* modules/video_chroma/i420_rgb_mmx.h: wrote an MMX intrinsics version of

    the conversion routines. Now that we no longer use inline asm, these
    modules do not generate non-PIC code with -fPIC.

diff --git a/modules/video_chroma/i420_rgb16.c b/modules/video_chroma/i420_rgb16.c
index ad14b51ebdf0117517588fc97f37201de6be0aa7..e2221941e7c22e1208049f6f9d6d3e9cdd1c9c70 100644 (file)
--- a/modules/video_chroma/i420_rgb16.c
+++ b/modules/video_chroma/i420_rgb16.c
@@ -34,6 +34,9 @@
 #if defined (MODULE_NAME_IS_i420_rgb)
 #   include "i420_rgb_c.h"
 #elif defined (MODULE_NAME_IS_i420_rgb_mmx)
+#   if defined(HAVE_MMX_INTRINSICS)
+#       include <mmintrin.h>
+#   endif
 #   include "i420_rgb_mmx.h"
 #endif
 
@@ -278,6 +281,13 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
             /* 15bpp 5/5/5 */
             for ( i_x = p_vout->render.i_width / 8; i_x--; )
             {
+#   if defined (HAVE_MMX_INTRINSICS)
+                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+                INTRINSICS_INIT_16
+                INTRINSICS_YUV_MUL
+                INTRINSICS_YUV_ADD
+                INTRINSICS_UNPACK_15
+#   else
                 __asm__( MMX_INIT_16
                          : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
 
@@ -286,6 +296,7 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
                          MMX_YUV_ADD
                          MMX_UNPACK_15
                          : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#   endif
 
                 p_y += 8;
                 p_u += 4;
@@ -298,6 +309,13 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
             /* 16bpp 5/6/5 */
             for ( i_x = p_vout->render.i_width / 8; i_x--; )
             {
+#   if defined (HAVE_MMX_INTRINSICS)
+                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+                INTRINSICS_INIT_16
+                INTRINSICS_YUV_MUL
+                INTRINSICS_YUV_ADD
+                INTRINSICS_UNPACK_16
+#   else
                 __asm__( MMX_INIT_16
                          : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
 
@@ -306,6 +324,7 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
                          MMX_YUV_ADD
                          MMX_UNPACK_16
                          : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#   endif
 
                 p_y += 8;
                 p_u += 4;
@@ -319,6 +338,11 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
          * at least we have all the pixels */
         if( i_rewind )
         {
+#if defined (MODULE_NAME_IS_i420_rgb_mmx)
+#   if defined (HAVE_MMX_INTRINSICS)
+            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+#   endif
+#endif
             p_y -= i_rewind;
             p_u -= i_rewind >> 1;
             p_v -= i_rewind >> 1;
@@ -329,26 +353,43 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
 #elif defined (MODULE_NAME_IS_i420_rgb_mmx)
+
+#   if defined (HAVE_MMX_INTRINSICS)
+            INTRINSICS_INIT_16
+#   else
             __asm__( MMX_INIT_16
                      : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#   endif
 
             if( p_vout->output.i_rmask == 0x7c00 )
             {
                 /* 15bpp 5/5/5 */
+#   if defined (HAVE_MMX_INTRINSICS)
+                INTRINSICS_YUV_MUL
+                INTRINSICS_YUV_ADD
+                INTRINSICS_UNPACK_15
+#   else
                 __asm__( ".align 8"
                          MMX_YUV_MUL
                          MMX_YUV_ADD
                          MMX_UNPACK_15
                          : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#   endif
             }
             else
             {
+#   if defined (HAVE_MMX_INTRINSICS)
+                INTRINSICS_YUV_MUL
+                INTRINSICS_YUV_ADD
+                INTRINSICS_UNPACK_16
+#   else
                 /* 16bpp 5/6/5 */
                 __asm__( ".align 8"
                          MMX_YUV_MUL
                          MMX_YUV_ADD
                          MMX_UNPACK_16
                          : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#   endif
             }
 
             p_y += 8;
@@ -453,6 +494,13 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
             CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
             CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
 #elif defined (MODULE_NAME_IS_i420_rgb_mmx)
+#   if defined (HAVE_MMX_INTRINSICS)
+            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+            INTRINSICS_INIT_32
+            INTRINSICS_YUV_MUL
+            INTRINSICS_YUV_ADD
+            INTRINSICS_UNPACK_32
+#   else
             __asm__( MMX_INIT_32
                      : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
 
