--- xvid.c	2002/07/09 01:48:49	1.29
+++ xvid.c	2002/09/04 22:07:54	1.35
@@ -3,6 +3,8 @@
  *  XVID MPEG-4 VIDEO CODEC
  *  - Native API implementation  -
  *
+ *  Copyright(C) 2001-2002 Peter Ross <pross@cs.rmit.edu.au>
+ *
  *  This program is an implementation of a part of one or more MPEG-4
  *  Video tools as specified in ISO/IEC 14496-2 standard.  Those intending
  *  to use this software module in hardware or software products are
@@ -28,19 +30,6 @@
  *
  ****************************************************************************/
 
-/*****************************************************************************
- *
- *  History
- *
- *	- 23.06.2002	added XVID_CPU_CHKONLY
- *  - 17.03.2002	Added interpolate8x8_halfpel_hv_xmm
- *  - 22.12.2001  API change: added xvid_init() - Isibaar
- *  - 16.12.2001	inital version; (c)2001 peter ross <pross@cs.rmit.edu.au>
- *
- *  $Id: xvid.c,v 1.29 2002/07/09 01:48:49 chenm001 Exp $
- *
- ****************************************************************************/
-
 #include "xvid.h"
 #include "decoder.h"
 #include "encoder.h"
@@ -52,11 +41,79 @@
 #include "utils/mem_transfer.h"
 #include "quant/quant_h263.h"
 #include "quant/quant_mpeg4.h"
+#include "motion/motion.h"
 #include "motion/sad.h"
 #include "utils/emms.h"
 #include "utils/timer.h"
 #include "bitstream/mbcoding.h"
 
+#if defined(ARCH_X86) && defined(EXPERIMENTAL_SSE2_CODE)
+
+#ifdef WIN32
+#include <windows.h>
+#else
+#include <signal.h>
+#include <setjmp.h>
+#endif
+
+
+#ifndef WIN32
+
+static jmp_buf mark;
+
+static void
+sigill_handler(int signal)
+{
+   longjmp(mark, 1);
+}
+#endif
+
+
+/*
+ * Calls the funcptr, and returns whether SIGILL (illegal instruction) was signalled
+ * Return values:
+ * -1 : could not determine
+ * 0  : SIGILL was *not* signalled
+ * 1  : SIGILL was signalled
+ */
+
+int
+sigill_check(void (*func)())
+{
+#ifdef WIN32
+	_try {
+		func();
+	}
+	_except(EXCEPTION_EXECUTE_HANDLER) {
+
+		if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
+			return 1;
+	}
+	return 0;
+#else
+    void * old_handler;
+    int jmpret;
+
+
+    old_handler = signal(SIGILL, sigill_handler);
+    if (old_handler == SIG_ERR)
+    {
+        return -1;
+    }
+
+    jmpret = setjmp(mark);
+    if (jmpret == 0)
+    {
+        func();
+    }
+
+    signal(SIGILL, old_handler);
+
+    return jmpret;
+#endif
+}
+#endif
+
 /*****************************************************************************
  * XviD Init Entry point
  *
@@ -88,21 +145,33 @@
 	/* Inform the client the core build - unused because we're still alpha */
 	init_param->core_build = 1000;
 
-	if ((init_param->cpu_flags & XVID_CPU_CHKONLY))
-	{
-		init_param->cpu_flags = check_cpu_features();
-		return XVID_ERR_OK;
-	}
-
 	/* Do we have to force CPU features  ? */
-	if ((init_param->cpu_flags & XVID_CPU_FORCE) > 0) {
+	if ((init_param->cpu_flags & XVID_CPU_FORCE)) {
+
 		cpu_flags = init_param->cpu_flags;
+
 	} else {
 
 		cpu_flags = check_cpu_features();
+
+#if defined(ARCH_X86) && defined(EXPERIMENTAL_SSE2_CODE)
+		if ((cpu_flags & XVID_CPU_SSE) && sigill_check(sse_os_trigger))
+			cpu_flags &= ~XVID_CPU_SSE;
+
+		if ((cpu_flags & XVID_CPU_SSE2) && sigill_check(sse2_os_trigger))
+			cpu_flags &= ~XVID_CPU_SSE2;
+#endif
+	}
+
+	if ((init_param->cpu_flags & XVID_CPU_CHKONLY))
+	{
 		init_param->cpu_flags = cpu_flags;
+		return XVID_ERR_OK;
 	}
 
+	init_param->cpu_flags = cpu_flags;
+
+
 	/* Initialize the function pointers */
 	idct_int32_init();
 	init_vlc_tables();
@@ -165,9 +234,12 @@
 	/* Functions used in motion estimation algorithms */
 	calc_cbp = calc_cbp_c;
 	sad16    = sad16_c;
-	sad16bi  = sad16bi_c;
 	sad8     = sad8_c;
+	sad16bi  = sad16bi_c;
+	sad8bi   = sad8bi_c;
 	dev16    = dev16_c;
+	
+	Halfpel8_Refine = Halfpel8_Refine_c;
 
 #ifdef ARCH_X86
 	if ((cpu_flags & XVID_CPU_MMX) > 0) {
@@ -221,10 +293,21 @@
 		calc_cbp = calc_cbp_mmx;
 		sad16    = sad16_mmx;
 		sad8     = sad8_mmx;
+		sad16bi = sad16bi_mmx;
+		sad8bi  = sad8bi_mmx;
 		dev16    = dev16_mmx;
 
 	}
 
+	/* these 3dnow functions are faster than mmx, but slower than xmm. */
+	if ((cpu_flags & XVID_CPU_3DNOW) > 0) {
+
+		/* ME functions */
+		sad16bi = sad16bi_3dn;
+		sad8bi  = sad8bi_3dn;
+	}
+
+
 	if ((cpu_flags & XVID_CPU_MMXEXT) > 0) {
 
 		/* Inverse DCT */
@@ -247,8 +330,9 @@
 
 		/* ME functions */
 		sad16 = sad16_xmm;
-		sad16bi = sad16bi_xmm;
 		sad8  = sad8_xmm;
+		sad16bi = sad16bi_xmm;
+		sad8bi  = sad8bi_xmm;
 		dev16 = dev16_xmm;
 
 	}
@@ -296,6 +380,7 @@
 	  sad16bi = sad16bi_ia64;
 	  sad8 = sad8_ia64;
 	  dev16 = dev16_ia64;
+	  Halfpel8_Refine = Halfpel8_Refine_ia64;
 	  quant_intra = quant_intra_ia64;
 	  dequant_intra = dequant_intra_ia64;
 	  quant_inter = quant_inter_ia64;
@@ -377,12 +462,6 @@
 {
 	switch (opt) {
 	case XVID_ENC_ENCODE:
-#ifdef BFRAMES
-		if (((Encoder *) handle)->mbParam.max_bframes >= 0)
-		return encoder_encode_bframes((Encoder *) handle, (XVID_ENC_FRAME *) param1,
-							  (XVID_ENC_STATS *) param2);
-		else 
-#endif
 		return encoder_encode((Encoder *) handle, (XVID_ENC_FRAME *) param1,
 							  (XVID_ENC_STATS *) param2);