Saturation in compiler generated Stickify

[AlexandreEichenberger]

This algorithm adds the equivalent of this (per William Jones) to the stickify compiler generated code.

static const __vector float min_vec = {-8573157376.F, -8573157376.F, -8573157376.F, -8573157376.F};
static const __vector float max_vec = {8573157376.F, 8573157376.F, 8573157376.F, 8573157376.F};

float* a = ...;
float* b = ...;
int16_t* out = ...;

__vector float* vec_a = (__vector float*)a;
__vector float* vec_b = (__vector float*)b;
__vector short* vec_out = (__vector short*)out;

*vec_out = vec_round_from_fp32(
    vec_max(vec_min(*vec_a, max_vec), min_vec),
    vec_max(vec_min(*vec_b, max_vec), min_vec),
    0);

When running on z16 with Roberta base 11 and 6x384 in sequential mode

zdnn
  zhigh.Stick, 85, 0.5742549, 48.8116667, 5.6%

compiler gen code without saturation
  zhigh.Stick, 85, 0.3593294, 30.5430000, 3.7%

compiler gen WITH saturation
  zhigh.Stick, 85, 0.4358196, 37.0446667, 4.4%

As shown, saturation runs still faster than zdnn based code.

Checked the asm generated, pattern use vector compare and vector select, which I believe is the best pattern at this time for floating point numbers

    81e6:	e7 91 10 00 24 ea 	vfchesb	%v9,%v17,%v1                    <<<<==== vadd
    81ec:	e7 20 20 00 21 8d 	vsel	%v2,%v0,%v2,%v18                   <<<<==== vshuffle

[AlexandreEichenberger]

Updated the code to gen the min/max directly.

update performance:

zdnn
  zhigh.Stick, 85, 0.5742549, 48.8116667, 5.6%

compiler gen code without saturation
  zhigh.Stick, 85, 0.3593294, 30.5430000, 3.7%

compiler gen WITH saturation
  zhigh.Stick, 85, 0.3676314, 31.2486667, 3.8%

For a performance reduction of 3% (compared to compiler generated without saturation), and still 55% faster than zDNN.

Verified that it generates the right code

    8050:	e7 11 00 40 2c ee 	vfminsb	%v17,%v17,%v0,4                 <<<<==== vadd
...
    8062:	e7 22 10 40 20 ef 	vfmaxsb	%v2,%v2,%v1,4                   <<<<==== vadd

[tungld]

LGTM. Thanks for a quick update!

[jenkins-droid]

Jenkins Linux s390x Build #15143 [push] Saturation in compiler g... started at 22:33

[jenkins-droid]

Jenkins Linux amd64 Build #15138 [push] Saturation in compiler g... started at 21:33

[jenkins-droid]

Jenkins Linux ppc64le Build #14168 [push] Saturation in compiler g... started at 22:44

[jenkins-droid]

Jenkins Linux amd64 Build #15138 [push] Saturation in compiler g... passed after 1 hr 13 min

[jenkins-droid]

Jenkins Linux s390x Build #15143 [push] Saturation in compiler g... passed after 1 hr 39 min

[jenkins-droid]

Jenkins Linux ppc64le Build #14168 [push] Saturation in compiler g... passed after 2 hr 9 min