I am currently experimenting with unbounded descriptor arrays and ultimately want to use bindless descriptors. I wrote a test application, that draws 100.000 simple object instances, where the transform for each object is pulled from a buffer array.
#pragma pack_matrix(row_major)
struct VertexData
{
float4 Position : SV_POSITION;
};
struct VertexInput
{
float3 Position : POSITION;
};
struct CameraData
{
float4x4 ViewProjection;
};
struct InstanceData
{
float4x4 Transform;
};
ConstantBuffer<CameraData> camera : register(b0, space0);
StructuredBuffer<InstanceData> instanceBuffer[] : register(t0, space1);
VertexData main(in VertexInput input, uint id : SV_InstanceID)
{
VertexData vertex;
InstanceData instance = instanceBuffer[id].Load(0);
float4 position = mul(float4(input.Position, 1.0), instance.Transform);
vertex.Position = mul(position, camera.ViewProjection);
return vertex;
}
I am using DXC to compile this shader to both, SPIR-V and DXIL. Under Vulkan, the draw call takes roughly 2ms, however under D3D, it takes around 10ms. I tried eliminating the buffer access by hashing the instance id and generating the transform for each object this way and (as expected) the draw calls were equally fast. I then tried using a ByteAddressBuffer instead of a StructuredBuffer and it did not show any difference in Vulkan (which is expected, since they are essentially the same in SPIR-V), but in D3D performance dropped even further to around 40ms per draw call. Now I am wondering what could cause such behavior. I tried looking at the IL, but I can't really tell how this compares to each other.
What could cause this discrepancy? Is there any way to profile this? All debuggers I tried just showed me the GPU times, which I mentioned above, but not more. Also some NSight Graphics features are not supported on my GPU (1080 Ti), so I could not test this compeltely.
