/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x4-psimd.c | 55 const psimd_f32 vk0x4567 = psimd_load_f32(w + 12); in xnn_f32_dwconv_ukernel_up8x4__psimd() local 57 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x4__psimd()
|
D | up8x4-psimd-acc2.c | 55 const psimd_f32 vk0x4567 = psimd_load_f32(w + 12); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local 57 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
|
D | up8x4-sse.c | 55 const __m128 vk0x4567 = _mm_load_ps(w + 12); in xnn_f32_dwconv_ukernel_up8x4__sse() local 57 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); in xnn_f32_dwconv_ukernel_up8x4__sse()
|
D | up8x4-sse-acc2.c | 55 const __m128 vk0x4567 = _mm_load_ps(w + 12); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local 57 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
|
D | up8x9-psimd.c | 65 const psimd_f32 vk0x4567 = psimd_load_f32(w + 12); in xnn_f32_dwconv_ukernel_up8x9__psimd() local 67 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd()
|
D | up8x9-neon-acc2.c | 63 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local 65 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
|
D | up8x9-neonfma-acc2.c | 63 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local 65 vacc4567p0 = vfmaq_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
|
D | up8x9-sse-acc2.c | 65 const __m128 vk0x4567 = _mm_load_ps(w + 12); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local 67 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
|
D | up8x9-neon.c | 63 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local 65 vacc4567p0 = vmlaq_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x9__neon()
|
D | up8x9-sse.c | 65 const __m128 vk0x4567 = _mm_load_ps(w + 12); in xnn_f32_dwconv_ukernel_up8x9__sse() local 67 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); in xnn_f32_dwconv_ukernel_up8x9__sse()
|
D | up8x9-psimd-acc2.c | 65 const psimd_f32 vk0x4567 = psimd_load_f32(w + 12); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local 67 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
|
D | up8x9-neonfma.c | 63 const float32x4_t vk0x4567 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local 65 vacc4567p0 = vfmaq_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
|
D | up8x25-psimd.c | 97 const psimd_f32 vk0x4567 = psimd_load_f32(w + 12); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 99 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd()
|
D | up8x25-psimd-acc2.c | 97 const psimd_f32 vk0x4567 = psimd_load_f32(w + 12); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 99 vacc4567p0 = psimd_qfma_f32(vacc4567p0, vi0x4567, vk0x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
|
D | up8x25-sse.c | 97 const __m128 vk0x4567 = _mm_load_ps(w + 12); in xnn_f32_dwconv_ukernel_up8x25__sse() local 99 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse()
|
D | up8x25-sse-acc2.c | 97 const __m128 vk0x4567 = _mm_load_ps(w + 12); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 99 vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
|