Re: [Qemu-devel] [PATCH 00/14] softfloat: Use POSIX integer types - benchmarked

[Andreas Färber]

Am 20.01.2012 14:05, schrieb Peter Maydell:
> On 16 January 2012 00:46, Andreas Färber <address@hidden> wrote:
>> For a loop count of 100,000 and 5 runs I got the following results:
>>
>>  current:        138.9-204.1 Whetstone-MIPS
>>  [u]int*_t:      185.2-188.7 Whetstone-MIPS
>>  [u]int_fast*_t: 285.7-294.1 Whetstone-MIPS
>>
>>  Toshiba AC100:  833.3-909.1 Whetstone-MIPS
>>
>> These results seem to indicate that the "fast" POSIX types are indeed
>> somewhat faster, both compared to exact-size POSIX types and to the
>> current state.
> 
> OTOH I did a run of scimark2 and got:
> current tree:
> **                                                              **
> ** SciMark2 Numeric Benchmark, see http://math.nist.gov/scimark **
> ** for details. (Results can be submitted to address@hidden)     **
> **                                                              **
> Using       2.00 seconds min time per kenel.
> Composite Score:           12.98
> FFT             Mflops:     7.66    (N=1024)
> SOR             Mflops:    19.49    (100 x 100)
> MonteCarlo:     Mflops:     6.12
> Sparse matmult  Mflops:    15.34    (N=1000, nz=5000)
> LU              Mflops:    16.28    (M=100, N=100)
> 
> with patches (yours and mine):
> **                                                              **
> ** SciMark2 Numeric Benchmark, see http://math.nist.gov/scimark **
> ** for details. (Results can be submitted to address@hidden)     **
> **                                                              **
> Using       2.00 seconds min time per kenel.
> Composite Score:           11.87
> FFT             Mflops:     7.12    (N=1024)
> SOR             Mflops:    17.66    (100 x 100)
> MonteCarlo:     Mflops:     5.75
> Sparse matmult  Mflops:    14.03    (N=1000, nz=5000)
> LU              Mflops:    14.81    (M=100, N=100)

One difference between our test environments comes to mind: I had tested
only typedefs for the int types, whereas my series also converts flag.
The fixes for lm32, sparc and qemu-tool shouldn't matter here. Your
patches "degrade" some variables from fast types to exact types of course.

Anyway, here's my Whetstone and CoreMark results for
520c0d8d2772ccc9f9275bd934e13ec9b15774e4 (target-sparc: Fix mixup of
uint64 and uint64_t) plus our patches. The margin has shrunk for
Whetstone, and for CoreMark I see a slight degradation of the max.
(I also note a slight Whetstone degradation between Natty and Oneiric.)

Have you tried benchmarking after our preceding patches but before the
actual fast conversion for comparison?

Andreas


master:

C Converted Double Precision Whetstones: 200.0 MIPS
C Converted Double Precision Whetstones: 204.1 MIPS

CoreMark 1.0 : 1287.747086 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1336.094595 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1339.943722 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap


master + PMM + fast:

C Converted Double Precision Whetstones: 204.1 MIPS
C Converted Double Precision Whetstones: 208.3 MIPS

CoreMark 1.0 : 1297.690112 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1299.629606 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1309.071868 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1315.270288 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1318.913216 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1319.870653 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 1321.527686 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap


AC100 (Ubuntu Oneiric):

C Converted Double Precision Whetstones: 769.2 MIPS
C Converted Double Precision Whetstones: 833.3 MIPS

CoreMark 1.0 : 2257.506208 / GCC4.6.1 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 2303.086135 / GCC4.6.1 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 2326.122354 / GCC4.6.1 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 2349.624060 / GCC4.6.1 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 2350.360389 / GCC4.6.1 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap


Pandaboard (openSUSE Factory):

C Converted Double Precision Whetstones: 833.3 MIPS

CoreMark 1.0 : 2304.855562 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 2305.209774 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 2306.273063 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap
CoreMark 1.0 : 2306.627710 / GCC4.6.2 -O2 -static -DPERFORMANCE_RUN=1
-lrt / Heap


Results are sorted, duplicates removed.

whetstone.c was compiled with -mfloat-abi=hard this time, using GCC
(SUSE Linux) 4.6.2, and so was CoreMark except on the AC100 with GCC
(Ubuntu/Linaro 4.6.1-9ubuntu3) 4.6.1.

coremark.exe was run with parameters 0x0 0x0 0x66 0.

-- 
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg