Index: jddctmgr.c
===================================================================
RCS file: /cvsroot/mozilla/jpeg/jddctmgr.c,v
retrieving revision 3.3
diff -u -r3.3 jddctmgr.c
--- jddctmgr.c	29 Apr 2002 23:27:33 -0000	3.3
+++ jddctmgr.c	31 Mar 2005 00:02:24 -0000
@@ -78,6 +78,7 @@
 #endif
 #endif
 
+#ifndef QUEK15
 GLOBAL(void)
 jpeg_idct_islow_sse2 (
 	j_decompress_ptr cinfo, 
@@ -85,7 +86,15 @@
 	JCOEFPTR coef_block,
 	JSAMPARRAY output_buf, 
 	JDIMENSION output_col);
-
+#else
+GLOBAL(void)
+jpeg_idct_ifast_sse2 (
+	j_decompress_ptr cinfo, 
+	jpeg_component_info * compptr,
+	JCOEFPTR coef_block,
+	JSAMPARRAY output_buf, 
+	JDIMENSION output_col);
+#endif
 
 /*
  * Prepare for an output pass.
@@ -128,13 +137,18 @@
 #ifdef HAVE_SSE2_INTEL_MNEMONICS
 		if(SSE2Available == 1)
 		{
+#ifndef QUEK15
 			method_ptr = jpeg_idct_islow_sse2;
 			method = JDCT_ISLOW;
+#else
+			method_ptr = jpeg_idct_ifast_sse2;
+			method = JDCT_IFAST;
+#endif ! QUEK15
 		}
 		else
 		{
-			method_ptr = jpeg_idct_islow;
-			method = JDCT_ISLOW;
+			method_ptr = jpeg_idct_ifast;
+			method = JDCT_IFAST;
 		}
 #else
 		method_ptr = jpeg_idct_islow;
@@ -148,8 +162,13 @@
 #ifdef HAVE_SSE2_INTEL_MNEMONICS
 		if (SSE2Available==1) 
 		{
+#ifndef QUEK15
 			method_ptr = jpeg_idct_islow_sse2;
 			method = JDCT_ISLOW;
+#else
+			method_ptr = jpeg_idct_ifast_sse2;
+			method = JDCT_ISLOW;
+#endif ! QUEK15
 		}
 		else
 		{
@@ -200,9 +219,63 @@
 	 * coefficients, but are stored as ints to ensure access efficiency.
 	 */
 	ISLOW_MULT_TYPE * ismtbl = (ISLOW_MULT_TYPE *) compptr->dct_table;
-	for (i = 0; i < DCTSIZE2; i++) {
-	  ismtbl[i] = (ISLOW_MULT_TYPE) qtbl->quantval[i];
+
+        /* We're moving 64 words. Use something fast. MGM Summer 2004. */
+        /* If we were really smart, we'd just point to the original    */
+        /* tables if that's possible. We might want to see if these    */
+        /* numbers change or if they're always the same.               */
+
+        void * input = (void *) qtbl->quantval;
+
+        __asm{
+            mov      eax, ismtbl
+            mov      ecx, input
+            movlps   xmm0, QWORD PTR [ecx]
+            movhps   xmm0, QWORD PTR [ecx+8]
+            movlps   xmm1, QWORD PTR [ecx+16]
+            movhps   xmm1, QWORD PTR [ecx+24]
+            movlps   xmm2, QWORD PTR [ecx+32]
+            movhps   xmm2, QWORD PTR [ecx+40]
+            movlps   xmm3, QWORD PTR [ecx+48]
+            movhps   xmm3, QWORD PTR [ecx+56]
+            movlps   QWORD PTR [eax], xmm0
+            movhps   QWORD PTR [eax+8], xmm0
+            movlps   QWORD PTR [eax+16], xmm1
+            movhps   QWORD PTR [eax+24], xmm1
+            movlps   QWORD PTR [eax+32], xmm2
+            movhps   QWORD PTR [eax+40], xmm2
+            movlps   QWORD PTR [eax+48], xmm3
+            movhps   QWORD PTR [eax+56], xmm3
+            movlps   xmm0, QWORD PTR [ecx+64]
+            movhps   xmm0, QWORD PTR [ecx+72]
+            movlps   xmm1, QWORD PTR [ecx+80]
+            movhps   xmm1, QWORD PTR [ecx+88]
+            movlps   xmm2, QWORD PTR [ecx+96]
+            movhps   xmm2, QWORD PTR [ecx+104]
+            movlps   xmm3, QWORD PTR [ecx+112]
+            movhps   xmm3, QWORD PTR [ecx+120]
+            movlps   QWORD PTR [eax+64], xmm0
+            movhps   QWORD PTR [eax+72], xmm0
+            movlps   QWORD PTR [eax+80], xmm1
+            movhps   QWORD PTR [eax+88], xmm1
+            movlps   QWORD PTR [eax+96], xmm2
+            movhps   QWORD PTR [eax+104], xmm2
+            movlps   QWORD PTR [eax+112], xmm3
+            movhps   QWORD PTR [eax+120], xmm3
+            }
+
+/*
+	for (i = 0; i < DCTSIZE2; i+=8) {
+	  ismtbl[i]   = (ISLOW_MULT_TYPE) qtbl->quantval[i];
+	  ismtbl[i+1] = (ISLOW_MULT_TYPE) qtbl->quantval[i+1];
+	  ismtbl[i+2] = (ISLOW_MULT_TYPE) qtbl->quantval[i+2];
+	  ismtbl[i+3] = (ISLOW_MULT_TYPE) qtbl->quantval[i+3];
+	  ismtbl[i+4] = (ISLOW_MULT_TYPE) qtbl->quantval[i+4];
+	  ismtbl[i+5] = (ISLOW_MULT_TYPE) qtbl->quantval[i+5];
+	  ismtbl[i+6] = (ISLOW_MULT_TYPE) qtbl->quantval[i+6];
+	  ismtbl[i+7] = (ISLOW_MULT_TYPE) qtbl->quantval[i+7];
 	}
+*/
       }
       break;
 #endif
@@ -216,6 +289,9 @@
 	 * For integer operation, the multiplier table is to be scaled by
 	 * IFAST_SCALE_BITS.
 	 */
+
+	/* Optimize this!!!!!!!!!!!!!!! */
+
 	IFAST_MULT_TYPE * ifmtbl = (IFAST_MULT_TYPE *) compptr->dct_table;
 #define CONST_BITS 14
 	static const INT16 aanscales[DCTSIZE2] = {
@@ -295,8 +371,22 @@
   for (ci = 0, compptr = cinfo->comp_info; ci < cinfo->num_components;
        ci++, compptr++) {
     /* Allocate and pre-zero a multiplier table for each component */
-    compptr->dct_table =
-      (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+
+#ifdef HAVE_SSE2_INTEL_MNEMONICS
+    if (SSE2Available == 1) {
+
+      /* For SSE2, align this table so that we can use */
+      /* high-performance parallel multiply in jidctint.c. */
+
+      compptr->dct_table =
+	(*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
+				  SIZEOF(multiplier_table)+16);
+      compptr->dct_table = (((unsigned int) compptr->dct_table) & 0xFFFFFFF0 ) + 16;
+    }
+    else
+#endif
+      compptr->dct_table =
+        (*cinfo->mem->alloc_small) ((j_common_ptr) cinfo, JPOOL_IMAGE,
 				  SIZEOF(multiplier_table));
     MEMZERO(compptr->dct_table, SIZEOF(multiplier_table));
     /* Mark multiplier table not yet set up for any method */