Revisiting the Business Card Raytracer

https://fabiensanglard.net/revisiting_the_businesscard_raytracer/index.html

by vanni • 6 years ago

176 52 6 years ago

ntry• 6 years ago

101,000ms to 150ms is a phenomenal speedup. Props

To be fair, the thing he started with was totally unreadable and fits on a business card and the thing he ended up with was readable, but didn't fit. It isn't apples-to-apples.

sp3322• 6 years ago

The original is 1,287 bytes after removing comments and line endings where possible. After just a bit of search-and-replace to shrink variable names, I got the new one down to 2,373 bytes including the same error checking and output at the end. So 1.8 business cards, which is not too bad.

And no I haven't tried to compile it. https://pastebin.com/LDRd6U4e

bigcheesegs• 6 years ago

Opening the binary with Binary Ninja revealed that clang had already managed to leverage the SSE registers.

X86-64 uses SSE registers for all floating point operations. I'm not sure that the author realized that they were looking at an -O0 binary. -O0 does not do vectorization (or anything else for that matter).

slavik815• 6 years ago

Looking at it on Godbolt, it doesn't really leverage SSE on -O3, either. You can get a reasonable grasp of whether it's using SSE effectively or not just by looking at the instruction names.

mulss: multiplication of a single single-precision floating point value.

mulsd: multiplication of a single double-precision floating point value.

mulps: multiplication of a packed group of single-precision floating point values.

mulpd: multiplication of a packed group of double-precision floating point values.

If you're mostly seeing -ps suffixes only on moves and shuffles, you're looking at code that is not being vectorized. (And, actually, if you're seeing a lot of shuffles, that's also a good sign its not well-vectorized.)

Incidentally, if you're seeing unexpected -sd suffixes, those are often due to unintended conversions between float and double. They can have a noticeable effect on performance, especially if you end up calling the double versions of math functions (as they often use iterative algorithms that need more iterations to achieve double-precision).

I'm linking GCC output, because it's simpler to follow, but you see more or less the same struggle with Clang.

https://godbolt.org/z/XtVqsU

gorgoiler2• 6 years ago

Off topic: I teach compilers in high school and godbolt.org looks amazing, thanks for the link!

K0nserv2• 6 years ago

This post is really topical for me. I spent hours yesterday trying to write explicit SIMD code[0] for my the vector dot product in my raytracer and all I managed to do was slow the code down about 20-30%.

The code generated by Rust from the naive solution uses ss instructions mostly whereas my two tries using `mm_dp_ps` and `mm_mul_ps` and `mm_hadd_ps` where both significantly slower even though it results in fewer instructions. I suspect that the issue is that for a single dot product the overhead of loading in and out of mm128 registers is more cost than it's worth.

Naive Rust version output

    .cfi_startproc
    pushq %rbp
    .cfi_def_cfa_offset 16
    .cfi_offset %rbp, -16
    movq  %rsp, %rbp
    .cfi_def_cfa_register %rbp
    vmovss  (%rdi), %xmm0
    vmulss  (%rsi), %xmm0, %xmm0
    vmovsd  4(%rdi), %xmm1
    vmovsd  4(%rsi), %xmm2
    vmulps  %xmm2, %xmm1, %xmm1
    vaddss  %xmm1, %xmm0, %xmm0
    vmovshdup %xmm1, %xmm1
    vaddss  %xmm1, %xmm0, %xmm0
    popq  %rbp
    retq

My handwritten version with `mm_mul_ps` and `mm_hadd_ps`

    .cfi_startproc
    pushq %rbp
    .cfi_def_cfa_offset 16
    .cfi_offset %rbp, -16
    movq  %rsp, %rbp
    .cfi_def_cfa_register %rbp
    vmovaps (%rdi), %xmm0
    vmulps  (%rsi), %xmm0, %xmm0
    vhaddps %xmm0, %xmm0, %xmm0
    vhaddps %xmm0, %xmm0, %xmm0
    popq  %rbp
    retq

Intuatively it feels like my version should be faster but it isn't. In this code I changed the the struct from 3 f32 components to an array with 4 f32 elements to avoid having to create the array during computation itself, the code also requires specific alignment not to segfault which I guess might also affected performance.

0: https://github.com/k0nserv/rusttracer/commits/SIMD-mm256-dp-...

moosedev2• 6 years ago

Yeah, use of SSE registers does not imply SIMD, since x87 is gone in x86-64, so even scalar FP has to use SSE registers. The asm snippets for v::operator*() in the "Optimization level 1" section use scalar SSE arithmetic only (mulss). (There's some use of movaps to move data around, but it's a stretch to call that SIMD.)

I think the "leverage" sentence you quoted and the "with SIMD taken care of" one shortly after are maybe a bit misleading, since the asm snippets there don't really demonstrate SIMD.

amluto2• 6 years ago

since x87 is gone in x86-64, so even scalar FP has to use SSE registers.

No, it’s still there. What’s actually going on is that all x86-64 CPUs support SSE2, so there is little reason to use x87 in 64-bit code.

(You can use it for 80-bit precision. OTOH, for most purposes, 80-bit precision is actively harmful, and x87 is an incredible mess, so almost no one wants it.)

fabiensanglard• 6 years ago

You are correct. Mārtiņš Možeiko pointed out that I had been too hasty when the article came out (https://twitter.com/mmozeiko/status/1257574246462570497). To conclude SIMD is leveraged when XMM registers are used is wrong. What I should have looked for are packed instructions.

mappu• 6 years ago

It would be interesting to see how an ispc version performs, if it is able to extract any more CPU parallelism.

lonk• 6 years ago

If you increase resolution you can put more code on business card :P

mjcohen• 6 years ago

4 point font!

Y_Y• 6 years ago

If you make your business cards out of silicon wafer I can get you 10nm resolution.

blondin• 6 years ago

i love what fabien is doing with his website!

also been experimenting with pure html with an itsy-bitsy amount of css. for months now i wondered how to display code without involving javascript.

that textarea is so perfect! and i bet you when you copy and paste into word or your todo list application they won't even try to be "smart" about knowing what "rich text" is...

that's very cool.

GuiA2• 6 years ago

i am glad that there is a small village of indomitable developers holding out against tens of megabytes of reactive functional progress modern javascript frameworks.

thank you

wetmore• 6 years ago

There is a place for both.

nucleardog• 6 years ago

Yeah, damn, I thought I'd done well getting my page+css+font down to ~100kB (css/font cached after first load, so ~2kB on subsequent pages) but his site is tiny. Even with his CSS inlined to save a request the entire page is 15kB gzipped.

Kwantuum• 6 years ago

that textarea is so perfect!

There are no textarea on that page. The code sections are using a <pre>

danielscrubs• 6 years ago

Well there goes my weekend. :)

I also tried to optimise the code, and got great speed increases with just constexpr the vector methods and could quickly see that rand was problematic and then Fabien releases this post with nvcc that are another level. Really great blog post!

tomsmeding• 6 years ago

Nice work on the GPU programming, and the multicore before that, but I'm mystified why going from -O0 to -O3 is named an "optimisation". All respect for Fabien, but running code that's supposed to run faster than a snail (and if you're not debugging and require -O0 for reasonable output) implies -O2 or -O3. (In practice, -O3 often doesn't give much performance over -O2, despite increasing compile times.)

The initial time is not 101.8 seconds, it's 11.6 seconds.

slavik815• 6 years ago

You could also add -ffast-math, which loosens the rules for floating point optimizations. For example, it would allow the compiler to turn floating point divisions into multiplications by an inverse, and to group operations more efficiently even if doing so would slightly affect rounding. It also rounds denormal numbers down to zero, which can greatly improve performance on a lot of hardware.

-march=native may also be useful, as it would allow the compiler to use newer CPU instructions, and tune the generated code to your hardware. That would make the program less portable, but it's not like CUDA is portable either.

My machine matches those numbers surprisingly closely. With -O0 it took 89.6s. With -O3, it took 11.7s. With -Ofast (which combines -O3 and -ffast-math), it took 10.6s. With -Ofast -march=native, it took 8.9s. I would expect those gains to extrapolate to the multi-threaded version, maybe pushing it down to 1 second without any further work. (Note: I'm using GCC on Ubuntu 18.04 with a Haswell i7. Your mileage may vary.)

amelius2• 6 years ago

Offtopic, but I think that languages should have special float types that trigger the use of fast math. That way, a programmer can better control which parts of a program are done with approximate floating point operations.

Natasha35Khan• 6 years ago

Agree with you

Aleeshakhan786• 6 years ago

Me too

mianos• 6 years ago

I wonder if the tool-chain would be better under Linux? It is kind if funny the way Windows development has always been a hassle. Mscvars.bat and such has been there for at least 20 years.

fegu• 6 years ago

Almost into passable frame rate territory. Next version could be business card VR:)

ectoplasmaboiii• 6 years ago

A Ray-Tracer in 7 Lines of K: http://nsl.com/k/ray/ray.k

jagged-chisel• 6 years ago

7? Just remove those comments and the last few newlines and it’ll be one line.

I don’t know K, but it looks like it uses semicolons to end statements. It’s “cheating” on line count to compress statements by just removing \n.

After all, how many “lines” are in the business card raytracer? 4.

rwmj• 6 years ago

Is there a mistake in the original code on the right hand side? I get:

  card.cpp:16:2: error: ‘g’ was not declared in this scope
     16 | <g;p)t=p,n=v(0,0,1),m=1;for(i k=19;k--;)
        |  ^

Edit: Yes there is. The ‘<g;’ seems like it should have been the single character ‘<’, perhaps a corrupted HTML escape.

joeraut• 6 years ago

Yep, I got the same error. The version here [0] works fine.

[0] https://fabiensanglard.net/postcard_pathtracer/

a_e_k2• 6 years ago

(Original author of this code[0] here.) Yes, that looks like a corrupted copy. The line should be:

    <p)t=p,n=v(0,0,1),m=1;for(i k=19;k--;)

[0] http://www.cs.utah.edu/~aek/code/card.cpp

rrss• 6 years ago

This was really fun to read, thanks fsanglard.

This is correlated with the warning nvcc issued. Because the raytracer uses recursion, it uses a lot of stacks. So much actually that the SM cannot keep more than a few alive.

Stack frame size / "local memory" size doesn't actually directly limit occupancy. There's a list of the limiters here: https://docs.nvidia.com/gameworks/content/developertools/des.... I'm not sure why the achieved occupancy went up after removing the recursion, but I'd guess it was something like the compiler was able to reduce register usage.

Revisiting the Business Card Raytracer

Related Stories