/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x9-psimd.c | 127 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__psimd() local 201 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__psimd() local 251 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
|
D | up8x9-neon.c | 111 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local 165 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_ukernel_up8x9__neon() local 211 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__neon() local
|
D | up8x9-sse.c | 127 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x9__sse() local 201 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x9__sse() local 251 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x9__sse() local
|
D | up8x9-neonfma.c | 111 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local 165 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local 211 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
|
D | up4x9-neon.c | 88 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local 134 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local
|
D | up4x9-neonfma.c | 88 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local 134 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local
|
D | up4x9-sse.c | 104 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x9__sse() local 154 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x9__sse() local
|
D | up4x9-psimd.c | 104 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x9__psimd() local 154 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x9__psimd() local
|
D | up8x9-neon-acc2.c | 111 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local 168 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local 216 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
|
D | up8x9-neonfma-acc2.c | 111 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local 168 const float32x4_t vk7x0123 = vld1q_f32(w + 60); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local 216 const float32x4_t vk7x0123 = vld1q_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-sse-acc2.c | 127 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local 204 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local 256 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
|
D | up8x9-psimd-acc2.c | 127 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local 204 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local 256 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
|
D | up4x9-psimd-acc2.c | 104 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local 156 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
|
D | up4x9-sse-acc2.c | 104 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local 156 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
|
D | up4x9-neonfma-acc2.c | 88 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local 136 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
|
D | up4x9-neon-acc2.c | 88 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local 136 const float32x4_t vk7x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
|
D | up4x25-sse.c | 136 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x25__sse() local 282 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x25__sse() local
|
D | up4x25-psimd.c | 136 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x25__psimd() local 282 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x25__psimd() local
|
D | up4x25-psimd-acc2.c | 136 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local 284 const psimd_f32 vk7x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local
|
D | up4x25-sse-acc2.c | 136 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() local 284 const __m128 vk7x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() local
|
D | up8x25-sse.c | 159 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x25__sse() local 377 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x25__sse() local 523 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x25__sse() local
|
D | up8x25-psimd.c | 159 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 377 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 523 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x25__psimd() local
|
D | up8x25-sse-acc2.c | 159 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 380 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 528 const __m128 vk7x0123 = _mm_load_ps(w + 64); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local
|
D | up8x25-psimd-acc2.c | 159 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 380 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 528 const psimd_f32 vk7x0123 = psimd_load_f32(w + 64); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local
|