Home
last modified time | relevance | path

Searched refs:vrsqrtx2 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-vsqrt/gen/
Davx512f-nr1fma1adj-x48.c37 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48() local
43 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
44 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48()
Dfma3-nr1fma1adj-x24.c38 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24() local
44 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24()
45 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24()
Davx512f-nr1fma1adj-x64.c38 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64() local
45 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
46 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64()
Dfma3-nr1fma1adj-x32.c39 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32() local
46 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
47 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32()
Davx512f-nr1fma1adj-x80.c39 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80() local
47 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
48 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80()
Dfma3-nr1fma1adj-x40.c40 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40() local
48 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
49 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40()
Davx512f-nr1fma1adj-x96.c40 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96() local
49 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
50 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96()
Dfma3-nr1fma1adj-x48.c41 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48() local
50 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
51 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48()
Davx512f-nr1fma1adj-x112.c41 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112() local
51 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
52 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
Dfma3-nr1fma1adj-x56.c42 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56() local
52 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
53 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x56()
Davx512f-nr1fma1adj-x128.c42 const __m512 vrsqrtx2 = _mm512_rsqrt14_ps(vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128() local
53 __m512 vsqrtx2 = _mm512_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
54 __m512 vhalfrsqrtx2 = _mm512_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
Dfma3-nr1fma1adj-x64.c43 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64() local
54 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()
55 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64()