Assembly version of put_pixels. This is currently the function that

takes the most time, and it allows for more efficient unaligned access
and better control over memory latencies.

Originally committed as revision 711 to svn://svn.ffmpeg.org/ffmpeg/trunk

2 files changed, 121 insertions, 2 deletions

diff --git a/libavcodec/alpha/dsputil_alpha.c b/libavcodec/alpha/dsputil_alpha.c
index 5e1aa20933..942eef780c 100644
--- a/libavcodec/alpha/dsputil_alpha.c
+++ b/libavcodec/alpha/dsputil_alpha.c
@@ -22,6 +22,8 @@
 
 void simple_idct_axp(DCTELEM *block);
 
+void put_pixels_axp_asm(uint8_t *block, const uint8_t *pixels,
+			int line_size, int h);
 void put_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels,
 				int line_size);
 void add_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels, 
@@ -232,12 +234,12 @@ static inline UINT64 avg4_no_rnd(UINT64 l1, UINT64 l2, UINT64 l3, UINT64 l4)
 
 void dsputil_init_alpha(void)
 {
-    put_pixels_tab[0] = put_pixels_axp;
+    put_pixels_tab[0] = put_pixels_axp_asm;
     put_pixels_tab[1] = put_pixels_x2_axp;
     put_pixels_tab[2] = put_pixels_y2_axp;
     put_pixels_tab[3] = put_pixels_xy2_axp;
 
-    put_no_rnd_pixels_tab[0] = put_pixels_axp;
+    put_no_rnd_pixels_tab[0] = put_pixels_axp_asm;
     put_no_rnd_pixels_tab[1] = put_no_rnd_pixels_x2_axp;
     put_no_rnd_pixels_tab[2] = put_no_rnd_pixels_y2_axp;
     put_no_rnd_pixels_tab[3] = put_no_rnd_pixels_xy2_axp;
diff --git a/libavcodec/alpha/dsputil_alpha_asm.S b/libavcodec/alpha/dsputil_alpha_asm.S
index a0d9562089..5349e443cd 100644
--- a/libavcodec/alpha/dsputil_alpha_asm.S
+++ b/libavcodec/alpha/dsputil_alpha_asm.S
@@ -44,6 +44,123 @@
         .text
 
 /************************************************************************
+ * void put_pixels_axp_asm(uint8_t *block, const uint8_t *pixels,
+ *                         int line_size, int h)
+ */
+        .align 6
+        .globl put_pixels_axp_asm
+        .ent put_pixels_axp_asm
+put_pixels_axp_asm:
+        .frame sp, 0, ra
+        .prologue 0
+
+#ifdef HAVE_GPROF
+        lda     AT, _mcount
+        jsr     AT, (AT), _mcount
+#endif
+
+        and     a1, 7, t0
+        beq     t0, $aligned
+
+        .align 4
+$unaligned:
+        ldq_u   t0, 0(a1)
+        ldq_u   t1, 8(a1)
+        addq    a1, a2, a1
+        nop
+
+        ldq_u   t2, 0(a1)
+        ldq_u   t3, 8(a1)
+        addq    a1, a2, a1
+        nop
+
+	ldq_u   t4, 0(a1)
+        ldq_u   t5, 8(a1)
+        addq    a1, a2, a1
+        nop
+
+        ldq_u   t6, 0(a1)
+        ldq_u   t7, 8(a1)
+        extql   t0, a1, t0
+        addq    a1, a2, a1
+
+        extqh   t1, a1, t1
+        addq    a0, a2, t8
+        extql   t2, a1, t2
+        addq    t8, a2, t9
+
+        extqh   t3, a1, t3
+        addq    t9, a2, ta
+        extql   t4, a1, t4
+        or      t0, t1, t0
+
+        extqh   t5, a1, t5
+        or      t2, t3, t2
+        extql   t6, a1, t6
+        or      t4, t5, t4
+
+        extqh   t7, a1, t7
+        or      t6, t7, t6
+        stq     t0, 0(a0)
+        stq     t2, 0(t8)
+
+        stq     t4, 0(t9)
+        subq    a3, 4, a3
+        stq     t6, 0(ta)
+        addq    ta, a2, a0
+
+        bne     a3, $unaligned
+        ret
+
+        .align 4
+$aligned:
+        ldq     t0, 0(a1)
+        addq    a1, a2, a1
+        ldq     t1, 0(a1)
+        addq    a1, a2, a1
+
+        ldq     t2, 0(a1)
+        addq    a1, a2, a1
+        ldq     t3, 0(a1)
+        addq    a1, a2, a1
+
+        ldq     t4, 0(a1)
+        addq    a1, a2, a1
+        ldq     t5, 0(a1)
+        addq    a1, a2, a1
+
+        ldq     t6, 0(a1)
+        addq    a1, a2, a1
+        ldq     t7, 0(a1)
+        addq    a1, a2, a1
+
+        addq    a0, a2, t8
+        stq     t0, 0(a0)
+        addq    t8, a2, t9
+        stq     t1, 0(t8)
+
+        addq    t9, a2, ta
+        stq     t2, 0(t9)
+        addq    ta, a2, tb
+        stq     t3, 0(ta)
+
+        addq    tb, a2, tc
+        stq     t4, 0(tb)
+        addq    tc, a2, td
+        stq     t5, 0(tc)
+
+        addq    td, a2, te
+        stq     t6, 0(td)
+        addq    te, a2, a0
+        stq     t7, 0(te)
+
+        subq    a3, 8, a3
+        bne     a3, $aligned
+
+        ret
+        .end put_pixels_axp_asm
+
+/************************************************************************
  * void put_pixels_clamped_mvi_asm(const DCTELEM *block, uint8_t *pixels, 
  *                                 int line_size)
  */