yum-mirror/slang

Making it easier to work with shaders

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

16-Bit-DogAddition of `Load`/`Store` coherent operations (#8395)1e0908bd7

master
branchesmasterall refs →
2.5 KiB56 linesraw
1//TEST:SIMPLE(filecheck=SPIRV):-stage compute -entry computeMain -target spirv -capability vk_mem_model
2//TEST(compute):COMPARE_COMPUTE(filecheck-buffer=CHECK):-vk -emit-spirv-directly -capability vk_mem_model -output-using-type
3
4// Tests if we optimize redundant load's correctly
5
6//TEST_INPUT:ubuffer(data=[0 0 0], stride=4):out,name=outputBuffer
7RWStructuredBuffer<int> outputBuffer;
8//TEST_INPUT:ubuffer(data=[0 0 0 11 10], stride=4),name=buffer
9uniform int* buffer;
10
11[numthreads(2, 1, 1)]
12void computeMain(uint3 group_thread_id: SV_GroupThreadID)
13{
14    Ptr<int, Access::ReadWrite, AddressSpace::Device> ptr = __getAddress(buffer[0]);
15    
16    // Consider the load from this store-load pattern as redundant since 
17    // the load is a sub-set memory-scope of the memory-scope of the store.
18    // Invocation == Invocation.
19    *ptr = 8;
20    outputBuffer[0] = loadCoherent<4, MemoryScope::Invocation>(ptr);
21    // CHECK: 8
22    // SPIRV: OpStore %ptr %int_8
23    // SPIRV-NOT: OpLoad
24    // SPIRV: %[[#OUTPUT_BUFFER1:]] = OpAccessChain {{.*}} %outputBuffer %{{.*}} %int_0
25    // SPIRV: OpStore %[[#OUTPUT_BUFFER1]] %int_8
26
27    // Consider the load from this store-load pattern as redundant since 
28    // the load is a sub-set memory-scope of the memory-scope of the store.
29    // Device > Workgroup.
30    let offset1 = ptr + 1;
31    storeCoherent<4, MemoryScope::Device>(offset1, 9);
32    outputBuffer[1] = loadCoherent<4, MemoryScope::Workgroup>(offset1);
33    // CHECK-NEXT: 9
34    // SPIRV: %[[#PTR_OFFSET:]] = OpPtrAccessChain {{.*}} %ptr %int_1
35    // SPIRV: OpStore %[[#PTR_OFFSET]] %int_9
36    // SPIRV-NOT: OpLoad
37    // SPIRV: %[[#OUTPUT_BUFFER2:]] = OpAccessChain {{.*}} %outputBuffer %{{.*}} %int_1
38    // SPIRV: OpStore %[[#OUTPUT_BUFFER2]] %int_9
39
40    // Consider the following store-load pattern as not redundant since the data stored
41    // may not be the same data that will be loaded if Workgroup-scope contains
42    // different data than the Subgroup-scope.
43    // Subgroup < Workgroup.
44    let offset2 = ptr + 2;
45    storeCoherent<4, MemoryScope::Subgroup>(offset2, buffer[3]);
46    if(group_thread_id.x == 1)
47    {
48        storeCoherent<4, MemoryScope::Invocation>(offset2, buffer[4]);
49        let result = loadCoherent<4, MemoryScope::Workgroup>(offset2);
50        outputBuffer[2] = (result == 11 || result == 10) ? 12 : 0;
51    }
52    // CHECK-NEXT: 12
53    // SPIRV: OpStore {{.*}}MakePointerAvailable{{.*}} 4 %int_3
54    // SPIRV: OpStore {{.*}}MakePointerAvailable{{.*}} 4 %int_4
55    // SPIRV: OpLoad {{.*}}MakePointerVisible{{.*}} 4 %int_2
56}