On Mon, Mar 19, 2018 at 1:30 PM, Eric Biggers <ebiggers3@xxxxxxxxx> wrote:
> On Mon, Mar 19, 2018 at 12:39:55PM -0700, Andiry Xu wrote:
>> On Sun, Mar 11, 2018 at 12:22 PM, Eric Biggers <ebiggers3@xxxxxxxxx> wrote:
>> > On Sun, Mar 11, 2018 at 02:00:13PM +0200, Nikolay Borisov wrote:
>> >> [Adding Herbert Xu to CC since he is the maintainer of the crypto subsys
>> >> maintainer]
>> >>
>> >> On 10.03.2018 20:17, Andiry Xu wrote:
>> >> <snip>
>> >>
>> >> > +static inline u32 nova_crc32c(u32 crc, const u8 *data, size_t len)
>> >> > +{
>> >> > + u8 *ptr = (u8 *) data;
>> >> > + u64 acc = crc; /* accumulator, crc32c value in lower 32b */
>> >> > + u32 csum;
>> >> > +
>> >> > + /* x86 instruction crc32 is part of SSE-4.2 */
>> >> > + if (static_cpu_has(X86_FEATURE_XMM4_2)) {
>> >> > + /* This inline assembly implementation should be equivalent
>> >> > + * to the kernel's crc32c_intel_le_hw() function used by
>> >> > + * crc32c(), but this performs better on test machines.
>> >> > + */
>> >> > + while (len > 8) {
>> >> > + asm volatile(/* 64b quad words */
>> >> > + "crc32q (%1), %0"
>> >> > + : "=r" (acc)
>> >> > + : "r" (ptr), "0" (acc)
>> >> > + );
>> >> > + ptr += 8;
>> >> > + len -= 8;
>> >> > + }
>> >> > +
>> >> > + while (len > 0) {
>> >> > + asm volatile(/* trailing bytes */
>> >> > + "crc32b (%1), %0"
>> >> > + : "=r" (acc)
>> >> > + : "r" (ptr), "0" (acc)
>> >> > + );
>> >> > + ptr++;
>> >> > + len--;
>> >> > + }
>> >> > +
>> >> > + csum = (u32) acc;
>> >> > + } else {
>> >> > + /* The kernel's crc32c() function should also detect and use the
>> >> > + * crc32 instruction of SSE-4.2. But calling in to this function
>> >> > + * is about 3x to 5x slower than the inline assembly version on
>> >> > + * some test machines.
>> >>
>> >> That is really odd. Did you try to characterize why this is the case? Is
>> >> it purely the overhead of dispatching to the correct backend function?
>> >> That's a rather big performance hit.
>> >>
>> >> > + */
>> >> > + csum = crc32c(crc, data, len);
>> >> > + }
>> >> > +
>> >> > + return csum;
>> >> > +}
>> >> > +
>> >
>> > Are you sure that CONFIG_CRYPTO_CRC32C_INTEL was enabled during your tests and
>> > that the accelerated version was being called? Or, perhaps CRC32C_PCL_BREAKEVEN
>> > (defined in arch/x86/crypto/crc32c-intel_glue.c) needs to be adjusted. Please
>> > don't hack around performance problems like this; if they exist, they need to be
>> > fixed for everyone.
>> >
>>
>> I have performed the crc32c test on a Xeon X5647 at 2.93GHz, 14G DDR3
>> memory at 1066MHz platform.
>> You are right that enabling CONFIG_CRYPTO_CRC32C_INTEL improves the
>> performance significantly. nova_crc32c() is still slightly faster than
>> crc32c() with the flag enabled.
>>
>> Result numbers are follows: data size in bytes, latency in ns, column
>> 3 is crc32c() with CONFIG_CRYPTO_CRC32C_INTEL enabled and column 4
>> disabled.
>>
>> data size (bytes) nova_crc32c() crc32c() -enabled
>> crc32c() -disabled
>> 64 19 21 56
>> 128 28 29 99
>> 256 46 43 182
>> 512 82 149 354
>> 1024 157 232 728
>> 2048 305 415 1440
>> 4096 603 725 2869
>>
>
> Probably CRC32C_PCL_BREAKEVEN needs to be adjusted for that CPU, as I suggested
> may be the case; notice that your measured speeds are about the same before 512
> (CRC32C_PCL_BREAKEVEN) bytes, but the crypto API version is slower at >= 512
> bytes. It would be possible to set the breakeven point in
> crc32c_intel_mod_init() depending on the CPU. Again, if the performance is not
> good enough you need to fix it for everyone, not hack around it.
>
We verify that by setting CRC32C_PCL_BREAKEVEN to 8192, the
performance difference between nova_crc32c() and kernel's crc32c() is
negligible. Thanks for the comments, and I will use kernel's crc32c()
in the next version.
Thanks,
Andiry
