| /external/XNNPACK/src/f32-vsqrt/gen/ |
| D | avx-sqrt-x8.c | 45 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vsqrt_ukernel__avx_sqrt_x8() local
47 _mm_storeu_ps(y, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x8()
48 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vsqrt_ukernel__avx_sqrt_x8()
52 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x8()
53 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x8()
57 _mm_store_ss(y, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x8()
|
| D | avx-sqrt-x16.c | 57 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vsqrt_ukernel__avx_sqrt_x16() local
59 _mm_storeu_ps(y, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x16()
60 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vsqrt_ukernel__avx_sqrt_x16()
64 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x16()
65 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x16()
69 _mm_store_ss(y, vy_lo); in xnn_f32_vsqrt_ukernel__avx_sqrt_x16()
|
| /external/XNNPACK/src/f32-vrnd/gen/ |
| D | vrndu-avx-x8.c | 48 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndu_ukernel__avx_x8() local
50 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndu_ukernel__avx_x8()
51 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndu_ukernel__avx_x8()
55 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndu_ukernel__avx_x8()
56 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndu_ukernel__avx_x8()
60 _mm_store_ss(y, vy_lo); in xnn_f32_vrndu_ukernel__avx_x8()
|
| D | vrndne-avx-x8.c | 48 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndne_ukernel__avx_x8() local
50 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndne_ukernel__avx_x8()
51 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndne_ukernel__avx_x8()
55 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndne_ukernel__avx_x8()
56 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndne_ukernel__avx_x8()
60 _mm_store_ss(y, vy_lo); in xnn_f32_vrndne_ukernel__avx_x8()
|
| D | vrndd-avx-x8.c | 48 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndd_ukernel__avx_x8() local
50 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndd_ukernel__avx_x8()
51 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndd_ukernel__avx_x8()
55 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndd_ukernel__avx_x8()
56 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndd_ukernel__avx_x8()
60 _mm_store_ss(y, vy_lo); in xnn_f32_vrndd_ukernel__avx_x8()
|
| D | vrndz-avx-x8.c | 48 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndz_ukernel__avx_x8() local
50 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndz_ukernel__avx_x8()
51 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndz_ukernel__avx_x8()
55 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndz_ukernel__avx_x8()
56 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndz_ukernel__avx_x8()
60 _mm_store_ss(y, vy_lo); in xnn_f32_vrndz_ukernel__avx_x8()
|
| D | vrndz-avx-x16.c | 60 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndz_ukernel__avx_x16() local
62 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndz_ukernel__avx_x16()
63 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndz_ukernel__avx_x16()
67 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndz_ukernel__avx_x16()
68 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndz_ukernel__avx_x16()
72 _mm_store_ss(y, vy_lo); in xnn_f32_vrndz_ukernel__avx_x16()
|
| D | vrndd-avx-x16.c | 60 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndd_ukernel__avx_x16() local
62 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndd_ukernel__avx_x16()
63 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndd_ukernel__avx_x16()
67 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndd_ukernel__avx_x16()
68 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndd_ukernel__avx_x16()
72 _mm_store_ss(y, vy_lo); in xnn_f32_vrndd_ukernel__avx_x16()
|
| D | vrndne-avx-x16.c | 60 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndne_ukernel__avx_x16() local
62 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndne_ukernel__avx_x16()
63 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndne_ukernel__avx_x16()
67 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndne_ukernel__avx_x16()
68 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndne_ukernel__avx_x16()
72 _mm_store_ss(y, vy_lo); in xnn_f32_vrndne_ukernel__avx_x16()
|
| D | vrndu-avx-x16.c | 60 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vrndu_ukernel__avx_x16() local
62 _mm_storeu_ps(y, vy_lo); in xnn_f32_vrndu_ukernel__avx_x16()
63 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vrndu_ukernel__avx_x16()
67 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vrndu_ukernel__avx_x16()
68 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vrndu_ukernel__avx_x16()
72 _mm_store_ss(y, vy_lo); in xnn_f32_vrndu_ukernel__avx_x16()
|
| /external/XNNPACK/src/f32-vunary/gen/ |
| D | vsqr-avx-x8.c | 49 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vsqr_ukernel__avx_x8() local
51 _mm_storeu_ps(y, vy_lo); in xnn_f32_vsqr_ukernel__avx_x8()
52 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vsqr_ukernel__avx_x8()
56 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vsqr_ukernel__avx_x8()
57 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vsqr_ukernel__avx_x8()
61 _mm_store_ss(y, vy_lo); in xnn_f32_vsqr_ukernel__avx_x8()
|
| D | vabs-avx-x8.c | 50 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vabs_ukernel__avx_x8() local
52 _mm_storeu_ps(y, vy_lo); in xnn_f32_vabs_ukernel__avx_x8()
53 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vabs_ukernel__avx_x8()
57 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vabs_ukernel__avx_x8()
58 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vabs_ukernel__avx_x8()
62 _mm_store_ss(y, vy_lo); in xnn_f32_vabs_ukernel__avx_x8()
|
| D | vneg-avx-x8.c | 50 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vneg_ukernel__avx_x8() local
52 _mm_storeu_ps(y, vy_lo); in xnn_f32_vneg_ukernel__avx_x8()
53 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vneg_ukernel__avx_x8()
