blob: 2108d562c26c0d172815b7683c15aca65f7defd1 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
|
//TEST:SIMPLE(filecheck=CUDA): -target cuda -entry compute_main -stage compute
//TEST:SIMPLE(filecheck=PTX): -target ptx -entry compute_main -stage compute
//TEST:SIMPLE(filecheck=SPV): -target spirv
// Check that we can specialize buffer loads through bindless handles, and
// do not load big struct elements into registers unnecessarily.
struct Bottom
{
float bigArray[1024];
float bottomGetValue(int index) { return bigArray[index]; }
}
struct Middle
{
Bottom bottom;
float middleGetValue(int index) { return bottom.bottomGetValue(index); }
}
struct Top
{
StructuredBuffer<Middle>.Handle middle;
// Calling `middleGetValue` on `middle[0]` should not causing the entire `Middle`
// struct to be loaded into registers. Instead, we should be able to specialize
// `middleGetValue` to take a `StructuredBuffer<Middle>.Handle` and an `int`
// index, and recursively specialize `bottomGetValue` to only load the `Bottom.bigArray[index]` element.
float topGetValue(int index) { return middle[0].middleGetValue(index); }
}
struct Root
{
Top top;
}
ConstantBuffer<Root> cb;
RWStructuredBuffer<float> outputBuffer;
// SPV: OpEntryPoint
// SPV-NOT: OpLoad %Middle
// SPV: %[[REG:[A-Za-z0-9_]+]] = OpLoad %float
// SPV: OpStore {{.*}} %[[REG]]
// Check that the generated CUDA code contains a specialized `bottomGetValue` function that has
// the complete parameter list to access the `bigArray` element directly, without needing to load
// the entire `Bottom` struct from the caller.
//
// CUDA-DAG: __device__ float Bottom_bottomGetValue{{.*}}(StructuredBuffer<Middle{{.*}}> {{.*}}, int {{.*}}, int {{.*}})
// PTX: compute_main
[shader("compute")]
[numthreads(1, 1, 1)]
void compute_main(uint3 tid: SV_DispatchThreadID)
{
outputBuffer[0] = cb.top.topGetValue(0);
}
|
