Revision 50e7d9797d9bf4b98a056d5df128c24dde6e78bd authored by Yong He on 23 March 2023, 23:59:02 UTC, committed by GitHub on 23 March 2023, 23:59:02 UTC
* Fix optimization pass not converging. * Fix. * Fix tests. --------- Co-authored-by: Yong He <yhe@nvidia.com>
1 parent 85f0058
buffer-layout.slang
// buffer-layout.slang
// The goal of this test is to make sure that constant and structured
// buffers obey the rules we expect of the each target API.
//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj
//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -dx12 -shaderobj
//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj
//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -cpu -shaderobj
//TEST_INPUT: ubuffer(data=[0 0 0 0], stride=4):out,name=outputBuffer
RWStructuredBuffer<int> outputBuffer;
struct A
{
float x;
float y;
}
struct S
{
// The first field in a struct isn't going to be that
// interesting, because it will always get offset zero,
// so we just use this to establish a poorly-aligned
// starting point for the next field.
//
// offset size alignment
//
// 0 4 4
//
float z;
// The `std140` and D3D constant buffer ruless both
// ensure a minimum of 16-byte alignment on `struct`
// types, but differ in that D3D does not round up
// the total size of a type to its alignment.
//
// The `std430` and structured buffer rules don't
// perform any over-alignment on `struct` types and
// instead align them using the "natural" rules one
// might expect of, e.g., a C compiler.
//
// offset size alignment
//
// cbuffer 16 8 16
// std140 16 16 16
//
// struct 4 8 4
// std430 4 8 4
//
A a;
// Now we insert an ordinary `int` field just as
// a way to probe the offset so far.
//
// offset size alignment
//
// cbuffer 24 4 4
// std140 32 4 4
//
// struct 12 4 4
// std430 12 4 4
//
int b;
// As our next stress-test case, we will insert an
// array with elements that aren't a multiple of
// 16 bytes in size.
//
// The contant/uniform buffer rules will set the
// array stride to a multiple of 16 bytes in this case.
// The only difference between D3D rules and `std140`
// here is that D3D does not round up the size to
// the alignment.
//
// The structured/std430 rules don't do anything
// to over-align an array, so it is laid out relatively
// naturally, but note that D3D still follows its rule
// of not letting a vector "straddle" a 16-byte boundary,
// even if it doesn't bump up the alignment of
// vector types.
//
// offset size alignment
//
// cbuffer 32 24 16
// std140 48 32 32
//
// struct 16 16 4
// std430 16 16 8
//
float2 c[2];
// Now we put in one more ordinary `int` field
// just to probe the offset computed so far.
// offset size alignment
//
// cbuffer 56 4 4
// std140 80 4 4
//
// struct 32 4 4
// std430 32 4 4
//
int d;
}
//TEST_INPUT:cbuffer(data=[0 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]):name=cb
ConstantBuffer<S> cb;
//TEST_INPUT:ubuffer(data=[0 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],stride=4):name=sb
RWStructuredBuffer<S> sb;
int test(int val)
{
val = val+1;
val = val*16 + cb.b;
val = val*256 + cb.d;
val = val*256 + sb[0].b;
val = val*256 + sb[0].d;
return val;
}
[numthreads(4, 1, 1)]
void computeMain(
int3 dispatchThreadID : SV_DispatchThreadID)
{
int tid = dispatchThreadID.x;
int inVal = tid;
int outVal = test(inVal);
outputBuffer[tid] = outVal;
}
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...