view libmpeg2/libmpeg2_changes.diff @ 31685:31b6397e3b28

Another try at fixing swscale on win64, as per r31153. Don't change paramater passing, but instead use casts. Shouldn't affect asm output on anything other than win64. libswscale should work on win64 now. The rest of ffmpeg still isn't win64 compatible due to the issue of xmm clobbers, but swscale doesn't use any SSE. Patch by Anton Mitrofanov <BugMaster AT narod DOT ru>.

--- libmpeg2/cpu_accel.c	2006-06-16 20:12:26.000000000 +0200
+++ libmpeg2/cpu_accel.c	2006-06-16 20:12:50.000000000 +0200
@@ -29,9 +33,13 @@
 #include "attributes.h"
 #include "mpeg2_internal.h"

+#include "cpudetect.h"
+
-#if defined(ARCH_X86) || defined(ARCH_X86_64)
+#if ARCH_X86 || ARCH_X86_64
 static inline uint32_t arch_accel (uint32_t accel)
 {
+/* Use MPlayer CPU detection instead of libmpeg2 variant. */
+#if 0
     if (accel & (MPEG2_ACCEL_X86_3DNOW | MPEG2_ACCEL_X86_MMXEXT))
 	accel |= MPEG2_ACCEL_X86_MMX;

@@ -124,6 +132,21 @@
 #endif /* ACCEL_DETECT */

     return accel;
+
+#else /* 0 */
+    accel = 0;
+    if (gCpuCaps.hasMMX)
+        accel |= MPEG2_ACCEL_X86_MMX;
+    if (gCpuCaps.hasSSE2)
+	accel |= MPEG2_ACCEL_X86_SSE2;
+    if (gCpuCaps.hasMMX2)
+	accel |= MPEG2_ACCEL_X86_MMXEXT;
+    if (gCpuCaps.has3DNow)
+	accel |= MPEG2_ACCEL_X86_3DNOW;
+
+    return accel;
+
+#endif /* 0 */
 }
 #endif /* ARCH_X86 || ARCH_X86_64 */

@@ -127,7 +150,7 @@
 }
 #endif /* ARCH_X86 || ARCH_X86_64 */

-#if defined(ACCEL_DETECT) && (defined(ARCH_PPC) || defined(ARCH_SPARC))
+#if defined(ACCEL_DETECT) && (ARCH_PPC || ARCH_SPARC)
 #include <signal.h>
 #include <setjmp.h>

@@ -146,7 +169,7 @@
 }
 #endif /* ACCEL_DETECT && (ARCH_PPC || ARCH_SPARC) */

-#ifdef ARCH_PPC
+#if ARCH_PPC
 static uint32_t arch_accel (uint32_t accel)
 {
 #ifdef ACCEL_DETECT
@@ -183,7 +206,7 @@
 }
 #endif /* ARCH_PPC */

-#ifdef ARCH_SPARC
+#if ARCH_SPARC
 static uint32_t arch_accel (uint32_t accel)
 {
     if (accel & MPEG2_ACCEL_SPARC_VIS2)
@@ -229,7 +252,7 @@
 }
 #endif /* ARCH_SPARC */

-#ifdef ARCH_ALPHA
+#if ARCH_ALPHA
 static inline uint32_t arch_accel (uint32_t accel)
 {
     if (accel & MPEG2_ACCEL_ALPHA_MVI)
@@ -253,7 +276,7 @@

 uint32_t mpeg2_detect_accel (uint32_t accel)
 {
-#if defined (ARCH_X86) || defined (ARCH_X86_64) || defined (ARCH_PPC) || defined (ARCH_ALPHA) || defined (ARCH_SPARC)
+#if ARCH_X86 || ARCH_X86_64 || ARCH_PPC || ARCH_ALPHA || ARCH_SPARC
     accel = arch_accel (accel);
 #endif
     return accel;
Index: libmpeg2/cpu_state.c
===================================================================
--- libmpeg2/cpu_state.c	(revision 28324)
+++ libmpeg2/cpu_state.c	(revision 28325)
@@ -29,21 +29,21 @@
 #include "mpeg2.h"
 #include "attributes.h"
 #include "mpeg2_internal.h"
-#if defined(ARCH_X86) || defined(ARCH_X86_64)
+#if ARCH_X86 || ARCH_X86_64
 #include "mmx.h"
 #endif

 void (* mpeg2_cpu_state_save) (cpu_state_t * state) = NULL;
 void (* mpeg2_cpu_state_restore) (cpu_state_t * state) = NULL;

-#if defined(ARCH_X86) || defined(ARCH_X86_64)
+#if ARCH_X86 || ARCH_X86_64
 static void state_restore_mmx (cpu_state_t * state)
 {
     emms ();
 }
 #endif

-#ifdef ARCH_PPC
+#if ARCH_PPC
 #if defined(__APPLE_CC__)	/* apple */
 #define LI(a,b) "li r" #a "," #b "\n\t"
 #define STVX0(a,b,c) "stvx v" #a ",0,r" #c "\n\t"
@@ -115,12 +115,12 @@

 void mpeg2_cpu_state_init (uint32_t accel)
 {
-#if defined(ARCH_X86) || defined(ARCH_X86_64)
+#if ARCH_X86 || ARCH_X86_64
     if (accel & MPEG2_ACCEL_X86_MMX) {
 	mpeg2_cpu_state_restore = state_restore_mmx;
     }
 #endif
-#ifdef ARCH_PPC
+#if ARCH_PPC
     if (accel & MPEG2_ACCEL_PPC_ALTIVEC) {
 	mpeg2_cpu_state_save = state_save_altivec;
 	mpeg2_cpu_state_restore = state_restore_altivec;
--- libmpeg2/header.c	2006-06-16 20:12:26.000000000 +0200
+++ libmpeg2/header.c	2006-06-16 20:12:50.000000000 +0200
@@ -872,6 +879,7 @@
 	mpeg2dec->scaled[idx] = decoder->q_scale_type;
 	for (i = 0; i < 32; i++) {
 	    k = decoder->q_scale_type ? non_linear_scale[i] : (i << 1);
+	    decoder->quantizer_scales[i] = k;
 	    for (j = 0; j < 64; j++)
 		decoder->quantizer_prescale[idx][i][j] =
 		    k * mpeg2dec->quantizer_matrix[idx][j];
--- libmpeg2/idct.c	(revision 26652)
+++ libmpeg2/idct.c	(working copy)
@@ -235,34 +239,40 @@

 void mpeg2_idct_init (uint32_t accel)
 {
-#ifdef ARCH_X86
+#if HAVE_SSE2
     if (accel & MPEG2_ACCEL_X86_SSE2) {
 	mpeg2_idct_copy = mpeg2_idct_copy_sse2;
 	mpeg2_idct_add = mpeg2_idct_add_sse2;
 	mpeg2_idct_mmx_init ();
-    } else if (accel & MPEG2_ACCEL_X86_MMXEXT) {
+    } else
+#elif HAVE_MMX2
+    if (accel & MPEG2_ACCEL_X86_MMXEXT) {
 	mpeg2_idct_copy = mpeg2_idct_copy_mmxext;
 	mpeg2_idct_add = mpeg2_idct_add_mmxext;
 	mpeg2_idct_mmx_init ();
-    } else if (accel & MPEG2_ACCEL_X86_MMX) {
+    } else
+#elif HAVE_MMX
+    if (accel & MPEG2_ACCEL_X86_MMX) {
 	mpeg2_idct_copy = mpeg2_idct_copy_mmx;
 	mpeg2_idct_add = mpeg2_idct_add_mmx;
 	mpeg2_idct_mmx_init ();
     } else
 #endif
-#ifdef ARCH_PPC
+#if HAVE_ALTIVEC
     if (accel & MPEG2_ACCEL_PPC_ALTIVEC) {
 	mpeg2_idct_copy = mpeg2_idct_copy_altivec;
 	mpeg2_idct_add = mpeg2_idct_add_altivec;
 	mpeg2_idct_altivec_init ();
     } else
 #endif
-#ifdef ARCH_ALPHA
+#if HAVE_MVI
     if (accel & MPEG2_ACCEL_ALPHA_MVI) {
 	mpeg2_idct_copy = mpeg2_idct_copy_mvi;
 	mpeg2_idct_add = mpeg2_idct_add_mvi;
 	mpeg2_idct_alpha_init ();
-    } else if (accel & MPEG2_ACCEL_ALPHA) {
+    } else
+#elif ARCH_ALPHA
+    if (accel & MPEG2_ACCEL_ALPHA) {
 	int i;

 	mpeg2_idct_copy = mpeg2_idct_copy_alpha;
Index: libmpeg2/idct_alpha.c
===================================================================
--- libmpeg2/idct_alpha.c	(revision 28324)
+++ libmpeg2/idct_alpha.c	(revision 28325)
@@ -24,7 +24,7 @@

 #include "config.h"

-#ifdef ARCH_ALPHA
+#if ARCH_ALPHA

 #include <stdlib.h>
 #include <inttypes.h>
Index: libmpeg2/idct_altivec.c
===================================================================
--- libmpeg2/idct_altivec.c	(revision 28324)
+++ libmpeg2/idct_altivec.c	(revision 28325)
@@ -23,7 +23,7 @@

 #include "config.h"

-#ifdef ARCH_PPC
+#if ARCH_PPC

 #ifdef HAVE_ALTIVEC_H
 #include <altivec.h>
Index: libmpeg2/idct_mmx.c
===================================================================
--- libmpeg2/idct_mmx.c	(revision 28324)
+++ libmpeg2/idct_mmx.c	(revision 28325)
@@ -23,7 +23,7 @@

 #include "config.h"

-#if defined(ARCH_X86) || defined(ARCH_X86_64)
+#if ARCH_X86 || ARCH_X86_64

 #include <inttypes.h>

--- libmpeg2/motion_comp.c	2006-06-16 20:12:26.000000000 +0200
+++ libmpeg2/motion_comp.c	2006-06-16 20:12:50.000000000 +0200
@@ -33,34 +37,40 @@

 void mpeg2_mc_init (uint32_t accel)
 {
-#ifdef ARCH_X86
+#if HAVE_MMX2
     if (accel & MPEG2_ACCEL_X86_MMXEXT)
 	mpeg2_mc = mpeg2_mc_mmxext;
-    else if (accel & MPEG2_ACCEL_X86_3DNOW)
+    else
+#endif
+#if HAVE_AMD3DNOW
+    if (accel & MPEG2_ACCEL_X86_3DNOW)
 	mpeg2_mc = mpeg2_mc_3dnow;
-    else if (accel & MPEG2_ACCEL_X86_MMX)
+    else
+#endif
+#if HAVE_MMX
+    if (accel & MPEG2_ACCEL_X86_MMX)
 	mpeg2_mc = mpeg2_mc_mmx;
     else
 #endif
-#ifdef ARCH_PPC
+#if HAVE_ALTIVEC
     if (accel & MPEG2_ACCEL_PPC_ALTIVEC)
 	mpeg2_mc = mpeg2_mc_altivec;
     else
 #endif
-#ifdef ARCH_ALPHA
+#if ARCH_ALPHA
     if (accel & MPEG2_ACCEL_ALPHA)
 	mpeg2_mc = mpeg2_mc_alpha;
     else
 #endif
-#ifdef ARCH_SPARC
+#if HAVE_VIS
     if (accel & MPEG2_ACCEL_SPARC_VIS)
 	mpeg2_mc = mpeg2_mc_vis;
     else
 #endif
-#ifdef ARCH_ARM
+#if ARCH_ARM
-    if (accel & MPEG2_ACCEL_ARM) {
+    if (accel & MPEG2_ACCEL_ARM)
 	mpeg2_mc = mpeg2_mc_arm;
-    } else
+    else
 #endif
 	mpeg2_mc = mpeg2_mc_c;
 }
Index: libmpeg2/motion_comp_alpha.c
===================================================================
--- libmpeg2/motion_comp_alpha.c	(revision 28324)
+++ libmpeg2/motion_comp_alpha.c	(revision 28325)
@@ -22,7 +22,7 @@

 #include "config.h"

-#ifdef ARCH_ALPHA
+#if ARCH_ALPHA

 #include <inttypes.h>

Index: libmpeg2/motion_comp_altivec.c
===================================================================
--- libmpeg2/motion_comp_altivec.c	(revision 28324)
+++ libmpeg2/motion_comp_altivec.c	(revision 28325)
@@ -23,7 +23,7 @@

 #include "config.h"

-#ifdef ARCH_PPC
+#if ARCH_PPC

 #ifdef HAVE_ALTIVEC_H
 #include <altivec.h>
Index: libmpeg2/motion_comp_arm.c
===================================================================
--- libmpeg2/motion_comp_arm.c	(revision 28324)
+++ libmpeg2/motion_comp_arm.c	(revision 28325)
@@ -22,7 +22,7 @@

 #include "config.h"

-#ifdef ARCH_ARM
+#if ARCH_ARM

 #include <inttypes.h>

Index: libmpeg2/motion_comp_mmx.c
===================================================================
--- libmpeg2/motion_comp_mmx.c	(revision 28324)
+++ libmpeg2/motion_comp_mmx.c	(revision 28325)
@@ -23,7 +23,7 @@

 #include "config.h"

-#if defined(ARCH_X86) || defined(ARCH_X86_64)
+#if ARCH_X86 || ARCH_X86_64

 #include <inttypes.h>

Index: libmpeg2/motion_comp_vis.c
===================================================================
--- libmpeg2/motion_comp_vis.c	(revision 28324)
+++ libmpeg2/motion_comp_vis.c	(revision 28325)
@@ -22,7 +22,7 @@

 #include "config.h"

-#ifdef ARCH_SPARC
+#if ARCH_SPARC

 #include <inttypes.h>

--- libmpeg2/mpeg2_internal.h	2006-06-16 20:12:26.000000000 +0200
+++ libmpeg2/mpeg2_internal.h	2006-06-16 20:12:50.000000000 +0200
@@ -152,6 +156,11 @@

     /* XXX: stuff due to xine shit */
     int8_t q_scale_type;
+
+    int quantizer_scales[32];
+    int quantizer_scale;
+    char* quant_store;
+    int quant_stride;
 };

 typedef struct {
@@ -226,7 +238,7 @@
 };

 typedef struct {
-#ifdef ARCH_PPC
+#if ARCH_PPC
     uint8_t regv[12*16];
 #endif
     int dummy;
--- libmpeg2/slice.c	2006-06-16 20:12:26.000000000 +0200
+++ libmpeg2/slice.c	2006-06-16 20:12:50.000000000 +0200
@@ -142,6 +146,7 @@

     quantizer_scale_code = UBITS (bit_buf, 5);
     DUMPBITS (bit_buf, bits, 5);
+    decoder->quantizer_scale = decoder->quantizer_scales[quantizer_scale_code];

     decoder->quantizer_matrix[0] =
 	decoder->quantizer_prescale[0][quantizer_scale_code];
@@ -1564,6 +1569,24 @@

 #define NEXT_MACROBLOCK							\
 do {									\
+    if(decoder->quant_store) {						\
+	if (decoder->picture_structure == TOP_FIELD)			\
+	    decoder->quant_store[2 * decoder->quant_stride		\
+				 * (decoder->v_offset >> 4)		\
+				 + (decoder->offset >> 4)]		\
+		= decoder->quantizer_scale;				\
+	else if (decoder->picture_structure == BOTTOM_FIELD)		\
+	    decoder->quant_store[2 * decoder->quant_stride		\
+				 * (decoder->v_offset >> 4)		\
+				 + decoder->quant_stride		\
+				 + (decoder->offset >> 4)]		\
+		= decoder->quantizer_scale;				\
+	else								\
+	    decoder->quant_store[decoder->quant_stride			\
+				 * (decoder->v_offset >> 4)		\
+				 + (decoder->offset >> 4)]		\
+		= decoder->quantizer_scale;				\
+    }									\
     decoder->offset += 16;						\
     if (decoder->offset == decoder->width) {				\
 	do { /* just so we can use the break statement */		\