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 <liboilclasses.h>
sl@0: #include <liboilfunction.h>
sl@0: #include <emmintrin.h>
sl@0: #include <xmmintrin.h>
sl@0: 
sl@0: #define SSE_FUNCTION __attribute__((force_align_arg_pointer))
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: add_f32_sse (float *dest, float *src1, float *src2, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1++ + *src2++;
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm1 = _mm_loadu_ps(src2);
sl@0:     xmm0 = _mm_add_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:     src2 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1++ + *src2++;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (add_f32_sse, add_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: add_f64_sse2 (double *dest, double *src1, double *src2, int n)
sl@0: {
sl@0:   __m128d xmm0, xmm1;
sl@0:   while (((long)dest & 15) && (0 < n)) {
sl@0:     *dest++ = *src1++ + *src2++;
sl@0:     n--;
sl@0:   }
sl@0:   while (1 < n) {
sl@0:     xmm0 = _mm_loadu_pd(src1);
sl@0:     xmm1 = _mm_loadu_pd(src2);
sl@0:     xmm0 = _mm_add_pd(xmm0, xmm1);
sl@0:     _mm_store_pd(dest, xmm0);
sl@0:     dest += 2;
sl@0:     src1 += 2;
sl@0:     src2 += 2;
sl@0:     n -= 2;
sl@0:   }
sl@0:   while (0 < n) {
sl@0:     *dest++ = *src1++ + *src2++;
sl@0:     n--;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (add_f64_sse2, add_f64, OIL_IMPL_FLAG_SSE2);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: add_f64_sse2_unroll (double *dest, double *src1, double *src2, int n)
sl@0: {
sl@0:   __m128d xmm0, xmm1;
sl@0:   while (((long)dest & 15) && (0 < n)) {
sl@0:     *dest++ = *src1++ + *src2++;
sl@0:     n--;
sl@0:   }
sl@0:   while (3 < n) {
sl@0:     xmm0 = _mm_loadu_pd(src1);
sl@0:     xmm1 = _mm_loadu_pd(src2);
sl@0:     xmm0 = _mm_add_pd(xmm0, xmm1);
sl@0:     _mm_store_pd(dest, xmm0);
sl@0: 
sl@0:     xmm0 = _mm_loadu_pd(src1+2);
sl@0:     xmm1 = _mm_loadu_pd(src2+2);
sl@0:     xmm0 = _mm_add_pd(xmm0, xmm1);
sl@0:     _mm_store_pd(dest+2, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:     src2 += 4;
sl@0:     n -= 4;
sl@0:   }
sl@0:   while (1 < n) {
sl@0:     xmm0 = _mm_loadu_pd(src1);
sl@0:     xmm1 = _mm_loadu_pd(src2);
sl@0:     xmm0 = _mm_add_pd(xmm0, xmm1);
sl@0:     _mm_store_pd(dest, xmm0);
sl@0:     dest += 2;
sl@0:     src1 += 2;
sl@0:     src2 += 2;
sl@0:     n -= 2;
sl@0:   }
sl@0:   while (0 < n) {
sl@0:     *dest++ = *src1++ + *src2++;
sl@0:     n--;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (add_f64_sse2_unroll, add_f64, OIL_IMPL_FLAG_SSE2);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: subtract_f32_sse (float *dest, float *src1, float *src2, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1++ - *src2++;
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm1 = _mm_loadu_ps(src2);
sl@0:     xmm0 = _mm_sub_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:     src2 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1++ - *src2++;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (subtract_f32_sse, subtract_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: multiply_f32_sse (float *dest, float *src1, float *src2, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1++ * *src2++;
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm1 = _mm_loadu_ps(src2);
sl@0:     xmm0 = _mm_mul_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:     src2 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1++ * *src2++;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (multiply_f32_sse, multiply_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: divide_f32_sse (float *dest, float *src1, float *src2, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1++ / *src2++;
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm1 = _mm_loadu_ps(src2);
sl@0:     xmm0 = _mm_div_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:     src2 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1++ / *src2++;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (divide_f32_sse, divide_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: minimum_f32_sse (float *dest, float *src1, float *src2, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1 < *src2 ? *src1 : *src2;
sl@0:     src1++;
sl@0:     src2++;
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm1 = _mm_loadu_ps(src2);
sl@0:     xmm0 = _mm_min_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:     src2 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1 < *src2 ? *src1 : *src2;
sl@0:     src1++;
sl@0:     src2++;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (minimum_f32_sse, minimum_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: maximum_f32_sse (float *dest, float *src1, float *src2, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1 > *src2 ? *src1 : *src2;
sl@0:     src1++;
sl@0:     src2++;
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm1 = _mm_loadu_ps(src2);
sl@0:     xmm0 = _mm_max_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:     src2 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1 > *src2 ? *src1 : *src2;
sl@0:     src1++;
sl@0:     src2++;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (maximum_f32_sse, maximum_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: inverse_f32_sse (float *dest, float *src1, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = 1.0 / *src1++;
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     /* While _mm_rcp_ps sounds promising, the results it gives are rather
sl@0:      * different from the 1.0 / src1 reference implementation, so do that.
