| /external/XNNPACK/src/f32-vsqrt/gen/ |
| D | avx512f-nr1fma1adj-x32.c | 34 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32() local
37 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
38 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32()
|
| D | fma3-nr1fma1adj-x16.c | 35 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16() local
38 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16()
39 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16()
|
| D | avx512f-nr1fma1adj-x48.c | 35 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48() local
39 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
40 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
|
| D | fma3-nr1fma1adj-x24.c | 36 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24() local
40 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24()
41 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24()
|
| D | avx512f-nr1fma1adj-x64.c | 36 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64() local
41 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
42 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
|
| D | fma3-nr1fma1adj-x32.c | 37 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32() local
42 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
43 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
|
| D | avx512f-nr1fma1adj-x80.c | 37 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80() local
43 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
44 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
|
| D | fma3-nr1fma1adj-x40.c | 38 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40() local
44 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
45 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
|
| D | avx512f-nr1fma1adj-x96.c | 38 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96() local
45 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
46 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
|
| D | fma3-nr1fma1adj-x48.c | 39 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48() local
46 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
47 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
|
| D | avx512f-nr1fma1adj-x112.c | 39 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112() local
47 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
48 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
|
| D | fma3-nr1fma1adj-x56.c | 40 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56() local
48 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
49 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
|
| D | avx512f-nr1fma1adj-x128.c | 40 const __m512 vrsqrtx0 = _mm512_rsqrt14_ps(vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128() local
49 __m512 vsqrtx0 = _mm512_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
50 __m512 vhalfrsqrtx0 = _mm512_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
|
| D | fma3-nr1fma1adj-x64.c | 41 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64() local
50 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
51 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
|