1/*
2 * Copyright 2014 The Android Open Source Project
3 *
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
6 */
7
8#ifndef SkMath_opts_SSE2_DEFINED
9#define SkMath_opts_SSE2_DEFINED
10
11#include <emmintrin.h>
12
13// Because no _mm_div_epi32() in SSE2, we use float division to emulate.
14// When using this function, make sure a and b don't exceed float's precision.
15static inline __m128i shim_mm_div_epi32(const __m128i& a, const __m128i& b) {
16    __m128 x = _mm_cvtepi32_ps(a);
17    __m128 y = _mm_cvtepi32_ps(b);
18    return _mm_cvttps_epi32(_mm_div_ps(x, y));
19}
20
21// Portable version of SkSqrtBits is in SkMath.cpp.
22static inline __m128i SkSqrtBits_SSE2(const __m128i& x, int count) {
23    __m128i root =  _mm_setzero_si128();
24    __m128i remHi = _mm_setzero_si128();
25    __m128i remLo = x;
26    __m128i one128 = _mm_set1_epi32(1);
27
28    do {
29        root = _mm_slli_epi32(root, 1);
30
31        remHi = _mm_or_si128(_mm_slli_epi32(remHi, 2),
32                             _mm_srli_epi32(remLo, 30));
33        remLo = _mm_slli_epi32(remLo, 2);
34
35        __m128i testDiv = _mm_slli_epi32(root, 1);
36        testDiv = _mm_add_epi32(testDiv, _mm_set1_epi32(1));
37
38        __m128i cmp = _mm_cmplt_epi32(remHi, testDiv);
39        __m128i remHi1 = _mm_and_si128(cmp, remHi);
40        __m128i root1 = _mm_and_si128(cmp, root);
41        __m128i remHi2 = _mm_andnot_si128(cmp, _mm_sub_epi32(remHi, testDiv));
42        __m128i root2 = _mm_andnot_si128(cmp, _mm_add_epi32(root, one128));
43
44        remHi = _mm_or_si128(remHi1, remHi2);
45        root = _mm_or_si128(root1, root2);
46    } while (--count >= 0);
47
48    return root;
49}
50
51#endif // SkMath_opts_SSE2_DEFINED
52