/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x4-psimd.c | 54 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd() local 104 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd() local 138 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
|
D | up8x4-sse.c | 54 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse() local 104 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse() local 138 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse() local
|
D | up8x4-psimd-acc2.c | 54 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local 107 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local 143 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
|
D | up8x4-sse-acc2.c | 54 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local 107 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local 143 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
|
D | up4x4-psimd.c | 52 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd() local 86 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd() local
|
D | up4x4-sse.c | 52 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse() local 86 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse() local
|
D | up4x4-psimd-acc2.c | 52 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local 88 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local
|
D | up4x4-sse-acc2.c | 52 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local 88 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local
|
D | up8x9-psimd.c | 64 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd() local 159 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd() local 223 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
|
D | up8x9-neon.c | 62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local 137 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neon() local 183 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neon() local
|
D | up8x9-sse.c | 64 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse() local 159 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse() local 223 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse() local
|
D | up8x9-neonfma.c | 62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local 137 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local 183 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
|
D | up4x9-neon.c | 60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local 106 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local
|
D | up4x9-neonfma.c | 60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local 106 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local
|
D | up4x9-sse.c | 62 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse() local 126 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse() local
|
D | up4x9-psimd.c | 62 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd() local 126 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd() local
|
D | up8x9-neon-acc2.c | 62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local 140 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local 188 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
|
D | up8x9-neonfma-acc2.c | 62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local 140 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local 188 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
|
D | up8x9-sse-acc2.c | 64 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local 162 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local 228 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
|
D | up8x9-psimd-acc2.c | 64 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local 162 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local 228 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
|
D | up4x9-psimd-acc2.c | 62 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local 128 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
|
D | up4x9-sse-acc2.c | 62 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local 128 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
|
D | up4x9-neonfma-acc2.c | 60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local 108 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
|
D | up4x9-neon-acc2.c | 60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local 108 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
|
D | up4x25-sse.c | 94 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x25__sse() local 254 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x25__sse() local
|