| /external/XNNPACK/src/f32-vsqrt/gen/ |
| D | avx512f-nr1fma1adj-x32.c | 51 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32() local
54 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
|
| D | fma3-nr1fma1adj-x16.c | 52 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16() local
55 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16()
|
| D | avx512f-nr1fma1adj-x48.c | 58 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48() local
62 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
|
| D | fma3-nr1fma1adj-x24.c | 59 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24() local
63 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24()
|
| D | avx512f-nr1fma1adj-x64.c | 65 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64() local
70 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
|
| D | fma3-nr1fma1adj-x32.c | 66 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32() local
71 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
|
| D | avx512f-nr1fma1adj-x80.c | 72 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80() local
78 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
|
| D | fma3-nr1fma1adj-x40.c | 73 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40() local
79 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
|
| D | avx512f-nr1fma1adj-x96.c | 79 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96() local
86 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
|
| D | fma3-nr1fma1adj-x48.c | 80 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48() local
87 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
|
| D | avx512f-nr1fma1adj-x112.c | 86 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112() local
94 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
|
| D | fma3-nr1fma1adj-x56.c | 87 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56() local
95 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
|
| D | avx512f-nr1fma1adj-x128.c | 93 const __m512 vadjustment1 = _mm512_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128() local
102 const __m512 vy1 = _mm512_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
|
| D | fma3-nr1fma1adj-x64.c | 94 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64() local
103 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
|