Mercurial > mplayer.hg
changeset 2516:9ef4fa15b780
More elegant solution
| author | nick |
|---|---|
| date | Sun, 28 Oct 2001 17:15:54 +0000 |
| parents | 622cd80caaba |
| children | 3d507ef1e3ed |
| files | libvo/mmx.h mmx_defs.h postproc/rgb2rgb.c postproc/rgb2rgb_template.c |
| diffstat | 4 files changed, 66 insertions(+), 780 deletions(-) [+] |
line wrap: on
line diff
--- a/libvo/mmx.h Sun Oct 28 16:57:04 2001 +0000 +++ b/libvo/mmx.h Sun Oct 28 17:15:54 2001 +0000 @@ -1,706 +1,1 @@ -/* mmx.h - - MultiMedia eXtensions GCC interface library for IA32. - - To use this library, simply include this header file - and compile with GCC. You MUST have inlining enabled - in order for mmx_ok() to work; this can be done by - simply using -O on the GCC command line. - - Compiling with -DMMX_TRACE will cause detailed trace - output to be sent to stderr for each mmx operation. - This adds lots of code, and obviously slows execution to - a crawl, but can be very useful for debugging. - - THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY - EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT - LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY - AND FITNESS FOR ANY PARTICULAR PURPOSE. - - 1997-99 by H. Dietz and R. Fisher - - Notes: - It appears that the latest gas has the pand problem fixed, therefore - I'll undefine BROKEN_PAND by default. -*/ - -#ifndef _MMX_H -#define _MMX_H - - -/* Warning: at this writing, the version of GAS packaged - with most Linux distributions does not handle the - parallel AND operation mnemonic correctly. If the - symbol BROKEN_PAND is defined, a slower alternative - coding will be used. If execution of mmxtest results - in an illegal instruction fault, define this symbol. -*/ -#undef BROKEN_PAND - - -/* The type of an value that fits in an MMX register - (note that long long constant values MUST be suffixed - by LL and unsigned long long values by ULL, lest - they be truncated by the compiler) -*/ -typedef union { - long long q; /* Quadword (64-bit) value */ - unsigned long long uq; /* Unsigned Quadword */ - int d[2]; /* 2 Doubleword (32-bit) values */ - unsigned int ud[2]; /* 2 Unsigned Doubleword */ - short w[4]; /* 4 Word (16-bit) values */ - unsigned short uw[4]; /* 4 Unsigned Word */ - char b[8]; /* 8 Byte (8-bit) values */ - unsigned char ub[8]; /* 8 Unsigned Byte */ - float s[2]; /* Single-precision (32-bit) value */ -} __attribute__ ((aligned (8))) mmx_t; /* On an 8-byte (64-bit) boundary */ - - - -/* Function to test if multimedia instructions are supported... -*/ -inline extern int -mm_support(void) -{ - /* Returns 1 if MMX instructions are supported, - 3 if Cyrix MMX and Extended MMX instructions are supported - 5 if AMD MMX and 3DNow! instructions are supported - 0 if hardware does not support any of these - */ - register int rval = 0; - - __asm__ __volatile__ ( - /* See if CPUID instruction is supported ... */ - /* ... Get copies of EFLAGS into eax and ecx */ - "pushf\n\t" - "popl %%eax\n\t" - "movl %%eax, %%ecx\n\t" - - /* ... Toggle the ID bit in one copy and store */ - /* to the EFLAGS reg */ - "xorl $0x200000, %%eax\n\t" - "push %%eax\n\t" - "popf\n\t" - - /* ... Get the (hopefully modified) EFLAGS */ - "pushf\n\t" - "popl %%eax\n\t" - - /* ... Compare and test result */ - "xorl %%eax, %%ecx\n\t" - "testl $0x200000, %%ecx\n\t" - "jz NotSupported1\n\t" /* CPUID not supported */ - - - /* Get standard CPUID information, and - go to a specific vendor section */ - "movl $0, %%eax\n\t" - "cpuid\n\t" - - /* Check for Intel */ - "cmpl $0x756e6547, %%ebx\n\t" - "jne TryAMD\n\t" - "cmpl $0x49656e69, %%edx\n\t" - "jne TryAMD\n\t" - "cmpl $0x6c65746e, %%ecx\n" - "jne TryAMD\n\t" - "jmp Intel\n\t" - - /* Check for AMD */ - "\nTryAMD:\n\t" - "cmpl $0x68747541, %%ebx\n\t" - "jne TryCyrix\n\t" - "cmpl $0x69746e65, %%edx\n\t" - "jne TryCyrix\n\t" - "cmpl $0x444d4163, %%ecx\n" - "jne TryCyrix\n\t" - "jmp AMD\n\t" - - /* Check for Cyrix */ - "\nTryCyrix:\n\t" - "cmpl $0x69727943, %%ebx\n\t" - "jne NotSupported2\n\t" - "cmpl $0x736e4978, %%edx\n\t" - "jne NotSupported3\n\t" - "cmpl $0x64616574, %%ecx\n\t" - "jne NotSupported4\n\t" - /* Drop through to Cyrix... */ - - - /* Cyrix Section */ - /* See if extended CPUID level 80000001 is supported */ - /* The value of CPUID/80000001 for the 6x86MX is undefined - according to the Cyrix CPU Detection Guide (Preliminary - Rev. 1.01 table 1), so we'll check the value of eax for - CPUID/0 to see if standard CPUID level 2 is supported. - According to the table, the only CPU which supports level - 2 is also the only one which supports extended CPUID levels. - */ - "cmpl $0x2, %%eax\n\t" - "jne MMXtest\n\t" /* Use standard CPUID instead */ - - /* Extended CPUID supported (in theory), so get extended - features */ - "movl $0x80000001, %%eax\n\t" - "cpuid\n\t" - "testl $0x00800000, %%eax\n\t" /* Test for MMX */ - "jz NotSupported5\n\t" /* MMX not supported */ - "testl $0x01000000, %%eax\n\t" /* Test for Ext'd MMX */ - "jnz EMMXSupported\n\t" - "movl $1, %0:\n\n\t" /* MMX Supported */ - "jmp Return\n\n" - "EMMXSupported:\n\t" - "movl $3, %0:\n\n\t" /* EMMX and MMX Supported */ - "jmp Return\n\t" - - - /* AMD Section */ - "AMD:\n\t" - - /* See if extended CPUID is supported */ - "movl $0x80000000, %%eax\n\t" - "cpuid\n\t" - "cmpl $0x80000000, %%eax\n\t" - "jl MMXtest\n\t" /* Use standard CPUID instead */ - - /* Extended CPUID supported, so get extended features */ - "movl $0x80000001, %%eax\n\t" - "cpuid\n\t" - "testl $0x00800000, %%edx\n\t" /* Test for MMX */ - "jz NotSupported6\n\t" /* MMX not supported */ - "testl $0x80000000, %%edx\n\t" /* Test for 3DNow! */ - "jnz ThreeDNowSupported\n\t" - "movl $1, %0:\n\n\t" /* MMX Supported */ - "jmp Return\n\n" - "ThreeDNowSupported:\n\t" - "movl $5, %0:\n\n\t" /* 3DNow! and MMX Supported */ - "jmp Return\n\t" - - - /* Intel Section */ - "Intel:\n\t" - - /* Check for MMX */ - "MMXtest:\n\t" - "movl $1, %%eax\n\t" - "cpuid\n\t" - "testl $0x00800000, %%edx\n\t" /* Test for MMX */ - "jz NotSupported7\n\t" /* MMX Not supported */ - "movl $1, %0:\n\n\t" /* MMX Supported */ - "jmp Return\n\t" - - /* Nothing supported */ - "\nNotSupported1:\n\t" - "#movl $101, %0:\n\n\t" - "\nNotSupported2:\n\t" - "#movl $102, %0:\n\n\t" - "\nNotSupported3:\n\t" - "#movl $103, %0:\n\n\t" - "\nNotSupported4:\n\t" - "#movl $104, %0:\n\n\t" - "\nNotSupported5:\n\t" - "#movl $105, %0:\n\n\t" - "\nNotSupported6:\n\t" - "#movl $106, %0:\n\n\t" - "\nNotSupported7:\n\t" - "#movl $107, %0:\n\n\t" - "movl $0, %0:\n\n\t" - - "Return:\n\t" - : "=a" (rval) - : /* no input */ - : "eax", "ebx", "ecx", "edx" - ); - - /* Return */ - return(rval); -} - -/* Function