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
|
//TEST_CATEGORY(wave, compute)
//DISABLE_TEST:COMPARE_COMPUTE_EX:-cpu -compute -shaderobj
//DISABLE_TEST:COMPARE_COMPUTE_EX:-slang -compute -shaderobj
//TEST:COMPARE_COMPUTE_EX:-slang -compute -dx12 -profile cs_6_0 -shaderobj -render-feature hardware-device
// Disabled on VK because glsl can't do WaveReadLaneAt on matrix.
//DISABLE_TEST(vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj
//TEST:COMPARE_COMPUTE_EX:-cuda -compute -capability cuda_sm_7_0 -shaderobj
//TEST_INPUT:ubuffer(data=[0 0 0 0], stride=4):out,name outputBuffer
RWStructuredBuffer<int> outputBuffer;
// Note from HLSL: `The input lane index must be uniform across the wave.`.
// The same restriction applies to glsl/SPIR-V 1.5
// So we are going to use the input buffer to achieve this.
//TEST_INPUT:ubuffer(data=[1 2 3 0], stride=4):name inputBuffer
RWStructuredBuffer<int> inputBuffer;
[numthreads(4, 1, 1)]
void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
{
int idx = int(dispatchThreadID.x);
int value = 0;
for (int i = 0; i < 4; ++i)
{
// Scalar
// The landId is 'dynamic' but it also uniform across the wave (as required by spec)
const int laneId = inputBuffer[i];
value += WaveReadLaneAt(idx, laneId);
// vector
{
float2 v = float2(idx + 1, idx + 2);
float2 readValue = WaveReadLaneAt(v, (laneId + 1) & 3);
value += int(readValue[0] + readValue[1]);
}
// matrix
{
matrix<int, 2, 2> v = matrix<int, 2, 2>(idx, idx - 1, idx * 3, idx - 2);
matrix<int, 2, 2> readValue = WaveReadLaneAt(v, (laneId - 1) & 3);
value += int(readValue[0][0] + readValue[0][1] + readValue[1][0] + readValue[1][1]);
}
}
outputBuffer[idx] = value;
}
|
