Wednesday, February 5, 2014

freedreno: new compiler

Complementing the hw binning support which landed earlier this year, and is now enabled by default, I've recently pushed the initial round of new-compiler work to mesa.  Initially I was going to keep it on a branch until I had a chance to sort out a better register allocation (RA) algorithm, but the improved instruction scheduling fixed so many bugs that I decided it should be merged in it's current form.

Or explained another way, ever since fedora updated to supertuxkart 0.8.1, about half the tracks had rendering problems and/or triggered gpu hangs.  The new compiler fixed all those problems (and more).  And I like supertuxkart :-)


The original a3xx compiler was more of a simple TGSI translator.  It translated each TGSI opcode into a simple sequence of one or more native instructions.  There was a fixed (per-shader) mapping between TGSI INPUT, OUTPUT, and TEMP vec4 register files to the native (flat) scalar register file.  A not-insignificant part of the code was relatively generic, in concept but not implementation, lowering of TGSI opcodes that relate more closely to old ARB shader instructions, (SCS - Sine Cosine, LIT - Light Coefficients, etc) than the instruction set of any modern GPU.

The simple TGSI translator approach works fine with simple shader ISA's.  It worked ok for a2xx, other than slightly suboptimal register usage.  But the problem is that a3xx (and a4xx) is not such a simple instruction set architecture.  In particular, the instruction scheduling required that the compiler be aware of the shader instruction pipeline(s).  

This was obvious pretty early on in the reverse engineering stage.  But in the early days of the gallium a3xx support, there were too many other things to do... spending the needed time on the compiler then was not really an option.  Instead the "use lots of nop's and hope for the best" strategy was employed.

And while it worked as a stop-gap solution, it turns out that there are a lot of edge cases where "hope for the best" does not really work out that well in practice.  After debugging a number of rendering bugs and piglit failures which all traced back to instruction scheduling problems, it was becoming clear that it was time for a more permanent solution.

In with the new:

First thing I wanted to do before adding a lot more complexity is to rip out a bunch of code.  With that in mind I implemented a generic TGSI lowering pass, to replace about a dozen opcodes with sequences of equivalent simpler instructions.  This probably should be made configurable and moved to util, I think most of the lowerings would be useful to other gallium drivers.

Once the handling of the now unneeded TGSI opcodes was removed, I copied fd3_compiler to fd3_compiler_old.  Originally the plan was to remove this before pushing upstream.  I just wanted a way to compare the results from the original compiler to the new compiler to help during testing and debugging.  But currently shaders with relative addressing need to fall back to the old compiler, so it stays for now.

The next step was to turn ir3 (the a3xx IR), which originates from the fdre-a3xx shader assembler into something more useful.  The approach I settled on (mostly to ease the transition) was to add a few extra "meta-instructions" to hold some additional information which would be needed in later passes, including Φ (Phi) instructions where a result depends on flow control.  Plus a few extra instruction and register flags, the important one being IR3_REG_SSA, used for src register nodes to indicate that the register node points to the dependent instruction.  Now what used to be the compiler (well, roughly 2/3rds of it) is the front-end.  Instead of producing a linear sequence of instructions fed directly to the assembler/codegen, the frontend is now generating a graph of instructions modified by subsequent passes until we have something suitable for codegen.

For each output, we keep the pointer to the instruction which generates that value (at the scalar level), which in turn has the pointer to the instructions generating it's srcs/inputs, and so on.  As before, the front end is generating sequences of scalar instructions for each (written) component in a TGSI vector instruction.  Although now instructions whose result is not used simply has nobody pointing to them so they naturally vanish.

At the same time, mostly to preserve my sanity while debugging, but partially also to make nifty pictures, I implemented an "ir3 dumper" which would dump out the graph in .dot syntax:

The first pass eliminates some redundant moves (some of which come from the front end, some from TGSI itself).  Probably the front end could be a bit more clever about not inserting unneeded moves, but since TGSI has separate INPUT/OUTPUT/TEMP register files, there will always be some extra moves which need eliminating.

After that, I calculate a "depth" for each instruction, where the depth is the number of instruction cycles/slots required to compute that value:

    dd(instr, n): depth(instr->src[n]) + delay(instr->src[n], instr)
    depth(instr): 1 + max(dd(instr, 0), ..., dd(instr, N))

where delay(p,c) gives the required number of instruction slots between an instruction which produces a value and an instruction which consumes a value.

The depth is used for scheduling.  The short version of how it works is to recursively schedule output instructions with the greatest depth until no more instructions can be scheduled (more delay slots needed).  For instructions with multiple inputs/srcs, the unscheduled src instruction with the greatest depth is scheduled first.  Once we hit a point where there are some delay slots to fill, we switch to the next deepest output, and so on until the needed delay slots are filled.  If there are no instructions that can be scheduled, then we insert nop's.

Once the graph is scheduled, we have a linear sequence of instructions, at which point we do RA.  I won't say too much about that now, since it is already a long post and I'll probably change the algorithm.  It is worth noting that some register assignment algorithms can coalesce unneeded moves.  Although moves factor into the scheduling decisions for the a3xx ISA, so I'm not really sure that this is too useful me.

The end result, thanks to a combination of removal of scalar instructions to calculate TGSI vec4 register components which are unused, plus removal of unnecessary moves, plus scheduling other instructions rather than filling with no-op's everywhere, for non trivial shaders it is not uncommon to see the compiler use ~33% the number of instructions, and half the number of registers.


Validating compilers is hard.  Piglit has a number of tests to exercise relatively specific features.  But with games, it isn't always the case that an incorrect shader produces (visually) incorrect results.  And visually incorrect results are not always straightforward to trace back to the problem.  Ie. games typically have many shaders, many draw calls, tracking down the problematic draw and it's shaders is not always easy.

So I wrote a very simplistic emulator for testing the output of the compiler.  I captured the TGSI dumps of all the shaders from various apps (ST_DEBUG=tgsi).  The test app would assemble the TGSI, feed into both the old and new compiler, then run same sets of randomized inputs through the resulting shaders and compare outputs.

There are a few cases where differing output is expected, since the new compiler has slightly more well defined undefined behaviour for shaders that use uninitialized values... to avoid invalid pointers in the graph produced by the front-end, uninitialized values get a 'mov Rdst, immed{0.0}' instruction.  So there are some cases where the resulting shader needs to be manually validated.  But in general this let me test (and debug) the new compiler with 100's of shaders in a relatively short amount of time.


So the obvious question, what does this all mean in terms of performance?  Well, start with the easy results, es2gears[1]:
  • original compiler: ~435fps
  • new compiler: ~539fps
With supertuxkart, the result is a bit easier to show in pictures.  Part of the problem is that the tracks that are heavy enough on the GPU to not be purely CPU limited, didn't actually work before with the original compiler.  That plus, as far as I know, there is no simple benchmark mode which spits out a number at the end, as with xonotic.  So I used the trace points + timechart approach, mentioned in a previous post.

    supertuxkart -f --track fortmagma --profile-laps=1

I manually took one second long captures, in as close to the same spot as possible (just after light turns green):

    ./perf timechart record -a -g -o sleep 1

In this case I was running on an apq8074/a330 device, fwiw.  Our starting point is:

Then once hw binning is in place, we are starting to look more CPU limited than anything:

And with addition of new compiler, the GPU is idle more of the time, but since the GPU is no longer the bottleneck (on the less demanding tracks) there isn't too much change in framerate:

Still, it could help power if the GPU can shut off sooner, and other levels which push the GPU harder benefit.

With binning plus improved compiler, there should not be any more huge performance gaps compared to the blob compiler.  Without linux blob drivers, there is no way to make a real apples to apples comparison, but remaining things that could be improved should be a few percent here and there.  Which is a good thing.  There are still plenty of missing features and undiscovered bugs, I'm sure.  But I'm hopefully that we can at least have things in good shape for a3xx before the first a4xx devices ship ;-)

[1] Windowed apps would benefit somewhat from XA support in DDX, avoiding stall for GPU to complete before sw blit (memcpy) to front buffer.. but the small default window size for 'gears means that hw binning does not have much impact.  The remaining figures are for fullscreen 1280x720.