| /external/XNNPACK/src/x32-zip/ |
| D | x2-psimd.c | 28 const psimd_u32 vy = psimd_load_u32(y); in xnn_x32_zip_x2_ukernel__psimd() local
41 const psimd_u32 vy = psimd_load2_u32(y); in xnn_x32_zip_x2_ukernel__psimd() local
49 const uint32_t vy = *y; in xnn_x32_zip_x2_ukernel__psimd() local
|
| D | x2-sse2.c | 28 const __m128i vy = _mm_loadu_si128((const __m128i*) y); in xnn_x32_zip_x2_ukernel__sse2() local
41 const __m128i vy = _mm_loadl_epi64((const __m128i*) y); in xnn_x32_zip_x2_ukernel__sse2() local
49 const uint32_t vy = *y; in xnn_x32_zip_x2_ukernel__sse2() local
|
| D | x3-sse2.c | 31 const __m128 vy = _mm_loadu_ps(y); in xnn_x32_zip_x3_ukernel__sse2() local
63 const __m128 vy = _mm_castpd_ps(_mm_load_sd((const double*) y)); in xnn_x32_zip_x3_ukernel__sse2() local
82 const __m128 vy = _mm_load_ss(y); in xnn_x32_zip_x3_ukernel__sse2() local
|
| D | x4-sse2.c | 30 const __m128i vy = _mm_loadu_si128((const __m128i*) y); in xnn_x32_zip_x4_ukernel__sse2() local
58 const __m128i vy = _mm_loadl_epi64((const __m128i*) y); in xnn_x32_zip_x4_ukernel__sse2() local
77 const uint32_t vy = *y; in xnn_x32_zip_x4_ukernel__sse2() local
|
| D | x4-psimd.c | 30 const psimd_u32 vy = psimd_load_u32(y); in xnn_x32_zip_x4_ukernel__psimd() local
58 const psimd_u32 vy = psimd_load2_u32(y); in xnn_x32_zip_x4_ukernel__psimd() local
77 const uint32_t vy = *y; in xnn_x32_zip_x4_ukernel__psimd() local
|
| /external/XNNPACK/src/x8-zip/ |
| D | x2-sse2.c | 27 const __m128i vy = _mm_loadu_si128((const __m128i*) y); in xnn_x8_zip_x2_ukernel__sse2() local
39 const __m128i vy = _mm_loadu_si128((const __m128i*) ((uintptr_t) y + address_increment)); in xnn_x8_zip_x2_ukernel__sse2() local
49 const uint8_t vy = *y++; in xnn_x8_zip_x2_ukernel__sse2() local
|
| D | x4-sse2.c | 29 const __m128i vy = _mm_loadu_si128((const __m128i*) y); in xnn_x8_zip_x4_ukernel__sse2() local
53 const __m128i vy = _mm_loadu_si128((const __m128i*) ((uintptr_t) y + address_increment)); in xnn_x8_zip_x4_ukernel__sse2() local
73 const uint8_t vy = *y++; in xnn_x8_zip_x4_ukernel__sse2() local
|
| /external/XNNPACK/src/f32-clamp/ |
| D | sse.c | 29 const __m128 vy = _mm_min_ps(_mm_max_ps(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__sse() local
37 __m128 vy = _mm_min_ps(_mm_max_ps(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__sse() local
|
| D | psimd.c | 29 const psimd_f32 vy = psimd_min_f32(psimd_max_f32(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__psimd() local
37 psimd_f32 vy = psimd_min_f32(psimd_max_f32(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__psimd() local
|
| D | avx512f.c | 30 const __m512 vy = _mm512_min_ps(_mm512_max_ps(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__avx512f() local
44 const __m512 vy = _mm512_min_ps(_mm512_max_ps(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__avx512f() local
|
| D | avx.c | 31 const __m256 vy = _mm256_min_ps(_mm256_max_ps(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__avx() local
43 const __m256 vy = _mm256_min_ps(_mm256_max_ps(vx, voutput_min), voutput_max); in xnn_f32_clamp_ukernel__avx() local
|
| D | neon.c | 35 float32x4_t vy = vminq_f32(vx, voutput_max); in xnn_f32_clamp_ukernel__neon() local
43 float32x4_t vy = vminq_f32(vx, voutput_max); in xnn_f32_clamp_ukernel__neon() local
|
| /external/XNNPACK/src/f32-vbinary/gen/ |
| D | vaddc-avx512f-x16.c | 50 __m512 vy = _mm512_add_ps(va, vb); in xnn_f32_vaddc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_add_ps(va, vb); in xnn_f32_vaddc_ukernel__avx512f_x16() local
|
| D | vrsubc-avx512f-x16.c | 50 __m512 vy = _mm512_sub_ps(vb, va); in xnn_f32_vrsubc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_sub_ps(vb, va); in xnn_f32_vrsubc_ukernel__avx512f_x16() local
|
| D | vrdivc-avx512f-x16.c | 50 __m512 vy = _mm512_div_ps(vb, va); in xnn_f32_vrdivc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_div_ps(vb, va); in xnn_f32_vrdivc_ukernel__avx512f_x16() local
|
| D | vminc-avx512f-x16.c | 50 __m512 vy = _mm512_min_ps(va, vb); in xnn_f32_vminc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_min_ps(va, vb); in xnn_f32_vminc_ukernel__avx512f_x16() local
|
| D | vsubc-avx512f-x16.c | 50 __m512 vy = _mm512_sub_ps(va, vb); in xnn_f32_vsubc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_sub_ps(va, vb); in xnn_f32_vsubc_ukernel__avx512f_x16() local
|
| D | vmulc-avx512f-x16.c | 50 __m512 vy = _mm512_mul_ps(va, vb); in xnn_f32_vmulc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_mul_ps(va, vb); in xnn_f32_vmulc_ukernel__avx512f_x16() local
|
| D | vdivc-avx512f-x16.c | 50 __m512 vy = _mm512_div_ps(va, vb); in xnn_f32_vdivc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_div_ps(va, vb); in xnn_f32_vdivc_ukernel__avx512f_x16() local
|
| D | vmaxc-avx512f-x16.c | 50 __m512 vy = _mm512_max_ps(va, vb); in xnn_f32_vmaxc_ukernel__avx512f_x16() local
65 __m512 vy = _mm512_max_ps(va, vb); in xnn_f32_vmaxc_ukernel__avx512f_x16() local
|
| D | vrsubc-avx512f-x32.c | 55 __m512 vy = _mm512_sub_ps(vb, va); in xnn_f32_vrsubc_ukernel__avx512f_x32() local
70 __m512 vy = _mm512_sub_ps(vb, va); in xnn_f32_vrsubc_ukernel__avx512f_x32() local
|
| D | vmulc-avx-x8.c | 51 __m256 vy = _mm256_mul_ps(va, vb); in xnn_f32_vmulc_ukernel__avx_x8() local
64 __m256 vy = _mm256_mul_ps(va, vb); in xnn_f32_vmulc_ukernel__avx_x8() local
|
| D | vaddc-avx-x8.c | 51 __m256 vy = _mm256_add_ps(va, vb); in xnn_f32_vaddc_ukernel__avx_x8() local
64 __m256 vy = _mm256_add_ps(va, vb); in xnn_f32_vaddc_ukernel__avx_x8() local
|
| D | vmin-avx512f-x16.c | 55 __m512 vy = _mm512_min_ps(va, vb); in xnn_f32_vmin_ukernel__avx512f_x16() local
71 __m512 vy = _mm512_min_ps(va, vb); in xnn_f32_vmin_ukernel__avx512f_x16() local
|
| D | vminc-avx512f-x32.c | 55 __m512 vy = _mm512_min_ps(va, vb); in xnn_f32_vminc_ukernel__avx512f_x32() local
70 __m512 vy = _mm512_min_ps(va, vb); in xnn_f32_vminc_ukernel__avx512f_x32() local
|