to test if mmx instructions are supported... -*/ -inline extern int -mmx_ok(void) -{ - /* Returns 1 if MMX instructions are supported, 0 otherwise */ - return ( mm_support() & 0x1 ); -} - - -/* Helper functions for the instruction macros that follow... - (note that memory-to-register, m2r, instructions are nearly - as efficient as register-to-register, r2r, instructions; - however, memory-to-memory instructions are really simulated - as a convenience, and are only 1/3 as efficient) -*/ -#ifdef MMX_TRACE - -/* Include the stuff for printing a trace to stderr... -*/ - -#include <stdio.h> - -#define mmx_i2r(op, imm, reg) \ - { \ - mmx_t mmx_trace; \ - mmx_trace.uq = (imm); \ - printf(#op "_i2r(" #imm "=0x%08x%08x, ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ ("movq %%" #reg ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#reg "=0x%08x%08x) => ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ (#op " %0, %%" #reg \ - : /* nothing */ \ - : "X" (imm)); \ - __asm__ __volatile__ ("movq %%" #reg ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#reg "=0x%08x%08x\n", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - } - -#define mmx_m2r(op, mem, reg) \ - { \ - mmx_t mmx_trace; \ - mmx_trace = (mem); \ - printf(#op "_m2r(" #mem "=0x%08x%08x, ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ ("movq %%" #reg ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#reg "=0x%08x%08x) => ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ (#op " %0, %%" #reg \ - : /* nothing */ \ - : "X" (mem)); \ - __asm__ __volatile__ ("movq %%" #reg ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#reg "=0x%08x%08x\n", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - } - -#define mmx_r2m(op, reg, mem) \ - { \ - mmx_t mmx_trace; \ - __asm__ __volatile__ ("movq %%" #reg ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#op "_r2m(" #reg "=0x%08x%08x, ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - mmx_trace = (mem); \ - printf(#mem "=0x%08x%08x) => ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ (#op " %%" #reg ", %0" \ - : "=X" (mem) \ - : /* nothing */ ); \ - mmx_trace = (mem); \ - printf(#mem "=0x%08x%08x\n", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - } - -#define mmx_r2r(op, regs, regd) \ - { \ - mmx_t mmx_trace; \ - __asm__ __volatile__ ("movq %%" #regs ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#op "_r2r(" #regs "=0x%08x%08x, ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ ("movq %%" #regd ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#regd "=0x%08x%08x) => ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ (#op " %" #regs ", %" #regd); \ - __asm__ __volatile__ ("movq %%" #regd ", %0" \ - : "=X" (mmx_trace) \ - : /* nothing */ ); \ - printf(#regd "=0x%08x%08x\n", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - } - -#define mmx_m2m(op, mems, memd) \ - { \ - mmx_t mmx_trace; \ - mmx_trace = (mems); \ - printf(#op "_m2m(" #mems "=0x%08x%08x, ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - mmx_trace = (memd); \ - printf(#memd "=0x%08x%08x) => ", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - __asm__ __volatile__ ("movq %0, %%mm0\n\t" \ - #op " %1, %%mm0\n\t" \ - "movq %%mm0, %0" \ - : "=X" (memd) \ - : "X" (mems)); \ - mmx_trace = (memd); \ - printf(#memd "=0x%08x%08x\n", \ - mmx_trace.d[1], mmx_trace.d[0]); \ - } - -#else - -/* These macros are a lot simpler without the tracing... -*/ - -#define mmx_i2r(op, imm, reg) \ - __asm__ __volatile__ (#op " %0, %%" #reg \ - : /* nothing */ \ - : "X" (imm) ) - -#define mmx_m2r(op, mem, reg) \ - __asm__ __volatile__ (#op " %0, %%" #reg \ - : /* nothing */ \ - : "X" (mem)) - -#define mmx_r2m(op, reg, mem) \ - __asm__ __volatile__ (#op " %%" #reg ", %0" \ - : "=X" (mem) \ - : /* nothing */ ) - -#define mmx_r2r(op, regs, regd) \ - __asm__ __volatile__ (#op " %" #regs ", %" #regd) - -#define mmx_m2m(op, mems, memd) \ - __asm__ __volatile__ ("movq %0, %%mm0\n\t" \ - #op " %1, %%mm0\n\t" \ - "movq %%mm0, %0" \ - : "=X" (memd) \ - : "X" (mems)) - -#endif - - -/* 1x64 MOVe Quadword - (this is both a load and a store... - in fact, it is the only way to store) -*/ -#define movq_m2r(var, reg) mmx_m2r(movq, var, reg) -#define movq_r2m(reg, var) mmx_r2m(movq, reg, var) -#define movq_r2r(regs, regd) mmx_r2r(movq, regs, regd) -#define movq(vars, vard) \ - __asm__ __volatile__ ("movq %1, %%mm0\n\t" \ - "movq %%mm0, %0" \ - : "=X" (vard) \ - : "X" (vars)) - - -/* 1x32 MOVe Doubleword - (like movq, this is both load and store... - but is most useful for moving things between - mmx registers and ordinary registers) -*/ -#define movd_m2r(var, reg) mmx_m2r(movd, var, reg) -#define movd_r2m(reg, var) mmx_r2m(movd, reg, var) -#define movd_r2r(regs, regd) mmx_r2r(movd, regs, regd) -#define movd(vars, vard) \ - __asm__ __volatile__ ("movd %1, %%mm0\n\t" \ - "movd %%mm0, %0" \ - : "=X" (vard) \ - : "X" (vars)) - - -/* 2x32, 4x16, and 8x8 Parallel ADDs -*/ -#define paddd_m2r(var, reg) mmx_m2r(paddd, var, reg) -#define paddd_r2r(regs, regd) mmx_r2r(paddd, regs, regd) -#define paddd(vars, vard) mmx_m2m(paddd, vars, vard) - -#define paddw_m2r(var, reg) mmx_m2r(paddw, var, reg) -#define paddw_r2r(regs, regd) mmx_r2r(paddw, regs, regd) -#define paddw(vars, vard) mmx_m2m(paddw, vars, vard) - -#define paddb_m2r(var, reg) mmx_m2r(paddb, var, reg) -#define paddb_r2r(regs, regd) mmx_r2r(paddb, regs, regd) -#define paddb(vars, vard) mmx_m2m(paddb, vars, vard) - - -/* 4x16 and 8x8 Parallel ADDs using Saturation arithmetic -*/ -#define paddsw_m2r(var, reg) mmx_m2r(paddsw, var, reg) -#define paddsw_r2r(regs, regd) mmx_r2r(paddsw, regs, regd) -#define paddsw(vars, vard) mmx_m2m(paddsw, vars, vard) - -#define paddsb_m2r(var, reg) mmx_m2r(paddsb, var, reg) -#define paddsb_r2r(regs, regd) mmx_r2r(paddsb, regs, regd) -#define paddsb(vars, vard) mmx_m2m(paddsb, vars, vard) - - -/* 4x16 and 8x8 Parallel ADDs using Unsigned Saturation arithmetic -*/ -#define paddusw_m2r(var, reg) mmx_m2r(paddusw, var, reg) -#define paddusw_r2r(regs, regd) mmx_r2r(paddusw, regs, regd) -#define paddusw(vars, vard) mmx_m2m(paddusw, vars, vard) - -#define paddusb_m2r(var, reg) mmx_m2r(paddusb, var, reg) -#define paddusb_r2r(regs, regd) mmx_r2r(paddusb, regs, regd) -#define paddusb(vars, vard) mmx_m2m(paddusb, vars, vard) - - -/* 2x32, 4x16, and 8x8 Parallel SUBs -*/ -#define psubd_m2r(var, reg) mmx_m2r(psubd, var, reg) -#define psubd_r2r(regs, regd) mmx_r2r(psubd, regs, regd) -#define psubd(vars, vard) mmx_m2m(psubd, vars, vard) - -#define psubw_m2r(var, reg) mmx_m2r(psubw, var, reg) -#define psubw_r2r(regs, regd) mmx_r2r(psubw, regs, regd) -#define psubw(vars, vard) mmx_m2m(psubw, vars, vard) - -#define psubb_m2r(var, reg) mmx_m2r(psubb, var, reg) -#define psubb_r2r(regs, regd) mmx_r2r(psubb, regs, regd) -#define psubb(vars, vard) mmx_m2m(psubb, vars, vard) - - -/* 4x16 and 8x8 Parallel SUBs using Saturation arithmetic -*/ -#define psubsw_m2r(var, reg) mmx_m2r(psubsw, var, reg) -#define psubsw_r2r(regs, regd) mmx_r2r(psubsw, regs, regd) -#define psubsw(vars, vard) mmx_m2m(psubsw, vars, vard) - -#define psubsb_m2r(var, reg) mmx_m2r(psubsb, var, reg) -#define psubsb_r2r(regs, regd) mmx_r2r(psubsb, regs, regd) -#define psubsb(vars, vard) mmx_m2m(psubsb, vars, vard) - - -/* 4x16 and 8x8 Parallel SUBs using Unsigned Saturation arithmetic -*/ -#define psubusw_m2r(var, reg) mmx_m2r(psubusw, var, reg) -#define psubusw_r2r(regs, regd) mmx_r2r(psubusw, regs, regd) -#define psubusw(vars, vard) mmx_m2m(psubusw, vars, vard) - -#define psubusb_m2r(var, reg) mmx_m2r(psubusb, var, reg) -#define psubusb_r2r(regs, regd) mmx_r2r(psubusb, regs, regd) -#define psubusb(vars, vard) mmx_m2m(psubusb, vars, vard) - - -/* 4x16 Parallel MULs giving Low 4x16 portions of results -*/ -#define pmullw_m2r(var, reg) mmx_m2r(pmullw, var, reg) -#define pmullw_r2r(regs, regd) mmx_r2r(pmullw, regs, regd) -#define pmullw(vars, vard) mmx_m2m(pmullw, vars, vard) - - -/* 4x16 Parallel MULs giving High 4x16 portions of results -*/ -#define pmulhw_m2r(var, reg) mmx_m2r(pmulhw, var, reg) -#define pmulhw_r2r(regs, regd) mmx_r2r(pmulhw, regs, regd) -#define pmulhw(vars, vard) mmx_m2m(pmulhw, vars, vard) - - -/* 4x16->2x32 Parallel Mul-ADD - (muls like pmullw, then adds adjacent 16-bit fields - in the multiply result to make the final 2x32 result) -*/ -#define pmaddwd_m2r(var, reg) mmx_m2r(pmaddwd, var, reg) -#define pmaddwd_r2r(regs, regd) mmx_r2r(pmaddwd, regs, regd) -#define pmaddwd(vars, vard) mmx_m2m(pmaddwd, vars, vard) - - -/* 1x64 bitwise AND -*/ -#ifdef BROKEN_PAND -#define pand_m2r(var, reg) \ - { \ - mmx_m2r(pandn, (mmx_t) -1LL, reg); \ - mmx_m2r(pandn, var, reg); \ - } -#define pand_r2r(regs, regd) \ - { \ - mmx_m2r(pandn, (mmx_t) -1LL, regd); \ - mmx_r2r(pandn, regs, regd) \ - } -#define pand(vars, vard) \ - { \ - movq_m2r(vard, mm0); \ - mmx_m2r(pandn, (mmx_t) -1LL, mm0); \ - mmx_m2r(pandn, vars, mm0); \ - movq_r2m(mm0, vard); \ - } -#else -#define pand_m2r(var, reg) mmx_m2r(pand, var, reg) -#define pand_r2r(regs, regd) mmx_r2r(pand, regs, regd) -#define pand(vars, vard) mmx_m2m(pand, vars, vard) -#endif - - -/* 1x64 bitwise AND with Not the destination -*/ -#define pandn_m2r(var, reg) mmx_m2r(pandn, var, reg) -#define pandn_r2r(regs, regd) mmx_r2r(pandn, regs, regd) -#define pandn(vars, vard) mmx_m2m(pandn, vars, vard) - - -/* 1x64 bitwise OR -*/ -#define por_m2r(var, reg) mmx_m2r(por, var, reg) -#define por_r2r(regs, regd) mmx_r2r(por, regs, regd) -#define por(vars, vard) mmx_m2m(por, vars, vard) - - -/* 1x64 bitwise eXclusive OR -*/ -#define pxor_m2r(var, reg) mmx_m2r(pxor, var, reg) -#define pxor_r2r(regs, regd) mmx_r2r(pxor, regs, regd) -#define pxor(vars, vard) mmx_m2m(pxor, vars, vard) - - -/* 2x32, 4x16, and 8x8 Parallel CoMPare for EQuality - (resulting fields are either 0 or -1) -*/ -#define pcmpeqd_m2r(var, reg) mmx_m2r(pcmpeqd, var, reg) -#define pcmpeqd_r2r(regs, regd) mmx_r2r(pcmpeqd, regs, regd) -#define