On Sat, 11 Dec 2021 at 20:44, Nikolay Sivov nsivov@codeweavers.com wrote:
@@ -419,9 +418,11 @@ static void parse_test_directive(struct shader_context *context, const char *lin hr = create_root_signature(context->c.device, &root_signature_desc, &context->c.root_signature); ok(hr == S_OK, "Failed to create root signature, hr %#x.\n", hr);
pso = create_pipeline_state(context->c.device, context->c.root_signature,
if (context->c.pso)ID3D12PipelineState_Release(context->c.pso);context->c.pso = create_pipeline_state(context->c.device, context->c.root_signature, context->c.render_target_desc.Format, NULL, &ps, NULL);
if (!pso)
if (!context->c.pso) return;
I suppose that's an improvement in the sense that it should work in practice for most of our existing tests, but it doesn't strike me as entirely correct either. In particular, the restriction is that we can't destroy the state object until all command lists referencing it have finished executing on the GPU. That mostly works out in practice because the typical pattern is to do a "draw", followed by a "probe", in which case the "probe" will wait for the command list to finish executing, but I could certainly imagine writing tests where we do multiple draws before probing.