]> git.sesse.net Git - ffmpeg/blobdiff - libavcodec/x86/mpegvideoenc_template.c
x86: build: Only compile FDCT code if MMX is enabled
[ffmpeg] / libavcodec / x86 / mpegvideoenc_template.c
index 195298d8f341c60456a99391042a0c62267a9045..4a17ff3da78328e2c794d518059521a9df58e954 100644 (file)
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
+#include <stdint.h>
+
+#include "libavutil/internal.h"
+#include "libavutil/x86/asm.h"
+#include "libavcodec/mpegvideo.h"
+
 #undef MMREG_WIDTH
 #undef MM
 #undef MOVQ
@@ -29,7 +35,7 @@
 #undef SAVE_SIGN
 #undef RESTORE_SIGN
 
-#if HAVE_SSE2
+#if COMPILE_TEMPLATE_SSE2
 #define MMREG_WIDTH "16"
 #define MM "%%xmm"
 #define MOVQ "movdqa"
@@ -48,7 +54,7 @@
 #define MMREG_WIDTH "8"
 #define MM "%%mm"
 #define MOVQ "movq"
-#if HAVE_MMXEXT
+#if COMPILE_TEMPLATE_MMXEXT
 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
 #define PMAX(a,b) \
@@ -74,7 +80,7 @@
 #endif
 #endif
 
-#if HAVE_SSSE3
+#if COMPILE_TEMPLATE_SSSE3
 #define SAVE_SIGN(a,b) \
             "movdqa "b", "a"            \n\t"\
             "pabsw  "b", "b"            \n\t"
@@ -92,7 +98,7 @@
 #endif
 
 static int RENAME(dct_quantize)(MpegEncContext *s,
-                            DCTELEM *block, int n,
+                            int16_t *block, int n,
                             int qscale, int *overflow)
 {
     x86_reg last_non_zero_p1;
@@ -103,7 +109,7 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
     assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
 
     //s->fdct (block);
-    RENAMEl(ff_fdct) (block); //cannot be anything else ...
+    RENAME_FDCT(ff_fdct)(block); // cannot be anything else ...
 
     if(s->dct_error_sum)
         s->denoise_dct(s, block);
@@ -168,7 +174,7 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
             : "+a" (last_non_zero_p1)
             : "r" (block+64), "r" (qmat), "r" (bias),
-              "r" (ff_inv_zigzag_direct16+64), "r" (temp_block+64)
+              "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
         );
@@ -202,7 +208,7 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
             : "+a" (last_non_zero_p1)
             : "r" (block+64), "r" (qmat+64), "r" (bias+64),
-              "r" (ff_inv_zigzag_direct16+64), "r" (temp_block+64)
+              "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
         );
@@ -212,7 +218,7 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
         SPREADW(MM"1")
         "psubusw "MM"1, "MM"4               \n\t"
         "packuswb "MM"4, "MM"4              \n\t"
-#if HAVE_SSE2
+#if COMPILE_TEMPLATE_SSE2
         "packuswb "MM"4, "MM"4              \n\t"
 #endif
         "movd "MM"4, %0                     \n\t" // *overflow
@@ -223,7 +229,7 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
     if(s->mb_intra) block[0]= level;
     else            block[0]= temp_block[0];
 
-    if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
+    if (s->idsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM) {
         if(last_non_zero_p1 <= 1) goto end;
         block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
         block[0x20] = temp_block[0x10];
@@ -267,50 +273,6 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
         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->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
-        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];