[PATCH 0/5] MR700: tests: Do some housekeeping for Vulkan implementations nobody loves.

[PATCH 1/5] tests: Mark the HS punned array test as buggy on MoltenVK.

[Giovanni Mascellani]

From: Giovanni Mascellani

MoltenVK doesn't support the Component decoration on tessellation
shaders.
---
 tests/d3d12.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tests/d3d12.c b/tests/d3d12.c
index 49b40a67a..3e75fada4 100644
--- a/tests/d3d12.c
+++ b/tests/d3d12.c
@@ -38139,6 +38139,7 @@ static void test_hull_shader_punned_array(void)
     transition_resource_state(command_list, context.render_target,
             D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE);

+    bug_if(is_mvk_device(context.device))
     check_sub_resource_uint(context.render_target, 0, queue, command_list, 0xcc996633, 0);

destroy_test_context(&context);

[PATCH 2/5] tests: Pop the test context when leaving its scope.

[Giovanni Mascellani]

From: Giovanni Mascellani

---
 tests/d3d12.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/tests/d3d12.c b/tests/d3d12.c
index 3e75fada4..ce907c966 100644
--- a/tests/d3d12.c
+++ b/tests/d3d12.c
@@ -32383,7 +32383,10 @@ static void test_resource_allocation_info(void)
             ok(infos1[i].SizeInBytes == sizes[i], "Got unexpected size %"PRIu64".\n", infos1[i].SizeInBytes);

if (!i)
+            {
+                vkd3d_test_pop_context();
                 continue;
+            }

ok(infos1[i].Offset - infos1[i - 1].Offset >= infos1[i - 1].SizeInBytes,
                     "Got unexpected prev size %"PRIu64", prev offset %"PRIu64", offset %"PRIu64".\n",

[PATCH 3/5] tests: Mark the coverage test as buggy on MoltenVK.

[Giovanni Mascellani]

From: Giovanni Mascellani

Apparently MoltenVK doesn't set the SampleMask built-in to the
fragment coverage mask.
---
 tests/d3d12.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tests/d3d12.c b/tests/d3d12.c
index ce907c966..7655c4e92 100644
--- a/tests/d3d12.c
+++ b/tests/d3d12.c
@@ -33454,6 +33454,7 @@ static void test_coverage(void)
         reset_command_list(command_list, context.allocator);
         transition_resource_state(command_list, texture,
                 D3D12_RESOURCE_STATE_UNORDERED_ACCESS, D3D12_RESOURCE_STATE_COPY_SOURCE);
+        bug_if(is_mvk_device(context.device))
         check_sub_resource_uint(texture, 0, queue, command_list, tests[i].sample_mask & 0xf, 0);

reset_command_list(command_list, context.allocator);

[PATCH 4/5] tests: Relax precision a bit for an interpolation test.

[Giovanni Mascellani]

From: Giovanni Mascellani

Otherwise the test fails on Intel, NVIDIA and llvmpipe.
---
 tests/hlsl/nointerpolation.shader_test | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/hlsl/nointerpolation.shader_test b/tests/hlsl/nointerpolation.shader_test
index 94584ffe2..d2b7d8a46 100644
--- a/tests/hlsl/nointerpolation.shader_test
+++ b/tests/hlsl/nointerpolation.shader_test
@@ -144,7 +144,7 @@ shader model >= 6.0

[test]
 draw triangle list 3
-probe (320,240) rgba (0.250390589, 0.499088615, 0.250520796, 1.0) 1
+probe (320,240) rgba (0.250390589, 0.499088615, 0.250520796, 1.0) 4

[require]
 shader model >= 4.0

[PATCH 5/5] tests: Clear the UAV before testing.

[Giovanni Mascellani]

From: Giovanni Mascellani

There is no guarantee that a resource is cleared upon creation. This
fixes the test on NVIDIA and llvmpipe.
---
 tests/hlsl_d3d12.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/tests/hlsl_d3d12.c b/tests/hlsl_d3d12.c
index b2273ffbc..b13242562 100644
--- a/tests/hlsl_d3d12.c
+++ b/tests/hlsl_d3d12.c
@@ -516,10 +516,16 @@ static void test_thread_id(void)

for (i = 0; i < 3; ++i)
     {
+        const UINT clear_value[4] = {0};
+
         textures[i] = create_default_texture3d(device, 16, 8, 8, 1, DXGI_FORMAT_R32G32B32A32_UINT,
                 D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS, D3D12_RESOURCE_STATE_UNORDERED_ACCESS);
         ID3D12Device_CreateUnorderedAccessView(device, textures[i], NULL, NULL,
                 get_cpu_descriptor_handle(&context, heap, i));
+        ID3D12GraphicsCommandList_ClearUnorderedAccessViewUint(command_list,
+                get_gpu_descriptor_handle(&context, heap, i),
+                get_cpu_descriptor_handle(&context, heap, i),
+                textures[i], clear_value, 0, NULL);
     }

cs_code = compile_shader(cs_source, "cs_5_0");

[PATCH 0/5] MR700: tests: Do some housekeeping for Vulkan implementations nobody loves. - approved

[Giovanni Mascellani (＠giomasce)]

This merge request was approved by Giovanni Mascellani.

[Henri Verbeet (＠hverbeet)]

> Subject: [PATCH 5/5] tests: Clear the UAV before testing.
>
> There is no guarantee that a resource is cleared upon creation. This
> fixes the test on NVIDIA and llvmpipe.

Does this need an UAV barrier between the clear and the dispatch?

[Giovanni Mascellani (＠giomasce)]

On Mon Mar 11 13:57:16 2024 +0000, Henri Verbeet wrote:
> > Subject: [PATCH 5/5] tests: Clear the UAV before testing.
> >
> > There is no guarantee that a resource is cleared upon creation. This
> > fixes the test on NVIDIA and llvmpipe.
>
> Does this need an UAV barrier between the clear and the dispatch?

Good point. I couldn't find any definitive answer, but from https://learn.microsoft.com/en-us/windows/win32/api/d3d12/ns-d3d12-d3d12_res... it seems that barriers must always used between pairs of write operations (even if the second operation doesn't read, or at least doesn't apparently read; the page rather gives the negative version of this: you're allowed to skip the barrier between two read-only operations). I couldn't find any formal definition of what is a "write operation", so it's not clear whether a clear operation is a write operation. For the avoidance of doubt I'll pretend it is a fix the MR.