1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsqrt/fma3-nr1fma1adj.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x8(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28
29 const __m256 vhalf = _mm256_broadcast_ss(¶ms->fma.half);
30 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
31 const __m256 vx = _mm256_loadu_ps(x);
32 x += 8;
33
34 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
35 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
36 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
37 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
38 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
39 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
40 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
41 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
42
43 _mm256_storeu_ps(y, vy);
44 y += 8;
45 }
46 if XNN_UNLIKELY(n != 0) {
47 assert(n >= 1 * sizeof(float));
48 assert(n <= 7 * sizeof(float));
49 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
50
51 const __m256 vx = _mm256_maskload_ps(x, vmask);
52
53 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
54 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
55 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
56 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
57 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
58 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
59 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
60 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
61
62 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
63 __m128 vy_lo = _mm256_castps256_ps128(vy);
64 if (n & (4 * sizeof(float))) {
65 _mm_storeu_ps(y, vy_lo);
66 vy_lo = _mm256_extractf128_ps(vy, 1);
67 y += 4;
68 }
69 if (n & (2 * sizeof(float))) {
70 _mm_storel_pi((__m64*) y, vy_lo);
71 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
72 y += 2;
73 }
74 if (n & (1 * sizeof(float))) {
75 _mm_store_ss(y, vy_lo);
76 }
77 }
78 }
79