[FFmpeg-cvslog] ac3dsp: simplify x86 versions of ac3_max_msb_abs_int16

[Justin Ruggles]

ffmpeg | branch: master | Justin Ruggles <justin.ruggles at gmail.com> | Mon May 14 15:56:39 2012 -0400| [95a98ab3f0439df82a907233f80a7404b987e838] | committer: Justin Ruggles

ac3dsp: simplify x86 versions of ac3_max_msb_abs_int16

Simplifies the code by using cpuflags and a new macro.
Also fixes the invalid use of the MMX2 pshufw operation in the MMX-only
function.

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=95a98ab3f0439df82a907233f80a7404b987e838
---

 libavcodec/x86/ac3dsp.asm   |   59 ++++++++++++++++++++++++++++---------------
 libavcodec/x86/ac3dsp_mmx.c |   10 ++++----
 2 files changed, 43 insertions(+), 26 deletions(-)

diff --git a/libavcodec/x86/ac3dsp.asm b/libavcodec/x86/ac3dsp.asm
index 1438811..f1e73d3 100644
--- a/libavcodec/x86/ac3dsp.asm
+++ b/libavcodec/x86/ac3dsp.asm
@@ -91,12 +91,36 @@ AC3_EXPONENT_MIN sse2
 ;        This is used for mmxext and sse2 because they have pminsw/pmaxsw.
 ;-----------------------------------------------------------------------------
 
-%macro AC3_MAX_MSB_ABS_INT16 2
-cglobal ac3_max_msb_abs_int16_%1, 2,2,5, src, len
+; logical 'or' of 4 or 8 words in an mmx or xmm register into the low word
+%macro OR_WORDS_HORIZ 2 ; src, tmp
+%if cpuflag(sse2)
+    movhlps     %2, %1
+    por         %1, %2
+    pshuflw     %2, %1, q0032
+    por         %1, %2
+    pshuflw     %2, %1, q0001
+    por         %1, %2
+%elif cpuflag(mmx2)
+    pshufw      %2, %1, q0032
+    por         %1, %2
+    pshufw      %2, %1, q0001
+    por         %1, %2
+%else ; mmx
+    movq        %2, %1
+    psrlq       %2, 32
+    por         %1, %2
+    movq        %2, %1
+    psrlq       %2, 16
+    por         %1, %2
+%endif
+%endmacro
+
+%macro AC3_MAX_MSB_ABS_INT16 1
+cglobal ac3_max_msb_abs_int16, 2,2,5, src, len
     pxor        m2, m2
     pxor        m3, m3
 .loop:
-%ifidn %2, min_max
+%ifidn %1, min_max
     mova        m0, [srcq]
     mova        m1, [srcq+mmsize]
     pminsw      m2, m0
@@ -104,7 +128,7 @@ cglobal ac3_max_msb_abs_int16_%1, 2,2,5, src, len
     pmaxsw      m3, m0
     pmaxsw      m3, m1
 %else ; or_abs
-%ifidn %1, mmx
+%if notcpuflag(ssse3)
     mova        m0, [srcq]
     mova        m1, [srcq+mmsize]
     ABS2        m0, m1, m3, m4
@@ -119,34 +143,27 @@ cglobal ac3_max_msb_abs_int16_%1, 2,2,5, src, len
     add       srcq, mmsize*2
     sub       lend, mmsize
     ja .loop
-%ifidn %2, min_max
+%ifidn %1, min_max
     ABS2        m2, m3, m0, m1
     por         m2, m3
 %endif
-%ifidn mmsize, 16
-    movhlps     m0, m2
-    por         m2, m0
-%endif
-    PSHUFLW     m0, m2, 0xe
-    por         m2, m0
-    PSHUFLW     m0, m2, 0x1
-    por         m2, m0
+    OR_WORDS_HORIZ m2, m0
     movd       eax, m2
     and        eax, 0xFFFF
     RET
 %endmacro
 
-INIT_MMX
+INIT_MMX mmx
 %define ABS2 ABS2_MMX
-%define PSHUFLW pshufw
-AC3_MAX_MSB_ABS_INT16 mmx, or_abs
+AC3_MAX_MSB_ABS_INT16 or_abs
+INIT_MMX mmx2
 %define ABS2 ABS2_MMX2
-AC3_MAX_MSB_ABS_INT16 mmxext, min_max
-INIT_XMM
-%define PSHUFLW pshuflw
-AC3_MAX_MSB_ABS_INT16 sse2, min_max
+AC3_MAX_MSB_ABS_INT16 min_max
+INIT_XMM sse2
+AC3_MAX_MSB_ABS_INT16 min_max
+INIT_XMM ssse3
 %define ABS2 ABS2_SSSE3
-AC3_MAX_MSB_ABS_INT16 ssse3, or_abs
+AC3_MAX_MSB_ABS_INT16 or_abs
 
 ;-----------------------------------------------------------------------------
 ; macro used for ff_ac3_lshift_int16() and ff_ac3_rshift_int32()
diff --git a/libavcodec/x86/ac3dsp_mmx.c b/libavcodec/x86/ac3dsp_mmx.c
index d6bb469..1a43183 100644
--- a/libavcodec/x86/ac3dsp_mmx.c
+++ b/libavcodec/x86/ac3dsp_mmx.c
@@ -27,10 +27,10 @@ extern void ff_ac3_exponent_min_mmx   (uint8_t *exp, int num_reuse_blocks, int n
 extern void ff_ac3_exponent_min_mmxext(uint8_t *exp, int num_reuse_blocks, int nb_coefs);
 extern void ff_ac3_exponent_min_sse2  (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
 
-extern int ff_ac3_max_msb_abs_int16_mmx   (const int16_t *src, int len);
-extern int ff_ac3_max_msb_abs_int16_mmxext(const int16_t *src, int len);
-extern int ff_ac3_max_msb_abs_int16_sse2  (const int16_t *src, int len);
-extern int ff_ac3_max_msb_abs_int16_ssse3 (const int16_t *src, int len);
+extern int ff_ac3_max_msb_abs_int16_mmx  (const int16_t *src, int len);
+extern int ff_ac3_max_msb_abs_int16_mmx2 (const int16_t *src, int len);
+extern int ff_ac3_max_msb_abs_int16_sse2 (const int16_t *src, int len);
+extern int ff_ac3_max_msb_abs_int16_ssse3(const int16_t *src, int len);
 
 extern void ff_ac3_lshift_int16_mmx (int16_t *src, unsigned int len, unsigned int shift);
 extern void ff_ac3_lshift_int16_sse2(int16_t *src, unsigned int len, unsigned int shift);
@@ -67,7 +67,7 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact)
     }
     if (mm_flags & AV_CPU_FLAG_MMX2 && HAVE_MMX2) {
         c->ac3_exponent_min = ff_ac3_exponent_min_mmxext;
-        c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmxext;
+        c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmx2;
     }
     if (mm_flags & AV_CPU_FLAG_SSE && HAVE_SSE) {
         c->float_to_fixed24 = ff_float_to_fixed24_sse;