idct permutation cleanup, idct can be selected per context now

fixing some threadunsafe code

Originally committed as revision 980 to svn://svn.ffmpeg.org/ffmpeg/trunk

1 files changed, 137 insertions, 26 deletions

diff --git a/libavcodec/i386/mpegvideo_mmx_template.c b/libavcodec/i386/mpegvideo_mmx_template.c
index 1eed906c63..8fda458a21 100644
--- a/libavcodec/i386/mpegvideo_mmx_template.c
+++ b/libavcodec/i386/mpegvideo_mmx_template.c
@@ -189,31 +189,143 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
         );
     }
 
-    if(s->mb_intra) temp_block[0]= level; //FIXME move afer permute
-        
-// last_non_zero_p1=64;       
-    /* permute for IDCT */
-    asm volatile(
-        "movl %0, %%eax			\n\t"
-	"pushl %%ebp			\n\t"
-	"movl %%esp, " MANGLE(esp_temp) "\n\t"
-	"1:				\n\t"
-	"movzbl (%1, %%eax), %%ebx	\n\t"
-	"movzbl 1(%1, %%eax), %%ebp	\n\t"
-	"movw (%2, %%ebx, 2), %%cx	\n\t"
-	"movw (%2, %%ebp, 2), %%sp	\n\t"
-	"movzbl " MANGLE(permutation) "(%%ebx), %%ebx\n\t"
-	"movzbl " MANGLE(permutation) "(%%ebp), %%ebp\n\t"
-	"movw %%cx, (%3, %%ebx, 2)	\n\t"
-	"movw %%sp, (%3, %%ebp, 2)	\n\t"
-	"addl $2, %%eax			\n\t"
-	" js 1b				\n\t"
-	"movl " MANGLE(esp_temp) ", %%esp\n\t"
-	"popl %%ebp			\n\t"
-	: 
-	: "g" (-last_non_zero_p1), "d" (zigzag_direct_noperm+last_non_zero_p1), "S" (temp_block), "D" (block)
-	: "%eax", "%ebx", "%ecx"
-	);
+    if(s->mb_intra) block[0]= level;
+    else            block[0]= temp_block[0];
+
+    if(s->idct_permutation[1]==8){
+        if(last_non_zero_p1 <= 1) goto end;
+        block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08]; 
+        block[0x20] = temp_block[0x10]; 
+        if(last_non_zero_p1 <= 4) goto end;
+        block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02]; 
+        block[0x09] = temp_block[0x03]; 
+        if(last_non_zero_p1 <= 7) goto end;
+        block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11]; 
+        block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20]; 
+        if(last_non_zero_p1 <= 11) goto end;
+        block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12]; 
+        block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04]; 
+        block[0x0C] = temp_block[0x05]; 
+        if(last_non_zero_p1 <= 16) goto end;
+        block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13]; 
+        block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21]; 
+        block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30]; 
+        block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22]; 
+        if(last_non_zero_p1 <= 24) goto end;
+        block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14]; 
+        block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06]; 
+        block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E]; 
+        block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C]; 
+        if(last_non_zero_p1 <= 32) goto end;
+        block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A]; 
+        block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38]; 
+        block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32]; 
+        block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24]; 
+        if(last_non_zero_p1 <= 40) goto end;
+        block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16]; 
+        block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17]; 
+        block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25]; 
+        block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33]; 
+        if(last_non_zero_p1 <= 48) goto end;
+        block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B]; 
+        block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D]; 
+        block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F]; 
+        block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E]; 
+        if(last_non_zero_p1 <= 56) goto end;
+        block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C]; 
+        block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36]; 
+        block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37]; 
+        block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
+    }else if(s->idct_permutation[1]==4){
+        if(last_non_zero_p1 <= 1) goto end;
+        block[0x04] = temp_block[0x01]; 
+        block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10]; 
+        if(last_non_zero_p1 <= 4) goto end;
+        block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02]; 
+        block[0x05] = temp_block[0x03]; 
+        if(last_non_zero_p1 <= 7) goto end;
+        block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11]; 
+        block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20]; 
+        if(last_non_zero_p1 <= 11) goto end;
+        block[0x1C] = temp_block[0x19]; 
+        block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B]; 
+        block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05]; 
+        if(last_non_zero_p1 <= 16) goto end;
+        block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13]; 
+        block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21]; 
+        block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30]; 
+        block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22]; 
+        if(last_non_zero_p1 <= 24) goto end;
+        block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14]; 
+        block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06]; 
+        block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E]; 
+        block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C]; 
+        if(last_non_zero_p1 <= 32) goto end;
+        block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A]; 
+        block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38]; 
+        block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32]; 
+        block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24]; 
+        if(last_non_zero_p1 <= 40) goto end;
+        block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16]; 
+        block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17]; 
+        block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25]; 
+        block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33]; 
+        if(last_non_zero_p1 <= 48) goto end;
+        block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B]; 
+        block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D]; 
+            block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F]; 
+        block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E]; 
+        if(last_non_zero_p1 <= 56) goto end;
+        block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C]; 
+        block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36]; 
+        block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37]; 
+        block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
+    }else{
+        if(last_non_zero_p1 <= 1) goto end;
+        block[0x01] = temp_block[0x01]; 
+        block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10]; 
+        if(last_non_zero_p1 <= 4) goto end;
+        block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02]; 
+        block[0x03] = temp_block[0x03]; 
+        if(last_non_zero_p1 <= 7) goto end;
+        block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11]; 
+        block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20]; 
+        if(last_non_zero_p1 <= 11) goto end;
+        block[0x19] = temp_block[0x19]; 
+        block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B]; 
+        block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05]; 
+        if(last_non_zero_p1 <= 16) goto end;
+        block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13]; 
+        block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21]; 
+        block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30]; 
+        block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22]; 
+        if(last_non_zero_p1 <= 24) goto end;
+        block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14]; 
+        block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06]; 
+        block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E]; 
+        block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C]; 
+        if(last_non_zero_p1 <= 32) goto end;
+        block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A]; 
+        block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38]; 
+        block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32]; 
+        block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24]; 
+        if(last_non_zero_p1 <= 40) goto end;
+        block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16]; 
+        block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17]; 
+        block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25]; 
+        block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33]; 
+        if(last_non_zero_p1 <= 48) goto end;
+        block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B]; 
+        block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D]; 
+        block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F]; 
+        block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E]; 
+        if(last_non_zero_p1 <= 56) goto end;
+        block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C]; 
+        block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36]; 
+        block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37]; 
+        block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
+    }
+    end:
 /*
     for(i=0; i<last_non_zero_p1; i++)
     {
@@ -221,7 +333,6 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
        block[block_permute_op(j)]= temp_block[j];
     }
 */
-//block_permute(block);
 
     return last_non_zero_p1 - 1;
 }