pcmpeqd(vars, vard) mmx_m2m(pcmpeqd, vars, vard) - -#define pcmpeqw_m2r(var, reg) mmx_m2r(pcmpeqw, var, reg) -#define pcmpeqw_r2r(regs, regd) mmx_r2r(pcmpeqw, regs, regd) -#define pcmpeqw(vars, vard) mmx_m2m(pcmpeqw, vars, vard) - -#define pcmpeqb_m2r(var, reg) mmx_m2r(pcmpeqb, var, reg) -#define pcmpeqb_r2r(regs, regd) mmx_r2r(pcmpeqb, regs, regd) -#define pcmpeqb(vars, vard) mmx_m2m(pcmpeqb, vars, vard) - - -/* 2x32, 4x16, and 8x8 Parallel CoMPare for Greater Than - (resulting fields are either 0 or -1) -*/ -#define pcmpgtd_m2r(var, reg) mmx_m2r(pcmpgtd, var, reg) -#define pcmpgtd_r2r(regs, regd) mmx_r2r(pcmpgtd, regs, regd) -#define pcmpgtd(vars, vard) mmx_m2m(pcmpgtd, vars, vard) - -#define pcmpgtw_m2r(var, reg) mmx_m2r(pcmpgtw, var, reg) -#define pcmpgtw_r2r(regs, regd) mmx_r2r(pcmpgtw, regs, regd) -#define pcmpgtw(vars, vard) mmx_m2m(pcmpgtw, vars, vard) - -#define pcmpgtb_m2r(var, reg) mmx_m2r(pcmpgtb, var, reg) -#define pcmpgtb_r2r(regs, regd) mmx_r2r(pcmpgtb, regs, regd) -#define pcmpgtb(vars, vard) mmx_m2m(pcmpgtb, vars, vard) - - -/* 1x64, 2x32, and 4x16 Parallel Shift Left Logical -*/ -#define psllq_i2r(imm, reg) mmx_i2r(psllq, imm, reg) -#define psllq_m2r(var, reg) mmx_m2r(psllq, var, reg) -#define psllq_r2r(regs, regd) mmx_r2r(psllq, regs, regd) -#define psllq(vars, vard) mmx_m2m(psllq, vars, vard) - -#define pslld_i2r(imm, reg) mmx_i2r(pslld, imm, reg) -#define pslld_m2r(var, reg) mmx_m2r(pslld, var, reg) -#define pslld_r2r(regs, regd) mmx_r2r(pslld, regs, regd) -#define pslld(vars, vard) mmx_m2m(pslld, vars, vard) - -#define psllw_i2r(imm, reg) mmx_i2r(psllw, imm, reg) -#define psllw_m2r(var, reg) mmx_m2r(psllw, var, reg) -#define psllw_r2r(regs, regd) mmx_r2r(psllw, regs, regd) -#define psllw(vars, vard) mmx_m2m(psllw, vars, vard) - - -/* 1x64, 2x32, and 4x16 Parallel Shift Right Logical -*/ -#define psrlq_i2r(imm, reg) mmx_i2r(psrlq, imm, reg) -#define psrlq_m2r(var, reg) mmx_m2r(psrlq, var, reg) -#define psrlq_r2r(regs, regd) mmx_r2r(psrlq, regs, regd) -#define psrlq(vars, vard) mmx_m2m(psrlq, vars, vard) - -#define psrld_i2r(imm, reg) mmx_i2r(psrld, imm, reg) -#define psrld_m2r(var, reg) mmx_m2r(psrld, var, reg) -#define psrld_r2r(regs, regd) mmx_r2r(psrld, regs, regd) -#define psrld(vars, vard) mmx_m2m(psrld, vars, vard) - -#define psrlw_i2r(imm, reg) mmx_i2r(psrlw, imm, reg) -#define psrlw_m2r(var, reg) mmx_m2r(psrlw, var, reg) -#define psrlw_r2r(regs, regd) mmx_r2r(psrlw, regs, regd) -#define psrlw(vars, vard) mmx_m2m(psrlw, vars, vard) - - -/* 2x32 and 4x16 Parallel Shift Right Arithmetic -*/ -#define psrad_i2r(imm, reg) mmx_i2r(psrad, imm, reg) -#define psrad_m2r(var, reg) mmx_m2r(psrad, var, reg) -#define psrad_r2r(regs, regd) mmx_r2r(psrad, regs, regd) -#define psrad(vars, vard) mmx_m2m(psrad, vars, vard) - -#define psraw_i2r(imm, reg) mmx_i2r(psraw, imm, reg) -#define psraw_m2r(var, reg) mmx_m2r(psraw, var, reg) -#define psraw_r2r(regs, regd) mmx_r2r(psraw, regs, regd) -#define psraw(vars, vard) mmx_m2m(psraw, vars, vard) - - -/* 2x32->4x16 and 4x16->8x8 PACK and Signed Saturate - (packs source and dest fields into dest in that order) -*/ -#define packssdw_m2r(var, reg) mmx_m2r(packssdw, var, reg) -#define packssdw_r2r(regs, regd) mmx_r2r(packssdw, regs, regd) -#define packssdw(vars, vard) mmx_m2m(packssdw, vars, vard) - -#define