Home
last modified time | relevance | path

Searched refs:vhalfrsqrtx3 (Results 1 – 10 of 10) sorted by relevance

/external/XNNPACK/src/f32-vsqrt/gen/
Davx512f-nr1fma1adj-x64.c48 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64() local
53 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
61 vhalfrsqrtx3 = _mm512_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
72 const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
Dfma3-nr1fma1adj-x32.c49 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32() local
54 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
62 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
73 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
Davx512f-nr1fma1adj-x80.c50 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80() local
57 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
66 vhalfrsqrtx3 = _mm512_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
80 const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
Dfma3-nr1fma1adj-x40.c51 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40() local
58 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
67 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
81 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
Davx512f-nr1fma1adj-x96.c52 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96() local
61 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
71 vhalfrsqrtx3 = _mm512_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
88 const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
Dfma3-nr1fma1adj-x48.c53 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48() local
62 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
72 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
89 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
Davx512f-nr1fma1adj-x112.c54 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112() local
65 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
76 vhalfrsqrtx3 = _mm512_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
96 const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
Dfma3-nr1fma1adj-x56.c55 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56() local
66 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
77 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
97 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
Davx512f-nr1fma1adj-x128.c56 __m512 vhalfrsqrtx3 = _mm512_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128() local
69 const __m512 vresidual3 = _mm512_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
81 vhalfrsqrtx3 = _mm512_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
104 const __m512 vy3 = _mm512_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
Dfma3-nr1fma1adj-x64.c57 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64() local
70 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
82 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
105 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()