Searched refs:vk24x0123 (Results 1 – 8 of 8) sorted by relevance
/external/XNNPACK/src/f32-dwconv/gen/ |
D | up4x25-psimd-acc2.c | 238 const psimd_f32 vk24x0123 = psimd_load_f32(w + 100); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local 239 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() 352 const psimd_f32 vk24x0123 = psimd_load_f32(w + 100); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local 353 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
|
D | up4x25-psimd.c | 238 const psimd_f32 vk24x0123 = psimd_load_f32(w + 100); in xnn_f32_dwconv_ukernel_up4x25__psimd() local 239 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd() 350 const psimd_f32 vk24x0123 = psimd_load_f32(w + 100); in xnn_f32_dwconv_ukernel_up4x25__psimd() local 351 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd()
|
D | up4x25-sse.c | 238 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_ukernel_up4x25__sse() local 239 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse() 350 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_ukernel_up4x25__sse() local 351 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse()
|
D | up4x25-sse-acc2.c | 238 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() local 239 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() 352 const __m128 vk24x0123 = _mm_load_ps(w + 100); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() local 353 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
|
D | up8x25-psimd.c | 312 const psimd_f32 vk24x0123 = psimd_load_f32(w + 200); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 314 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd() 479 const psimd_f32 vk24x0123 = psimd_load_f32(w + 200); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 480 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd() 591 const psimd_f32 vk24x0123 = psimd_load_f32(w + 200); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 592 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd()
|
D | up8x25-psimd-acc2.c | 312 const psimd_f32 vk24x0123 = psimd_load_f32(w + 200); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 314 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() 482 const psimd_f32 vk24x0123 = psimd_load_f32(w + 200); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 483 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() 596 const psimd_f32 vk24x0123 = psimd_load_f32(w + 200); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 597 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi24x0123, vk24x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
|
D | up8x25-sse.c | 312 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__sse() local 314 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse() 479 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__sse() local 480 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse() 591 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__sse() local 592 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse()
|
D | up8x25-sse-acc2.c | 312 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 314 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() 482 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 483 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() 596 const __m128 vk24x0123 = _mm_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 597 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi24x0123, vk24x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
|