yum-mirror/slang

Making it easier to work with shaders

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

kaizhangNVFix regression in partial specialization of existential arguments (#6818)ee8a91e22

master
branchesmasterall refs →
1.8 KiB63 linesraw
1//TEST:SIMPLE(filecheck=CHECK): -target spirv
2
3// This is a test that checks that we can apply partial specialization to a function
4// we won't specialize the function parameters too aggressively. Instead, we will specialize
5// the parameters at the same time of specializing the arguments. Otherwise, we could lose
6// the chance to specialize the argument.
7//
8// In this test, `matrix_vector_interfaces` will be fully specialized, otherwise the compile
9// will fail because we don't allow opaque type in the existential type. So as long as the target
10// spirv code can be generated, we are good.
11
12// CHECK: %main
13public interface ITensor<T : IDifferentiable, let D : int>
14{
15    public T get(int idx);
16
17}
18
19public interface IRWTensor<T : IDifferentiable, let D : int> : ITensor<T, D>
20{
21}
22
23
24public struct RWTensor<T : IDifferentiable, let D : int> : IRWTensor<T, D>
25{
26    public RWStructuredBuffer<T> buffer;
27    public T get(int idx) { return buffer[idx]; }
28}
29
30public struct GradInOutTensor<T : IDifferentiable, let D : int> : IRWTensor<T, D>
31{
32    public RWTensor<T, D> primal;
33    public T get(int idx) { return primal.get(idx); }
34}
35
36struct CallData
37{
38    GradInOutTensor<float, 3> weights;
39    GradInOutTensor<float, 2> biases;
40    RWStructuredBuffer<float> _result;
41}
42ParameterBlock<CallData> call_data;
43
44float matrix_vector_interfaces(ITensor<float, 2> weights, ITensor<float, 1> biases)
45{
46    return weights.get(0);
47}
48
49[shader("compute")]
50[numthreads(1, 1, 1)]
51void main(uint3 dispatchThreadID: SV_DispatchThreadID)
52{
53    float _result;
54    GradInOutTensor<float, 2> weights;
55    GradInOutTensor<float, 1> biases;
56
57    weights.primal.buffer = call_data.weights.primal.buffer;
58    biases.primal.buffer = call_data.biases.primal.buffer;
59
60    _result = matrix_vector_interfaces(weights, biases);
61
62    call_data._result[0] = _result;
63}