/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x9-psimd.c | 109 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__psimd() local 189 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__psimd() local 243 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
|
D | up8x9-neon.c | 97 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local 157 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_ukernel_up8x9__neon() local 203 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__neon() local
|
D | up8x9-sse.c | 109 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x9__sse() local 189 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x9__sse() local 243 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x9__sse() local
|
D | up8x9-neonfma.c | 97 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local 157 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local 203 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
|
D | up4x9-neon.c | 80 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local 126 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local
|
D | up4x9-neonfma.c | 80 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local 126 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local
|
D | up4x9-sse.c | 92 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x9__sse() local 146 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x9__sse() local
|
D | up4x9-psimd.c | 92 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x9__psimd() local 146 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x9__psimd() local
|
D | up8x9-neon-acc2.c | 97 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local 160 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local 208 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
|
D | up8x9-neonfma-acc2.c | 97 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local 160 const float32x4_t vk5x0123 = vld1q_f32(w + 44); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local 208 const float32x4_t vk5x0123 = vld1q_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-sse-acc2.c | 109 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local 192 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local 248 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
|
D | up8x9-psimd-acc2.c | 109 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local 192 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local 248 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
|
D | up4x9-psimd-acc2.c | 92 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local 148 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
|
D | up4x9-sse-acc2.c | 92 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local 148 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
|
D | up4x9-neonfma-acc2.c | 80 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local 128 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
|
D | up4x9-neon-acc2.c | 80 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local 128 const float32x4_t vk5x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
|
D | up4x25-sse.c | 124 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x25__sse() local 274 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x25__sse() local
|
D | up4x25-psimd.c | 124 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x25__psimd() local 274 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x25__psimd() local
|
D | up4x25-psimd-acc2.c | 124 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local 276 const psimd_f32 vk5x0123 = psimd_load_f32(w + 24); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local
|
D | up4x25-sse-acc2.c | 124 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() local 276 const __m128 vk5x0123 = _mm_load_ps(w + 24); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() local
|
D | up8x25-sse.c | 141 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x25__sse() local 365 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x25__sse() local 515 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x25__sse() local
|
D | up8x25-psimd.c | 141 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 365 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x25__psimd() local 515 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x25__psimd() local
|
D | up8x25-sse-acc2.c | 141 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 368 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local 520 const __m128 vk5x0123 = _mm_load_ps(w + 48); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local
|
D | up8x25-psimd-acc2.c | 141 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 368 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local 520 const psimd_f32 vk5x0123 = psimd_load_f32(w + 48); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local
|