Code cleanup: some reshuffling of SIMD defines moving more code to util_optimization.h.

This commit is contained in:
Brecht Van Lommel 2014-01-15 15:11:50 +01:00
parent ebbb6adf32
commit 8af782ad22
6 changed files with 94 additions and 69 deletions

View File

@ -20,7 +20,6 @@
/* CPU Kernel Interface */
#include "util_types.h"
#include "util_optimization.h"
CCL_NAMESPACE_BEGIN

View File

@ -17,16 +17,16 @@
/* Optimized CPU kernel entry points. This file is compiled with SSE2
* optimization flags and nearly all functions inlined, while kernel.cpp
* is compiled without for other CPU's. */
#include "util_optimization.h"
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
/* SSE optimization disabled for now on 32 bit, see bug #36316 */
#if !(defined(__GNUC__) && (defined(i386) || defined(_M_IX86)))
#define __KERNEL_SSE2__
#endif
#include "util_optimization.h"
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
#include "kernel.h"
#include "kernel_compat_cpu.h"
#include "kernel_math.h"

View File

@ -17,10 +17,6 @@
/* Optimized CPU kernel entry points. This file is compiled with SSE3/SSSE3
* optimization flags and nearly all functions inlined, while kernel.cpp
* is compiled without for other CPU's. */
#include "util_optimization.h"
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
/* SSE optimization disabled for now on 32 bit, see bug #36316 */
#if !(defined(__GNUC__) && (defined(i386) || defined(_M_IX86)))
@ -29,6 +25,10 @@
#define __KERNEL_SSSE3__
#endif
#include "util_optimization.h"
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
#include "kernel.h"
#include "kernel_compat_cpu.h"
#include "kernel_math.h"

View File

@ -17,10 +17,6 @@
/* Optimized CPU kernel entry points. This file is compiled with SSE3/SSSE3
* optimization flags and nearly all functions inlined, while kernel.cpp
* is compiled without for other CPU's. */
#include "util_optimization.h"
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
/* SSE optimization disabled for now on 32 bit, see bug #36316 */
#if !(defined(__GNUC__) && (defined(i386) || defined(_M_IX86)))
@ -30,6 +26,10 @@
#define __KERNEL_SSE41__
#endif
#include "util_optimization.h"
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
#include "kernel.h"
#include "kernel_compat_cpu.h"
#include "kernel_math.h"

View File

@ -14,8 +14,31 @@
* limitations under the License
*/
#ifndef __UTIL_OPTIMIZATION_H__
#define __UTIL_OPTIMIZATION_H__
#ifndef __KERNEL_GPU__
/* x86
*
* Compile a regular, SSE2 and SSE3 kernel. */
#if defined(i386) || defined(_M_IX86)
#define WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
#define WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
#endif
/* x86-64
*
* Compile a regular (includes SSE2), SSE3 and SSE 4.1 kernel. */
#if defined(__x86_64__) || defined(_M_X64)
/* SSE2 is always available on x86-64 CPUs, so auto enable */
#define __KERNEL_SSE2__
/* no SSE2 kernel on x86-64, part of regular kernel */
#define WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
#define WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
@ -27,9 +50,60 @@
#endif
#if defined(i386) || defined(_M_IX86)
/* SSE Experiment
*
* This is disabled code for an experiment to use SSE types globally for types
* such as float3 and float4. Currently this gives an overall slowdown. */
#define WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
#define WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
#if 0
#define __KERNEL_SSE__
#ifndef __KERNEL_SSE2__
#define __KERNEL_SSE2__
#endif
#ifndef __KERNEL_SSE3__
#define __KERNEL_SSE3__
#endif
#ifndef __KERNEL_SSSE3__
#define __KERNEL_SSSE3__
#endif
#ifndef __KERNEL_SSE4__
#define __KERNEL_SSE4__
#endif
#endif
/* SSE Intrinsics includes
*
* We assume __KERNEL_SSEX__ flags to have been defined at this point */
/* SSE intrinsics headers */
#ifndef FREE_WINDOWS64
#ifdef __KERNEL_SSE2__
#include <xmmintrin.h> /* SSE 1 */
#include <emmintrin.h> /* SSE 2 */
#endif
#ifdef __KERNEL_SSE3__
#include <pmmintrin.h> /* SSE 3 */
#endif
#ifdef __KERNEL_SSSE3__
#include <tmmintrin.h> /* SSSE 3 */
#endif
#ifdef __KERNEL_SSE41__
#include <smmintrin.h> /* SSE 4.1 */
#endif
#else
/* MinGW64 has conflicting declarations for these SSE headers in <windows.h>.
* Since we can't avoid including <windows.h>, better only include that */
#include <windows.h>
#endif
#endif
#endif /* __UTIL_OPTIMIZATION_H__ */

View File

@ -57,67 +57,19 @@
#endif
/* SIMD Types */
/* Standard Integer Types */
#ifndef __KERNEL_GPU__
#define __KERNEL_SSE2__
/* not enabled, globally applying it gives slowdown, only for testing. */
#if 0
#define __KERNEL_SSE__
#ifndef __KERNEL_SSE2__
#define __KERNEL_SSE2__
#endif
#ifndef __KERNEL_SSE3__
#define __KERNEL_SSE3__
#endif
#ifndef __KERNEL_SSSE3__
#define __KERNEL_SSSE3__
#endif
#ifndef __KERNEL_SSE4__
#define __KERNEL_SSE4__
#endif
#endif
/* SSE2 is always available on x86_64 CPUs, so auto enable */
#if defined(__x86_64__) && !defined(__KERNEL_SSE2__)
#define __KERNEL_SSE2__
#endif
/* SSE intrinsics headers */
#ifndef FREE_WINDOWS64
#ifdef __KERNEL_SSE2__
#include <xmmintrin.h> /* SSE 1 */
#include <emmintrin.h> /* SSE 2 */
#endif
#ifdef __KERNEL_SSE3__
#include <pmmintrin.h> /* SSE 3 */
#endif
#ifdef __KERNEL_SSSE3__
#include <tmmintrin.h> /* SSSE 3 */
#endif
#ifdef __KERNEL_SSE41__
#include <smmintrin.h> /* SSE 4.1 */
#endif
#else
/* MinGW64 has conflicting declarations for these SSE headers in <windows.h>.
* Since we can't avoid including <windows.h>, better only include that */
#include <windows.h>
#endif
/* int8_t, uint16_t, and friends */
#ifndef _WIN32
#include <stdint.h>
#endif
/* SIMD Types */
#include "util_optimization.h"
#endif
CCL_NAMESPACE_BEGIN