Home
last modified time | relevance | path

Searched refs:vacc4567p1 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-psimd-acc2.c108 psimd_f32 vacc4567p1 = psimd_mul_f32(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2() local
126 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi3x4567, vk3x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
144 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi5x4567, vk5x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
162 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi7x4567, vk7x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
180 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi9x4567, vk9x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
198 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi11x4567, vk11x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
216 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi13x4567, vk13x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
234 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi15x4567, vk15x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
252 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi17x4567, vk17x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
270 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi19x4567, vk19x4567); in xnn_f32_dwconv_ukernel_up8x25__psimd_acc2()
[all …]
Dup8x25-sse-acc2.c108 __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2() local
126 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi3x4567, vk3x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
144 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi5x4567, vk5x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
162 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi7x4567, vk7x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
180 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi9x4567, vk9x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
198 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi11x4567, vk11x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
216 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi13x4567, vk13x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
234 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi15x4567, vk15x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
252 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi17x4567, vk17x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
270 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi19x4567, vk19x4567)); in xnn_f32_dwconv_ukernel_up8x25__sse_acc2()
[all …]
Dup8x9-neon-acc2.c72 float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2() local
86 vacc4567p1 = vmlaq_f32(vacc4567p1, vi3x4567, vk3x4567); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
100 vacc4567p1 = vmlaq_f32(vacc4567p1, vi5x4567, vk5x4567); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
114 vacc4567p1 = vmlaq_f32(vacc4567p1, vi7x4567, vk7x4567); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
125 vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1); in xnn_f32_dwconv_ukernel_up8x9__neon_acc2()
Dup8x9-neonfma-acc2.c72 float32x4_t vacc4567p1 = vmulq_f32(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2() local
86 vacc4567p1 = vfmaq_f32(vacc4567p1, vi3x4567, vk3x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
100 vacc4567p1 = vfmaq_f32(vacc4567p1, vi5x4567, vk5x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
114 vacc4567p1 = vfmaq_f32(vacc4567p1, vi7x4567, vk7x4567); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
125 vacc4567p0 = vaddq_f32(vacc4567p0, vacc4567p1); in xnn_f32_dwconv_ukernel_up8x9__neonfma_acc2()
Dup8x9-sse-acc2.c76 __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2() local
94 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi3x4567, vk3x4567)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
112 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi5x4567, vk5x4567)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
130 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi7x4567, vk7x4567)); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
145 vacc4567p0 = _mm_add_ps(vacc4567p0, vacc4567p1); in xnn_f32_dwconv_ukernel_up8x9__sse_acc2()
Dup8x9-psimd-acc2.c76 psimd_f32 vacc4567p1 = psimd_mul_f32(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2() local
94 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi3x4567, vk3x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
112 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi5x4567, vk5x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
130 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi7x4567, vk7x4567); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
145 vacc4567p0 = psimd_add_f32(vacc4567p0, vacc4567p1); in xnn_f32_dwconv_ukernel_up8x9__psimd_acc2()
Dup8x4-psimd-acc2.c66 psimd_f32 vacc4567p1 = psimd_mul_f32(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2() local
84 vacc4567p1 = psimd_qfma_f32(vacc4567p1, vi3x4567, vk3x4567); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
90 vacc4567p0 = psimd_add_f32(vacc4567p0, vacc4567p1); in xnn_f32_dwconv_ukernel_up8x4__psimd_acc2()
Dup8x4-sse-acc2.c66 __m128 vacc4567p1 = _mm_mul_ps(vi1x4567, vk1x4567); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2() local
84 vacc4567p1 = _mm_add_ps(vacc4567p1, _mm_mul_ps(vi3x4567, vk3x4567)); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()
90 vacc4567p0 = _mm_add_ps(vacc4567p0, vacc4567p1); in xnn_f32_dwconv_ukernel_up8x4__sse_acc2()