16419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org/* 26419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org * Copyright 2014 The Android Open Source Project 36419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org * 46419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org * Use of this source code is governed by a BSD-style license that can be 56419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org * found in the LICENSE file. 66419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org */ 76419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 86419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org#ifndef SkMath_opts_SSE2_DEFINED 96419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org#define SkMath_opts_SSE2_DEFINED 106419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 116419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org#include <emmintrin.h> 126419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 13e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.org// Because no _mm_div_epi32() in SSE2, we use float division to emulate. 14e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.org// When using this function, make sure a and b don't exceed float's precision. 15e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.orgstatic inline __m128i shim_mm_div_epi32(const __m128i& a, const __m128i& b) { 16e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.org __m128 x = _mm_cvtepi32_ps(a); 17e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.org __m128 y = _mm_cvtepi32_ps(b); 18e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.org return _mm_cvttps_epi32(_mm_div_ps(x, y)); 19e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.org} 20e1ba93ee01aa7df27197189ab4d82a7d5387dc8acommit-bot@chromium.org 216419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org// Portable version of SkSqrtBits is in SkMath.cpp. 226419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.orgstatic inline __m128i SkSqrtBits_SSE2(const __m128i& x, int count) { 236419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i root = _mm_setzero_si128(); 246419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i remHi = _mm_setzero_si128(); 256419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i remLo = x; 266419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i one128 = _mm_set1_epi32(1); 276419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 286419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org do { 296419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org root = _mm_slli_epi32(root, 1); 306419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 316419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org remHi = _mm_or_si128(_mm_slli_epi32(remHi, 2), 326419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org _mm_srli_epi32(remLo, 30)); 336419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org remLo = _mm_slli_epi32(remLo, 2); 346419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 356419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i testDiv = _mm_slli_epi32(root, 1); 366419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org testDiv = _mm_add_epi32(testDiv, _mm_set1_epi32(1)); 376419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 386419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i cmp = _mm_cmplt_epi32(remHi, testDiv); 396419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i remHi1 = _mm_and_si128(cmp, remHi); 406419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i root1 = _mm_and_si128(cmp, root); 416419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i remHi2 = _mm_andnot_si128(cmp, _mm_sub_epi32(remHi, testDiv)); 426419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org __m128i root2 = _mm_andnot_si128(cmp, _mm_add_epi32(root, one128)); 436419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 446419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org remHi = _mm_or_si128(remHi1, remHi2); 456419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org root = _mm_or_si128(root1, root2); 466419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org } while (--count >= 0); 476419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 486419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org return root; 496419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org} 506419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org 516419a5edb8af999623e018b09c3dd88ec371d2e6commit-bot@chromium.org#endif // SkMath_opts_SSE2_DEFINED 52