/external/XNNPACK/src/f32-vbinary/gen/ |
D | vrsubc-minmax-avx512f-x32.c | 42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vb, vaGHIJKLMNOPQRSTUV); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32() local 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32() 49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32() 52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrsubc_minmax_ukernel__avx512f_x32()
|
D | vsubc-minmax-avx512f-x32.c | 42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32() local 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32() 49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32() 52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsubc_minmax_ukernel__avx512f_x32()
|
D | vaddc-minmax-avx512f-x32.c | 42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_add_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() local 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() 49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32() 52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vaddc_minmax_ukernel__avx512f_x32()
|
D | vmulc-minmax-avx512f-x32.c | 42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32() local 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32() 49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32() 52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vmulc_minmax_ukernel__avx512f_x32()
|
D | vrdivc-minmax-avx512f-x32.c | 42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_div_ps(vb, vaGHIJKLMNOPQRSTUV); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32() local 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32() 49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32() 52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrdivc_minmax_ukernel__avx512f_x32()
|
D | vdivc-minmax-avx512f-x32.c | 42 __m512 vyGHIJKLMNOPQRSTUV = _mm512_div_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32() local 46 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32() 49 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32() 52 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vdivc_minmax_ukernel__avx512f_x32()
|
D | vadd-minmax-avx512f-x32.c | 45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_add_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vadd_minmax_ukernel__avx512f_x32() local 49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vadd_minmax_ukernel__avx512f_x32() 52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vadd_minmax_ukernel__avx512f_x32() 55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
|
D | vmul-minmax-avx512f-x32.c | 45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vmul_minmax_ukernel__avx512f_x32() local 49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vmul_minmax_ukernel__avx512f_x32() 52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vmul_minmax_ukernel__avx512f_x32() 55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
|
D | vsub-minmax-avx512f-x32.c | 45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vsub_minmax_ukernel__avx512f_x32() local 49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vsub_minmax_ukernel__avx512f_x32() 52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vsub_minmax_ukernel__avx512f_x32() 55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
|
D | vdiv-minmax-avx512f-x32.c | 45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_div_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32() local 49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32() 52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32() 55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
|
D | vsqrdiffc-avx512f-x32.c | 40 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32() local 43 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32() 47 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiffc_ukernel__avx512f_x32()
|
D | vsqrdiff-avx512f-x32.c | 43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiff_ukernel__avx512f_x32() local 46 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiff_ukernel__avx512f_x32() 50 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
|
D | vminc-avx512f-x32.c | 40 __m512 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vminc_ukernel__avx512f_x32() local 45 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vminc_ukernel__avx512f_x32()
|
D | vmaxc-avx512f-x32.c | 40 __m512 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vaGHIJKLMNOPQRSTUV, vb); in xnn_f32_vmaxc_ukernel__avx512f_x32() local 45 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vmaxc_ukernel__avx512f_x32()
|
D | vmin-avx512f-x32.c | 43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vmin_ukernel__avx512f_x32() local 48 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vmin_ukernel__avx512f_x32()
|
D | vmax-avx512f-x32.c | 43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV); in xnn_f32_vmax_ukernel__avx512f_x32() local 48 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vmax_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-vrnd/gen/ |
D | vrndz-avx512f-x32.c | 35 const __m512 vyGHIJKLMNOPQRSTUV = _mm512_roundscale_ps(vxGHIJKLMNOPQRSTUV, _MM_FROUND_TO_ZERO); in xnn_f32_vrndz_ukernel__avx512f_x32() local 38 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrndz_ukernel__avx512f_x32()
|
D | vrndne-avx512f-x32.c | 35 …const __m512 vyGHIJKLMNOPQRSTUV = _mm512_roundscale_ps(vxGHIJKLMNOPQRSTUV, _MM_FROUND_TO_NEAREST_I… in xnn_f32_vrndne_ukernel__avx512f_x32() local 38 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrndne_ukernel__avx512f_x32()
|
D | vrndd-avx512f-x32.c | 35 … const __m512 vyGHIJKLMNOPQRSTUV = _mm512_roundscale_ps(vxGHIJKLMNOPQRSTUV, _MM_FROUND_TO_NEG_INF); in xnn_f32_vrndd_ukernel__avx512f_x32() local 38 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrndd_ukernel__avx512f_x32()
|
D | vrndu-avx512f-x32.c | 35 … const __m512 vyGHIJKLMNOPQRSTUV = _mm512_roundscale_ps(vxGHIJKLMNOPQRSTUV, _MM_FROUND_TO_POS_INF); in xnn_f32_vrndu_ukernel__avx512f_x32() local 38 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vrndu_ukernel__avx512f_x32()
|
/external/XNNPACK/src/f32-vunary/gen/ |
D | vsqr-avx512f-x32.c | 36 const __m512 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vxGHIJKLMNOPQRSTUV, vxGHIJKLMNOPQRSTUV); in xnn_f32_vsqr_ukernel__avx512f_x32() local 39 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vsqr_ukernel__avx512f_x32()
|
D | vabs-avx512f-x32.c | 37 const __m512i vyGHIJKLMNOPQRSTUV = _mm512_and_epi32(vxGHIJKLMNOPQRSTUV, vnonsign_mask); in xnn_f32_vabs_ukernel__avx512f_x32() local 40 _mm512_storeu_si512(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vabs_ukernel__avx512f_x32()
|
D | vneg-avx512f-x32.c | 37 const __m512i vyGHIJKLMNOPQRSTUV = _mm512_xor_epi32(vxGHIJKLMNOPQRSTUV, vsign_mask); in xnn_f32_vneg_ukernel__avx512f_x32() local 40 _mm512_storeu_si512(y + 16, vyGHIJKLMNOPQRSTUV); in xnn_f32_vneg_ukernel__avx512f_x32()
|