Avoid splitting of loops with many iterations when using shader model 3.

Avoid splitting of loops with many iterations when using shader model 3.

This is only necessary for shader model 2 and it can take the D3DX shader compiler a long time to unroll all the resulting loops. The D3DX compiler does not emit an error if a loop has more than 256 iterations with shader model 3.

Also, fix issue where loop discontinuity detection was being performed for vertex shaders. This is not necessary because vertex shaders do not have gradient operations.

[apatrick1, 2013-01-18 19:53:11]

PTAL. This drastically improves the compile time of some shaders that have loops
with many iterations.

[kbr1, 2013-01-18 21:14:47]

LGTM

[apatrick1, 2013-01-18 21:18:04]

https://codereview.appspot.com/7139055/diff/1007/src/compiler/OutputHLSL.cpp
File src/compiler/OutputHLSL.cpp (right):

https://codereview.appspot.com/7139055/diff/1007/src/compiler/OutputHLSL.cpp#...
src/compiler/OutputHLSL.cpp:102: mContainsLoopDiscontinuity =
mContext.shaderType == SH_FRAGMENT_SHADER &&
containsLoopDiscontinuity(mContext.treeRoot);
Sorry, this...

https://codereview.appspot.com/7139055/diff/1007/src/compiler/OutputHLSL.cpp#...
src/compiler/OutputHLSL.cpp:1782: if (mContainsLoopDiscontinuity &&
!mInsideDiscontinuousLoop)
... and this are part of the discontinuity detection patch.
https://codereview.appspot.com/7131057/. Fix coming.

[nicolas, 2013-01-21 21:55:25]

When disabling loop splitting of a 2000 iteration loop with simple arithmetic, I
get the following HLSL compiler warning and error:

warning X3569: loop executes for more than 255 iterations (maximum for this
shader target), forcing loop to unroll
error X3511: unable to unroll loop, loop does not appear to terminate in a
timely manner (1024 iterations)

This is with Direct3D 9 and Shader Model 3. So it appears that the HLSL compiler
won't unroll a loop for more than 1024 iterations. If I reduce the loop count to
1000, it compiles correctly. What's more, the assembly actually shows a loop of
1000 iterations, but I still get warning X3569.

However, when disabling loop splitting the acos and asin tests from the Khronos
GTF fail because they do cause loop unrolling and this in turn causes it to use
more than the available number of constant registers.

So supporting high loop iteration counts is very hard to achieve with Shader
Model 3. I'm afraid this might require adjusting people's expectations. ES2
demands loops to be unrollable, and compilation may fail due to running out of
hardware resources.

[kbr1, 2013-01-22 20:32:03]

Could you email us your test case offline?

Recently two shaders were brought to our attention which worked fine on SM3 with
loop splitting disabled, but either failed to compile, or took too long to
compile, with it turned on. One was turned into a conformance test:
https://www.khronos.org/registry/webgl/sdk/tests/conformance/glsl/misc/large-...
. Do you think that that test is reasonable?

It seems that either with high iteration counts or a lot of work being done in
loops without a very high iteration count, loop splitting is causing the D3D
shader compiler to take a very long time.

Do you have other suggestions? What about, on SM3, attempting the compile with
loop splitting disabled, and turning it on and re-translating and re-compiling
if that fails?


On 2013/01/21 21:55:25, nicolas wrote:
> When disabling loop splitting of a 2000 iteration loop with simple arithmetic,
I
> get the following HLSL compiler warning and error:
> 
> warning X3569: loop executes for more than 255 iterations (maximum for this
> shader target), forcing loop to unroll
> error X3511: unable to unroll loop, loop does not appear to terminate in a
> timely manner (1024 iterations)
> 
> This is with Direct3D 9 and Shader Model 3. So it appears that the HLSL
compiler
> won't unroll a loop for more than 1024 iterations. If I reduce the loop count
to
> 1000, it compiles correctly. What's more, the assembly actually shows a loop
of
> 1000 iterations, but I still get warning X3569.
> 
> However, when disabling loop splitting the acos and asin tests from the
Khronos
> GTF fail because they do cause loop unrolling and this in turn causes it to
use
> more than the available number of constant registers.
> 
> So supporting high loop iteration counts is very hard to achieve with Shader
> Model 3. I'm afraid this might require adjusting people's expectations. ES2
> demands loops to be unrollable, and compilation may fail due to running out of
> hardware resources.