Mercurial > sdl-ios-xcode
view src/hermes/x86_main.asm @ 1542:a8bf1aa21020
Fixed bug #15
SDL_blit_A.mmx-speed.patch.txt --
Speed improvements and a bugfix for the current GCC inline mmx
asm code:
- Changed some ops and removed some resulting useless ones.
- Added some instruction parallelism (some gain)
The resulting speed on my Xeon improved upto 35% depending on
the function (measured in fps).
- Fixed a bug where BlitRGBtoRGBSurfaceAlphaMMX() was
setting the alpha component on the destination surfaces (to
opaque-alpha) even when the surface had none.
SDL_blit_A.mmx-msvc.patch.txt --
MSVC mmx intrinsics version of the same GCC asm code.
MSVC compiler tries to parallelize the code and to avoid
register stalls, but does not always do a very good job.
Per-surface blending MSVC functions run quite a bit faster
than their pure-asm counterparts (upto 55% faster for 16bit
ones), but the per-pixel blending runs somewhat slower than asm.
- BlitRGBtoRGBSurfaceAlphaMMX and BlitRGBtoRGBPixelAlphaMMX (and all
variants) can now also handle formats other than (A)RGB8888. Formats
like RGBA8888 and some quite exotic ones are allowed -- like
RAGB8888, or actually anything having channels aligned on 8bit
boundary and full 8bit alpha (for per-pixel alpha blending).
The performance cost of this change is virtually 0 for per-surface
alpha blending (no extra ops inside the loop) and a single non-MMX
op inside the loop for per-pixel blending. In testing, the per-pixel
alpha blending takes a ~2% performance hit, but it still runs much
faster than the current code in CVS. If necessary, a separate function
with this functionality can be made.
This code requires Processor Pack for VC6.
author | Sam Lantinga <slouken@libsdl.org> |
---|---|
date | Wed, 15 Mar 2006 15:39:29 +0000 |
parents | 3202d727bb4b |
children | 393092a3ebf6 |
line wrap: on
line source
; ; x86 format converters for HERMES ; Some routines Copyright (c) 1998 Christian Nentwich (brn@eleet.mcb.at) ; This source code is licensed under the GNU LGPL ; ; Please refer to the file COPYING.LIB contained in the distribution for ; licensing conditions ; ; Most routines are (c) Glenn Fiedler (ptc@gaffer.org), used with permission ; BITS 32 GLOBAL _ConvertX86 GLOBAL _x86return GLOBAL _Hermes_X86_CPU SECTION .text ;; _ConvertX86: ;; [ESP+8] ConverterInfo* ;; -------------------------------------------------------------------------- ;; ConverterInfo (ebp+..) ;; 0: void *s_pixels ;; 4: int s_width ;; 8: int s_height ;; 12: int s_add ;; 16: void *d_pixels ;; 20: int d_width ;; 24: int d_height ;; 28: int d_add ;; 32: void (*converter_function)() ;; 36: int32 *lookup _ConvertX86: push ebp mov ebp,esp ; Save the registers used by the blitters, necessary for optimized code pusha mov eax,[ebp+8] cmp dword [eax+4],BYTE 0 je endconvert mov ebp,eax mov esi,[ebp+0] mov edi,[ebp+16] y_loop: mov ecx,[ebp+4] jmp [ebp+32] _x86return: add esi,[ebp+12] add edi,[ebp+28] dec dword [ebp+8] jnz y_loop ; Restore the registers used by the blitters, necessary for optimized code popa pop ebp endconvert: ret ;; Hermes_X86_CPU returns the CPUID flags in eax _Hermes_X86_CPU: pushfd pop eax mov ecx,eax xor eax,040000h push eax popfd pushfd pop eax xor eax,ecx jz .L1 ; Processor is 386 push ecx popfd mov eax,ecx xor eax,200000h push eax popfd pushfd pop eax xor eax,ecx je .L1 push ebx mov eax,1 cpuid mov eax,edx pop ebx .L1: ret %ifidn __OUTPUT_FORMAT__,elf section .note.GNU-stack noalloc noexec nowrite progbits %endif