[Mesa-dev] [RFC] nir/opt_peephole_ffma: Bypass fusion if any operand of fadd and fmul is a const

[Matt Turner]

On Fri, Sep 18, 2015 at 12:49 AM, Eduardo Lima Mitev <elima at igalia.com> wrote:
> When both fadd and fmul instructions have at least one operand that is a
> constant and it is only used once, the total number of instructions can
> be reduced from 3 (1 ffma + 2 load_const) to 2 (1 fmul + 1 fadd); because
> the constants will be progagated as immediate operands of fmul and fadd.
>
> This patch modifies opt_peephole_ffma pass to detect this situation and
> bails-out fusing fmul+fadd into ffma.
>
> As shown in shader-db results below, it seems to help a good bunch. However,
> there are some caveats:
>
> * It seems i965 specific, so I'm not sure if modifying the NIR pass
> directly is desired, as opposed to moving this to the backend.
>
> * There are still a high number of HURTs, but these could be reduced by being
> more specific in the conditions to bailout.
>
> total instructions in shared programs: 1683959 -> 1677447 (-0.39%)
> instructions in affected programs:     604918 -> 598406 (-1.08%)
> helped:                                4633
> HURT:                                  804
> GAINED:                                0
> LOST:                                  0
> ---

Interesting -- yeah, I've thought about doing this as well. It was
more difficult before because with GLSL IR (where I was trying to do
it) it wasn't possible to determine if the constant was used by
multiple 3-src instructions. Actually, your check might be able to be
more refined to consider only uses of 3-src instructions.

But that's getting kind of hardware-specific.

Perhaps another approach would be to modify the
opt_combine_constants() pass to split MADs under some circumstances --
e.g., it accounts for the only use of a constant we would otherwise
have to promote. But of course we don't have that pass for the vec4
backend.

In the mean time, I've sent a related patch that may be of interest:
"[PATCH] nir: Don't fuse fmul into ffma if used by more than 4 fadds."

This patch, applied on top of mine gives these results on Haswell:

Total:
total instructions in shared programs: 6595563 -> 6584885 (-0.16%)
instructions in affected programs:     1183608 -> 1172930 (-0.90%)
helped:                                8074
HURT:                                  842
GAINED:                                4

FS:
total instructions in shared programs: 4863484 -> 4859884 (-0.07%)
instructions in affected programs:     554042 -> 550442 (-0.65%)
helped:                                3072
HURT:                                  38
GAINED:                                4

VS:
total instructions in shared programs: 1729224 -> 1722146 (-0.41%)
instructions in affected programs:     629566 -> 622488 (-1.12%)
total loops in shared programs:        221 -> 221 (0.00%)
helped:                                5002
HURT:                                  804

Another thing to consider for the vec4 backend is that vec4 uniforms
have to be unpacked for use by 3-src instructions (see the
VEC4_OPCODE_UNPACK_UNIFORM opcode). We CSE the unpacking operations,
but they often do account for increases in instruction counts.