@@ -461,6 +509,7 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
                      MMX_YUV_ADD
                      MMX_UNPACK_32
                      : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#   endif
 
             p_y += 8;
             p_u += 4;
@@ -473,6 +522,11 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
          * at least we have all the pixels */
         if( i_rewind )
         {
+#if defined (MODULE_NAME_IS_i420_rgb_mmx)
+#   if defined (HAVE_MMX_INTRINSICS)
+            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+#   endif
+#endif
             p_y -= i_rewind;
             p_u -= i_rewind >> 1;
             p_v -= i_rewind >> 1;
@@ -483,6 +537,12 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
             CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
             CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
 #elif defined (MODULE_NAME_IS_i420_rgb_mmx)
+#   if defined (HAVE_MMX_INTRINSICS)
+            INTRINSICS_INIT_32
+            INTRINSICS_YUV_MUL
+            INTRINSICS_YUV_ADD
+            INTRINSICS_UNPACK_32
+#   else
             __asm__( MMX_INIT_32
                      : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
 
@@ -491,6 +551,7 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
                      MMX_YUV_ADD
                      MMX_UNPACK_32
                      : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#   endif
 
             p_y += 8;
             p_u += 4;

diff --git a/modules/video_chroma/i420_rgb_mmx.h b/modules/video_chroma/i420_rgb_mmx.h
index ef03260ac1cb1465a1bcb94c65ad16c1ac720a67..e1e32405d02bd009fc8c0378fb1f82851b280fa0 100644 (file)
--- a/modules/video_chroma/i420_rgb_mmx.h
+++ b/modules/video_chroma/i420_rgb_mmx.h
@@ -53,6 +53,13 @@ movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
 #movl      $0, (%3)         # cache preload for image                       \n\
 "
 
+#define INTRINSICS_INIT_16 \
+    mm0 = (__m64)(uint64_t)*(uint32_t *)p_u; \
+    mm1 = (__m64)(uint64_t)*(uint32_t *)p_v; \
+    mm4 = (__m64)(uint64_t)0; \
+    mm6 = (__m64)*(uint64_t *)p_y; \
+    /* *(uint16_t *)p_buffer = 0; */
+
 #define MMX_INIT_16_GRAY "                                                  \n\
 movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
 #movl      $0, (%3)         # cache preload for image                       \n\
@@ -66,6 +73,13 @@ pxor      %%mm4, %%mm4      # zero mm4                                      \n\
 movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
 "
 
+#define INTRINSICS_INIT_32 \
+    mm0 = (__m64)(uint64_t)*(uint32_t *)p_u; \
+    *(uint16_t *)p_buffer = 0; \
+    mm1 = (__m64)(uint64_t)*(uint32_t *)p_v; \
+    mm4 = (__m64)(uint64_t)0; \
+    mm6 = (__m64)*(uint64_t *)p_y;
+
 /*
  * Do the multiply part of the conversion for even and odd pixels,
  * register usage:
@@ -101,6 +115,30 @@ pmulhw    mmx_Y_coeff, %%mm6    # Mul 4 Y even    00 y6 00 y4 00 y2 00 y0   \n\
 pmulhw    mmx_Y_coeff, %%mm7    # Mul 4 Y odd     00 y7 00 y5 00 y3 00 y1   \n\
 "
 
+#define INTRINSICS_YUV_MUL \
+    mm0 = _mm_unpacklo_pi8(mm0, mm4); \
+    mm1 = _mm_unpacklo_pi8(mm1, mm4); \
+    mm0 = _mm_subs_pi16(mm0, (__m64)mmx_80w); \
+    mm1 = _mm_subs_pi16(mm1, (__m64)mmx_80w); \
+    mm0 = _mm_slli_pi16(mm0, 3); \
+    mm1 = _mm_slli_pi16(mm1, 3); \
+    mm2 = mm0; \
+    mm3 = mm1; \
+    mm2 = _mm_mulhi_pi16(mm2, (__m64)mmx_U_green); \
+    mm3 = _mm_mulhi_pi16(mm3, (__m64)mmx_V_green); \
+    mm0 = _mm_mulhi_pi16(mm0, (__m64)mmx_U_blue); \
+    mm1 = _mm_mulhi_pi16(mm1, (__m64)mmx_V_red); \
+    mm2 = _mm_adds_pi16(mm2, mm3); \
+    \
+    mm6 = _mm_subs_pu8(mm6, (__m64)mmx_10w); \
+    mm7 = mm6; \
+    mm6 = _mm_and_si64(mm6, (__m64)mmx_00ffw); \
+    mm7 = _mm_srli_pi16(mm7, 8); \
+    mm6 = _mm_slli_pi16(mm6, 3); \
+    mm7 = _mm_slli_pi16(mm7, 3); \
+    mm6 = _mm_mulhi_pi16(mm6, (__m64)mmx_Y_coeff); \
+    mm7 = _mm_mulhi_pi16(mm7, (__m64)mmx_Y_coeff);
+
 /*
  * Do the addition part of the conversion for even and odd pixels,
  * register usage:
@@ -137,6 +175,29 @@ punpcklbw %%mm4, %%mm1          #                 R7 R6 R5 R4 R3 R2 R1 R0   \n\
 punpcklbw %%mm5, %%mm2          #                 G7 G6 G5 G4 G3 G2 G1 G0   \n\
 "
 
+#define INTRINSICS_YUV_ADD \
+    mm3 = mm0; \
+    mm4 = mm1; \
+    mm5 = mm2; \
+    mm0 = _mm_adds_pi16(mm0, mm6); \
+    mm3 = _mm_adds_pi16(mm3, mm7); \
+    mm1 = _mm_adds_pi16(mm1, mm6); \
+    mm4 = _mm_adds_pi16(mm4, mm7); \
+    mm2 = _mm_adds_pi16(mm2, mm6); \
+    mm5 = _mm_adds_pi16(mm5, mm7); \
+    \
+    mm0 = _mm_packs_pu16(mm0, mm0); \
+    mm1 = _mm_packs_pu16(mm1, mm1); \
+    mm2 = _mm_packs_pu16(mm2, mm2); \
+    \
+    mm3 = _mm_packs_pu16(mm3, mm3); \
+    mm4 = _mm_packs_pu16(mm4, mm4); \
+    mm5 = _mm_packs_pu16(mm5, mm5); \
+    \
+    mm0 = _mm_unpacklo_pi8(mm0, mm3); \
+    mm1 = _mm_unpacklo_pi8(mm1, mm4); \
+    mm2 = _mm_unpacklo_pi8(mm2, mm5);
+
 /*
  * Grayscale case, only use Y
  */
@@ -215,6 +276,31 @@ movd      4(%2), %%mm1          # Load 4 Cr       __ __ __ __ v3 v2 v1 v0   \n\
 movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
 "
 
+#define INTRINSICS_UNPACK_15 \
+    mm0 = _mm_and_si64(mm0, (__m64)mmx_mask_f8); \
+    mm0 = _mm_srli_pi16(mm0, 3); \
+    mm2 = _mm_and_si64(mm2, (__m64)mmx_mask_f8); \
+    mm1 = _mm_and_si64(mm1, (__m64)mmx_mask_f8); \
+    mm1 = _mm_srli_pi16(mm1, 1); \
+    mm4 = (__m64)(uint64_t)0; \
+    mm5 = mm0; \
+    mm7 = mm2; \
+    \
+    mm2 = _mm_unpacklo_pi8(mm2, mm4); \
+    mm0 = _mm_unpacklo_pi8(mm0, mm1); \
+    mm2 = _mm_slli_pi16(mm2, 2); \
+    mm0 = _mm_or_si64(mm0, mm2); \
+    mm6 = (__m64)*(uint64_t *)(p_y + 8); \
+    *(uint64_t *)p_buffer = (uint64_t)mm0; \
+    \
+    mm7 = _mm_unpackhi_pi8(mm7, mm4); \
+    mm5 = _mm_unpackhi_pi8(mm5, mm1); \
+    mm7 = _mm_slli_pi16(mm7, 2); \
+    mm0 = (__m64)(uint64_t)*(uint32_t *)(p_u + 4); \
+    mm5 = _mm_or_si64(mm5, mm7); \
+    mm1 = (__m64)(uint64_t)*(uint32_t *)(p_v + 4); \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
+
 /*
  * convert RGB plane to RGB 16 bits,
  * mm0 -> B, mm1 -> R, mm2 -> G,
@@ -250,6 +336,30 @@ movd      4(%2), %%mm1          # Load 4 Cr       __ __ __ __ v3 v2 v1 v0   \n\
 movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
 "
 
+#define INTRINSICS_UNPACK_16 \
+    mm0 = _mm_and_si64(mm0, (__m64)mmx_mask_f8); \
+    mm2 = _mm_and_si64(mm2, (__m64)mmx_mask_fc); \
+    mm1 = _mm_and_si64(mm1, (__m64)mmx_mask_f8); \
+    mm0 = _mm_srli_pi16(mm0, 3); \
+    mm4 = (__m64)(uint64_t)0; \
+    mm5 = mm0; \
+    mm7 = mm2; \
+    \
+    mm2 = _mm_unpacklo_pi8(mm2, mm4); \
+    mm0 = _mm_unpacklo_pi8(mm0, mm1); \
+    mm2 = _mm_slli_pi16(mm2, 3); \
+    mm0 = _mm_or_si64(mm0, mm2); \
+    mm6 = (__m64)*(uint64_t *)(p_y + 8); \
+    *(uint64_t *)p_buffer = (uint64_t)mm0; \
+    \
+    mm7 = _mm_unpackhi_pi8(mm7, mm4); \
+    mm5 = _mm_unpackhi_pi8(mm5, mm1); \
+    mm7 = _mm_slli_pi16(mm7, 3); \
+    mm0 = (__m64)(uint64_t)*(uint32_t *)(p_u + 4); \
+    mm5 = _mm_or_si64(mm5, mm7); \
+    mm1 = (__m64)(uint64_t)*(uint32_t *)(p_v + 4); \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
+
 /*
  * convert RGB plane to RGB packed format,
  * mm0 -> B, mm1 -> R, mm2 -> G, mm3 -> 0,
@@ -286,3 +396,26 @@ movq      %%mm4, 24(%3) # Store ARGB7 ARGB6                                 \n\
 #movq      8(%0), %%mm6  # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0           \n\
 "
 
+#define INTRINSICS_UNPACK_32 \
+    mm3 = (__m64)(uint64_t)0; \
+    mm6 = mm0; \
+    mm7 = mm1; \
+    mm4 = mm0; \
+    mm5 = mm1; \
+    mm6 = _mm_unpacklo_pi8(mm6, mm2); \
+    mm7 = _mm_unpacklo_pi8(mm7, mm3); \
+    mm6 = _mm_unpacklo_pi16(mm6, mm7); \
+    *(uint64_t *)p_buffer = (uint64_t)mm6; \
+    mm6 = mm0; \
+    mm6 = _mm_unpacklo_pi8(mm6, mm2); \
+    mm6 = _mm_unpackhi_pi16(mm6, mm7); \
+    *(uint64_t *)(p_buffer + 2) = (uint64_t)mm6; \
+    mm4 = _mm_unpackhi_pi8(mm4, mm2); \
+    mm5 = _mm_unpackhi_pi8(mm5, mm3); \
+    mm4 = _mm_unpacklo_pi16(mm4, mm5); \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm4; \
+    mm4 = mm0; \
+    mm4 = _mm_unpackhi_pi8(mm4, mm2); \
+    mm4 = _mm_unpackhi_pi16(mm4, mm5); \
+    *(uint64_t *)(p_buffer + 6) = (uint64_t)mm4; \
+