sl@0:      */
sl@0:     xmm0 = _mm_set_ps1(1.0);
sl@0:     xmm1 = _mm_loadu_ps(src1);
sl@0:     xmm0 = _mm_div_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = 1.0 / *src1++;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (inverse_f32_sse, inverse_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: negative_f32_sse (float *dest, float *src1, int n)
sl@0: {
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = -(*src1++);
sl@0:   }
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0, xmm1;
sl@0:     xmm0 = _mm_setzero_ps();
sl@0:     xmm1 = _mm_loadu_ps(src1);
sl@0:     xmm0 = _mm_sub_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = -(*src1++);
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (negative_f32_sse, negative_f32, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: scalaradd_f32_ns_sse (float *dest, float *src1, float *val, int n)
sl@0: {
sl@0:   __m128 xmm1;
sl@0: 
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1++ + *val;
sl@0:   }
sl@0:   xmm1 = _mm_load_ps1(val);
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm0 = _mm_add_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1++ + *val;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (scalaradd_f32_ns_sse, scalaradd_f32_ns, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: scalarmultiply_f32_ns_sse (float *dest, float *src1, float *val, int n)
sl@0: {
sl@0:   __m128 xmm1;
sl@0: 
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1++ * *val;
sl@0:   }
sl@0:   xmm1 = _mm_load_ps1(val);
sl@0:   for (; n >= 4; n -= 4) {
sl@0:     __m128 xmm0;
sl@0:     xmm0 = _mm_loadu_ps(src1);
sl@0:     xmm0 = _mm_mul_ps(xmm0, xmm1);
sl@0:     _mm_store_ps(dest, xmm0);
sl@0:     dest += 4;
sl@0:     src1 += 4;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1++ * *val;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (scalarmultiply_f32_ns_sse, scalarmultiply_f32_ns, OIL_IMPL_FLAG_SSE);
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: scalarmultiply_f64_ns_sse2 (double *dest, double *src1, double *val, int n)
sl@0: {
sl@0:   __m128d xmm1;
sl@0: 
sl@0:   /* Initial operations to align the destination pointer */
sl@0:   for (; ((long)dest & 15) && (n > 0); n--) {
sl@0:     *dest++ = *src1++ * *val;
sl@0:   }
sl@0:   xmm1 = _mm_load_pd1(val);
sl@0:   for (; n >= 2; n -= 2) {
sl@0:     __m128d xmm0;
sl@0:     xmm0 = _mm_loadu_pd(src1);
sl@0:     xmm0 = _mm_mul_pd(xmm0, xmm1);
sl@0:     _mm_store_pd(dest, xmm0);
sl@0:     dest += 2;
sl@0:     src1 += 2;
sl@0:   }
sl@0:   for (; n > 0; n--) {
sl@0:     *dest++ = *src1++ * *val;
sl@0:   }
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (scalarmultiply_f64_ns_sse2, scalarmultiply_f64_ns, OIL_IMPL_FLAG_SSE2);
sl@0: 
sl@0: 
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_add_f32_sse, add_f32() {
sl@0: 		return &_oil_function_impl_add_f32_sse, add_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_add_f64_sse2, add_f64() {
sl@0: 		return &_oil_function_impl_add_f64_sse2, add_f64;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_add_f64_sse2_unroll, add_f64() {
sl@0: 		return &_oil_function_impl_add_f64_sse2_unroll, add_f64;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_subtract_f32_sse, subtract_f32() {
sl@0: 		return &_oil_function_impl_subtract_f32_sse, subtract_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_multiply_f32_sse, multiply_f32() {
sl@0: 		return &_oil_function_impl_multiply_f32_sse, multiply_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_divide_f32_sse, divide_f32() {
sl@0: 		return &_oil_function_impl_divide_f32_sse, divide_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_minimum_f32_sse, minimum_f32() {
sl@0: 		return &_oil_function_impl_minimum_f32_sse, minimum_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_maximum_f32_sse, maximum_f32() {
sl@0: 		return &_oil_function_impl_maximum_f32_sse, maximum_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_inverse_f32_sse, inverse_f32() {
sl@0: 		return &_oil_function_impl_inverse_f32_sse, inverse_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_negative_f32_sse, negative_f32() {
sl@0: 		return &_oil_function_impl_negative_f32_sse, negative_f32;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_scalaradd_f32_ns_sse, scalaradd_f32_ns() {
sl@0: 		return &_oil_function_impl_scalaradd_f32_ns_sse, scalaradd_f32_ns;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_scalarmultiply_f32_ns_sse, scalarmultiply_f32_ns() {
sl@0: 		return &_oil_function_impl_scalarmultiply_f32_ns_sse, scalarmultiply_f32_ns;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_scalarmultiply_f64_ns_sse2, scalarmultiply_f64_ns() {
sl@0: 		return &_oil_function_impl_scalarmultiply_f64_ns_sse2, scalarmultiply_f64_ns;
sl@0: }
sl@0: #endif
sl@0: