[Pixman] [PATCH 5/9 repost] armv7: Use prefetch for small-width images too

[Ben Avison]

After discovering that the ARMv6 optimised fast paths often out-performed
the ARMv7 ones on a Cortex-A7, particularly on the RT benchmark, I found
that the problem was due to the fact that the ARMv7 macros didn't attempt
any sort of prefetch for small images (fewer than pixblock_size * 2 pixels
across).

Since a pixblock is chosen to be no larger than a cacheline, and is in many
cases smaller, it seemed a reasonable compromise to avoid adding a lot of
complexity by simply doing one prefetch for the start of a pixel row when
starting to process the preceding one, and that is what this patch does.

I compared the effect of using LDRB (which is what is currently used at the
end of each long pixel row) against PLD for each of the source and
destination buffers for a selection of common operations: src_8888_8888,
over_8888_8888 and add_8888_8888, and in each case PLD of both buffers was
the most beneficial. PLDW didn't make any measurable difference.

The overall effect of this patch on the three operations is as follows
(L1, L2 and M tests can be ignored because they're known not to involve the
use of short rows):

src_8888_8888

    Before          After
    Mean   StdDev   Mean   StdDev  Confidence  Change
HT  60.8   0.1      61.1   0.1     100.0%      +0.6%
VT  61.0   0.3      62.6   0.2     100.0%      +2.6%
R   45.5   0.2      46.2   0.2     100.0%      +1.5%
RT  19.8   0.0      21.4   0.0     100.0%      +7.8%

over_8888_8888

    Before          After
    Mean   StdDev   Mean   StdDev  Confidence  Change
HT  40.2   0.1      40.7   0.4     100.0%      +1.0%
VT  35.5   0.2      37.9   0.3     100.0%      +6.7%
R   32.8   0.0      33.8   0.3     100.0%      +3.0%
RT  12.9   0.0      15.6   0.2     100.0%      +21.4%

add_8888_8888

    Before          After
    Mean   StdDev   Mean   StdDev  Confidence  Change
HT  51.0   0.6      51.9   0.5     100.0%      +1.7%
VT  44.0   0.4      46.8   0.5     100.0%      +6.3%
R   39.6   0.5      41.0   0.4     100.0%      +3.5%
RT  15.2   0.2      18.0   0.2     100.0%      +18.5%

Signed-off-by: Ben Avison <bavison at riscosopen.org>
---
 pixman/pixman-arm-neon-asm.h |    9 +++++++++
 1 files changed, 9 insertions(+), 0 deletions(-)

diff --git a/pixman/pixman-arm-neon-asm.h b/pixman/pixman-arm-neon-asm.h
index 03257cc..a116e47 100644
--- a/pixman/pixman-arm-neon-asm.h
+++ b/pixman/pixman-arm-neon-asm.h
@@ -881,6 +881,15 @@ local skip1
  * nor prefetch are used.
  */
 8:
+.if src_bpp_shift >= 0
+    PF pld, [SRC, SRC_STRIDE, lsl #src_bpp_shift]
+.endif
+.if dst_r_bpp != 0
+    PF pld, [DST_R, DST_STRIDE, lsl #dst_bpp_shift]
+.endif
+.if mask_bpp_shift >= 0
+    PF pld, [MASK, MASK_STRIDE, lsl #mask_bpp_shift]
+.endif
     /* Process exactly pixblock_size pixels if needed */
     tst         W, #pixblock_size
     beq         1f
-- 
1.7.5.4