57 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vneg_ukernel__avx_x8()
58 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vneg_ukernel__avx_x8()
62 _mm_store_ss(y, vy_lo); in xnn_f32_vneg_ukernel__avx_x8()
|
| D | vsqr-avx-x16.c | 59 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vsqr_ukernel__avx_x16() local
61 _mm_storeu_ps(y, vy_lo); in xnn_f32_vsqr_ukernel__avx_x16()
62 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vsqr_ukernel__avx_x16()
66 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vsqr_ukernel__avx_x16()
67 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vsqr_ukernel__avx_x16()
71 _mm_store_ss(y, vy_lo); in xnn_f32_vsqr_ukernel__avx_x16()
|
| D | vabs-avx-x16.c | 60 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vabs_ukernel__avx_x16() local
62 _mm_storeu_ps(y, vy_lo); in xnn_f32_vabs_ukernel__avx_x16()
63 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vabs_ukernel__avx_x16()
67 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vabs_ukernel__avx_x16()
68 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vabs_ukernel__avx_x16()
72 _mm_store_ss(y, vy_lo); in xnn_f32_vabs_ukernel__avx_x16()
|
| /external/XNNPACK/src/f32-vbinary/gen/ |
| D | vminc-avx-x8.c | 56 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vminc_ukernel__avx_x8() local
58 _mm_storeu_ps(y, vy_lo); in xnn_f32_vminc_ukernel__avx_x8()
59 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vminc_ukernel__avx_x8()
63 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vminc_ukernel__avx_x8()
64 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vminc_ukernel__avx_x8()
68 _mm_store_ss(y, vy_lo); in xnn_f32_vminc_ukernel__avx_x8()
|
| D | vmaxc-avx-x8.c | 56 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vmaxc_ukernel__avx_x8() local
58 _mm_storeu_ps(y, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x8()
59 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vmaxc_ukernel__avx_x8()
63 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x8()
64 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x8()
68 _mm_store_ss(y, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x8()
|
| D | vmin-avx-x8.c | 59 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vmin_ukernel__avx_x8() local
61 _mm_storeu_ps(y, vy_lo); in xnn_f32_vmin_ukernel__avx_x8()
62 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vmin_ukernel__avx_x8()
66 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vmin_ukernel__avx_x8()
67 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vmin_ukernel__avx_x8()
71 _mm_store_ss(y, vy_lo); in xnn_f32_vmin_ukernel__avx_x8()
|
| D | vmax-avx-x8.c | 59 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vmax_ukernel__avx_x8() local
61 _mm_storeu_ps(y, vy_lo); in xnn_f32_vmax_ukernel__avx_x8()
62 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vmax_ukernel__avx_x8()
66 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vmax_ukernel__avx_x8()
67 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vmax_ukernel__avx_x8()
71 _mm_store_ss(y, vy_lo); in xnn_f32_vmax_ukernel__avx_x8()
|
| D | vsqrdiffc-avx-x8.c | 58 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vsqrdiffc_ukernel__avx_x8() local
60 _mm_storeu_ps(y, vy_lo); in xnn_f32_vsqrdiffc_ukernel__avx_x8()
61 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vsqrdiffc_ukernel__avx_x8()
65 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vsqrdiffc_ukernel__avx_x8()
66 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vsqrdiffc_ukernel__avx_x8()
70 _mm_store_ss(y, vy_lo); in xnn_f32_vsqrdiffc_ukernel__avx_x8()
|
| D | vsqrdiff-avx-x8.c | 61 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vsqrdiff_ukernel__avx_x8() local
63 _mm_storeu_ps(y, vy_lo); in xnn_f32_vsqrdiff_ukernel__avx_x8()
64 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vsqrdiff_ukernel__avx_x8()
68 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vsqrdiff_ukernel__avx_x8()
69 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vsqrdiff_ukernel__avx_x8()
73 _mm_store_ss(y, vy_lo); in xnn_f32_vsqrdiff_ukernel__avx_x8()
|
| D | vsubc-minmax-avx-x8.c | 63 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vsubc_minmax_ukernel__avx_x8() local
65 _mm_storeu_ps(y, vy_lo); in xnn_f32_vsubc_minmax_ukernel__avx_x8()
66 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vsubc_minmax_ukernel__avx_x8()
70 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vsubc_minmax_ukernel__avx_x8()
71 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vsubc_minmax_ukernel__avx_x8()
75 _mm_store_ss(y, vy_lo); in xnn_f32_vsubc_minmax_ukernel__avx_x8()
|
| D | vmaxc-avx-x16.c | 67 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vmaxc_ukernel__avx_x16() local
69 _mm_storeu_ps(y, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x16()
70 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vmaxc_ukernel__avx_x16()
74 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x16()
75 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x16()
79 _mm_store_ss(y, vy_lo); in xnn_f32_vmaxc_ukernel__avx_x16()
|
| D | vdivc-minmax-avx-x8.c | 63 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vdivc_minmax_ukernel__avx_x8() local
65 _mm_storeu_ps(y, vy_lo); in xnn_f32_vdivc_minmax_ukernel__avx_x8()
66 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vdivc_minmax_ukernel__avx_x8()
70 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vdivc_minmax_ukernel__avx_x8()
71 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vdivc_minmax_ukernel__avx_x8()
75 _mm_store_ss(y, vy_lo); in xnn_f32_vdivc_minmax_ukernel__avx_x8()
|
| D | vminc-avx-x16.c | 67 __m128 vy_lo = _mm256_castps256_ps128(vy); in xnn_f32_vminc_ukernel__avx_x16() local
69 _mm_storeu_ps(y, vy_lo); in xnn_f32_vminc_ukernel__avx_x16()
70 vy_lo = _mm256_extractf128_ps(vy, 1); in xnn_f32_vminc_ukernel__avx_x16()
74 _mm_storel_pi((__m64*) y, vy_lo); in xnn_f32_vminc_ukernel__avx_x16()
75 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); in xnn_f32_vminc_ukernel__avx_x16()
79 _mm_store_ss(y, vy_lo); in xnn_f32_vminc_ukernel__avx_x16()
|