| Age | Commit message (Collapse) | Author |
|---|
|
This patch fixes a bug in the memcpy that doesn't save/restore
the FR12-FR15 registers (callee save registers in ST40 ABI) while
copying many cache lines with FPU in single paired precision mode
and by using all FPU registers (DR and XD).
Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Reviewed-by: Carmelo Amoroso <carmelo.amoroso@st.com>
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
|
|
After exploring different prefetch distance-degree combinations
in this new update of the memcpy function, a new loop has been added
for moving many cache lines with an aggressive prefetching schema.
Prefetch has been removed when move few cache line aligned blocks.
As final result, this memcpy gives us the same performances for small
sizes (we already had!) and better numbers for big copies.
In case of SH4-300 CPU Series, benchmarks show a gain of ~20% for sizes
from 4KiB to 256KiB.
In case of the SH4-200, there is a gain of ~40% for sizes bigger than
32KiB.
Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
|
|
With this patch the movca.l instruction is used within the memset.
The current memset implementation only uses the FPU and there is
an real gain for all the sizes.
Adding the movca.l instruction numbers always are better than the generic code.
There is a big gain for size greater than 64 KiB but number are worst for 4-32KiB
sizes compared with the implementation without movca.l.
Time Memory Bandwidth (Mbytes)
-------------------------------------------------
Generic SH4 SH4
(FPU) (FPU+movca.l)
-------------------------------------------------
512 1143 1998 1596
1 KiB 1273 2567 1915
2 KiB 1350 2993 2128
4-32KiB 1391 3262 2252
64KiB-16MiB 170 186 *830*
Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
|
|
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
|
|
This patch disables SH-4 optimizations that rely on the FPU when
building for variants that don't have an FPU, such as SH-4AL.
Signed-off-by: Andrew Stubbs <ams@codesourcery.com>
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
|
|
This optimization is based on prefetching and 64bit data transfer via FPU
(only for the little endianess)
Tests shows that:
----------------------------------------
Memory bandwidth | Gain
| sh4-300 | sh4-200
----------------------------------------
512 bytes to 16KiB | ~20% | ~25%
from 32KiB to 16MiB | ~190% | ~5%
----------------------------------------
Signed-off-by: Austin Foxley <austinf@cetoncorp.com>
Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
|
|
Signed-off-by: Hideo Saito <saito@densan.co.jp>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Signed-off-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
See Linux Kernel commit:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=e08b954c9a140f2062649faec72514eb505f18c3
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
|
|
@STMicroelectronics).
This implementation is based on 'backward copying'.
Signed-off-by: Carmelo Amoroso <carmelo.amoroso@st.com>
|
