/aosp_15_r20/external/XNNPACK/src/f32-vsqrt/gen/ |
H A D | neonfma-nr2fma1adj-x40.c | 70 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 173 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 188 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local
|
H A D | neonfma-nr1rsqrts1fma1adj-x40.c | 103 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 177 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 192 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local
|
H A D | neonfma-nr2fma1adj-x4.c | 32 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x4() local 47 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x4() local
|
H A D | avx512f-nr1fma1adj-x16.c | 34 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16() local 54 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x16() local
|
H A D | fma3-nr1fma1adj-x8.c | 33 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8() local 52 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8() local
|
H A D | neonfma-nr1rsqrts1fma1adj-x4.c | 35 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4() local 50 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x4() local
|
H A D | avx512f-nr1fma1adj-x32.c | 65 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32() local 85 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x32() local
|
H A D | neonfma-nr2fma1adj-x8.c | 69 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() local 84 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() local
|
H A D | fma3-nr1fma1adj-x16.c | 64 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16() local 83 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x16() local
|
H A D | neonfma-nr2fma1adj-x12.c | 82 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() local 97 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() local
|
H A D | fma3-nr1fma1adj-x24.c | 74 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24() local 93 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x24() local
|
H A D | avx512f-nr1fma1adj-x48.c | 75 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48() local 95 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x48() local
|
H A D | neonfma-nr1rsqrts1fma1adj-x8.c | 73 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() local 88 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() local
|
H A D | avx512f-nr1fma1adj-x64.c | 85 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64() local 105 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x64() local
|
H A D | neonfma-nr2fma1adj-x16.c | 95 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 110 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local
|
H A D | fma3-nr1fma1adj-x32.c | 84 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32() local 103 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x32() local
|
H A D | neonfma-nr1rsqrts1fma1adj-x12.c | 86 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() local 101 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() local
|
H A D | fma3-nr1fma1adj-x40.c | 94 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40() local 113 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x40() local
|
H A D | avx512f-nr1fma1adj-x80.c | 95 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80() local 115 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x80() local
|
H A D | neonfma-nr2fma1adj-x20.c | 108 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 123 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local
|
H A D | avx512f-nr1fma1adj-x96.c | 105 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96() local 125 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x96() local
|
/aosp_15_r20/external/XNNPACK/src/f32-vsqrt/ |
H A D | neonfma-nr2fma1adj.c.in | 66 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); variable 81 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); variable
|
H A D | neonfma-nr1rsqrts1fma1adj.c.in | 71 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); variable 86 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx); variable
|
H A D | avx512f-nr1fma1adj.c.in | 66 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); variable 86 __m512 vsqrtx = _mm512_mul_ps(vrsqrtx, vx); variable
|
H A D | fma3-nr1fma1adj.c.in | 65 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); variable 84 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx); variable
|