diff -pruN libfame-0.9.1.orig/src/dct_mmx.h libfame-0.9.1/src/dct_mmx.h --- libfame-0.9.1.orig/src/dct_mmx.h 2008-08-02 23:18:19.164140803 +1000 +++ libfame-0.9.1/src/dct_mmx.h 2008-08-02 23:19:00.554142886 +1000 @@ -22,6 +22,9 @@ #define precision +extern FAME_ALIGNED short const _mmx_1[]; +extern FAME_ALIGNED short const _mmx_cos[]; + static void inline dct_aan_pass(dct_t *cache) { // register unsigned short const *mmx_cos = _mmx_cos; @@ -66,42 +69,42 @@ static void inline dct_aan_pass(dct_t *c #ifdef precision "psllw $0x01, %%mm5\n" /* precision(va0) += 1 bit */ #endif - "paddw " ASMSYM "_mmx_1, %%mm4\n" /* + 1 */ + "paddw (%2), %%mm4\n" /* + 1 */ // "pmulhw 16(%1), %%mm5\n" /* (v14+v16)*COS6 -> mm5 (va0) */ - "pmulhw " ASMSYM "_mmx_cos+16, %%mm5\n" /* (v14+v16)*COS6 -> mm5 (va0) */ + "pmulhw 16(%3), %%mm5\n" /* (v14+v16)*COS6 -> mm5 (va0) */ "" /* STEP 4 */ #ifdef precision "psllw $0x02, %%mm6\n" /* precision(v22) += 1 bit */ #else "psllw $0x01, %%mm6\n" /* */ #endif - "paddw " ASMSYM "_mmx_1, %%mm4\n" /* + 1 */ + "paddw (%2), %%mm4\n" /* + 1 */ // "pmulhw 8(%1), %%mm6\n" /* 2*v22*COS4/2 -> mm6 (v32)*/ - "pmulhw " ASMSYM "_mmx_cos+8, %%mm6\n" /* 2*v22*COS4/2 -> mm6 (v32)*/ + "pmulhw 8(%3), %%mm6\n" /* 2*v22*COS4/2 -> mm6 (v32)*/ #ifdef precision "psllw $0x02, %%mm2\n" /* precision(v15) += 1 bit */ #else "psllw $0x01, %%mm2\n" /* */ #endif - "paddw " ASMSYM "_mmx_1, %%mm4\n" /* + 1 */ + "paddw (%2), %%mm4\n" /* + 1 */ // "pmulhw 8(%1), %%mm2\n" /* 2*v15*COS4/2 -> mm2 (v35) */ - "pmulhw " ASMSYM "_mmx_cos+8, %%mm2\n" /* 2*v15*COS4/2 -> mm2 (v35) */ + "pmulhw 8(%3), %%mm2\n" /* 2*v15*COS4/2 -> mm2 (v35) */ #ifdef precision "psllw $0x02, %%mm4\n" /* precision(v14) += 1 bit */ #else "psllw $0x01, %%mm4\n" /* */ #endif - "paddw " ASMSYM "_mmx_1, %%mm4\n" /* + 1 */ + "paddw (%2), %%mm4\n" /* + 1 */ // "pmulhw 0(%1), %%mm4\n" /* 2 * v14 * -COS2/2 -> mm4 */ - "pmulhw " ASMSYM "_mmx_cos, %%mm4\n" /* 2 * v14 * -COS2/2 -> mm4 */ + "pmulhw (%3), %%mm4\n" /* 2 * v14 * -COS2/2 -> mm4 */ "psubsw %%mm5, %%mm4\n" /* v14*-COS2 - va0 -> mm4 (v34) */ #ifdef precision "psllw $0x01, %%mm1\n" /* precision(v16) += 1 bit */ #endif "psubsw %%mm1, %%mm5\n" /* va0 - v16 -> mm5 */ - "paddw " ASMSYM "_mmx_1, %%mm4\n" /* + 1 */ + "paddw (%2), %%mm4\n" /* + 1 */ // "pmulhw 24(%1), %%mm1\n" /* v16 * (COS8 - 1) -> mm1 */ - "pmulhw " ASMSYM "_mmx_cos+24, %%mm1\n" /* v16 * (COS8 - 1) -> mm1 */ + "pmulhw 24(%3), %%mm1\n" /* v16 * (COS8 - 1) -> mm1 */ "psubsw %%mm5, %%mm1\n" /* v16 * COS8 - va0 -> mm1 (v36)*/ "" /* STEP 5 */ "movq 0x70(%0), %%mm0\n" /* retrieve v07 -> mm0 */ @@ -138,8 +141,8 @@ static void inline dct_aan_pass(dct_t *c "movq %%mm0, 0x30(%0)\n" /* store line 3 */ "movq %%mm4, 0x50(%0)\n" /* store line 5 */ "movq %%mm2, 0x70(%0)\n" /* store line 7 */ - : "=r"(cache)/*, "=r"(mmx_cos)*/ - : "0"(cache)/*, "1"(mmx_cos)*/ + : "=r"(cache) + : "0"(cache), "r"(_mmx_1), "r"(_mmx_cos) : "memory"); } diff -pruN libfame-0.9.1.orig/src/dequantize_mmx.h libfame-0.9.1/src/dequantize_mmx.h --- libfame-0.9.1.orig/src/dequantize_mmx.h 2008-08-02 23:18:19.164140803 +1000 +++ libfame-0.9.1/src/dequantize_mmx.h 2008-08-02 23:19:00.555141217 +1000 @@ -27,8 +27,8 @@ "pmullw 0x" #x "8(%3), %%mm5\n" /* premultiply for iDCT */ \ "psrlw $0x0b, %%mm4\n" /* keep 5 bits */ \ "psrlw $0x0b, %%mm5\n" /* keep 5 bits */ \ - "paddw " ASMSYM "_mmx_1, %%mm4\n" /* + 1 */ \ - "paddw " ASMSYM "_mmx_1, %%mm5\n" /* + 1 */ \ + "paddw (%8), %%mm4\n" /* + 1 */ \ + "paddw (%8), %%mm5\n" /* + 1 */ \ "psrlw $0x01, %%mm4\n" /* keep 4 bits rounded */ \ "psrlw $0x01, %%mm5\n" /* keep 4 bits rounded */ \ "psllw $0x04, %%mm0\n" /* multiply by 16 for iDCT */ \ @@ -107,7 +107,7 @@ static void inline dequantize_intra_glob DEQUANTIZE_GLOBAL_MISMATCH_CONTROL() DEQUANTIZE_PRESCALE_STEP(7) : "=r"(block), "=r"(dqmatrix), "=r"(cache), "=r"(psmatrix) - : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix) + : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix), "r"(_mmx_1) : "memory"); asm volatile("movd %%mm6, %0\n" /* export mismatch */ @@ -160,8 +160,8 @@ static void inline dequantize_intra_loca "pcmpeqw %%mm7, %%mm3\n" /* invert sign */ \ "paddw %%mm2, %%mm0\n" /* sub 1 if >0 */ \ "paddw %%mm3, %%mm1\n" /* sub 1 if >0 */ \ - "por " ASMSYM "_mmx_1, %%mm0\n" /* or 1 */ \ - "por " ASMSYM "_mmx_1, %%mm1\n" /* or 1 */ \ + "por (%8), %%mm0\n" /* or 1 */ \ + "por (%8), %%mm1\n" /* or 1 */ \ "pand %%mm4, %%mm0\n" /* [0-3]=0 if [0-3] was zero */ \ "pand %%mm5, %%mm1\n" /* [4-7]=0 if [4-7] was zero */ @@ -184,7 +184,7 @@ static void inline dequantize_intra_loca DEQUANTIZE_INTRA_LOCAL_STEP(7) DEQUANTIZE_PRESCALE_STEP(7) : "=r"(block), "=r"(dqmatrix), "=r"(cache), "=r"(psmatrix) - : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix) + : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix), "r"(_mmx_1) : "memory"); } @@ -256,7 +256,7 @@ static void inline dequantize_inter_glob /* resetting the accumulator when the block is coded intra */ DEQUANTIZE_PRESCALE_STEP(7) : "=r"(block), "=r"(dqmatrix), "=r"(cache), "=r"(psmatrix) - : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix) + : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix), "r"(_mmx_1) : "memory"); asm volatile("movd %%mm6, %0\n" /* export mismatch */ @@ -324,8 +324,8 @@ static void inline dequantize_inter_loca "pcmpeqw %%mm7, %%mm3\n" /* invert sign */ \ "paddw %%mm2, %%mm0\n" /* sub 1 if >0 */ \ "paddw %%mm3, %%mm1\n" /* sub 1 if >0 */ \ - "por " ASMSYM "_mmx_1, %%mm0\n" /* or 1 */ \ - "por " ASMSYM "_mmx_1, %%mm1\n" /* or 1 */ \ + "por (%8), %%mm0\n" /* or 1 */ \ + "por (%8), %%mm1\n" /* or 1 */ \ "pand %%mm4, %%mm0\n" /* [0-3]=0 if [0-3] was zero */ \ "pand %%mm5, %%mm1\n" /* [4-7]=0 if [4-7] was zero */ @@ -348,6 +348,6 @@ static void inline dequantize_inter_loca DEQUANTIZE_INTER_LOCAL_STEP(7) DEQUANTIZE_PRESCALE_STEP(7) : "=r"(block), "=r"(dqmatrix), "=r"(cache), "=r"(psmatrix) - : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix) + : "0"(block), "1"(dqmatrix), "2"(cache), "3"(psmatrix), "r"(_mmx_1) : "memory"); } diff -pruN libfame-0.9.1.orig/src/fame_syntax_mpeg1.c libfame-0.9.1/src/fame_syntax_mpeg1.c --- libfame-0.9.1.orig/src/fame_syntax_mpeg1.c 2008-08-02 23:18:19.167141250 +1000 +++ libfame-0.9.1/src/fame_syntax_mpeg1.c 2008-08-02 23:19:00.556140239 +1000 @@ -469,89 +469,6 @@ static void mpeg1_block_intra(fame_synta fast_bitbuffer_write(data, shift, table[v+255].code, table[v+255].length); /* encode AC coefficients */ -#if defined(HAS_BSWAP) - { - unsigned long dummy1, dummy2; - - /* Note: - movsx mpeg1_table_clip+4096(, %%eax ,2), %%eax - has been replaced by - movw mpeg1_table_clip+4096(, %%eax ,2), %%ax - movsx %%ax, %%eax - because the first instruction failed on a PIII!! (wrong sign extension) - whereas it worked well on my P75 :) - */ - /* Ok, a bit of explanations for a couple of tricks: - The DC value of block is already coded and stored in v so we can use it to store something. - We add one index to the zigzag table so that after coding block[63] we go to index 0. There - we need to escape the zero counting loop (1), what we ensure by putting a non-zero value in - the DC coefficient. Then we can test for index == 0 to exit. - Now this non-zero value is a bit special :) - In order to have one more 'half' register, we store sp value (16 less significant bit of the - 32 bit register esp) *plus one* in the DC coefficient. Since the stack is aligned at an - address multiple of 4 bytes (at least), we are sure that sp != 0xffff and thus sp+1 will - never be zero. We then retrieve sp at the end for it is needed by 'pop' instructions. - */ - /* TODO : echange the role of edx and esp */ - __asm__ __volatile__ ("pushl %%ebx\n" /* save ebx */ - "pushl %%ebp\n" /* save stack pointer */ - "inc %%sp\n" /* make sure sp != 0 */ - "movw %%sp, (%%edx)\n" /* store sp+1 in DC ;) */ - "movl %%esi, %%ebp\n" /* ebp = vlc_table */ - "xorl %%eax, %%eax\n" /* eax = 0 */ - "movl $" ASMSYM "mpeg1_zigzag_table+1, %%esi\n" /*esi = zigzag*/ - "lea 1(%%esi), %%ebx\n" /* ebx = zigzag_table+1*/ - "neg %%ebx\n" /* ebx = -(esi+1) */ - ".p2align 4,,7\n" /* align for jump */ - "0: xorw %%sp, %%sp\n" /* sp = 0 */ - "1: movb (%%esi), %%al\n" /* eax = index in block*/ - "incl %%esi\n" /* (faster than lodsb) */ - "addw (%%edx, %%eax, 2), %%sp\n" /* sp = unzig */ - "jz 1b\n" /* coeff == 0 then loop*/ - "orl %%eax, %%eax\n" /* index == 0 then quit*/ - "jz 2f\n" /* (faster than jcxz) */ - "movsx %%sp, %%eax\n" /* extend sign */ - "movw " ASMSYM "mpeg1_table_clip_data+4096(, %%eax ,2), %%ax\n" /*clip*/ - "movsx %%ax, %%eax\n" /* extend sign */ - "addl %%esi, %%ebx\n" /* ebx = run */ - "shll $7, %%eax\n" /* eax *= 128(indexing)*/ - "lea (%%eax, %%ebx, 2), %%eax\n" /*eax = 2 * offset*/ - "lea (%%ebp, %%eax, 4), %%ebx\n" /* ebx = &vlc */ - "movl (%%ebx), %%eax\n" /* eax = code */ - "addl 4(%%ebx), %%ecx\n" /* ecx = shift+=length */ - "xorl %%ebx, %%ebx\n" /* ebx = 0 */ - "shrd %%cl, %%eax, %%ebx\n" /* adjust code to fit */ - "shr %%cl, %%eax\n" /* adjust code to fit */ - "bswap %%eax\n" /* reverse byte order of code */ - "bswap %%ebx\n" /* reverse byte order of code */ - "or %%eax, (%%edi)\n" /* put first 32 bits */ - "movl %%ecx, %%eax\n" /* eax = shift + length*/ - "shrl $5, %%eax\n" /* get dword increment */ - "andl $31, %%ecx\n" /* mask shift */ - "lea (%%edi, %%eax, 4), %%edi\n"/* data+=(ecx>32)*/ - "orl %%ebx, (%%edi)\n" /* put last 32 bits */ - "xorl %%eax, %%eax\n" /* eax = 0 */ - "lea 1(%%esi), %%ebx\n" /* ebx = esi + 1 (last)*/ - "neg %%ebx\n" /* ebx = -(esi + 1) */ - "jmp 0b\n" /* loop */ - "2:\n" - "movw (%%edx), %%sp\n" /* retrieve sp+1 */ - "dec %%sp\n" /* restore esp */ - "popl %%ebp\n" /* reload stack pointer*/ - "popl %%ebx\n" /* reload ebx */ - : "=c"(shift), - "=a"(dummy1), - "=d"(block), - "=D"(data), - "=S"(dummy2) - : "d"(block), - "c"(shift), - "D"(data), - "S"(syntax_mpeg1->vlc_table) - : "memory"); - block[0] = v; /* restore DC value */ - } -#else { short i; unsigned long last; @@ -573,7 +490,6 @@ static void mpeg1_block_intra(fame_synta } } } -#endif /* HAS_BSWAP */ /* mark end of block */ fast_bitbuffer_write(data, shift, 2, 2); diff -pruN libfame-0.9.1.orig/src/half_mmx.h libfame-0.9.1/src/half_mmx.h --- libfame-0.9.1.orig/src/half_mmx.h 2008-08-02 23:18:19.161140026 +1000 +++ libfame-0.9.1/src/half_mmx.h 2008-08-02 23:19:00.556140239 +1000 @@ -68,8 +68,8 @@ static void inline mmx_interpolate(unsig "paddw %%mm5, %%mm6\n" /* mm6 = ref00+ref10+ref11+1-r 4-7*/ "psrlw $1, %%mm4\n" /* divide by 2 */ "psrlw $1, %%mm5\n" /* divide by 2 */ - "paddw " ASMSYM "_mmx_one, %%mm3\n" /* add 1 */ - "paddw " ASMSYM "_mmx_one, %%mm6\n" /* add 1 */ + "paddw (%8), %%mm3\n" /* add 1 */ + "paddw (%8), %%mm6\n" /* add 1 */ "packuswb %%mm5, %%mm4\n" /* pack to byte and saturate */ "movq 1(%3), %%mm1\n" /* mm1 = [ref+1] */ "movq %%mm1, %%mm2\n" /* mm2 = mm1 */ @@ -87,7 +87,7 @@ static void inline mmx_interpolate(unsig "movl 12(%0), %3\n" /* %3 = ref[3] */ "movq %%mm3, (%3)\n" /* store in frame */ : "=r"(ref), "=r"(pitch), "=r"(rc), "=r"(dummy) - : "0"(ref), "1"(pitch), "2"(rc), "3"(dummy) + : "0"(ref), "1"(pitch), "2"(rc), "3"(dummy), "r"(_mmx_one) : "memory"); } diff -pruN libfame-0.9.1.orig/src/half_sse.h libfame-0.9.1/src/half_sse.h --- libfame-0.9.1.orig/src/half_sse.h 2008-08-02 23:18:19.160143502 +1000 +++ libfame-0.9.1/src/half_sse.h 2008-08-02 23:19:00.557139019 +1000 @@ -71,8 +71,8 @@ static void inline mmx_interpolate_signe "paddw %%mm5, %%mm6\n" /* mm6 = ref00+ref10+ref11+1-r 4-7*/ "psrlw $1, %%mm4\n" /* divide by 2 */ "psrlw $1, %%mm5\n" /* divide by 2 */ - "paddw " ASMSYM "_mmx_one, %%mm3\n" /* add 1 */ - "paddw " ASMSYM "_mmx_one, %%mm6\n" /* add 1 */ + "paddw (%8), %%mm3\n" /* add 1 */ + "paddw (%8), %%mm6\n" /* add 1 */ "packuswb %%mm5, %%mm4\n" /* pack to byte and saturate */ "movq 1(%3), %%mm1\n" /* mm1 = [ref+1] */ "movq %%mm1, %%mm2\n" /* mm2 = mm1 */ @@ -90,7 +90,7 @@ static void inline mmx_interpolate_signe "movl 12(%0), %3\n" /* %3 = ref[3] */ "movq %%mm3, (%3)\n" /* store in frame */ : "=r"(ref), "=r"(pitch), "=r"(rc), "=r"(dummy) - : "0"(ref), "1"(pitch), "2"(rc), "3"(dummy) + : "0"(ref), "1"(pitch), "2"(rc), "3"(dummy), "r"(_mmx_one) : "memory"); } diff -pruN libfame-0.9.1.orig/src/idct_mmx.h libfame-0.9.1/src/idct_mmx.h --- libfame-0.9.1.orig/src/idct_mmx.h 2008-08-02 23:18:19.160143502 +1000 +++ libfame-0.9.1/src/idct_mmx.h 2008-08-02 23:19:00.557139019 +1000 @@ -18,6 +18,10 @@ */ /*************************** MMX accelerated iDCT ****************************/ +extern FAME_ALIGNED short const _mmx_1[]; +extern FAME_ALIGNED short const _mmx_cos[]; +extern FAME_ALIGNED short const _mmx_icos[]; + static void inline idct_aan_pass(dct_t * block) { // register unsigned short const *mmx_icos = _mmx_icos; @@ -65,9 +69,9 @@ static void inline idct_aan_pass(dct_t * block[row*8+6] = v45; - v71, v11, v44, v65, v24 - */ "psllw $0x02, %%mm5\n" /* adjust v22 for multiply */ - "paddw " ASMSYM "_mmx_1, %%mm5\n" /* + 1 for rounding */ + "paddw (%2), %%mm5\n" /* + 1 for rounding */ // "pmulhw 8(%1), %%mm5\n" /* 4*v15*ICOS4/4 -> mm5 (v23) */ - "pmulhw " ASMSYM "_mmx_icos+8, %%mm5\n" /* 4*v15*ICOS4/4 -> mm5 (v23)*/ + "pmulhw 8(%3), %%mm5\n" /* 4*v15*ICOS4/4 -> mm5 (v23)*/ "psubsw %%mm4, %%mm5\n" /* v23 - v62 -> mm5 (v24) */ "movq %%mm3, %%mm6\n" /* v44 -> mm6 */ "paddsw %%mm5, %%mm6\n" /* v44 + v24 -> mm6 (v45) */ @@ -125,25 +129,25 @@ static void inline idct_aan_pass(dct_t * block[row*8+4] += v55; - - */ "psllw $0x02, %%mm0\n" /* adjust v12 for multiply */ - "paddw " ASMSYM "_mmx_1, %%mm0\n" /* + 1 for rounding */ + "paddw (%2), %%mm0\n" /* + 1 for rounding */ // "pmulhw 8(%1), %%mm0\n" /* 4*v12*ICOS4/4 -> mm0 (v13) */ - "pmulhw " ASMSYM "_mmx_icos+8, %%mm0\n" /* 4*v12*ICOS4/4 -> mm0 (v13) */ + "pmulhw 8(%3), %%mm0\n" /* 4*v12*ICOS4/4 -> mm0 (v13) */ "movq %%mm2, %%mm6\n" /* v51 -> mm6 */ "psubsw %%mm1, %%mm6\n" /* v51 - v71 -> mm6 (va2) */ "psllw $0x03, %%mm2\n" /* adjust v51 for multiply */ - "paddw " ASMSYM "_mmx_1, %%mm2\n" /* + 1 for rounding */ + "paddw (%2), %%mm2\n" /* + 1 for rounding */ /* should add another one here but it seems to look better without */ // "pmulhw 16(%1), %%mm2\n" /* 8*v51*ICOS6/8 -> mm2 (v53) */ - "pmulhw " ASMSYM "_mmx_icos+16, %%mm2\n" /* 8*v51*ICOS6/8 -> mm2 (v53) */ + "pmulhw 16(%3), %%mm2\n" /* 8*v51*ICOS6/8 -> mm2 (v53) */ "psllw $0x02, %%mm1\n" /* adjust v71 for multiply */ - "paddw " ASMSYM "_mmx_1, %%mm1\n" /* + 1 for rounding */ + "paddw (%2), %%mm1\n" /* + 1 for rounding */ /* should add another one here but it seems to look better without */ // "pmulhw 0(%1), %%mm1\n" /* 4*v71*ICOS2/4 -> mm1 (v73) */ - "pmulhw " ASMSYM "_mmx_icos, %%mm1\n" /* 4*v71*ICOS2/4 -> mm1 (v73) */ + "pmulhw (%3), %%mm1\n" /* 4*v71*ICOS2/4 -> mm1 (v73) */ "psllw $0x01, %%mm6\n" /* adjust va2 for multiply */ - "paddw " ASMSYM "_mmx_1, %%mm6\n" /* + 1 for rounding */ + "paddw (%2), %%mm6\n" /* + 1 for rounding */ // "pmulhw 24(%1), %%mm6\n" /* 2*v12*ICOS8/2 -> mm6 (va3) */ - "pmulhw " ASMSYM "_mmx_icos+24, %%mm6\n" /* 2*v12*ICOS8/2 -> mm6 (va3) */ + "pmulhw 24(%3), %%mm6\n" /* 2*v12*ICOS8/2 -> mm6 (va3) */ "psubsw %%mm6, %%mm2\n" /* v53 - va3 -> mm2 (v54) */ "psubsw %%mm6, %%mm1\n" /* v73 - va3 -> mm1 (v74) */ "psubsw %%mm3, %%mm1\n" /* v74 - v32 -> mm3 (v75) */ @@ -167,8 +171,8 @@ static void inline idct_aan_pass(dct_t * "paddsw %%mm0, %%mm7\n" /* v65 + v55 -> mm7 */ "movq %%mm6, 0x30(%0)\n" /* mm6 -> line 3 */ "movq %%mm7, 0x40(%0)\n" /* mm7 -> line 4 */ - : "=r"(block)/*, "=r"(mmx_icos)*/ - : "0"(block)/*, "1"(mmx_icos)*/ + : "=r"(block) + : "0"(block), "r"(_mmx_1), "r"(_mmx_icos) : "memory"); }