Mercurial > sdl-ios-xcode
view src/atomic/SDL_atomic.c @ 5053:b5b42be9333c
Fixed bug #1026
Vittorio Giovara 2010-07-16 19:09:28 PDT
i was reading SDL_renderer_gles and i noticed that every time we there
is some gl call the gl state is modified with a couple of
glEnableClientState()/glDisableClientState.
While this is completely fine for desktops systems, this is a major
performace kill on mobile devices, right where opengles is
implemented.
Normal practice in this case is to update the glstate once, keep it
always the same and disable/enable other states only in very special
occasions.
On the web there's plenty of documentation (on the top of my head
http://developer.apple.com/iphone/library/documentation/3DDrawing/Conceptual/OpenGLES_ProgrammingGuide/Performance/Performance.html#//apple_ref/doc/uid/TP40008793-CH105-SW5
) and i personally tried this.
I modified my code and got a 10 fps boost, then modified SDL_render_gles and
shifted from 40 fps to 50 fps alone -- considering that i started from ~30fps i
got an 80% performance increase with this technique.
I have attached a dif of my changes, hope that it will be included in
mainstream.
author | Sam Lantinga <slouken@libsdl.org> |
---|---|
date | Wed, 19 Jan 2011 23:56:16 -0800 |
parents | 0c72ae7b7cb2 |
children | dceec93471e7 |
line wrap: on
line source
/* SDL - Simple DirectMedia Layer Copyright (C) 1997-2010 Sam Lantinga This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License as published by the Free Software Foundation; either version 2.1 of the License, or (at your option) any later version. This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details. You should have received a copy of the GNU Lesser General Public License along with this library; if not, write to the Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA Sam Lantinga slouken@libsdl.org */ #include "SDL_stdinc.h" #include "SDL_atomic.h" /* Note that we undefine the atomic operations here, in case they are defined as compiler intrinsics while building SDL but the library user doesn't have that compiler. That way we always have a working set of atomic operations built into the library. */ /* If any of the operations are not provided then we must emulate some of them. That means we need a nice implementation of spin locks that avoids the "one big lock" problem. We use a vector of spin locks and pick which one to use based on the address of the operand of the function. To generate the index of the lock we first shift by 3 bits to get rid on the zero bits that result from 32 and 64 bit allignment of data. We then mask off all but 5 bits and use those 5 bits as an index into the table. Picking the lock this way insures that accesses to the same data at the same time will go to the same lock. OTOH, accesses to different data have only a 1/32 chance of hitting the same lock. That should pretty much eliminate the chances of several atomic operations on different data from waiting on the same "big lock". If it isn't then the table of locks can be expanded to a new size so long as the new size is a power of two. Contributed by Bob Pendleton, bob@pendleton.com */ static SDL_SpinLock locks[32]; static __inline__ void enterLock(void *a) { uintptr_t index = ((((uintptr_t)a) >> 3) & 0x1f); SDL_AtomicLock(&locks[index]); } static __inline__ void leaveLock(void *a) { uintptr_t index = ((((uintptr_t)a) >> 3) & 0x1f); SDL_AtomicUnlock(&locks[index]); } #undef SDL_AtomicSet int SDL_AtomicSet(SDL_atomic_t *a, int value) { int oldvalue; enterLock(a); oldvalue = a->value; a->value = value; leaveLock(a); return oldvalue; } #undef SDL_AtomicGet int SDL_AtomicGet(SDL_atomic_t *a) { /* Assuming integral reads on this platform, we're safe here since the functions that set the variable have the necessary memory barriers. */ return a->value; } #undef SDL_AtomicAdd int SDL_AtomicAdd(SDL_atomic_t *a, int value) { int oldvalue; enterLock(a); oldvalue = a->value; a->value += value; leaveLock(a); return oldvalue; } #undef SDL_AtomicIncRef void SDL_AtomicIncRef(SDL_atomic_t *a) { SDL_AtomicAdd(a, 1); } #undef SDL_AtomicDecRef SDL_bool SDL_AtomicDecRef(SDL_atomic_t *a) { return SDL_AtomicAdd(a, -1) == 1; } #undef SDL_AtomicCAS SDL_bool SDL_AtomicCAS(SDL_atomic_t *a, int oldval, int newval) { SDL_bool retval = SDL_FALSE; enterLock(a); if (a->value == oldval) { a->value = newval; retval = SDL_TRUE; } leaveLock(a); return retval; } #undef SDL_AtomicSetPtr void SDL_AtomicSetPtr(void** a, void* value) { void *prevval; do { prevval = *a; } while (!SDL_AtomicCASPtr(a, prevval, value)); } #undef SDL_AtomicGetPtr void* SDL_AtomicGetPtr(void** a) { /* Assuming integral reads on this platform, we're safe here since the functions that set the pointer have the necessary memory barriers. */ return *a; } #undef SDL_AtomicCASPtr SDL_bool SDL_AtomicCASPtr(void **a, void *oldval, void *newval) { SDL_bool retval = SDL_FALSE; enterLock(a); if (*a == oldval) { *a = newval; retval = SDL_TRUE; } leaveLock(a); return retval; } /* vi: set ts=4 sw=4 expandtab: */