Merge commit '5169e688956be3378adb3b16a93962fe0048f1c9'

* commit '5169e688956be3378adb3b16a93962fe0048f1c9':
  dsputil: Propagate bit depth information to all (sub)init functions

Conflicts:
	libavcodec/arm/dsputil_init_arm.c
	libavcodec/arm/dsputil_init_armv5te.c
	libavcodec/arm/dsputil_init_armv6.c
	libavcodec/arm/dsputil_init_neon.c
	libavcodec/dsputil.c
	libavcodec/dsputil.h
	libavcodec/ppc/dsputil_ppc.c
	libavcodec/x86/dsputil_init.c

Merged-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/arm/dsputil_arm.h

@@ -24,8 +24,11 @@
 #include "libavcodec/avcodec.h"
 #include "libavcodec/dsputil.h"
 
-void ff_dsputil_init_armv5te(DSPContext *c, AVCodecContext *avctx);
-void ff_dsputil_init_armv6(DSPContext *c, AVCodecContext *avctx);
-void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx);
+void ff_dsputil_init_armv5te(DSPContext *c, AVCodecContext *avctx,
+                             unsigned high_bit_depth);
+void ff_dsputil_init_armv6(DSPContext *c, AVCodecContext *avctx,
+                           unsigned high_bit_depth);
+void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx,
+                          unsigned high_bit_depth);
 
 #endif /* AVCODEC_ARM_DSPUTIL_ARM_H */

libavcodec/arm/dsputil_init_arm.c

@@ -64,14 +64,15 @@ static void simple_idct_arm_add(uint8_t *dest, int line_size, int16_t *block)
     ff_add_pixels_clamped(block, dest, line_size);
 }
 
-av_cold void ff_dsputil_init_arm(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_arm(DSPContext *c, AVCodecContext *avctx,
+                                 unsigned high_bit_depth)
 {
     int cpu_flags = av_get_cpu_flags();
 
     ff_put_pixels_clamped = c->put_pixels_clamped;
     ff_add_pixels_clamped = c->add_pixels_clamped;
 
-    if (!avctx->lowres && avctx->bits_per_raw_sample <= 8) {
+    if (!avctx->lowres && !high_bit_depth) {
         if (avctx->idct_algo == FF_IDCT_AUTO ||
             avctx->idct_algo == FF_IDCT_ARM) {
             c->idct_put              = j_rev_dct_arm_put;
@@ -89,9 +90,9 @@ av_cold void ff_dsputil_init_arm(DSPContext *c, AVCodecContext *avctx)
     c->add_pixels_clamped = ff_add_pixels_clamped_arm;
 
     if (have_armv5te(cpu_flags))
-        ff_dsputil_init_armv5te(c, avctx);
+        ff_dsputil_init_armv5te(c, avctx, high_bit_depth);
     if (have_armv6(cpu_flags))
-        ff_dsputil_init_armv6(c, avctx);
+        ff_dsputil_init_armv6(c, avctx, high_bit_depth);
     if (have_neon(cpu_flags))
-        ff_dsputil_init_neon(c, avctx);
+        ff_dsputil_init_neon(c, avctx, high_bit_depth);
 }

libavcodec/arm/dsputil_init_armv5te.c

@@ -29,9 +29,10 @@ void ff_simple_idct_armv5te(int16_t *data);
 void ff_simple_idct_put_armv5te(uint8_t *dest, int line_size, int16_t *data);
 void ff_simple_idct_add_armv5te(uint8_t *dest, int line_size, int16_t *data);
 
-av_cold void ff_dsputil_init_armv5te(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_armv5te(DSPContext *c, AVCodecContext *avctx,
+                                     unsigned high_bit_depth)
 {
-    if (!avctx->lowres && avctx->bits_per_raw_sample <= 8 &&
+    if (!avctx->lowres && !high_bit_depth &&
         (avctx->idct_algo == FF_IDCT_AUTO ||
          avctx->idct_algo == FF_IDCT_SIMPLEARMV5TE)) {
         c->idct_put              = ff_simple_idct_put_armv5te;

libavcodec/arm/dsputil_init_armv6.c

@@ -52,17 +52,17 @@ int ff_sse16_armv6(void *s, uint8_t *blk1, uint8_t *blk2,
 int ff_pix_norm1_armv6(uint8_t *pix, int line_size);
 int ff_pix_sum_armv6(uint8_t *pix, int line_size);
 
-av_cold void ff_dsputil_init_armv6(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_armv6(DSPContext *c, AVCodecContext *avctx,
+                                   unsigned high_bit_depth)
 {
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
-    if (!avctx->lowres && avctx->bits_per_raw_sample <= 8 &&
-        (avctx->idct_algo == FF_IDCT_AUTO ||
-         avctx->idct_algo == FF_IDCT_SIMPLEARMV6)) {
-        c->idct_put              = ff_simple_idct_put_armv6;
-        c->idct_add              = ff_simple_idct_add_armv6;
-        c->idct                  = ff_simple_idct_armv6;
-        c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
+    if (!avctx->lowres && !high_bit_depth) {
+        if (avctx->idct_algo == FF_IDCT_AUTO ||
+            avctx->idct_algo == FF_IDCT_SIMPLEARMV6) {
+            c->idct_put              = ff_simple_idct_put_armv6;
+            c->idct_add              = ff_simple_idct_add_armv6;
+            c->idct                  = ff_simple_idct_armv6;
+            c->idct_permutation_type = FF_LIBMPEG2_IDCT_PERM;
+        }
     }
     c->add_pixels_clamped = ff_add_pixels_clamped_armv6;
 

libavcodec/arm/dsputil_init_neon.c

@@ -47,11 +47,10 @@ int32_t ff_scalarproduct_int16_neon(const int16_t *v1, const int16_t *v2, int le
 int32_t ff_scalarproduct_and_madd_int16_neon(int16_t *v1, const int16_t *v2,
                                              const int16_t *v3, int len, int mul);
 
-av_cold void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_neon(DSPContext *c, AVCodecContext *avctx,
+                                  unsigned high_bit_depth)
 {
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
-    if (!avctx->lowres && avctx->bits_per_raw_sample <= 8) {
+    if (!avctx->lowres && !high_bit_depth) {
         if (avctx->idct_algo == FF_IDCT_AUTO ||
             avctx->idct_algo == FF_IDCT_SIMPLENEON) {
             c->idct_put              = ff_simple_idct_put_neon;

libavcodec/bfin/dsputil_init.c

@@ -147,18 +147,14 @@ static int bfin_pix_abs8_xy2(void *c, uint8_t *blk1, uint8_t *blk2,
  * 2.64s    2/20 same sman.mp4 decode only
  */
 
-av_cold void ff_dsputil_init_bfin(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_bfin(DSPContext *c, AVCodecContext *avctx,
+                                  unsigned high_bit_depth)
 {
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
     c->diff_pixels = ff_bfin_diff_pixels;
 
     c->put_pixels_clamped = ff_bfin_put_pixels_clamped;
     c->add_pixels_clamped = ff_bfin_add_pixels_clamped;
 
-    if (!high_bit_depth)
-        c->get_pixels = ff_bfin_get_pixels;
-
     c->clear_blocks = bfin_clear_blocks;
 
     c->pix_sum   = ff_bfin_pix_sum;
@@ -182,7 +178,9 @@ av_cold void ff_dsputil_init_bfin(DSPContext *c, AVCodecContext *avctx)
     c->sse[1] = ff_bfin_sse8;
     c->sse[2] = ff_bfin_sse4;
 
-    if (avctx->bits_per_raw_sample <= 8) {
+    if (!high_bit_depth) {
+        c->get_pixels = ff_bfin_get_pixels;
+
         if (avctx->dct_algo == FF_DCT_AUTO)
             c->fdct = ff_bfin_fdct;
 

libavcodec/dsputil.c

@@ -2884,6 +2884,8 @@ int ff_check_alignment(void)
 
 av_cold void ff_dsputil_init(DSPContext *c, AVCodecContext *avctx)
 {
+    const unsigned high_bit_depth = avctx->bits_per_raw_sample > 8;
+
     ff_check_alignment();
 
 #if CONFIG_ENCODERS
@@ -3127,13 +3129,13 @@ av_cold void ff_dsputil_init(DSPContext *c, AVCodecContext *avctx)
     if (ARCH_ALPHA)
         ff_dsputil_init_alpha(c, avctx);
     if (ARCH_ARM)
-        ff_dsputil_init_arm(c, avctx);
+        ff_dsputil_init_arm(c, avctx, high_bit_depth);
     if (ARCH_BFIN)
-        ff_dsputil_init_bfin(c, avctx);
+        ff_dsputil_init_bfin(c, avctx, high_bit_depth);
     if (ARCH_PPC)
-        ff_dsputil_init_ppc(c, avctx);
+        ff_dsputil_init_ppc(c, avctx, high_bit_depth);
     if (ARCH_X86)
-        ff_dsputil_init_x86(c, avctx);
+        ff_dsputil_init_x86(c, avctx, high_bit_depth);
 
     ff_init_scantable_permutation(c->idct_permutation,
                                   c->idct_permutation_type);

libavcodec/dsputil.h

@@ -355,10 +355,14 @@ int ff_check_alignment(void);
 void ff_set_cmp(DSPContext *c, me_cmp_func *cmp, int type);
 
 void ff_dsputil_init_alpha(DSPContext* c, AVCodecContext *avctx);
-void ff_dsputil_init_arm(DSPContext *c, AVCodecContext *avctx);
-void ff_dsputil_init_bfin(DSPContext *c, AVCodecContext *avctx);
-void ff_dsputil_init_ppc(DSPContext *c, AVCodecContext *avctx);
-void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx);
+void ff_dsputil_init_arm(DSPContext *c, AVCodecContext *avctx,
+                         unsigned high_bit_depth);
+void ff_dsputil_init_bfin(DSPContext *c, AVCodecContext *avctx,
+                          unsigned high_bit_depth);
+void ff_dsputil_init_ppc(DSPContext *c, AVCodecContext *avctx,
+                         unsigned high_bit_depth);
+void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx,
+                         unsigned high_bit_depth);
 
 void ff_dsputil_init_dwt(DSPContext *c);
 

libavcodec/ppc/dsputil_altivec.c

@@ -926,10 +926,9 @@ static int hadamard8_diff16_altivec(/* MpegEncContext */ void *s, uint8_t *dst,
     return score;
 }
 
-av_cold void ff_dsputil_init_altivec(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_altivec(DSPContext *c, AVCodecContext *avctx,
+                                     unsigned high_bit_depth)
 {
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
     c->pix_abs[0][1] = sad16_x2_altivec;
     c->pix_abs[0][2] = sad16_y2_altivec;
     c->pix_abs[0][3] = sad16_xy2_altivec;

libavcodec/ppc/dsputil_altivec.h

@@ -38,7 +38,8 @@ void ff_gmc1_altivec(uint8_t *dst, uint8_t *src, int stride, int h,
 void ff_idct_put_altivec(uint8_t *dest, int line_size, int16_t *block);
 void ff_idct_add_altivec(uint8_t *dest, int line_size, int16_t *block);
 
-void ff_dsputil_init_altivec(DSPContext *c, AVCodecContext *avctx);
+void ff_dsputil_init_altivec(DSPContext *c, AVCodecContext *avctx,
+                             unsigned high_bit_depth);
 void ff_int_init_altivec(DSPContext *c, AVCodecContext *avctx);
 
 #endif /* AVCODEC_PPC_DSPUTIL_ALTIVEC_H */

libavcodec/ppc/dsputil_ppc.c

@@ -125,11 +125,10 @@ static long check_dcbzl_effect(void)
     return count;
 }
 
-av_cold void ff_dsputil_init_ppc(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_ppc(DSPContext *c, AVCodecContext *avctx,
+                                 unsigned high_bit_depth)
 {
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
     int mm_flags = av_get_cpu_flags();
-
     // common optimizations whether AltiVec is available or not
     if (!high_bit_depth) {
         switch (check_dcbzl_effect()) {
@@ -145,25 +144,25 @@ av_cold void ff_dsputil_init_ppc(DSPContext *c, AVCodecContext *avctx)
     }
 
     if (PPC_ALTIVEC(mm_flags)) {
-        ff_dsputil_init_altivec(c, avctx);
+        ff_dsputil_init_altivec(c, avctx, high_bit_depth);
         ff_int_init_altivec(c, avctx);
         c->gmc1 = ff_gmc1_altivec;
 
+        if (!high_bit_depth) {
 #if CONFIG_ENCODERS
-        if (avctx->bits_per_raw_sample <= 8 &&
-            (avctx->dct_algo == FF_DCT_AUTO ||
-             avctx->dct_algo == FF_DCT_ALTIVEC)) {
-            c->fdct = ff_fdct_altivec;
-        }
+            if (avctx->dct_algo == FF_DCT_AUTO ||
+                avctx->dct_algo == FF_DCT_ALTIVEC) {
+                c->fdct = ff_fdct_altivec;
+            }
 #endif //CONFIG_ENCODERS
-
-        if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
+          if (avctx->lowres == 0) {
             if ((avctx->idct_algo == FF_IDCT_AUTO) ||
                 (avctx->idct_algo == FF_IDCT_ALTIVEC)) {
                 c->idct_put              = ff_idct_put_altivec;
                 c->idct_add              = ff_idct_add_altivec;
                 c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
             }
+          }
         }
     }
 }

libavcodec/x86/dsputil_init.c

@@ -521,11 +521,9 @@ do {                                                                         \
 } while (0)
 
 static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx,
-                                     int cpu_flags)
+                                     int cpu_flags, unsigned high_bit_depth)
 {
 #if HAVE_MMX_INLINE
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
@@ -549,11 +547,9 @@ static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx,
 }
 
 static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
-                                        int cpu_flags)
+                                        int cpu_flags, unsigned high_bit_depth)
 {
 #if HAVE_MMXEXT_INLINE
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX && avctx->lowres == 0) {
         c->idct_put = ff_idct_xvid_mmxext_put;
         c->idct_add = ff_idct_xvid_mmxext_add;
@@ -580,11 +576,9 @@ static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
 }
 
 static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx,
-                                     int cpu_flags)
+                                     int cpu_flags, unsigned high_bit_depth)
 {
 #if HAVE_SSE_INLINE
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
     c->vector_clipf = ff_vector_clipf_sse;
 
     /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
@@ -605,11 +599,9 @@ static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx,
 }
 
 static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
-                                      int cpu_flags)
+                                      int cpu_flags, unsigned high_bit_depth)
 {
 #if HAVE_SSE2_INLINE
-    const int high_bit_depth = avctx->bits_per_raw_sample > 8;
-
     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX && avctx->lowres == 0) {
         c->idct_put              = ff_idct_xvid_sse2_put;
         c->idct_add              = ff_idct_xvid_sse2_add;
@@ -631,7 +623,7 @@ static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
 }
 
 static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
-                                       int cpu_flags)
+                                       int cpu_flags, unsigned high_bit_depth)
 {
 #if HAVE_SSSE3_EXTERNAL
     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
@@ -645,14 +637,15 @@ static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
 }
 
 static av_cold void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
-                                      int cpu_flags)
+                                      int cpu_flags, unsigned high_bit_depth)
 {
 #if HAVE_SSE4_EXTERNAL
     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
 #endif /* HAVE_SSE4_EXTERNAL */
 }
 
-av_cold void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx,
+                                 unsigned high_bit_depth)
 {
     int cpu_flags = av_get_cpu_flags();
 
@@ -665,7 +658,7 @@ av_cold void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx)
 #if HAVE_INLINE_ASM
         const int idct_algo = avctx->idct_algo;
 
-        if (avctx->lowres == 0 && avctx->bits_per_raw_sample <= 8) {
+        if (avctx->lowres == 0 && !high_bit_depth) {
             if (idct_algo == FF_IDCT_AUTO || idct_algo == FF_IDCT_SIMPLEMMX) {
                 c->idct_put              = ff_simple_idct_put_mmx;
                 c->idct_add              = ff_simple_idct_add_mmx;
@@ -679,24 +672,24 @@ av_cold void ff_dsputil_init_x86(DSPContext *c, AVCodecContext *avctx)
         }
 #endif /* HAVE_INLINE_ASM */
 
-        dsputil_init_mmx(c, avctx, cpu_flags);
+        dsputil_init_mmx(c, avctx, cpu_flags, high_bit_depth);
     }
 
     if (X86_MMXEXT(cpu_flags))
-        dsputil_init_mmxext(c, avctx, cpu_flags);
+        dsputil_init_mmxext(c, avctx, cpu_flags, high_bit_depth);
 
     if (X86_SSE(cpu_flags))
-        dsputil_init_sse(c, avctx, cpu_flags);
+        dsputil_init_sse(c, avctx, cpu_flags, high_bit_depth);
 
     if (X86_SSE2(cpu_flags))
-        dsputil_init_sse2(c, avctx, cpu_flags);
+        dsputil_init_sse2(c, avctx, cpu_flags, high_bit_depth);
 
     if (EXTERNAL_SSSE3(cpu_flags))
-        dsputil_init_ssse3(c, avctx, cpu_flags);
+        dsputil_init_ssse3(c, avctx, cpu_flags, high_bit_depth);
 
     if (EXTERNAL_SSE4(cpu_flags))
-        dsputil_init_sse4(c, avctx, cpu_flags);
+        dsputil_init_sse4(c, avctx, cpu_flags, high_bit_depth);
 
     if (CONFIG_ENCODERS)
-        ff_dsputilenc_init_mmx(c, avctx);
+        ff_dsputilenc_init_mmx(c, avctx, high_bit_depth);
 }

libavcodec/x86/dsputil_x86.h

@@ -104,7 +104,8 @@
     "psubb "#regb", "#regr"             \n\t"                    \
     "psubb "#regd", "#regp"             \n\t"
 
-void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx);
+void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx,
+                            unsigned high_bit_depth);
 void ff_dsputil_init_pix_mmx(DSPContext *c, AVCodecContext *avctx);
 
 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,

libavcodec/x86/dsputilenc_mmx.c

@@ -987,16 +987,15 @@ hadamard_func(mmxext)
 hadamard_func(sse2)
 hadamard_func(ssse3)
 
-av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx)
+av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx,
+                                    unsigned high_bit_depth)
 {
     int cpu_flags = av_get_cpu_flags();
     const int dct_algo = avctx->dct_algo;
 
 #if HAVE_YASM
-    int bit_depth = avctx->bits_per_raw_sample;
-
     if (EXTERNAL_MMX(cpu_flags)) {
-        if (bit_depth <= 8)
+        if (!high_bit_depth)
             c->get_pixels = ff_get_pixels_mmx;
         c->diff_pixels = ff_diff_pixels_mmx;
         c->pix_sum     = ff_pix_sum16_mmx;
@@ -1004,13 +1003,13 @@ av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx)
     }
 
     if (EXTERNAL_SSE2(cpu_flags))
-        if (bit_depth <= 8)
+        if (!high_bit_depth)
             c->get_pixels = ff_get_pixels_sse2;
 #endif /* HAVE_YASM */
 
 #if HAVE_INLINE_ASM
     if (INLINE_MMX(cpu_flags)) {
-        if (avctx->bits_per_raw_sample <= 8 &&
+        if (!high_bit_depth &&
             (dct_algo == FF_DCT_AUTO || dct_algo == FF_DCT_MMX))
             c->fdct = ff_fdct_mmx;
 
@@ -1040,7 +1039,7 @@ av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx)
     }
 
     if (INLINE_MMXEXT(cpu_flags)) {
-        if (avctx->bits_per_raw_sample <= 8 &&
+        if (!high_bit_depth &&
             (dct_algo == FF_DCT_AUTO || dct_algo == FF_DCT_MMX))
             c->fdct = ff_fdct_mmxext;
 
@@ -1055,7 +1054,7 @@ av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx)
     }
 
     if (INLINE_SSE2(cpu_flags)) {
-        if (avctx->bits_per_raw_sample <= 8 &&
+        if (!high_bit_depth &&
             (dct_algo == FF_DCT_AUTO || dct_algo == FF_DCT_MMX))
             c->fdct = ff_fdct_sse2;