Mesa (main): panfrost: Enable more FP16 caps on Valhall

Tue Jun 21 22:55:23 UTC 2022

Module: Mesa
Branch: main
Commit: 7d84bb00dc4d75652e07d0a3d9f0663431f9bdc8
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=7d84bb00dc4d75652e07d0a3d9f0663431f9bdc8

Author: Alyssa Rosenzweig <alyssa at collabora.com>
Date:   Fri Jun 17 12:53:49 2022 -0400

panfrost: Enable more FP16 caps on Valhall

This brings the FP16 capabilities of Valhall to parity with Bifrost.
Supporting FP16 constant buffers in particular reduces ALU in a ton of GLES
shaders, so that's a nice win. FP16 derivatives get vectorized which is a big
win where that applies, but they are considerably less common.

The lost shaders are from enabling PIPE_SHADER_CAP_FP16_CONST_BUFFERS (these
shaders compile on Midgard but not on Bifrost). The shaders in question declare
the same uniform in linked vertex and fragment shaders with different
precisions. This is contrary to the GLSL ES specification, which states
precisions must match for default uniforms of linked shaders. All the lost
shaders are in 8 Ball Pool and Hill Climb Racing. As those are proprietary
games, if that becomes a problem in the future, drirc is the solution.

total instructions in shared programs: 2697897 -> 2674595 (-0.86%)
instructions in affected programs: 1019922 -> 996620 (-2.28%)
helped: 4838
HURT: 2599
helped stats (abs) min: 1.0 max: 52.0 x̄: 7.13 x̃: 5
helped stats (rel) min: 0.16% max: 46.51% x̄: 8.04% x̃: 5.33%
HURT stats (abs)   min: 1.0 max: 36.0 x̄: 4.30 x̃: 3
HURT stats (rel)   min: 0.17% max: 133.33% x̄: 10.53% x̃: 3.85%
95% mean confidence interval for instructions value: -3.32 -2.95
95% mean confidence interval for instructions %-change: -1.89% -1.22%
Instructions are helped.

total cycles in shared programs: 141764.61 -> 140602.88 (-0.82%)
cycles in affected programs: 5728.22 -> 4566.48 (-20.28%)
helped: 665
HURT: 89
helped stats (abs) min: 0.015625 max: 15.0 x̄: 1.75 x̃: 0
helped stats (rel) min: 0.30% max: 61.54% x̄: 11.17% x̃: 4.62%
HURT stats (abs)   min: 0.015625 max: 0.265625 x̄: 0.04 x̃: 0
HURT stats (rel)   min: 0.30% max: 66.67% x̄: 6.77% x̃: 1.94%
95% mean confidence interval for cycles value: -1.77 -1.31
95% mean confidence interval for cycles %-change: -10.11% -7.99%
Cycles are helped.

total fma in shared programs: 22577.56 -> 22575.91 (<.01%)
fma in affected programs: 2422.78 -> 2421.12 (-0.07%)
helped: 533
HURT: 653
helped stats (abs) min: 0.015625 max: 0.0625 x̄: 0.03 x̃: 0
helped stats (rel) min: 0.30% max: 50.00% x̄: 8.25% x̃: 1.35%
HURT stats (abs)   min: 0.015625 max: 0.125 x̄: 0.03 x̃: 0
HURT stats (rel)   min: 0.19% max: 100.00% x̄: 4.53% x̃: 2.08%
95% mean confidence interval for fma value: -0.00 0.00
95% mean confidence interval for fma %-change: -1.98% -0.44%
Inconclusive result (value mean confidence interval includes 0).

total cvt in shared programs: 14460.95 -> 14122.50 (-2.34%)
cvt in affected programs: 6159.02 -> 5820.56 (-5.50%)
helped: 4827
HURT: 2577
helped stats (abs) min: 0.015625 max: 0.796875 x̄: 0.11 x̃: 0
helped stats (rel) min: 0.20% max: 81.82% x̄: 17.78% x̃: 12.90%
HURT stats (abs)   min: 0.015625 max: 0.546875 x̄: 0.07 x̃: 0
HURT stats (rel)   min: 0.00% max: 600.00% x̄: 43.66% x̃: 13.04%
95% mean confidence interval for cvt value: -0.05 -0.04
95% mean confidence interval for cvt %-change: 2.28% 4.93%
Inconclusive result (value mean confidence interval and %-change mean confidence interval disagree).

