Skip to content

etnaviv: nir: improve uniform usage for ALU opc

Christian Gmeiner requested to merge austriancoder/mesa:nir-lower-alu-improvements into main

What does this MR do and why?

The current code in lower_alu(..) counts how many const values are used by one ALU opc. If there are used more then one the compiler tries to fix this issues by e.g. resolve with a single combined const src.

We are doing this as some GPUs only allow one const src per ISA instruction. But it is allowed to use the same const for multiple srcs.

Lets have a closer look at a real world shader:

impl main {
        /* preds: */
        vec1 32 ssa_0 = load_const (0x3f800000 = 1.000000)
        vec1 32 ssa_1 = load_const (0x00000000 = 0.000000)
        vec4 32 ssa_2 = intrinsic load_uniform (ssa_1) (base=0, range=1, dest_type=bool32 /*38*/)       /* u_var */
        vec1 32 ssa_4 = fmul ssa_2.x, ssa_2.y
        vec1 32 ssa_11 = load_const (0x00000000 = 0.000000)
        vec1 32 ssa_13 = seq ssa_2.w, ssa_11
        vec1 32 ssa_6 = fmul ssa_2.z, ssa_13
        vec1 32 ssa_7 = fmul ssa_4, ssa_6
        vec1 32 ssa_9 = deref_var &gl_FragColor (shader_out vec4)
        vec4 32 ssa_10 = vec4 ssa_7, ssa_7, ssa_7, ssa_0
        intrinsic store_deref (ssa_9, ssa_10) (wrmask=xyzw /*15*/, access=0)
        /* succs: block_1 */
        block block_1:
}

The current compiler transforms it to:

impl main {
        block block_0:
        /* preds: */
        vec1 32 ssa_0 = load_const (0x3f800000 = 1.000000)
        vec4 32 ssa_14 = load_const (0x00000000, 0x00000001, 0x00000002, 0x00000003) = (0.000000, 0.000000, 0.000000, 0.000000)
        vec2 32 ssa_15 = load_const (0x00000000, 0x00000001) = (0.000000, 0.000000)
        vec1 32 ssa_4 = fmul ssa_15.x, ssa_15.y
        vec2 32 ssa_16 = load_const (0x00000003, 0x00000000) = (0.000000, 0.000000)
        vec1 32 ssa_13 = seq ssa_16.x, ssa_16.y
        vec1 32 ssa_6 = fmul ssa_14.z, ssa_13
        vec1 32 ssa_7 = fmul ssa_4, ssa_6
        vec1 32 ssa_9 = deref_var &gl_FragColor (shader_out vec4)
        vec1 32 ssa_17 = mov ssa_0
        vec4 32 ssa_10 = vec4 ssa_7, ssa_7, ssa_7, ssa_17
        intrinsic store_deref (ssa_9, ssa_10) (wrmask=xyzw /*15*/, access=0)
        /* succs: block_1 */
        block block_1:
}

There is no need to create ssa_15 as we can use ssa_14 for the first fmul.

With this change the compiler creates the following shader:

impl main {
        block block_0:
        /* preds: */
        vec1 32 ssa_0 = load_const (0x3f800000 = 1.000000)
        vec4 32 ssa_14 = load_const (0x00000000, 0x00000001, 0x00000002, 0x00000003) = (0.000000, 0.000000, 0.000000, 0.000000)
        vec1 32 ssa_4 = fmul ssa_14.x, ssa_14.y
        vec2 32 ssa_15 = load_const (0x00000003, 0x00000000) = (0.000000, 0.000000)
        vec1 32 ssa_13 = seq ssa_15.x, ssa_15.y
        vec1 32 ssa_6 = fmul ssa_14.z, ssa_13
        vec1 32 ssa_7 = fmul ssa_4, ssa_6
        vec1 32 ssa_9 = deref_var &gl_FragColor (shader_out vec4)
        vec1 32 ssa_16 = mov ssa_0
        vec4 32 ssa_10 = vec4 ssa_7, ssa_7, ssa_7, ssa_16
        intrinsic store_deref (ssa_9, ssa_10) (wrmask=xyzw /*15*/, access=0)
        /* succs: block_1 */
        block block_1:
}

This change reduces immediate pressure and reduces spend CPU cycles.

No piglit or deqp regression seen.

shader-db results for GC2000:

total instructions in shared programs: 955128 -> 955128 (0.00%)
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0

total temps in shared programs: 85689 -> 85689 (0.00%)
temps in affected programs: 0 -> 0
helped: 0
HURT: 0

total immediates in shared programs: 155428 -> 155240 (-0.12%)
immediates in affected programs: 1840 -> 1652 (-10.22%)
helped: 34
HURT: 1
helped stats (abs) min: 4 max: 16 x̄: 5.65 x̃: 4
helped stats (rel) min: 2.94% max: 33.33% x̄: 16.92% x̃: 16.67%
HURT stats (abs)   min: 4 max: 4 x̄: 4.00 x̃: 4
HURT stats (rel)   min: 14.29% max: 14.29% x̄: 14.29% x̃: 14.29%
95% mean confidence interval for immediates value: -6.57 -4.17
95% mean confidence interval for immediates %-change: -19.83% -12.23%
Immediates are helped.

total loops in shared programs: 0 -> 0
loops in affected programs: 0 -> 0
helped: 0
HURT: 0

LOST:   0
GAINED: 0

Total CPU time (seconds): 102.55 -> 96.35 (-6.05%)

This MR was tested:

  • manually tested on my development machine(s) with the following:
  • repeatedly tested since there is a risk of introducing flakes
  • not applicable

Merge request reports