const psimd_f32 vk2x4567 = psimd_load_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local
117 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi2x4567, vk2x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
|
| D | up8x25-sse.c | 115 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_ukernel_up8x25__sse() local
117 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse()
|
| D | up8x25-sse-acc2.c | 115 const __m128 vk2x4567 = _mm_load_ps(w + 28); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local
117 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
|