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 →
2.6 KiB74 linesraw
1//TEST_CATEGORY(wave-mask, compute)
2//DISABLE_TEST:COMPARE_COMPUTE_EX:-cpu -compute -shaderobj
3//DISABLE_TEST:COMPARE_COMPUTE_EX:-slang -compute -shaderobj
4//TEST:COMPARE_COMPUTE_EX:-slang -compute -dx12 -profile cs_6_0 -shaderobj -render-feature hardware-device
5//TEST(vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -render-feature hardware-device
6//TEST:COMPARE_COMPUTE_EX:-cuda -compute -shaderobj
7
8//TEST_INPUT:ubuffer(data=[0 0 0 0  0 0 0 0  0 0 0 0  0 0 0 0], stride=4):out,name outputBuffer
9RWStructuredBuffer<int> outputBuffer;
10
11//TEST_INPUT:ubuffer(data=[3 10 2 -1  4 53 4 6  1 2 3 4  7 5 3 1], stride=4):name inputBuffer
12RWStructuredBuffer<int> inputBuffer;
13
14groupshared int sharedMem[32];
15
16int exclusivePrefixSum(WaveMask mask, int index, int waveLaneId, int originalValue, int elementCount)
17{
18    WaveMask localMask = WaveMaskBallot(mask, index < elementCount);
19    
20    sharedMem[index] = 0;
21    
22    if(waveLaneId < elementCount)
23    {
24        int temp = 0;
25        int val = originalValue;        
26        
27        for(int i = 1; i < elementCount; i += i)
28        {
29            int temp = WaveMaskShuffle(localMask, val, waveLaneId - i);            
30            if(waveLaneId >= i)
31            {
32                val += temp;
33            }
34        }        
35        
36        // Make it an exclusive prefix sum
37        val -= originalValue;
38        
39        // Write to shared memory
40        sharedMem[index] = val;
41        return val;
42    }    
43    
44    return 0;    
45}
46
47// It matters how kernels with WaveMask intrinsics are launched(!). 
48// TODO(JS):
49// If I launch with an numthreads amount that is not the size of the Wave on the device, then some 
50// lanes will not be executing at startup, and the kernel will have to know that is the case. 
51// This works currently though because the mask is only used
52// on CUDA, and it's Wave size is 32.
53[numthreads(32, 1, 1)]
54void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
55{
56    // Assumes all threads in the Wave are active at start.
57    WaveMask waveMask = ~WaveMask(0);
58    
59    int index = int(dispatchThreadID.x);
60    const int waveLaneId = int(WaveGetLaneIndex());
61    
62    const int value = inputBuffer[index];
63    const int elementCount = 9;
64    
65    exclusivePrefixSum(waveMask, index, waveLaneId, value, elementCount);
66    
67    // We don't read from any other lane, so we don't actually need any sync
68    //WaveMaskSharedSync(waveMask); 
69    
70    // It returns the result, but we are going to read from shared memory, to check that aspect worked
71    int prefixValue = sharedMem[index];
72    
73    outputBuffer[index] = prefixValue;
74}