The setup.py file for Python states:

        if (not cross_compiling and
                os.uname().machine == "x86_64" and
                sys.maxsize >  2**32):
            # Every x86_64 machine has at least SSE2.  Check for sys.maxsize
            # in case that kernel is 64-bit but userspace is 32-bit.
            blake2_macros.append(('BLAKE2_USE_SSE', '1'))

While the assertion about having SSE2 is true, it doesn't mean that it's worthwhile to use. I've tested pure (i.e. without SSSE3 and so on) on three different machines, getting the following results:

Athlon64 X2 (SSE2 is the best supported variant), 540 MiB of data:

SSE2: [5.189988004000043, 5.070812243997352]
ref:  [2.0161159170020255, 2.0475422790041193]

Core i3, same data file:

SSE2: [1.924425926999902, 1.92461746999993, 1.9298037500000191]
ref:  [1.7940209749999667, 1.7900855569999976, 1.7835538760000418]

Xeon E5630 server, 230 MiB data file:

SSE2: [0.7671358410007088, 0.7797677099879365, 0.7648976119962754]
ref:  [0.5784736709902063, 0.5717909929953748, 0.5717219939979259]

So in all the tested cases, pure SSE2 implementation is *slower* than the reference implementation. SSSE3 and other variants are faster and AFAIU they are enabled automatically based on CFLAGS, so it doesn't matter for most of the systems.

However, for old CPUs that do not support SSSE3, the choice of SSE2 makes the algorithm prohibitively slow -- it's 2.5 times slower than the reference implementation!

New changeset 1aa00ff383c43335e4a5044274617dbf59bc839e by Benjamin Peterson (Michał Górny) in branch 'master':
fixes bpo-31834: Use optimized code for BLAKE2 only with SSSE3+ (#4066)
https://github.com/python/cpython/commit/1aa00ff383c43335e4a5044274617dbf59bc839e

I'm pretty sure that your PR has disabled all SSE optimizations. AFAIK gcc does not enable SSE3 and SSE4 on X86_64 by default.

$ gcc -dM -E - < /dev/null | grep SSE
#define __SSE2_MATH__ 1
#define __SSE_MATH__ 1
#define __SSE2__ 1
#define __SSE__ 1

You have to set a compiler flag like -msse4

$ gcc -msse4 -dM -E - < /dev/null | grep SSE
#define __SSE4_1__ 1
#define __SSE4_2__ 1
#define __SSE2_MATH__ 1
#define __SSE_MATH__ 1
#define __SSE2__ 1
#define __SSSE3__ 1
#define __SSE__ 1
#define __SSE3__ 1

> AFAIK gcc does not enable SSE3 and SSE4 on X86_64 by default.

Linux now supports multiple variants of the same function, one variant per CPU type, the binding is done when a library is loaded. But I don't know how to implement that :-(

There is target_clones("sse4.1,avx") the function attribute in GCC for example. It compiles a function twice, once for generic CPU, once for SSE4.1.

See also ifunc: "indirect function", "CPU dispatch" or "function resolver".

On Tue, Oct 24, 2017, at 00:25, Christian Heimes wrote:
> 
> Christian Heimes <lists@cheimes.de> added the comment:
> 
> I'm pretty sure that your PR has disabled all SSE optimizations. AFAIK
> gcc does not enable SSE3 and SSE4 on X86_64 by default.
> 
> $ gcc -dM -E - < /dev/null | grep SSE
> #define __SSE2_MATH__ 1
> #define __SSE_MATH__ 1
> #define __SSE2__ 1
> #define __SSE__ 1

Before this patch, this would cause blake2b.c to use slow SSE2 only
instruction, though, right?

It seems to me this represents an improvement or the status quo in all
cases. With no extra GCC flags, the reference implementation is used
rather than a slow SSE2 implementation. If extra -m flags are in CFLAGS,
the fastest implementation for the target is used.

@tiran, can we close this again?