Home
last modified time | relevance | path

Searched defs:vk3x0123 (Results 1 – 25 of 32) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x4-psimd.c81 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
122 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
150 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
Dup8x4-sse.c81 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__sse() local
122 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__sse() local
150 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__sse() local
Dup8x4-psimd-acc2.c81 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
125 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
155 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
Dup8x4-sse-acc2.c81 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
125 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
155 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
Dup4x4-psimd.c70 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x4__psimd() local
98 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x4__psimd() local
Dup4x4-sse.c70 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x4__sse() local
98 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x4__sse() local
Dup4x4-psimd-acc2.c70 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local
100 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local
Dup4x4-sse-acc2.c70 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local
100 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local
Dup8x9-psimd.c91 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
177 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
235 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
Dup8x9-neon.c83 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local
149 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x9__neon() local
195 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__neon() local
Dup8x9-sse.c91 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__sse() local
177 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__sse() local
235 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__sse() local
Dup8x9-neonfma.c83 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
149 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
195 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
Dup4x9-neon.c72 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local
118 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local
Dup4x9-neonfma.c72 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local
118 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local
Dup4x9-sse.c80 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x9__sse() local
138 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x9__sse() local
Dup4x9-psimd.c80 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x9__psimd() local
138 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x9__psimd() local
Dup8x9-neon-acc2.c83 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
152 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
200 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
Dup8x9-neonfma-acc2.c83 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
152 const float32x4_t vk3x0123 = vld1q_f32(w + 28); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
200 const float32x4_t vk3x0123 = vld1q_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
Dup8x9-sse-acc2.c91 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
180 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
240 const __m128 vk3x0123 = _mm_load_ps(w + 32); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
Dup8x9-psimd-acc2.c91 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
180 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
240 const psimd_f32 vk3x0123 = psimd_load_f32(w + 32); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
Dup4x9-psimd-acc2.c80 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
140 const psimd_f32 vk3x0123 = psimd_load_f32(w + 16); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
Dup4x9-sse-acc2.c80 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
140 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
Dup4x9-neonfma-acc2.c72 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
120 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
Dup4x9-neon-acc2.c72 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
120 const float32x4_t vk3x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
Dup4x25-sse.c112 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x25__sse() local
266 const __m128 vk3x0123 = _mm_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up4x25__sse() local

12