From afb833d4e89c312460a4ab9ed6a7a8ca4ebbfe1c Mon Sep 17 00:00:00 2001 From: jtg Date: Thu, 19 Aug 1999 00:55:39 +0000 Subject: Initial revision --- src/mesa/x86/mmx_blend.S | 363 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 363 insertions(+) create mode 100644 src/mesa/x86/mmx_blend.S (limited to 'src/mesa/x86/mmx_blend.S') diff --git a/src/mesa/x86/mmx_blend.S b/src/mesa/x86/mmx_blend.S new file mode 100644 index 0000000000..7bd9ec3d24 --- /dev/null +++ b/src/mesa/x86/mmx_blend.S @@ -0,0 +1,363 @@ +#include "assyntax.h" + + +#if !defined(NASM_ASSEMBLER) && !defined(MASM_ASSEMBLER) +#define LLBL(a) .L ## a +#else +#define LLBL(a) a +#endif + + +SEG_TEXT + + +ALIGNTEXT16 +GLOBL GLNAME(gl_mmx_blend_transparency) + +GLNAME( gl_mmx_blend_transparency ): + PUSH_L ( EBP ) + MOV_L ( ESP, EBP ) + SUB_L ( CONST(52), ESP ) + PUSH_L ( EBX ) + MOV_L ( CONST(16711680), REGOFF(-8, EBP) ) + MOV_L ( CONST(16711680), REGOFF(-4, EBP) ) + MOV_L ( CONST(0), REGOFF(-16, EBP) ) + MOV_L ( CONST(-1), REGOFF(-12, EBP) ) + MOV_L ( CONST(-1), REGOFF(-24, EBP) ) + MOV_L ( CONST(0), REGOFF(-20, EBP) ) + MOV_L ( REGOFF(24, EBP), EAX ) + ADD_L ( CONST(4), EAX ) + MOV_L ( EAX, EDX ) + AND_L ( REGOFF(20, EBP), EDX ) + MOV_L ( EDX, EAX ) + AND_L ( CONST(4), EAX ) + CMP_L ( CONST(8), EAX ) + JNE ( LLBL(GMBT_2) ) + MOV_L ( REGOFF(20, EBP), EAX ) + ADD_L ( CONST(3), EAX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(EAX), DL ) + MOV_L ( EDX, REGOFF(-32, EBP) ) + MOV_L ( CONST(255), EAX ) + MOV_L ( EAX, EBX ) + SUB_L ( REGOFF(-32, EBP), EBX ) + MOV_L ( EBX, REGOFF(-36, EBP) ) + MOV_L ( REGOFF(20, EBP), EAX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(EAX), DL ) + MOV_L ( EDX, EAX ) + IMUL_L ( REGOFF(-32, EBP), EAX ) + MOV_L ( REGOFF(24, EBP), EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EDX ) + IMUL_L ( REGOFF(-36, EBP), EDX ) + ADD_L ( EDX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-40, EBP) ) + MOV_L ( REGOFF(20, EBP), EAX ) + INC_L ( EAX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(EAX), DL ) + MOV_L ( EDX, EAX ) + IMUL_L ( REGOFF(-32, EBP), EAX ) + MOV_L ( REGOFF(24, EBP), EDX ) + INC_L ( EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EDX ) + IMUL_L ( REGOFF(-36, EBP), EDX ) + ADD_L ( EDX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-44, EBP) ) + MOV_L ( REGOFF(20, EBP), EAX ) + ADD_L ( CONST(2), EAX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(EAX), DL ) + MOV_L ( EDX, EAX ) + IMUL_L ( REGOFF(-32, EBP), EAX ) + MOV_L ( REGOFF(24, EBP), EDX ) + ADD_L ( CONST(2), EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EDX ) + IMUL_L ( REGOFF(-36, EBP), EDX ) + ADD_L ( EDX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-48, EBP) ) + MOV_L ( REGOFF(20, EBP), EAX ) + ADD_L ( CONST(3), EAX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(EAX), DL ) + MOV_L ( EDX, EAX ) + IMUL_L ( REGOFF(-32, EBP), EAX ) + MOV_L ( REGOFF(24, EBP), EDX ) + ADD_L ( CONST(3), EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EDX ) + IMUL_L ( REGOFF(-36, EBP), EDX ) + ADD_L ( EDX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-52, EBP) ) + MOV_L ( REGOFF(20, EBP), EAX ) + MOV_B ( REGOFF(-40, EBP), DL ) + MOV_B ( DL, REGIND(EAX) ) + MOV_L ( REGOFF(20, EBP), EAX ) + INC_L ( EAX ) + MOV_B ( REGOFF(-44, EBP), DL ) + MOV_B ( DL, REGIND(EAX) ) + MOV_L ( REGOFF(20, EBP), EAX ) + ADD_L ( CONST(2), EAX ) + MOV_B ( REGOFF(-48, EBP), DL ) + MOV_B ( DL, REGIND(EAX) ) + MOV_L ( REGOFF(20, EBP), EAX ) + ADD_L ( CONST(3), EAX ) + MOV_B ( REGOFF(-52, EBP), DL ) + MOV_B ( DL, REGIND(EAX) ) + INC_L ( REGOFF(16, EBP) ) + ADD_L ( CONST(4), REGOFF(20, EBP) ) + ADD_L ( CONST(4), REGOFF(24, EBP) ) + DEC_L ( REGOFF(12, EBP) ) +LLBL( GMBT_2 ): + + CMP_L ( CONST(0), REGOFF(12, EBP) ) + JE ( LLBL(GMBT_3) ) + MOV_L ( CONST(0), REGOFF(-28, EBP) ) +ALIGNTEXT4 +LLBL( GMBT_4 ): + + MOV_L ( REGOFF(12, EBP), EDX ) + MOV_L ( EDX, EAX ) + SHR_L ( CONST(1), EAX ) + CMP_L ( EAX, REGOFF(-28, EBP) ) + JB ( LLBL(GMBT_7) ) + JMP ( LLBL(GMBT_5) ) +ALIGNTEXT16 +LLBL( GMBT_7 ): + + MOV_L ( REGOFF(-28, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,2), EDX ) + MOV_L ( REGOFF(16, EBP), EAX ) + CMP_B ( CONST(0), REGBI(EAX,EDX) ) + JE ( LLBL(GMBT_6) ) + MOV_L ( REGOFF(-28, EBP), EAX ) + MOV_L ( EAX, EDX ) + LEA_L ( REGDIS(0,EDX,8), ECX ) + MOV_L ( ECX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + MOV_L ( REGOFF(-28, EBP), EDX ) + MOV_L ( EDX, ECX ) + LEA_L ( REGDIS(0,ECX,8), EDX ) + MOV_L ( EDX, ECX ) + ADD_L ( REGOFF(24, EBP), ECX ) + + MOVQ ( REGIND(EAX), MM4 ) + PXOR ( MM5, MM5 ) + MOVQ ( MM4, MM1 ) + MOVQ ( REGIND(ECX), MM7 ) + PUNPCKLBW ( MM5, MM1 ) + MOVQ ( MM7, MM6 ) + MOVQ ( MM1, MM0 ) + PUNPCKLBW ( MM5, MM6 ) + MOVQ ( MM1, MM2 ) + PSRLQ ( CONST(48), MM0 ) + PUNPCKHBW ( MM5, MM4 ) + PACKSSDW ( MM0, MM0 ) + MOVQ ( MM0, MM3 ) + PUNPCKHBW ( MM5, MM7 ) + PSLLQ ( CONST(16), MM3 ) + POR ( REGOFF(-8, EBP), MM0 ) + PUNPCKLWD ( MM6, MM1 ) + PSUBW ( MM3, MM0 ) + PUNPCKHWD ( MM6, MM2 ) + MOVQ ( MM4, MM3 ) + PSRLQ ( CONST(48), MM3 ) + PACKSSDW ( MM3, MM3 ) + MOVQ ( MM3, MM6 ) + POR ( REGOFF(-8, EBP), MM3 ) + PSLLQ ( CONST(16), MM6 ) + PSUBW ( MM6, MM3 ) + MOVQ ( MM4, MM5 ) + PUNPCKLWD ( MM7, MM4 ) + PUNPCKHWD ( MM7, MM5 ) + PMADDWD ( MM0, MM1 ) + PMADDWD ( MM3, MM4 ) + PMADDWD ( MM0, MM2 ) + PMADDWD ( MM3, MM5 ) + PSRLD ( CONST(8), MM1 ) + PSRLD ( CONST(8), MM2 ) + PSRLD ( CONST(8), MM4 ) + PACKSSDW ( MM2, MM1 ) + PSRLD ( CONST(8), MM5 ) + PACKUSWB ( MM1, MM1 ) + PACKSSDW ( MM5, MM4 ) + PAND ( REGOFF(-24, EBP), MM1 ) + PACKUSWB ( MM4, MM4 ) + PAND ( REGOFF(-16, EBP), MM4 ) + POR ( MM1, MM4 ) + MOVQ ( MM4, REGIND(EAX) ) + + +LLBL( GMBT_8 ): + +LLBL( GMBT_6 ): + + INC_L ( REGOFF(-28, EBP) ) + JMP ( LLBL(GMBT_4) ) +ALIGNTEXT16 +LLBL( GMBT_5 ): + + + EMMS + +LLBL( GMBT_3 ): + + MOV_L ( REGOFF(12, EBP), EAX ) + AND_L ( CONST(1), EAX ) + TEST_L ( EAX, EAX ) + JE ( LLBL(GMBT_9) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-1, EAX), EDX ) + XOR_L ( EAX, EAX ) + MOV_B ( REGIND(EDX), AL ) + MOV_L ( EAX, REGOFF(-52, EBP) ) + MOV_L ( CONST(255), EAX ) + MOV_L ( EAX, EBX ) + SUB_L ( REGOFF(-52, EBP), EBX ) + MOV_L ( EBX, REGOFF(-48, EBP) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-4, EAX), EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EAX ) + IMUL_L ( REGOFF(-52, EBP), EAX ) + MOV_L ( REGOFF(12, EBP), EDX ) + LEA_L ( REGDIS(0,EDX,4), ECX ) + MOV_L ( ECX, EDX ) + ADD_L ( REGOFF(24, EBP), EDX ) + LEA_L ( REGOFF(-4, EDX), ECX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(ECX), DL ) + MOV_L ( EDX, ECX ) + IMUL_L ( REGOFF(-48, EBP), ECX ) + ADD_L ( ECX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-44, EBP) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-3, EAX), EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EAX ) + IMUL_L ( REGOFF(-52, EBP), EAX ) + MOV_L ( REGOFF(12, EBP), EDX ) + LEA_L ( REGDIS(0,EDX,4), ECX ) + MOV_L ( ECX, EDX ) + ADD_L ( REGOFF(24, EBP), EDX ) + LEA_L ( REGOFF(-3, EDX), ECX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(ECX), DL ) + MOV_L ( EDX, ECX ) + IMUL_L ( REGOFF(-48, EBP), ECX ) + ADD_L ( ECX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-40, EBP) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-2, EAX), EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EAX ) + IMUL_L ( REGOFF(-52, EBP), EAX ) + MOV_L ( REGOFF(12, EBP), EDX ) + LEA_L ( REGDIS(0,EDX,4), ECX ) + MOV_L ( ECX, EDX ) + ADD_L ( REGOFF(24, EBP), EDX ) + LEA_L ( REGOFF(-2, EDX), ECX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(ECX), DL ) + MOV_L ( EDX, ECX ) + IMUL_L ( REGOFF(-48, EBP), ECX ) + ADD_L ( ECX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-36, EBP) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-1, EAX), EDX ) + XOR_L ( ECX, ECX ) + MOV_B ( REGIND(EDX), CL ) + MOV_L ( ECX, EAX ) + IMUL_L ( REGOFF(-52, EBP), EAX ) + MOV_L ( REGOFF(12, EBP), EDX ) + LEA_L ( REGDIS(0,EDX,4), ECX ) + MOV_L ( ECX, EDX ) + ADD_L ( REGOFF(24, EBP), EDX ) + LEA_L ( REGOFF(-1, EDX), ECX ) + XOR_L ( EDX, EDX ) + MOV_B ( REGIND(ECX), DL ) + MOV_L ( EDX, ECX ) + IMUL_L ( REGOFF(-48, EBP), ECX ) + ADD_L ( ECX, EAX ) + MOV_L ( EAX, EBX ) + SAR_L ( CONST(8), EBX ) + MOV_L ( EBX, REGOFF(-32, EBP) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-4, EAX), EDX ) + MOV_B ( REGOFF(-44, EBP), AL ) + MOV_B ( AL, REGIND(EDX) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-3, EAX), EDX ) + MOV_B ( REGOFF(-40, EBP), AL ) + MOV_B ( AL, REGIND(EDX) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-2, EAX), EDX ) + MOV_B ( REGOFF(-36, EBP), AL ) + MOV_B ( AL, REGIND(EDX) ) + MOV_L ( REGOFF(12, EBP), EAX ) + LEA_L ( REGDIS(0,EAX,4), EDX ) + MOV_L ( EDX, EAX ) + ADD_L ( REGOFF(20, EBP), EAX ) + LEA_L ( REGOFF(-1, EAX), EDX ) + MOV_B ( REGOFF(-32, EBP), AL ) + MOV_B ( AL, REGIND(EDX) ) +LLBL( GMBT_9 ): + +LLBL( GMBT_1 ): + + MOV_L ( REGOFF(-56, EBP), EBX ) + MOV_L ( EBP, ESP ) + POP_L ( EBP ) + RET + + + + -- cgit v1.2.3