| /external/XNNPACK/src/f32-prelu/gen/ |
| D | sse41-2x8.c | 54 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse41_2x8() local
61 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse41_2x8()
66 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x8()
92 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse41_2x8() local
97 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse41_2x8()
100 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x8()
118 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse41_2x8() local
123 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse41_2x8()
126 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x8()
|
| D | psimd-2x8.c | 54 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_prelu_ukernel__psimd_2x8() local
61 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x8()
66 vacc0x0123 = psimd_signblend_f32(vi0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__psimd_2x8()
92 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_prelu_ukernel__psimd_2x8() local
97 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x8()
100 vacc0x0123 = psimd_signblend_f32(vi0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__psimd_2x8()
118 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_prelu_ukernel__psimd_2x8() local
123 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x8()
126 vacc0x0123 = psimd_signblend_f32(vi0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__psimd_2x8()
|
| D | neon-2x8.c | 53 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_prelu_ukernel__neon_2x8() local
58 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x8()
59 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0)); in xnn_f32_prelu_ukernel__neon_2x8()
67 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__neon_2x8()
90 const float32x4_t vi0x0123 = vld1q_f32(i0); in xnn_f32_prelu_ukernel__neon_2x8() local
95 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x8()
96 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0)); in xnn_f32_prelu_ukernel__neon_2x8()
100 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__neon_2x8()
115 const float32x4_t vi0x0123 = vld1q_f32(i0); in xnn_f32_prelu_ukernel__neon_2x8() local
120 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x8()
[all …]
|
| D | sse2-2x8.c | 54 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse2_2x8() local
61 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse2_2x8()
62 …8 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123))); in xnn_f32_prelu_ukernel__sse2_2x8()
70 …vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123)); in xnn_f32_prelu_ukernel__sse2_2x8()
96 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse2_2x8() local
101 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse2_2x8()
102 …8 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123))); in xnn_f32_prelu_ukernel__sse2_2x8()
106 …vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123)); in xnn_f32_prelu_ukernel__sse2_2x8()
124 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse2_2x8() local
129 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse2_2x8()
[all …]
|
| D | sse41-2x4.c | 53 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse41_2x4() local
58 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse41_2x4()
61 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x4()
79 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse41_2x4() local
84 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse41_2x4()
87 __m128 vacc0x0123 = _mm_blendv_ps(vi0x0123, vprod0x0123, vi0x0123); in xnn_f32_prelu_ukernel__sse41_2x4()
|
| D | psimd-2x4.c | 53 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_prelu_ukernel__psimd_2x4() local
58 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x4()
61 vacc0x0123 = psimd_signblend_f32(vi0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__psimd_2x4()
79 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_prelu_ukernel__psimd_2x4() local
84 psimd_f32 vacc0x0123 = psimd_mul_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__psimd_2x4()
87 vacc0x0123 = psimd_signblend_f32(vi0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__psimd_2x4()
|
| D | neon-2x4.c | 52 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_prelu_ukernel__neon_2x4() local
55 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x4()
56 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0)); in xnn_f32_prelu_ukernel__neon_2x4()
60 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__neon_2x4()
75 const float32x4_t vi0x0123 = vld1q_f32(i0); in xnn_f32_prelu_ukernel__neon_2x4() local
80 float32x4_t vacc0x0123 = vmulq_f32(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__neon_2x4()
81 const uint32x4_t vm0x0123 = vcltq_s32(vreinterpretq_s32_f32(vi0x0123), vmovq_n_s32(0)); in xnn_f32_prelu_ukernel__neon_2x4()
85 vacc0x0123 = vbslq_f32(vm0x0123, vacc0x0123, vi0x0123); in xnn_f32_prelu_ukernel__neon_2x4()
|
| D | sse2-2x4.c | 53 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse2_2x4() local
58 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse2_2x4()
59 …8 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123))); in xnn_f32_prelu_ukernel__sse2_2x4()
63 …vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123)); in xnn_f32_prelu_ukernel__sse2_2x4()
81 const __m128 vi0x0123 = _mm_loadu_ps(i0); in xnn_f32_prelu_ukernel__sse2_2x4() local
86 const __m128 vprod0x0123 = _mm_mul_ps(vi0x0123, vw0123); in xnn_f32_prelu_ukernel__sse2_2x4()
87 …8 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vi0x0123))); in xnn_f32_prelu_ukernel__sse2_2x4()
91 …vacc0x0123 = _mm_or_ps(_mm_and_ps(vprod0x0123, vmask0x0123), _mm_andnot_ps(vmask0x0123, vi0x0123)); in xnn_f32_prelu_ukernel__sse2_2x4()
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x4-psimd.c | 50 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
56 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd()
101 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
105 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd()
137 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up8x4__psimd() local
139 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x4__psimd()
|
| D | up8x4-psimd-acc2.c | 50 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
104 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
142 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
|
| D | up8x4-sse.c | 50 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
101 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
137 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
|
| D | up8x4-sse-acc2.c | 50 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
104 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
142 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
|
| D | up4x4-psimd.c | 49 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up4x4__psimd() local
53 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd()
85 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up4x4__psimd() local
87 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up4x4__psimd()
|
| D | up4x4-psimd-acc2.c | 49 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
87 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
|
| D | up4x4-sse.c | 49 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
85 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
|
| D | up4x4-sse-acc2.c | 49 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
87 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
|
| D | up8x9-psimd.c | 60 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
66 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd()
156 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
160 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd()
222 const psimd_f32 vi0x0123 = psimd_load_f32(i0); in xnn_f32_dwconv_ukernel_up8x9__psimd() local
224 vacc0123p0 = psimd_qfma_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__psimd()
|
| D | up8x9-neon-acc2.c | 60 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 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()
139 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 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()
187 const float32x4_t vi0x0123 = vld1q_f32(i0); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
189 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
|
| D | up8x9-neonfma-acc2.c | 60 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 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()
139 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 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()
187 const float32x4_t vi0x0123 = vld1q_f32(i0); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
189 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
|
| D | up8x9-sse-acc2.c | 60 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
159 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
227 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
|
| D | up8x9-neon.c | 60 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local
64 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon()
136 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_ukernel_up8x9__neon() local
138 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon()
182 const float32x4_t vi0x0123 = vld1q_f32(i0); in xnn_f32_dwconv_ukernel_up8x9__neon() local
184 vacc0123p0 = vmlaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neon()
|
| D | up8x9-sse.c | 60 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
156 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
222 const __m128 vi0x0123 = _mm_loadu_ps(i0); 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()
|
| D | up8x9-psimd-acc2.c | 60 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
159 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
227 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
|
| D | up8x9-neonfma.c | 60 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
64 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
136 const float32x4_t vi0x0123 = vld1q_f32(i0); i0 += 4; in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
138 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
182 const float32x4_t vi0x0123 = vld1q_f32(i0); in xnn_f32_dwconv_ukernel_up8x9__neonfma() local
184 vacc0123p0 = vfmaq_f32(vacc0123p0, vi0x0123, vk0x0123); in xnn_f32_dwconv_ukernel_up8x9__neonfma()
|
| D | up4x9-psimd-acc2.c | 59 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
127 const psimd_f32 vi0x0123 = psimd_load_f32(i0); 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()
|