packsswb_m2r(var, reg) mmx_m2r(packsswb, var, reg) -#define packsswb_r2r(regs, regd) mmx_r2r(packsswb, regs, regd) -#define packsswb(vars, vard) mmx_m2m(packsswb, vars, vard) - - -/* 4x16->8x8 PACK and Unsigned Saturate - (packs source and dest fields into dest in that order) -*/ -#define packuswb_m2r(var, reg) mmx_m2r(packuswb, var, reg) -#define packuswb_r2r(regs, regd) mmx_r2r(packuswb, regs, regd) -#define packuswb(vars, vard) mmx_m2m(packuswb, vars, vard) - - -/* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK Low - (interleaves low half of dest with low half of source - as padding in each result field) -*/ -#define punpckldq_m2r(var, reg) mmx_m2r(punpckldq, var, reg) -#define punpckldq_r2r(regs, regd) mmx_r2r(punpckldq, regs, regd) -#define punpckldq(vars, vard) mmx_m2m(punpckldq, vars, vard) - -#define punpcklwd_m2r(var, reg) mmx_m2r(punpcklwd, var, reg) -#define punpcklwd_r2r(regs, regd) mmx_r2r(punpcklwd, regs, regd) -#define punpcklwd(vars, vard) mmx_m2m(punpcklwd, vars, vard) - -#define punpcklbw_m2r(var, reg) mmx_m2r(punpcklbw, var, reg) -#define punpcklbw_r2r(regs, regd) mmx_r2r(punpcklbw, regs, regd) -#define punpcklbw(vars, vard) mmx_m2m(punpcklbw, vars, vard) - - -/* 2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK High - (interleaves high half of dest with high half of source - as padding in each result field) -*/ -#define punpckhdq_m2r(var, reg) mmx_m2r(punpckhdq, var, reg) -#define punpckhdq_r2r(regs, regd) mmx_r2r(punpckhdq, regs, regd) -#define punpckhdq(vars, vard) mmx_m2m(punpckhdq, vars, vard) - -#define punpckhwd_m2r(var, reg) mmx_m2r(punpckhwd, var, reg) -#define punpckhwd_r2r(regs, regd) mmx_r2r(punpckhwd, regs, regd) -#define punpckhwd(vars, vard) mmx_m2m(punpckhwd, vars, vard) - -#define punpckhbw_m2r(var, reg) mmx_m2r(punpckhbw, var, reg) -#define punpckhbw_r2r(regs, regd) mmx_r2r(punpckhbw, regs, regd) -#define punpckhbw(vars, vard) mmx_m2m(punpckhbw, vars, vard) - - -/* Empty MMx State - (used to clean-up when going from mmx to float use - of the registers that are shared by both; note that - there is no float-to-mmx operation needed, because - only the float tag word info is corruptible) -*/ -#ifdef MMX_TRACE - -#define emms() \ - { \ - printf("emms()\n"); \ - __asm__ __volatile__ ("emms"); \ - } - -#else - -#define emms() __asm__ __volatile__ ("emms") - -#endif - -#endif - +#include "../mmx.h" \ No newline at end of file
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/mmx_defs.h Sun Oct 28 17:15:54 2001 +0000 @@ -0,0 +1,55 @@ +/* + This file contains most common definitions to make MMX stuff + easy portable between different kinds of mmx clones + Written By Nick Kurshev <nickols_k@mail.ru> +*/ + +#ifndef MMXDEFS_INCLUDED +#define MMXDEFS_INCLUDED + +#include "config.h" + +#undef HAVE_MMX1 +#if defined(HAVE_MMX) && !defined(HAVE_MMX2) && !defined(HAVE_3DNOW) && !defined(HAVE_SSE) +#define HAVE_MMX1 +#endif + +#undef HAVE_K6_2PLUS +#if !defined( HAVE_MMX2 ) && defined( HAVE_3DNOW ) +#define HAVE_K6_2PLUS +#endif + +#ifdef HAVE_SSE2 +#define MMREG_SIZE 16 +#else +#define MMREG_SIZE 8 +#endif + +#ifdef HAVE_3DNOW +#define PREFETCH "prefetch" +#define PREFETCHW "prefetchw" +#elif defined ( HAVE_MMX2 ) +#define PREFETCH "prefetchnta" +#define PREFETCHW "prefetcht0" +#else +#define PREFETCH "/nop" +#define PREFETCHW "/nop" +#endif + +#ifdef HAVE_K6_2PLUS +/* On K6 femms is faster of emms. On K7 femms is directly mapped on emms. */ +#define EMMS "femms" +#else +#define EMMS "emms" +#endif + +#ifdef HAVE_MMX2 +#define MOVNTQ "movntq" +#define SFENCE "sfence" +#else +#define MOVNTQ "movq" +#define SFENCE "/nop" +#endif + + +#endif \ No newline at end of file
--- a/postproc/rgb2rgb.c Sun Oct 28 16:57:04 2001 +0000 +++ b/postproc/rgb2rgb.c Sun Oct 28 17:15:54 2001 +0000 @@ -2,30 +2,7 @@ #include "../config.h" #include "rgb2rgb.h" #include "mmx.h" - -#ifdef HAVE_3DNOW -#define PREFETCH "prefetch" -#define PREFETCHW "prefetchw" -#elif defined ( HAVE_MMX2 ) -#define PREFETCH "prefetchnta" -#define PREFETCHW "prefetcht0" -#endif - -#ifdef HAVE_3DNOW -#define EMMS "femms" -#else -#define EMMS "emms" -#endif - -#ifdef HAVE_MMX2 -#define MOVNTQ "movntq" -#else -#define MOVNTQ "movq" -#endif - -#ifdef HAVE_MMX2 -#define SFENCE "sfence" -#endif +#include "../mmx_defs.h" void rgb24to32(uint8_t *src,uint8_t *dst,uint32_t src_size) { @@ -38,21 +15,14 @@ #endif end = s + src_size; #ifdef HAVE_MMX -#ifdef PREFETCH - __asm __volatile( - PREFETCH" %0\n\t" - ::"m"(*s):"memory"); -#endif - mm_end = (uint8_t*)((((unsigned long)end)/16)*16); + __asm __volatile(PREFETCH" %0\n\t"::"m"(*s):"memory"); + mm_end = (uint8_t*)((((unsigned long)end)/(MMREG_SIZE*2))*(MMREG_SIZE*2)); __asm __volatile("movq %0, %%mm7"::"m"(mask32):"memory"); + if(mm_end == end) mm_end -= MMREG_SIZE*2; while(s < mm_end) { -#ifdef PREFETCH __asm __volatile( - PREFETCH" 32%0\n\t" - ::"m"(*s):"memory"); -#endif - __asm __volatile( + PREFETCH" 32%1\n\t" "movd %1, %%mm0\n\t" "movd 3%1, %%mm1\n\t" "movd 6%1, %%mm2\n\t" @@ -69,9 +39,7 @@ dest += 16; s += 12; } -#ifdef SFENCE __asm __volatile(SFENCE:::"memory"); -#endif __asm __volatile(EMMS:::"memory"); #endif while(s < end)
--- a/postproc/rgb2rgb_template.c Sun Oct 28 16:57:04 2001 +0000 +++ b/postproc/rgb2rgb_template.c Sun Oct 28 17:15:54 2001 +0000 @@ -2,30 +2,7 @@ #include "../config.h" #include "rgb2rgb.h" #include "mmx.h" - -#ifdef HAVE_3DNOW -#define PREFETCH "prefetch" -#define PREFETCHW "prefetchw" -#elif defined ( HAVE_MMX2 ) -#define PREFETCH "prefetchnta" -#define PREFETCHW "prefetcht0" -#endif - -#ifdef HAVE_3DNOW -#define EMMS "femms" -#else -#define EMMS "emms" -#endif - -#ifdef HAVE_MMX2 -#define MOVNTQ "movntq" -#else -#define MOVNTQ "movq" -#endif - -#ifdef HAVE_MMX2 -#define SFENCE "sfence" -#endif +#include "../mmx_defs.h" void rgb24to32(uint8_t *src,uint8_t *dst,uint32_t src_size) { @@ -38,21 +15,14 @@ #endif end = s + src_size; #ifdef HAVE_MMX -#ifdef PREFETCH - __asm __volatile( - PREFETCH" %0\n\t" - ::"m"(*s):"memory"); -#endif - mm_end = (uint8_t*)((((unsigned long)end)/16)*16); + __asm __volatile(PREFETCH" %0\n\t"::"m"(*s):"memory"); + mm_end = (uint8_t*)((((unsigned long)end)/(MMREG_SIZE*2))*(MMREG_SIZE*2)); __asm __volatile("movq %0, %%mm7"::"m"(mask32):"memory"); + if(mm_end == end) mm_end -= MMREG_SIZE*2; while(s < mm_end) { -#ifdef PREFETCH __asm __volatile( - PREFETCH" 32%0\n\t" - ::"m"(*s):"memory"); -#endif - __asm __volatile( + PREFETCH" 32%1\n\t" "movd %1, %%mm0\n\t" "movd 3%1, %%mm1\n\t" "movd 6%1, %%mm2\n\t" @@ -69,9 +39,7 @@ dest += 16; s += 12; } -#ifdef SFENCE __asm __volatile(SFENCE:::"memory"); -#endif __asm __volatile(EMMS:::"memory"); #endif while(s < end)