total sfu in shared programs: 7593.56 -> 7571.06 (-0.30%)
sfu in affected programs: 357.19 -> 334.69 (-6.30%)
helped: 149
HURT: 1
helped stats (abs) min: 0.0625 max: 0.25 x̄: 0.15 x̃: 0
helped stats (rel) min: 5.26% max: 36.36% x̄: 6.79% x̃: 5.56%
HURT stats (abs)   min: 0.0625 max: 0.0625 x̄: 0.06 x̃: 0
HURT stats (rel)   min: 3.57% max: 3.57% x̄: 3.57% x̃: 3.57%
95% mean confidence interval for sfu value: -0.16 -0.14
95% mean confidence interval for sfu %-change: -7.51% -5.93%
Sfu are helped.

total v in shared programs: 8722.62 -> 8722.31 (<.01%)
v in affected programs: 1.62 -> 1.31 (-19.23%)
helped: 2
HURT: 0

total ls in shared programs: 129666 -> 128494 (-0.90%)
ls in affected programs: 4163 -> 2991 (-28.15%)
helped: 192
HURT: 0
helped stats (abs) min: 1.0 max: 15.0 x̄: 6.10 x̃: 5
helped stats (rel) min: 4.35% max: 75.00% x̄: 30.23% x̃: 26.32%
95% mean confidence interval for ls value: -6.67 -5.54
95% mean confidence interval for ls %-change: -32.67% -27.79%
Ls are helped.

total quadwords in shared programs: 1461496 -> 1449768 (-0.80%)
quadwords in affected programs: 273592 -> 261864 (-4.29%)
helped: 1992
HURT: 687
helped stats (abs) min: 8.0 max: 24.0 x̄: 8.76 x̃: 8
helped stats (rel) min: 1.43% max: 50.00% x̄: 16.30% x̃: 11.11%
HURT stats (abs)   min: 8.0 max: 16.0 x̄: 8.31 x̃: 8
HURT stats (rel)   min: 1.92% max: 100.00% x̄: 36.39% x̃: 25.00%
95% mean confidence interval for quadwords value: -4.67 -4.08
95% mean confidence interval for quadwords %-change: -3.95% -1.62%
Quadwords are helped.

total threads in shared programs: 53496 -> 53551 (0.10%)
threads in affected programs: 112 -> 167 (49.11%)
helped: 74
HURT: 19
helped stats (abs) min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
helped stats (rel) min: 100.00% max: 100.00% x̄: 100.00% x̃: 100.00%
HURT stats (abs)   min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 50.00% max: 50.00% x̄: 50.00% x̃: 50.00%
95% mean confidence interval for threads value: 0.42 0.76
95% mean confidence interval for threads %-change: 56.83% 81.88%
Threads are helped.

total loops in shared programs: 128 -> 127 (-0.78%)
loops in affected programs: 1 -> 0
helped: 1
HURT: 0

total fills in shared programs: 684 -> 672 (-1.75%)
fills in affected programs: 160 -> 148 (-7.50%)
helped: 2
HURT: 0

Signed-off-by: Alyssa Rosenzweig <alyssa at collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17101>

---

 src/gallium/drivers/panfrost/pan_screen.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/gallium/drivers/panfrost/pan_screen.c b/src/gallium/drivers/panfrost/pan_screen.c
index b7cc255645c..9a365aa2b9d 100644
--- a/src/gallium/drivers/panfrost/pan_screen.c
+++ b/src/gallium/drivers/panfrost/pan_screen.c
@@ -431,11 +431,11 @@ panfrost_get_shader_param(struct pipe_screen *screen,
                 return !is_nofp16;
         case PIPE_SHADER_CAP_FP16_DERIVATIVES:
         case PIPE_SHADER_CAP_FP16_CONST_BUFFERS:
-                return pan_is_bifrost(dev) && !is_nofp16;
+                return dev->arch >= 6 && !is_nofp16;
         case PIPE_SHADER_CAP_INT16:
                 /* XXX: Advertise this CAP when a proper fix to lower_precision
                  * lands. GLSL IR validation failure in glmark2 -bterrain */
-                return pan_is_bifrost(dev) && !is_nofp16 && is_deqp;
+                return dev->arch >= 6 && !is_nofp16 && is_deqp;
 
         case PIPE_SHADER_CAP_INT64_ATOMICS:
         case PIPE_SHADER_CAP_DROUND_SUPPORTED:

