| /external/XNNPACK/src/f32-vbinary/gen/ |
| D | vrdivc-minmax-sse-x8.c | 41 __m128 vy0123 = _mm_div_ps(vb, va0123); in xnn_f32_vrdivc_minmax_ukernel__sse_x8() local
45 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrdivc_minmax_ukernel__sse_x8()
48 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vrdivc_minmax_ukernel__sse_x8()
51 _mm_storeu_ps(y, vy0123); in xnn_f32_vrdivc_minmax_ukernel__sse_x8()
59 __m128 vy0123 = _mm_div_ps(vb, va0123); in xnn_f32_vrdivc_minmax_ukernel__sse_x8() local
60 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrdivc_minmax_ukernel__sse_x8()
61 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vrdivc_minmax_ukernel__sse_x8()
62 _mm_storeu_ps(y, vy0123); in xnn_f32_vrdivc_minmax_ukernel__sse_x8()
68 __m128 vy0123 = _mm_div_ps(vb, va0123); in xnn_f32_vrdivc_minmax_ukernel__sse_x8() local
69 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrdivc_minmax_ukernel__sse_x8()
[all …]
|
| D | vmulc-minmax-sse-x8.c | 41 __m128 vy0123 = _mm_mul_ps(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__sse_x8() local
45 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__sse_x8()
48 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vmulc_minmax_ukernel__sse_x8()
51 _mm_storeu_ps(y, vy0123); in xnn_f32_vmulc_minmax_ukernel__sse_x8()
59 __m128 vy0123 = _mm_mul_ps(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__sse_x8() local
60 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__sse_x8()
61 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vmulc_minmax_ukernel__sse_x8()
62 _mm_storeu_ps(y, vy0123); in xnn_f32_vmulc_minmax_ukernel__sse_x8()
68 __m128 vy0123 = _mm_mul_ps(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__sse_x8() local
69 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__sse_x8()
[all …]
|
| D | vaddc-minmax-sse-x8.c | 41 __m128 vy0123 = _mm_add_ps(va0123, vb); in xnn_f32_vaddc_minmax_ukernel__sse_x8() local
45 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vaddc_minmax_ukernel__sse_x8()
48 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vaddc_minmax_ukernel__sse_x8()
51 _mm_storeu_ps(y, vy0123); in xnn_f32_vaddc_minmax_ukernel__sse_x8()
59 __m128 vy0123 = _mm_add_ps(va0123, vb); in xnn_f32_vaddc_minmax_ukernel__sse_x8() local
60 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vaddc_minmax_ukernel__sse_x8()
61 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vaddc_minmax_ukernel__sse_x8()
62 _mm_storeu_ps(y, vy0123); in xnn_f32_vaddc_minmax_ukernel__sse_x8()
68 __m128 vy0123 = _mm_add_ps(va0123, vb); in xnn_f32_vaddc_minmax_ukernel__sse_x8() local
69 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vaddc_minmax_ukernel__sse_x8()
[all …]
|
| D | vdivc-minmax-sse-x8.c | 41 __m128 vy0123 = _mm_div_ps(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__sse_x8() local
45 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__sse_x8()
48 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vdivc_minmax_ukernel__sse_x8()
51 _mm_storeu_ps(y, vy0123); in xnn_f32_vdivc_minmax_ukernel__sse_x8()
59 __m128 vy0123 = _mm_div_ps(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__sse_x8() local
60 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__sse_x8()
61 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vdivc_minmax_ukernel__sse_x8()
62 _mm_storeu_ps(y, vy0123); in xnn_f32_vdivc_minmax_ukernel__sse_x8()
68 __m128 vy0123 = _mm_div_ps(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__sse_x8() local
69 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__sse_x8()
[all …]
|
| D | vrsubc-minmax-sse-x8.c | 41 __m128 vy0123 = _mm_sub_ps(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x8() local
45 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__sse_x8()
48 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vrsubc_minmax_ukernel__sse_x8()
51 _mm_storeu_ps(y, vy0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x8()
59 __m128 vy0123 = _mm_sub_ps(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x8() local
60 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__sse_x8()
61 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vrsubc_minmax_ukernel__sse_x8()
62 _mm_storeu_ps(y, vy0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x8()
68 __m128 vy0123 = _mm_sub_ps(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x8() local
69 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__sse_x8()
[all …]
|
| D | vsubc-minmax-sse-x8.c | 41 __m128 vy0123 = _mm_sub_ps(va0123, vb); in xnn_f32_vsubc_minmax_ukernel__sse_x8() local
45 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vsubc_minmax_ukernel__sse_x8()
48 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vsubc_minmax_ukernel__sse_x8()
51 _mm_storeu_ps(y, vy0123); in xnn_f32_vsubc_minmax_ukernel__sse_x8()
59 __m128 vy0123 = _mm_sub_ps(va0123, vb); in xnn_f32_vsubc_minmax_ukernel__sse_x8() local
60 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vsubc_minmax_ukernel__sse_x8()
61 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vsubc_minmax_ukernel__sse_x8()
62 _mm_storeu_ps(y, vy0123); in xnn_f32_vsubc_minmax_ukernel__sse_x8()
68 __m128 vy0123 = _mm_sub_ps(va0123, vb); in xnn_f32_vsubc_minmax_ukernel__sse_x8() local
69 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vsubc_minmax_ukernel__sse_x8()
[all …]
|
| D | vmul-minmax-sse-x8.c | 44 __m128 vy0123 = _mm_mul_ps(va0123, vb0123); in xnn_f32_vmul_minmax_ukernel__sse_x8() local
48 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmul_minmax_ukernel__sse_x8()
51 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vmul_minmax_ukernel__sse_x8()
54 _mm_storeu_ps(y, vy0123); in xnn_f32_vmul_minmax_ukernel__sse_x8()
65 __m128 vy0123 = _mm_mul_ps(va0123, vb0123); in xnn_f32_vmul_minmax_ukernel__sse_x8() local
66 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmul_minmax_ukernel__sse_x8()
67 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vmul_minmax_ukernel__sse_x8()
68 _mm_storeu_ps(y, vy0123); in xnn_f32_vmul_minmax_ukernel__sse_x8()
75 __m128 vy0123 = _mm_mul_ps(va0123, vb0123); in xnn_f32_vmul_minmax_ukernel__sse_x8() local
76 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmul_minmax_ukernel__sse_x8()
[all …]
|
| D | vsub-minmax-sse-x8.c | 44 __m128 vy0123 = _mm_sub_ps(va0123, vb0123); in xnn_f32_vsub_minmax_ukernel__sse_x8() local
48 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vsub_minmax_ukernel__sse_x8()
51 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vsub_minmax_ukernel__sse_x8()
54 _mm_storeu_ps(y, vy0123); in xnn_f32_vsub_minmax_ukernel__sse_x8()
65 __m128 vy0123 = _mm_sub_ps(va0123, vb0123); in xnn_f32_vsub_minmax_ukernel__sse_x8() local
66 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vsub_minmax_ukernel__sse_x8()
67 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vsub_minmax_ukernel__sse_x8()
68 _mm_storeu_ps(y, vy0123); in xnn_f32_vsub_minmax_ukernel__sse_x8()
75 __m128 vy0123 = _mm_sub_ps(va0123, vb0123); in xnn_f32_vsub_minmax_ukernel__sse_x8() local
76 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vsub_minmax_ukernel__sse_x8()
[all …]
|
| D | vdiv-minmax-sse-x8.c | 44 __m128 vy0123 = _mm_div_ps(va0123, vb0123); in xnn_f32_vdiv_minmax_ukernel__sse_x8() local
48 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdiv_minmax_ukernel__sse_x8()
51 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vdiv_minmax_ukernel__sse_x8()
54 _mm_storeu_ps(y, vy0123); in xnn_f32_vdiv_minmax_ukernel__sse_x8()
65 __m128 vy0123 = _mm_div_ps(va0123, vb0123); in xnn_f32_vdiv_minmax_ukernel__sse_x8() local
66 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdiv_minmax_ukernel__sse_x8()
67 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vdiv_minmax_ukernel__sse_x8()
68 _mm_storeu_ps(y, vy0123); in xnn_f32_vdiv_minmax_ukernel__sse_x8()
75 __m128 vy0123 = _mm_div_ps(va0123, vb0123); in xnn_f32_vdiv_minmax_ukernel__sse_x8() local
76 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdiv_minmax_ukernel__sse_x8()
[all …]
|
| D | vadd-minmax-sse-x8.c | 44 __m128 vy0123 = _mm_add_ps(va0123, vb0123); in xnn_f32_vadd_minmax_ukernel__sse_x8() local
48 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vadd_minmax_ukernel__sse_x8()
51 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vadd_minmax_ukernel__sse_x8()
54 _mm_storeu_ps(y, vy0123); in xnn_f32_vadd_minmax_ukernel__sse_x8()
65 __m128 vy0123 = _mm_add_ps(va0123, vb0123); in xnn_f32_vadd_minmax_ukernel__sse_x8() local
66 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vadd_minmax_ukernel__sse_x8()
67 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vadd_minmax_ukernel__sse_x8()
68 _mm_storeu_ps(y, vy0123); in xnn_f32_vadd_minmax_ukernel__sse_x8()
75 __m128 vy0123 = _mm_add_ps(va0123, vb0123); in xnn_f32_vadd_minmax_ukernel__sse_x8() local
76 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vadd_minmax_ukernel__sse_x8()
[all …]
|
