sl@0: /* sl@0: * Copyright (c) 2005 sl@0: * Eric Anholt. All rights reserved. sl@0: * sl@0: * Redistribution and use in source and binary forms, with or without sl@0: * modification, are permitted provided that the following conditions sl@0: * are met: sl@0: * 1. Redistributions of source code must retain the above copyright sl@0: * notice, this list of conditions and the following disclaimer. sl@0: * 2. Redistributions in binary form must reproduce the above copyright sl@0: * notice, this list of conditions and the following disclaimer in the sl@0: * documentation and/or other materials provided with the distribution. sl@0: * sl@0: * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND sl@0: * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE sl@0: * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE sl@0: * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE sl@0: * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL sl@0: * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS sl@0: * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) sl@0: * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT sl@0: * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY sl@0: * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF sl@0: * SUCH DAMAGE. sl@0: */ sl@0: //Portions Copyright (c) 2008-2009 Nokia Corporation and/or its subsidiary(-ies). All rights reserved. sl@0: sl@0: #ifdef HAVE_CONFIG_H sl@0: #include "config.h" sl@0: #endif sl@0: #include sl@0: #include sl@0: #include sl@0: #include sl@0: sl@0: #define SSE_FUNCTION __attribute__((force_align_arg_pointer)) sl@0: sl@0: #define COMPOSITE_ADD(d,s) oil_clamp_255((d) + (s)) sl@0: sl@0: SSE_FUNCTION static void sl@0: composite_add_argb_sse (uint32_t *dest, const uint32_t *src, int n) sl@0: { sl@0: /* Initial operations to align the destination pointer */ sl@0: for (; ((long)dest & 15) && (n > 0); n--) { sl@0: uint32_t d = *dest, s = *src++; sl@0: sl@0: *dest++ = oil_argb( sl@0: COMPOSITE_ADD(oil_argb_A(d), oil_argb_A(s)), sl@0: COMPOSITE_ADD(oil_argb_R(d), oil_argb_R(s)), sl@0: COMPOSITE_ADD(oil_argb_G(d), oil_argb_G(s)), sl@0: COMPOSITE_ADD(oil_argb_B(d), oil_argb_B(s))); sl@0: } sl@0: for (; n >= 4; n -= 4) { sl@0: __m128i s; sl@0: s = _mm_loadu_si128((__m128i *)src); sl@0: *(__m128i *)dest = _mm_adds_epu8(s, *(__m128i *)dest); sl@0: src += 4; sl@0: dest += 4; sl@0: } sl@0: for (; n > 0; n--) { sl@0: uint32_t d = *dest, s = *src++; sl@0: sl@0: *dest++ = oil_argb( sl@0: COMPOSITE_ADD(oil_argb_A(d), oil_argb_A(s)), sl@0: COMPOSITE_ADD(oil_argb_R(d), oil_argb_R(s)), sl@0: COMPOSITE_ADD(oil_argb_G(d), oil_argb_G(s)), sl@0: COMPOSITE_ADD(oil_argb_B(d), oil_argb_B(s))); sl@0: } sl@0: } sl@0: OIL_DEFINE_IMPL_FULL (composite_add_argb_sse, composite_add_argb, sl@0: OIL_IMPL_FLAG_SSE2); sl@0: sl@0: SSE_FUNCTION static void sl@0: composite_add_argb_const_src_sse (uint32_t *dest, const uint32_t *src_1, int n) sl@0: { sl@0: __m128i s; sl@0: uint32_t val = *src_1; sl@0: sl@0: /* Initial operations to align the destination pointer */ sl@0: for (; ((long)dest & 15) && (n > 0); n--) { sl@0: uint32_t d = *dest; sl@0: sl@0: *dest++ = oil_argb( sl@0: COMPOSITE_ADD(oil_argb_A(d), oil_argb_A(val)), sl@0: COMPOSITE_ADD(oil_argb_R(d), oil_argb_R(val)), sl@0: COMPOSITE_ADD(oil_argb_G(d), oil_argb_G(val)), sl@0: COMPOSITE_ADD(oil_argb_B(d), oil_argb_B(val))); sl@0: } sl@0: s = _mm_set1_epi32(val); sl@0: for (; n >= 4; n -= 4) { sl@0: __m128i xmm0; sl@0: xmm0 = _mm_adds_epu8(s, *(__m128i *)dest); sl@0: _mm_store_si128((__m128i *)dest, xmm0); sl@0: dest += 4; sl@0: } sl@0: for (; n > 0; n--) { sl@0: uint32_t d = *dest; sl@0: sl@0: *dest++ = oil_argb( sl@0: COMPOSITE_ADD(oil_argb_A(d), oil_argb_A(val)), sl@0: COMPOSITE_ADD(oil_argb_R(d), oil_argb_R(val)), sl@0: COMPOSITE_ADD(oil_argb_G(d), oil_argb_G(val)), sl@0: COMPOSITE_ADD(oil_argb_B(d), oil_argb_B(val))); sl@0: } sl@0: } sl@0: OIL_DEFINE_IMPL_FULL (composite_add_argb_const_src_sse, sl@0: composite_add_argb_const_src, OIL_IMPL_FLAG_SSE2); sl@0: sl@0: SSE_FUNCTION static void sl@0: composite_add_u8_sse (uint8_t *dest, const uint8_t *src, int n) sl@0: { sl@0: /* Initial operations to align the destination pointer */ sl@0: for (; ((long)dest & 15) && (n > 0); n--) { sl@0: int x = (int)*dest + *src++; sl@0: if (x > 255) sl@0: x = 255; sl@0: *dest++ = x; sl@0: } sl@0: for (; n >= 16; n -= 16) { sl@0: __m128i d, s; sl@0: s = _mm_loadu_si128((__m128i *)src); sl@0: d = _mm_adds_epu8(s, *(__m128i *)dest); sl@0: _mm_store_si128((__m128i *)dest, d); sl@0: src += 16; sl@0: dest += 16; sl@0: } sl@0: for (; n > 0; n--) { sl@0: int x = (int)*dest + *src++; sl@0: if (x > 255) sl@0: x = 255; sl@0: *dest++ = x; sl@0: } sl@0: } sl@0: OIL_DEFINE_IMPL_FULL (composite_add_u8_sse, composite_add_u8, sl@0: OIL_IMPL_FLAG_SSE2); sl@0: sl@0: SSE_FUNCTION static void sl@0: composite_add_u8_const_src_sse (uint8_t *dest, const uint8_t *src_1, int n) sl@0: { sl@0: __m128i s; sl@0: int val = *src_1; sl@0: sl@0: /* Initial operations to align the destination pointer */ sl@0: for (; ((long)dest & 15) && (n > 0); n--) { sl@0: int x = *dest + val; sl@0: if (x > 255) sl@0: x = 255; sl@0: *dest++ = x; sl@0: } sl@0: s = _mm_set1_epi8(val); sl@0: for (; n >= 16; n -= 16) { sl@0: __m128i d; sl@0: d = _mm_adds_epu8(*(__m128i *)dest, s); sl@0: _mm_store_si128((__m128i *)dest, d); sl@0: dest += 16; sl@0: } sl@0: for (; n > 0; n--) { sl@0: int x = *dest + val; sl@0: if (x > 255) sl@0: x = 255; sl@0: *dest++ = x; sl@0: } sl@0: } sl@0: OIL_DEFINE_IMPL_FULL (composite_add_u8_const_src_sse, sl@0: composite_add_u8_const_src, OIL_IMPL_FLAG_SSE2); sl@0: sl@0: sl@0: #ifdef __SYMBIAN32__ sl@0: sl@0: OilFunctionImpl* __oil_function_impl_composite_add_argb_sse, composite_add_argb() { sl@0: return &_oil_function_impl_composite_add_argb_sse, composite_add_argb; sl@0: } sl@0: #endif sl@0: sl@0: #ifdef __SYMBIAN32__ sl@0: sl@0: OilFunctionImpl* __oil_function_impl_composite_add_argb_const_src_sse() { sl@0: return &_oil_function_impl_composite_add_argb_const_src_sse; sl@0: } sl@0: #endif sl@0: sl@0: #ifdef __SYMBIAN32__ sl@0: sl@0: OilFunctionImpl* __oil_function_impl_composite_add_u8_sse, composite_add_u8() { sl@0: return &_oil_function_impl_composite_add_u8_sse, composite_add_u8; sl@0: } sl@0: #endif sl@0: sl@0: #ifdef __SYMBIAN32__ sl@0: sl@0: OilFunctionImpl* __oil_function_impl_composite_add_u8_const_src_sse() { sl@0: return &_oil_function_impl_composite_add_u8_const_src_sse; sl@0: } sl@0: #endif sl@0: