Add texture support from CuTextures.jl

[maleadt]

I messed up and #206 got closed.

[maleadt]

I wonder if it would be valuable to try and expose this through array abstractions like broadcast. Since those typically require matching sizes, I could imagine something like warp(::CuTexture, dims...) creating a 'view' with changed dimensions so that we can have it interpolate. Or maybe warp(f, ::CuTexture) to apply a mapping to indices and, e.g., be able to do the sine warp example.

@ChrisRackauckas, you've expressed interest in this, what's your use case?

[codecov]

Codecov Report

Merging #209 into master will decrease coverage by 0.40%.
The diff coverage is 81.53%.

@@            Coverage Diff             @@
##           master     #209      +/-   ##
==========================================
- Coverage   80.53%   80.13%   -0.41%     
==========================================
  Files         152      154       +2     
  Lines        9911    10174     +263     
==========================================
+ Hits         7982     8153     +171     
- Misses       1929     2021      +92     

Impacted Files	Coverage Δ
src/CUDA.jl	100.00% <ø> (ø)
src/pointer.jl	78.04% <50.00%> (-9.05%)	⬇️
lib/cudadrv/memory.jl	85.85% <77.41%> (-7.95%)	⬇️
src/texture.jl	84.61% <84.61%> (ø)
test/texture.jl	90.74% <90.74%> (ø)
examples/wmma/high-level.jl	11.11% <0.00%> (-38.89%)	⬇️
examples/wmma/low-level.jl	14.28% <0.00%> (-35.72%)	⬇️
test/device/wmma.jl	0.00% <0.00%> (-7.41%)	⬇️
test/execution.jl	38.68% <0.00%> (-1.47%)	⬇️
test/device/cuda.jl	9.72% <0.00%> (-0.73%)	⬇️
... and 2 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 160c998...5f6e3c1. Read the comment docs.

[maleadt]

julia> a = CUDA.rand(2,2)
2×2 CuArray{Float32,2,Nothing}:
 0.148875   0.386196
 0.0170659  0.696161

julia> b = similar(a)
2×2 CuArray{Float32,2,Nothing}:
 0.0  0.0
 0.0  0.0

julia> b .= CuTexture(a)
2×2 CuArray{Float32,2,Nothing}:
 0.148875   0.148875
 0.0170659  0.0170659

julia> @device_code_llvm debuginfo=:none b .= CuTexture(a)
; PTX CompilerJob of kernel broadcast(CUDA.CuKernelContext, CuDeviceArray{Float32,2,CUDA.AS.Global}, Base.Broadcast.Broadcasted{Nothing,Tuple{Base.OneTo{Int64},Base.OneTo{Int64}},typeof(identity),Tuple{Base.Broadcast.Extruded{CuDeviceTexture{Float32,2,false},Tuple{Bool,Bool},Tuple{Int64,Int64}}}}) for sm_75

...
L152:                                             ; preds = %L146, %L141
  %31 = load i64, i64* %56, align 8
  %32 = call [4 x float] @llvm.nvvm.tex.unified.2d.v4f32.s32(i64 %31, i32 %57, i32 %28)
  %.fca.0.extract = extractvalue [4 x float] %32, 0
  %33 = getelementptr inbounds { [2 x i64], i64 }, { [2 x i64], i64 }* %0, i64 0, i32 1
  %34 = bitcast i64* %33 to float addrspace(1)**
  %35 = load float addrspace(1)*, float addrspace(1)** %34, align 8
  %36 = getelementptr float, float addrspace(1)* %35, i64 %27
  store float %.fca.0.extract, float addrspace(1)* %36, align 4
  br label %L33
...
}

[maleadt]

CI is green, so let's merge this. API is not final yet, it requires explicit use of CuTexture/CuTextureArray objects (more or less like @cdsousa designed it, but integrated with the rest of the stack). I'd like to try making CuArray compatible with it, dispatching on the inner buffer object.

[cdsousa]

Nice!!!

The issue above, got fixed? I've started looking at it and was able to test in my original code, where it was ok, but then I had no more time to continue.

[maleadt]

The issue above, got fixed? I've started looking at it and was able to test in my original code, where it was ok, but then I had no more time to continue.

Which issue? The tests are mostly the same as they were in your version.

[cdsousa]

The issue above, got fixed? I've started looking at it and was able to test in my original code, where it was ok, but then I had no more time to continue.

#209 (comment)

Wasn't broadcast assignment supposed to give the same as a copy?

[maleadt]

Ah, I didn't even see they mismatched! I was just happy it generated appropriate code. But it seems to work now:

julia> a = CUDA.rand(8,8)
8×8 CuArray{Float32,2,Nothing}:
 0.916239  0.976847   0.755844  0.66619   0.011888  0.213493   0.665336   0.132537
 0.547251  0.867846   0.316787  0.580942  0.953204  0.762209   0.0874605  0.0591269
 0.413512  0.374084   0.369481  0.642127  0.724004  0.0811296  0.657992   0.788099
 0.48726   0.417901   0.889191  0.795766  0.334781  0.387315   0.913375   0.881644
 0.943592  0.584336   0.455078  0.28139   0.591649  0.0605415  0.832624   0.409376
 0.57543   0.0464874  0.612669  0.451585  0.950189  0.647614   0.658178   0.89397
 0.987634  0.586837   0.71754   0.884229  0.501966  0.217039   0.520088   0.542805
 0.730895  0.838332   0.796267  0.295658  0.348608  0.727785   0.867374   0.255682

julia> b = similar(a)

julia> b .= CuTexture(a)
8×8 CuArray{Float32,2,Nothing}:
 0.916239  0.976847   0.755844  0.66619   0.011888  0.213493   0.665336   0.132537
 0.547251  0.867846   0.316787  0.580942  0.953204  0.762209   0.0874605  0.0591269
 0.413512  0.374084   0.369481  0.642127  0.724004  0.0811296  0.657992   0.788099
 0.48726   0.417901   0.889191  0.795766  0.334781  0.387315   0.913375   0.881644
 0.943592  0.584336   0.455078  0.28139   0.591649  0.0605415  0.832624   0.409376
 0.57543   0.0464874  0.612669  0.451585  0.950189  0.647614   0.658178   0.89397
 0.987634  0.586837   0.71754   0.884229  0.501966  0.217039   0.520088   0.542805
 0.730895  0.838332   0.796267  0.295658  0.348608  0.727785   0.867374   0.255682

julia> a == b
true

Although the exact example doesn't anymore:

julia> CuTexture(CUDA.rand(2,2))
ERROR: CUDA error: invalid argument (code 1, ERROR_INVALID_VALUE)

Anyway, CuTexture isn't really ready to be used with broadcast yet, but it's nice to see the essentials work already.