| D | vsqrdiffc-sse-x8.c | 39 __m128 vy0123 = _mm_sub_ps(va0123, vb); in xnn_f32_vsqrdiffc_ukernel__sse_x8() local
42 vy0123 = _mm_mul_ps(vy0123, vy0123); in xnn_f32_vsqrdiffc_ukernel__sse_x8()
46 _mm_storeu_ps(y, vy0123); in xnn_f32_vsqrdiffc_ukernel__sse_x8()
54 __m128 vy0123 = _mm_sub_ps(va0123, vb); in xnn_f32_vsqrdiffc_ukernel__sse_x8() local
55 vy0123 = _mm_mul_ps(vy0123, vy0123); in xnn_f32_vsqrdiffc_ukernel__sse_x8()
56 _mm_storeu_ps(y, vy0123); in xnn_f32_vsqrdiffc_ukernel__sse_x8()
62 __m128 vy0123 = _mm_sub_ps(va0123, vb); in xnn_f32_vsqrdiffc_ukernel__sse_x8() local
63 vy0123 = _mm_mul_ps(vy0123, vy0123); in xnn_f32_vsqrdiffc_ukernel__sse_x8()
65 _mm_storel_pi((__m64*) y, vy0123); in xnn_f32_vsqrdiffc_ukernel__sse_x8()
66 vy0123 = _mm_movehl_ps(vy0123, vy0123); in xnn_f32_vsqrdiffc_ukernel__sse_x8()
[all …]
|
| D | vsqrdiff-sse-x8.c | 42 __m128 vy0123 = _mm_sub_ps(va0123, vb0123); in xnn_f32_vsqrdiff_ukernel__sse_x8() local
45 vy0123 = _mm_mul_ps(vy0123, vy0123); in xnn_f32_vsqrdiff_ukernel__sse_x8()
49 _mm_storeu_ps(y, vy0123); in xnn_f32_vsqrdiff_ukernel__sse_x8()
60 __m128 vy0123 = _mm_sub_ps(va0123, vb0123); in xnn_f32_vsqrdiff_ukernel__sse_x8() local
61 vy0123 = _mm_mul_ps(vy0123, vy0123); in xnn_f32_vsqrdiff_ukernel__sse_x8()
62 _mm_storeu_ps(y, vy0123); in xnn_f32_vsqrdiff_ukernel__sse_x8()
69 __m128 vy0123 = _mm_sub_ps(va0123, vb0123); in xnn_f32_vsqrdiff_ukernel__sse_x8() local
70 vy0123 = _mm_mul_ps(vy0123, vy0123); in xnn_f32_vsqrdiff_ukernel__sse_x8()
72 _mm_storel_pi((__m64*) y, vy0123); in xnn_f32_vsqrdiff_ukernel__sse_x8()
73 vy0123 = _mm_movehl_ps(vy0123, vy0123); in xnn_f32_vsqrdiff_ukernel__sse_x8()
[all …]
|
| D | vrdivc-minmax-neon-x8.c | 39 float32x4_t vy0123 = vdivq_f32(vb, va0123); in xnn_f32_vrdivc_minmax_ukernel__neon_x8() local
43 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vrdivc_minmax_ukernel__neon_x8()
46 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vrdivc_minmax_ukernel__neon_x8()
49 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vrdivc_minmax_ukernel__neon_x8()
55 float32x4_t vy0123 = vdivq_f32(vb, va0123); in xnn_f32_vrdivc_minmax_ukernel__neon_x8() local
56 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vrdivc_minmax_ukernel__neon_x8()
57 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vrdivc_minmax_ukernel__neon_x8()
58 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vrdivc_minmax_ukernel__neon_x8()
63 float32x4_t vy0123 = vdivq_f32(vb, va0123); in xnn_f32_vrdivc_minmax_ukernel__neon_x8() local
64 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vrdivc_minmax_ukernel__neon_x8()
[all …]
|
| D | vaddc-minmax-neon-x8.c | 39 float32x4_t vy0123 = vaddq_f32(va0123, vb); in xnn_f32_vaddc_minmax_ukernel__neon_x8() local
43 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vaddc_minmax_ukernel__neon_x8()
46 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vaddc_minmax_ukernel__neon_x8()
49 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vaddc_minmax_ukernel__neon_x8()
55 float32x4_t vy0123 = vaddq_f32(va0123, vb); in xnn_f32_vaddc_minmax_ukernel__neon_x8() local
56 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vaddc_minmax_ukernel__neon_x8()
57 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vaddc_minmax_ukernel__neon_x8()
58 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vaddc_minmax_ukernel__neon_x8()
63 float32x4_t vy0123 = vaddq_f32(va0123, vb); in xnn_f32_vaddc_minmax_ukernel__neon_x8() local
64 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vaddc_minmax_ukernel__neon_x8()
[all …]
|
| D | vdivc-minmax-neon-x8.c | 39 float32x4_t vy0123 = vdivq_f32(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__neon_x8() local
43 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__neon_x8()
46 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vdivc_minmax_ukernel__neon_x8()
49 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vdivc_minmax_ukernel__neon_x8()
55 float32x4_t vy0123 = vdivq_f32(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__neon_x8() local
56 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__neon_x8()
57 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vdivc_minmax_ukernel__neon_x8()
58 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vdivc_minmax_ukernel__neon_x8()
63 float32x4_t vy0123 = vdivq_f32(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__neon_x8() local
64 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__neon_x8()
[all …]
|
| D | vsubc-minmax-neon-x8.c | 39 float32x4_t vy0123 = vsubq_f32(va0123, vb); in xnn_f32_vsubc_minmax_ukernel__neon_x8() local
43 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vsubc_minmax_ukernel__neon_x8()
46 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vsubc_minmax_ukernel__neon_x8()
49 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vsubc_minmax_ukernel__neon_x8()
55 float32x4_t vy0123 = vsubq_f32(va0123, vb); in xnn_f32_vsubc_minmax_ukernel__neon_x8() local
56 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vsubc_minmax_ukernel__neon_x8()
57 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vsubc_minmax_ukernel__neon_x8()
58 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vsubc_minmax_ukernel__neon_x8()
63 float32x4_t vy0123 = vsubq_f32(va0123, vb); in xnn_f32_vsubc_minmax_ukernel__neon_x8() local
64 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vsubc_minmax_ukernel__neon_x8()
[all …]
|
| D | vmulc-minmax-neon-x8.c | 39 float32x4_t vy0123 = vmulq_f32(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__neon_x8() local
43 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__neon_x8()
46 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vmulc_minmax_ukernel__neon_x8()
49 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vmulc_minmax_ukernel__neon_x8()
55 float32x4_t vy0123 = vmulq_f32(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__neon_x8() local
56 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__neon_x8()
57 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vmulc_minmax_ukernel__neon_x8()
58 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vmulc_minmax_ukernel__neon_x8()
63 float32x4_t vy0123 = vmulq_f32(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__neon_x8() local
64 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__neon_x8()
[all …]
|
| D | vrsubc-minmax-neon-x8.c | 39 float32x4_t vy0123 = vsubq_f32(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__neon_x8() local
43 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
46 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
49 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
55 float32x4_t vy0123 = vsubq_f32(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__neon_x8() local
56 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
57 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
58 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
63 float32x4_t vy0123 = vsubq_f32(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__neon_x8() local
64 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__neon_x8()
[all …]
|
| D | vdiv-minmax-neon-x8.c | 40 float32x4_t vy0123 = vdivq_f32(va0123, vb0123); in xnn_f32_vdiv_minmax_ukernel__neon_x8() local
44 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vdiv_minmax_ukernel__neon_x8()
47 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vdiv_minmax_ukernel__neon_x8()
50 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vdiv_minmax_ukernel__neon_x8()
57 float32x4_t vy0123 = vdivq_f32(va0123, vb0123); in xnn_f32_vdiv_minmax_ukernel__neon_x8() local
58 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vdiv_minmax_ukernel__neon_x8()
59 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vdiv_minmax_ukernel__neon_x8()
60 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vdiv_minmax_ukernel__neon_x8()
66 float32x4_t vy0123 = vdivq_f32(va0123, vb0123); in xnn_f32_vdiv_minmax_ukernel__neon_x8() local
67 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vdiv_minmax_ukernel__neon_x8()
[all …]
|
| D | vsub-minmax-neon-x8.c | 40 float32x4_t vy0123 = vsubq_f32(va0123, vb0123); in xnn_f32_vsub_minmax_ukernel__neon_x8() local
44 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vsub_minmax_ukernel__neon_x8()
47 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vsub_minmax_ukernel__neon_x8()
50 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vsub_minmax_ukernel__neon_x8()
57 float32x4_t vy0123 = vsubq_f32(va0123, vb0123); in xnn_f32_vsub_minmax_ukernel__neon_x8() local
58 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vsub_minmax_ukernel__neon_x8()
59 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vsub_minmax_ukernel__neon_x8()
60 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vsub_minmax_ukernel__neon_x8()
66 float32x4_t vy0123 = vsubq_f32(va0123, vb0123); in xnn_f32_vsub_minmax_ukernel__neon_x8() local
67 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vsub_minmax_ukernel__neon_x8()
[all …]
|
| D | vadd-minmax-neon-x8.c | 40 float32x4_t vy0123 = vaddq_f32(va0123, vb0123); in xnn_f32_vadd_minmax_ukernel__neon_x8() local
44 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vadd_minmax_ukernel__neon_x8()
47 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vadd_minmax_ukernel__neon_x8()
50 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vadd_minmax_ukernel__neon_x8()
57 float32x4_t vy0123 = vaddq_f32(va0123, vb0123); in xnn_f32_vadd_minmax_ukernel__neon_x8() local
58 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vadd_minmax_ukernel__neon_x8()
59 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vadd_minmax_ukernel__neon_x8()
60 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vadd_minmax_ukernel__neon_x8()
66 float32x4_t vy0123 = vaddq_f32(va0123, vb0123); in xnn_f32_vadd_minmax_ukernel__neon_x8() local
67 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vadd_minmax_ukernel__neon_x8()
[all …]
|
| D | vmul-minmax-neon-x8.c | 40 float32x4_t vy0123 = vmulq_f32(va0123, vb0123); in xnn_f32_vmul_minmax_ukernel__neon_x8() local
44 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vmul_minmax_ukernel__neon_x8()
47 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vmul_minmax_ukernel__neon_x8()
50 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vmul_minmax_ukernel__neon_x8()
57 float32x4_t vy0123 = vmulq_f32(va0123, vb0123); in xnn_f32_vmul_minmax_ukernel__neon_x8() local
58 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vmul_minmax_ukernel__neon_x8()
59 vy0123 = vminq_f32(vy0123, vy_max); in xnn_f32_vmul_minmax_ukernel__neon_x8()
60 vst1q_f32(y, vy0123); y += 4; in xnn_f32_vmul_minmax_ukernel__neon_x8()
66 float32x4_t vy0123 = vmulq_f32(va0123, vb0123); in xnn_f32_vmul_minmax_ukernel__neon_x8() local
67 vy0123 = vmaxq_f32(vy0123, vy_min); in xnn_f32_vmul_minmax_ukernel__neon_x8()
[all …]
|
| D | vrsubc-minmax-sse-x4.c | 40 __m128 vy0123 = _mm_sub_ps(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x4() local
43 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
45 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
47 _mm_storeu_ps(y, vy0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
53 __m128 vy0123 = _mm_sub_ps(vb, va0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x4() local
54 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
55 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
57 _mm_storel_pi((__m64*) y, vy0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
58 vy0123 = _mm_movehl_ps(vy0123, vy0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
62 _mm_store_ss(y, vy0123); in xnn_f32_vrsubc_minmax_ukernel__sse_x4()
|
| D | vdivc-minmax-sse-x4.c | 40 __m128 vy0123 = _mm_div_ps(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__sse_x4() local
43 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
45 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
47 _mm_storeu_ps(y, vy0123); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
53 __m128 vy0123 = _mm_div_ps(va0123, vb); in xnn_f32_vdivc_minmax_ukernel__sse_x4() local
54 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
55 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
57 _mm_storel_pi((__m64*) y, vy0123); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
58 vy0123 = _mm_movehl_ps(vy0123, vy0123); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
62 _mm_store_ss(y, vy0123); in xnn_f32_vdivc_minmax_ukernel__sse_x4()
|
| D | vmulc-minmax-sse-x4.c | 40 __m128 vy0123 = _mm_mul_ps(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__sse_x4() local
43 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
45 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
47 _mm_storeu_ps(y, vy0123); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
53 __m128 vy0123 = _mm_mul_ps(va0123, vb); in xnn_f32_vmulc_minmax_ukernel__sse_x4() local
54 vy0123 = _mm_max_ps(vy0123, vy_min); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
55 vy0123 = _mm_min_ps(vy0123, vy_max); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
57 _mm_storel_pi((__m64*) y, vy0123); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
58 vy0123 = _mm_movehl_ps(vy0123, vy0123); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
62 _mm_store_ss(y, vy0123); in xnn_f32_vmulc_minmax_ukernel__sse_x4()
|