yum-mirror/slang

Making it easier to work with shaders

git clone https://git.yummers.dev/yum-mirror/slang

James Helferty (NVIDIA)render-test: Change D3D12 default to sm_6_5 (#8320)f02b08490

master
branchesmasterall refs →
2.5 KiB85 linesraw
1//DISABLED_TEST(compute, vulkan):COMPARE_COMPUTE_EX():-vk -compute -shaderobj -output-using-type -render-features wave-ops
2//DISABLED_TEST(compute):COMPARE_COMPUTE_EX():-dx12 -profile sm_6_5 -compute -shaderobj -output-using-type -render-features wave-ops
3//TEST:SIMPLE(filecheck=CHECK):-target hlsl -profile cs_6_5 -entry computeMain -line-directive-mode none
4//TEST:SIMPLE(filecheck=CHECK):-target glsl -profile cs_6_5 -entry computeMain -line-directive-mode none
5
6//
7// This test checks that we fuse calls to saturated cooperation, even with different input values
8//
9
10//TEST_INPUT:ubuffer(data=[0 3 2 2], stride=4):out,name=outputBuffer
11RWStructuredBuffer<int> outputBuffer;
12
13static int count = 0;
14
15int coopAdd(float x, int i)
16{
17    count += i;
18    return int(x) * 2;
19}
20
21int fallback1(float x, int)
22{
23    count = -1;
24    return -1;
25}
26
27int coopMul(float x, float f)
28{
29    count *= int(f);
30    return int(x) * 2;
31}
32
33int fallback2(float x, float)
34{
35    count = -1;
36    return -1;
37}
38
39int coopExp(float x, int i)
40{
41    int c = count;
42    count = 1;
43    for(int j = 0; j < i; ++j)
44        count *= c;
45    return int(x) * 2;
46}
47
48// Make sure that we have enough invocations to saturate the first workgroup
49[numthreads(128, 1, 1)]
50void computeMain(uint tig : SV_GroupIndex)
51{
52    // The values we're cooperating over are {0, 2, 3}
53    // We track the number of sets evaluated in the "count" variable, and write
54    // that at index 0
55    //
56    // If these are not fused, then we'd expect count to be incremented three
57    // times then doubled three times. What we want to see is
58    // 0, +1, *2, **2, +1, *2, **2, +1, *2, **2 = 40804
59    // If they're not interleaved then we'll see
60    // 0, +1, +1, +1, *2, *2, *2, **2, **2, **2 = -1593835520 (having overflowed)
61    let i = tig < 4 ? float(outputBuffer[tig]) : 0;
62    let j = i + 1;
63    let k = j + 1;
64    let y = saturated_cooperation(coopAdd, fallback1, i, 1);
65    let m = 2.f;
66    let x = saturated_cooperation(coopMul, fallback2, j, m);
67    // Put some calculation between the calls to check that the fusion still takes place
68    let e = min(m*100, 2);
69    let z = saturated_cooperation(coopExp, fallback1, k, int(e));
70
71    // check that all the piping functions we invent during fusing is inlined away
72    // CHECK:      = coopAdd
73    // CHECK-NEXT: = coopMul
74    // there is a tuple pack here
75    // CHECK-NEXT: Tuple
76    // CHECK-NEXT: = coopExp
77
78    // CHECK:      = fallback1
79    // CHECK-NEXT: = fallback2
80    // CHECK-NEXT: Tuple
81    // CHECK-NEXT: = fallback1
82
83    if(tig < 4)
84        outputBuffer[tig] = tig == 0 ? count : z;
85}