Home
last modified time | relevance | path

Searched refs:vk0x0123 (Results 1 – 25 of 32) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x4-psimd.c54 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
56 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd()
104 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
105 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd()
138 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
139 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd()
Dup8x4-psimd-acc2.c54 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
56 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
107 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
108 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
143 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
144 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
Dup8x4-sse.c54 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse() local
56 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse()
104 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse() local
105 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse()
138 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse() local
139 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse()
Dup8x4-sse-acc2.c54 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
56 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
107 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
108 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
143 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
144 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
Dup4x4-psimd.c52 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd() local
53 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd()
86 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd() local
87 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd()
Dup4x4-psimd-acc2.c52 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local
53 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2()
88 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local
89 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2()
Dup4x4-sse.c52 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse() local
53 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x4__sse()
86 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse() local
87 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x4__sse()
Dup4x4-sse-acc2.c52 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local
53 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2()
88 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local
89 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2()
Dup8x9-psimd.c64 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
66 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd()
159 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
160 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd()
223 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
224 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd()
Dup8x9-neon-acc2.c62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
64 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
140 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
141 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
188 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
189 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
Dup8x9-neonfma-acc2.c62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
64 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
140 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
141 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
188 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
189 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
Dup8x9-sse-acc2.c64 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
66 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
162 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
163 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
228 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
229 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
Dup8x9-neon.c62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local
64 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon()
137 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neon() local
138 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon()
183 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neon() local
184 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon()
Dup8x9-sse.c64 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse() local
66 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse()
159 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse() local
160 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse()
223 const __m128 vk0x0123 = _mm_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__sse() local
224 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse()
Dup8x9-psimd-acc2.c64 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
66 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
162 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
163 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
228 const psimd_f32 vk0x0123 = psimd_load_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
229 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
Dup8x9-neonfma.c62 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
64 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
137 const float32x4_t vk0x0123 = vld1q_f32(w + 4); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
138 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
183 const float32x4_t vk0x0123 = vld1q_f32(w + 8); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
184 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
Dup4x9-psimd-acc2.c62 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
63 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
128 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
129 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
Dup4x9-sse-acc2.c62 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
63 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
128 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
129 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
Dup4x9-neon.c60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local
61 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neon()
106 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon() local
107 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neon()
Dup4x9-neonfma.c60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local
61 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma()
106 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma() local
107 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma()
Dup4x9-sse.c62 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse() local
63 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse()
126 const __m128 vk0x0123 = _mm_load_ps(w + 4); in xnn_f32_dwconv_ukernel_up4x9__sse() local
127 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse()
Dup4x9-psimd.c62 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd() local
63 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd()
126 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x9__psimd() local
127 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd()
Dup4x9-neonfma-acc2.c60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
61 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
108 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
109 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
Dup4x9-neon-acc2.c60 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
61 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
108 const float32x4_t vk0x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
109 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
Dup4x25-psimd-acc2.c94 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local
95 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
256 const psimd_f32 vk0x0123 = psimd_load_f32(w + 4); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local
257 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()

12