| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up4x25-psimd-acc2.c | 101 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2() local
113 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
125 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
137 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
149 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
161 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
173 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
185 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
197 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
209 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_ukernel_up4x25__psimd_acc2()
[all …]
|
| D | up4x25-sse-acc2.c | 101 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2() local
113 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
125 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
137 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
149 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
161 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
173 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
185 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
197 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
209 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_ukernel_up4x25__sse_acc2()
[all …]
|
| D | up8x25-psimd-acc2.c | 107 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local
125 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
143 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
161 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
179 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
197 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
215 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
233 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
251 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
269 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
[all …]
|
| D | up8x25-sse-acc2.c | 107 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local
125 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
143 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
161 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
179 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
197 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
215 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
233 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
251 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
269 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
[all …]
|
| D | up8x9-neon-acc2.c | 71 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
85 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
99 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
113 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
124 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
145 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
153 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
161 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
169 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
176 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
[all …]
|
| D | up8x9-neonfma-acc2.c | 71 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
85 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
99 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
113 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
124 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
145 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
153 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
161 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
169 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
176 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
[all …]
|
| D | up8x9-sse-acc2.c | 75 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
93 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
111 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
129 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
144 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
169 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
181 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
193 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
205 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
216 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
[all …]
|
| D | up8x9-psimd-acc2.c | 75 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
93 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
111 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
129 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
144 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
169 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
181 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
193 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
205 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
216 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
[all …]
|
| D | up4x9-psimd-acc2.c | 69 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
81 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
93 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
105 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
116 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
133 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2() local
141 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
149 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
157 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
164 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__psimd_acc2()
|
| D | up4x9-sse-acc2.c | 69 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
81 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
93 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
105 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
116 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
133 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2() local
141 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
149 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
157 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
164 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__sse_acc2()
|
| D | up4x9-neonfma-acc2.c | 65 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
73 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
81 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
89 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
96 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
113 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2() local
121 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
129 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
137 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
144 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__neonfma_acc2()
|
| D | up4x9-neon-acc2.c | 65 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
73 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
81 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
89 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
96 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
113 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2() local
121 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
129 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
137 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
144 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x9__neon_acc2()
|
| D | up8x4-psimd-acc2.c | 65 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
83 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
89 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
114 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
126 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
131 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
148 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
156 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
159 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
|
| D | up8x4-sse-acc2.c | 65 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
83 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
89 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
114 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
126 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
131 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
148 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
156 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
159 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
|
| D | up4x4-psimd-acc2.c | 59 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local
71 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2()
76 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2()
93 psimd_f32 vacc0123p1 = psimd_mul_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2() local
101 vacc0123p1 = psimd_qfma_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2()
104 vacc0123p0 = psimd_add_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x4__psimd_acc2()
|
| D | up4x4-sse-acc2.c | 59 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local
71 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2()
76 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2()
93 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2() local
101 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2()
104 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_ukernel_up4x4__sse_acc2()
|