intel/fs: Lower large local arrays to scratch

Shader-db results on Kaby Lake:

    total instructions in shared programs: 14929212 -> 14880028 (-0.33%)
    instructions in affected programs: 72428 -> 23244 (-67.91%)
    helped: 6
    HURT: 2
    helped stats (abs) min: 2165 max: 15981 x̄: 8590.00 x̃: 7624
    helped stats (rel) min: 56.06% max: 74.52% x̄: 67.55% x̃: 72.08%
    HURT stats (abs)   min: 1178 max: 1178 x̄: 1178.00 x̃: 1178
    HURT stats (rel)   min: 350.60% max: 361.35% x̄: 355.97% x̃: 355.97%
    95% mean confidence interval for instructions value: -11947.03 -348.97
    95% mean confidence interval for instructions %-change: -125.72% 202.37%
    Inconclusive result (%-change mean confidence interval includes 0).

    total cycles in shared programs: 368585300 -> 342557344 (-7.06%)
    cycles in affected programs: 28144921 -> 2116965 (-92.48%)
    helped: 6
    HURT: 2
    helped stats (abs) min: 1404978 max: 7766106 x̄: 4353922.00 x̃: 3890682
    helped stats (rel) min: 82.01% max: 95.57% x̄: 89.95% x̃: 92.28%
    HURT stats (abs)   min: 47778 max: 47798 x̄: 47788.00 x̃: 47788
    HURT stats (rel)   min: 278.20% max: 282.98% x̄: 280.59% x̃: 280.59%
    95% mean confidence interval for cycles value: -5900438.73 -606550.27
    95% mean confidence interval for cycles %-change: -140.79% 146.16%
    Inconclusive result (%-change mean confidence interval includes 0).

    total spills in shared programs: 9243 -> 8901 (-3.70%)
    spills in affected programs: 2718 -> 2376 (-12.58%)
    helped: 4
    HURT: 4

    total fills in shared programs: 21831 -> 10141 (-53.55%)
    fills in affected programs: 11804 -> 114 (-99.03%)
    helped: 6
    HURT: 2

    total sends in shared programs: 815912 -> 815912 (0.00%)
    sends in affected programs: 0 -> 0
    helped: 0
    HURT: 0

    LOST:   1
    GAINED: 3

The helped shaders are all compute shaders in Aztec Ruins.  There is
also a compute shader in synmark2 OglCSDof that's helped but it doesn't
show up in above shader-db results because it went from SIMD8 to SIMD16.
That shader improves enough to yield an 15-20% performance boost to the
benchmark as a whole on my KBL laptop.  The hurt shaders are a couple
shaders in Kerbal Space Program and a couple in Aztec Ruins.

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>

diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
index 547b60ddc0f7eefb381f200a9ace29f8b7dbbb8f..31bf25bb88ae5089b942c0232957b4ee8d302862 100644 (file)
--- a/src/intel/compiler/brw_nir.c
+++ b/src/intel/compiler/brw_nir.c
@@ -718,6 +718,25 @@ brw_preprocess_nir(const struct brw_compiler *compiler, nir_shader *nir,
 
    OPT(nir_lower_clip_cull_distance_arrays);
 
+   if (devinfo->gen >= 7 && is_scalar) {
+      /* TODO: Yes, we could in theory do this on gen6 and earlier.  However,
+       * that would require plumbing through support for these indirect
+       * scratch read/write messages with message registers and that's just a
+       * pain.  Also, the primary benefit of this is for compute shaders which
+       * won't run on gen6 and earlier anyway.
+       *
+       * The threshold of 128B was chosen semi-arbitrarily.  The idea is that
+       * 128B per channel on a SIMD8 program is 32 registers or 25% of the
+       * register file.  Any array that large is likely to cause pressure
+       * issues.  Also, this value is sufficiently high that the benchmarks
+       * known to suffer from large temporary array issues are helped but
+       * nothing else in shader-db is hurt except for maybe that one kerbal
+       * space program shader.
+       */
+      OPT(nir_lower_vars_to_scratch, nir_var_function_temp, 128,
+          glsl_get_natural_size_align_bytes);
+   }
+
    nir_variable_mode indirect_mask =
       brw_nir_no_indirect_mask(compiler, nir->info.stage);
    OPT(nir_lower_indirect_derefs, indirect_mask);