Mercurial > sdl-ios-xcode
view src/hermes/mmxp2_32.asm @ 1230:88c2d6aed428
From Mike Frysinger and/or Gentoo:
- libsdl-PIC-load-mmx-masks-from-stack.patch
this one may be a little controversial ... the fix here is again that you cant
reference the memory addresses like this to load into a mmx register, so the
way to do it is to push two 32bit words onto the stack, load the 64bit value
off of the stack into the mmx register, and then adjust the stack so that
it's back to normal.
author | Ryan C. Gordon <icculus@icculus.org> |
---|---|
date | Thu, 05 Jan 2006 15:25:19 +0000 |
parents | 2d6dc7de1145 |
children | 393092a3ebf6 |
line wrap: on
line source
; ; pII-optimised MMX format converters for HERMES ; Copyright (c) 1998 Christian Nentwich (c.nentwich@cs.ucl.ac.uk) ; and (c) 1999 Jonathan Matthew (jmatthew@uq.net.au) ; This source code is licensed under the GNU LGPL ; ; Please refer to the file COPYING.LIB contained in the distribution for ; licensing conditions ; ; COPYRIGHT NOTICE ; ; This file partly contains code that is (c) Intel Corporation, specifically ; the mode detection routine, and the converter to 15 bit (8 pixel ; conversion routine from the mmx programming tutorial pages). ; ; ; These routines aren't exactly pII optimised - it's just that as they ; are, they're terrible on p5 MMXs, but less so on pIIs. Someone needs to ; optimise them for p5 MMXs.. BITS 32 GLOBAL _ConvertMMXpII32_24RGB888 GLOBAL _ConvertMMXpII32_16RGB565 GLOBAL _ConvertMMXpII32_16BGR565 GLOBAL _ConvertMMXpII32_16RGB555 GLOBAL _ConvertMMXpII32_16BGR555 EXTERN _mmxreturn ;; Macros for conversion routines %macro _push_immq_mask 1 push dword %1 push dword %1 %endmacro %macro load_immq 2 _push_immq_mask %2 movq %1, [esp] %endmacro %macro pand_immq 2 _push_immq_mask %2 pand %1, [esp] %endmacro %define CLEANUP_IMMQ_LOADS(num) \ add esp, byte 8 * num %define mmx32_rgb888_mask 00ffffffh %define mmx32_rgb565_b 000000f8h %define mmx32_rgb565_g 0000fc00h %define mmx32_rgb565_r 00f80000h %define mmx32_rgb555_rb 00f800f8h %define mmx32_rgb555_g 0000f800h %define mmx32_rgb555_mul 20000008h %define mmx32_bgr555_mul 00082000h SECTION .text _ConvertMMXpII32_24RGB888: ; set up mm6 as the mask, mm7 as zero load_immq mm6, mmx32_rgb888_mask CLEANUP_IMMQ_LOADS(1) pxor mm7, mm7 mov edx, ecx ; save ecx and ecx, 0fffffffch ; clear lower two bits jnz .L1 jmp .L2 .L1: movq mm0, [esi] ; A R G B a r g b pand mm0, mm6 ; 0 R G B 0 r g b movq mm1, [esi+8] ; A R G B a r g b pand mm1, mm6 ; 0 R G B 0 r g b movq mm2, mm0 ; 0 R G B 0 r g b punpckhdq mm2, mm7 ; 0 0 0 0 0 R G B punpckldq mm0, mm7 ; 0 0 0 0 0 r g b psllq mm2, 24 ; 0 0 R G B 0 0 0 por mm0, mm2 ; 0 0 R G B r g b movq mm3, mm1 ; 0 R G B 0 r g b psllq mm3, 48 ; g b 0 0 0 0 0 0 por mm0, mm3 ; g b R G B r g b movq mm4, mm1 ; 0 R G B 0 r g b punpckhdq mm4, mm7 ; 0 0 0 0 0 R G B punpckldq mm1, mm7 ; 0 0 0 0 0 r g b psrlq mm1, 16 ; 0 0 0 R G B 0 r psllq mm4, 8 ; 0 0 0 0 R G B 0 por mm1, mm4 ; 0 0 0 0 R G B r movq [edi], mm0 add esi, BYTE 16 movd [edi+8], mm1 add edi, BYTE 12 sub ecx, BYTE 4 jnz .L1 .L2: mov ecx, edx and ecx, BYTE 3 jz .L4 .L3: mov al, [esi] mov bl, [esi+1] mov dl, [esi+2] mov [edi], al mov [edi+1], bl mov [edi+2], dl add esi, BYTE 4 add edi, BYTE 3 dec ecx jnz .L3 .L4: jmp _mmxreturn _ConvertMMXpII32_16RGB565: ; set up masks load_immq mm5, mmx32_rgb565_b load_immq mm6, mmx32_rgb565_g load_immq mm7, mmx32_rgb565_r CLEANUP_IMMQ_LOADS(3) mov edx, ecx shr ecx, 2 jnz .L1 jmp .L2 ; not necessary at the moment, but doesn't hurt (much) .L1: movq mm0, [esi] ; argb movq mm1, mm0 ; argb pand mm0, mm6 ; 00g0 movq mm3, mm1 ; argb pand mm1, mm5 ; 000b pand mm3, mm7 ; 0r00 pslld mm1, 2 ; 0 0 000000bb bbb00000 por mm0, mm1 ; 0 0 ggggggbb bbb00000 psrld mm0, 5 ; 0 0 00000ggg gggbbbbb movq mm4, [esi+8] ; argb movq mm2, mm4 ; argb pand mm4, mm6 ; 00g0 movq mm1, mm2 ; argb pand mm2, mm5 ; 000b pand mm1, mm7 ; 0r00 pslld mm2, 2 ; 0 0 000000bb bbb00000 por mm4, mm2 ; 0 0 ggggggbb bbb00000 psrld mm4, 5 ; 0 0 00000ggg gggbbbbb packuswb mm3, mm1 ; R 0 r 0 packssdw mm0, mm4 ; as above.. ish por mm0, mm3 ; done. movq [edi], mm0 add esi, 16 add edi, 8 dec ecx jnz .L1 .L2: mov ecx, edx and ecx, BYTE 3 jz .L4 .L3: mov al, [esi] mov bh, [esi+1] mov ah, [esi+2] shr al, 3 and eax, 0F81Fh ; BYTE? shr ebx, 5 and ebx, 07E0h ; BYTE? add eax, ebx mov [edi], al mov [edi+1], ah add esi, BYTE 4 add edi, BYTE 2 dec ecx jnz .L3 .L4: jmp _mmxreturn _ConvertMMXpII32_16BGR565: load_immq mm5, mmx32_rgb565_r load_immq mm6, mmx32_rgb565_g load_immq mm7, mmx32_rgb565_b CLEANUP_IMMQ_LOADS(3) mov edx, ecx shr ecx, 2 jnz .L1 jmp .L2 .L1: movq mm0, [esi] ; a r g b movq mm1, mm0 ; a r g b pand mm0, mm6 ; 0 0 g 0 movq mm3, mm1 ; a r g b pand mm1, mm5 ; 0 r 0 0 pand mm3, mm7 ; 0 0 0 b psllq mm3, 16 ; 0 b 0 0 psrld mm1, 14 ; 0 0 000000rr rrr00000 por mm0, mm1 ; 0 0 ggggggrr rrr00000 psrld mm0, 5 ; 0 0 00000ggg gggrrrrr movq mm4, [esi+8] ; a r g b movq mm2, mm4 ; a r g b pand mm4, mm6 ; 0 0 g 0 movq mm1, mm2 ; a r g b pand mm2, mm5 ; 0 r 0 0 pand mm1, mm7 ; 0 0 0 b psllq mm1, 16 ; 0 b 0 0 psrld mm2, 14 ; 0 0 000000rr rrr00000 por mm4, mm2 ; 0 0 ggggggrr rrr00000 psrld mm4, 5 ; 0 0 00000ggg gggrrrrr packuswb mm3, mm1 ; BBBBB000 00000000 bbbbb000 00000000 packssdw mm0, mm4 ; 00000GGG GGGRRRRR 00000GGG GGGRRRRR por mm0, mm3 ; BBBBBGGG GGGRRRRR bbbbbggg gggrrrrr movq [edi], mm0 add esi, BYTE 16 add edi, BYTE 8 dec ecx jnz .L1 .L2: and edx, BYTE 3 jz .L4 .L3: mov al, [esi+2] mov bh, [esi+1] mov ah, [esi] shr al, 3 and eax, 0F81Fh ; BYTE ? shr ebx, 5 and ebx, 07E0h ; BYTE ? add eax, ebx mov [edi], al mov [edi+1], ah add esi, BYTE 4 add edi, BYTE 2 dec edx jnz .L3 .L4: jmp _mmxreturn _ConvertMMXpII32_16BGR555: ; the 16BGR555 converter is identical to the RGB555 one, ; except it uses a different multiplier for the pmaddwd ; instruction. cool huh. load_immq mm7, mmx32_bgr555_mul jmp _convert_bgr555_cheat ; This is the same as the Intel version.. they obviously went to ; much more trouble to expand/coil the loop than I did, so theirs ; would almost certainly be faster, even if only a little. ; I did rename 'mmx32_rgb555_add' to 'mmx32_rgb555_mul', which is ; (I think) a more accurate name.. _ConvertMMXpII32_16RGB555: load_immq mm7, mmx32_rgb555_mul _convert_bgr555_cheat: load_immq mm6, mmx32_rgb555_g CLEANUP_IMMQ_LOADS(2) mov edx,ecx ; Save ecx and ecx,BYTE 0fffffff8h ; clear lower three bits jnz .L_OK jmp near .L2 .L_OK: movq mm2,[esi+8] movq mm0,[esi] movq mm3,mm2 pand_immq mm3, mmx32_rgb555_rb movq mm1,mm0 pand_immq mm1, mmx32_rgb555_rb pmaddwd mm3,mm7 CLEANUP_IMMQ_LOADS(2) pmaddwd mm1,mm7 pand mm2,mm6 .L1: movq mm4,[esi+24] pand mm0,mm6 movq mm5,[esi+16] por mm3,mm2 psrld mm3,6 por mm1,mm0 movq mm0,mm4 psrld mm1,6 pand_immq mm0, mmx32_rgb555_rb packssdw mm1,mm3 movq mm3,mm5 pmaddwd mm0,mm7 pand_immq mm3, mmx32_rgb555_rb pand mm4,mm6 movq [edi],mm1 pmaddwd mm3,mm7 add esi,BYTE 32 por mm4,mm0 pand mm5,mm6 psrld mm4,6 movq mm2,[esi+8] por mm5,mm3 movq mm0,[esi] psrld mm5,6 movq mm3,mm2 movq mm1,mm0 pand_immq mm3, mmx32_rgb555_rb packssdw mm5,mm4 pand_immq mm1, mmx32_rgb555_rb pand mm2,mm6 CLEANUP_IMMQ_LOADS(4) movq [edi+8],mm5 pmaddwd mm3,mm7 pmaddwd mm1,mm7 add edi,BYTE 16 sub ecx,BYTE 8 jz .L2 jmp .L1 .L2: mov ecx,edx and ecx,BYTE 7 jz .L4 .L3: mov ebx,[esi] add esi,BYTE 4 mov eax,ebx mov edx,ebx shr eax,3 shr edx,6 and eax,BYTE 0000000000011111b and edx, 0000001111100000b shr ebx,9 or eax,edx and ebx, 0111110000000000b or eax,ebx mov [edi],ax add edi,BYTE 2 dec ecx jnz .L3 .L4: jmp _mmxreturn %ifidn __OUTPUT_FORMAT__,elf section .note.GNU-stack noalloc noexec nowrite progbits %endif