[PATCH vkd3d 3/8] vkd3d: Do not disable depth-stencil tests for PSOs with invalid DSV format.
From: Józef Kucia jkucia@codeweavers.com
This is invalid usage according to the D3D12 validation layer. However, Shadow of the Tomb Raider uses PSOs with DSVFormat equal to DXGI_FORMAT_UNKNOWN and enabled depth-stencil tests. Moreover, the test in the next commit passes on Windows with AMD, Intel and Nvidia GPUs.
Signed-off-by: Józef Kucia jkucia@codeweavers.com --- libs/vkd3d/command.c | 14 +++++-- libs/vkd3d/state.c | 78 ++++++++++++++++++++++++++++---------- libs/vkd3d/vkd3d_private.h | 5 ++- 3 files changed, 72 insertions(+), 25 deletions(-)
diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c index 8c7057b36340..861c8405c3e6 100644 --- a/libs/vkd3d/command.c +++ b/libs/vkd3d/command.c @@ -2251,11 +2251,13 @@ static void d3d12_command_list_reset_state(struct d3d12_command_list *list, list->fb_width = 0; list->fb_height = 0; list->fb_layer_count = 0; + list->dsv_format = VK_FORMAT_UNDEFINED;
list->xfb_enabled = false;
list->current_framebuffer = VK_NULL_HANDLE; list->current_pipeline = VK_NULL_HANDLE; + list->pso_render_pass = VK_NULL_HANDLE; list->current_render_pass = VK_NULL_HANDLE;
memset(list->pipeline_bindings, 0, sizeof(list->pipeline_bindings)); @@ -2370,7 +2372,7 @@ static bool d3d12_command_list_update_current_framebuffer(struct d3d12_command_l fb_desc.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO; fb_desc.pNext = NULL; fb_desc.flags = 0; - fb_desc.renderPass = graphics->render_pass; + fb_desc.renderPass = list->pso_render_pass; fb_desc.attachmentCount = view_count; fb_desc.pAttachments = views; d3d12_command_list_get_fb_extent(list, &fb_desc.width, &fb_desc.height, &fb_desc.layers); @@ -2407,9 +2409,12 @@ static bool d3d12_command_list_update_current_pipeline(struct d3d12_command_list }
if (!(vk_pipeline = d3d12_pipeline_state_get_or_create_pipeline(list->state, - list->primitive_topology, list->strides))) + list->primitive_topology, list->strides, list->dsv_format, &list->pso_render_pass))) return false;
+ if (!list->pso_render_pass) + list->pso_render_pass = list->state->u.graphics.render_pass; + VK_CALL(vkCmdBindPipeline(list->vk_command_buffer, list->state->vk_bind_point, vk_pipeline)); list->current_pipeline = vk_pipeline;
@@ -2797,7 +2802,8 @@ static bool d3d12_command_list_begin_render_pass(struct d3d12_command_list *list if (list->current_render_pass != VK_NULL_HANDLE) return true;
- vk_render_pass = list->state->u.graphics.render_pass; + vk_render_pass = list->pso_render_pass; + assert(vk_render_pass);
begin_desc.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO; begin_desc.pNext = NULL; @@ -4373,11 +4379,13 @@ static void STDMETHODCALLTYPE d3d12_command_list_OMSetRenderTargets(ID3D12Graphi list->fb_width = max(list->fb_width, dsv_desc->width); list->fb_height = max(list->fb_height, dsv_desc->height); list->fb_layer_count = max(list->fb_layer_count, dsv_desc->layer_count); + list->dsv_format = dsv_desc->format; } else { WARN("DSV descriptor is not initialized.\n"); list->views[0] = VK_NULL_HANDLE; + list->dsv_format = VK_FORMAT_UNDEFINED; } }
diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c index b1f4ef9acac1..457f3d9bc84b 100644 --- a/libs/vkd3d/state.c +++ b/libs/vkd3d/state.c @@ -1236,6 +1236,7 @@ struct vkd3d_pipeline_key { D3D12_PRIMITIVE_TOPOLOGY topology; uint32_t strides[D3D12_IA_VERTEX_INPUT_RESOURCE_SLOT_COUNT]; + VkFormat dsv_format; };
struct vkd3d_compiled_pipeline @@ -1243,6 +1244,7 @@ struct vkd3d_compiled_pipeline struct list entry; struct vkd3d_pipeline_key key; VkPipeline vk_pipeline; + VkRenderPass vk_render_pass; };
/* ID3D12PipelineState */ @@ -2052,7 +2054,8 @@ bool d3d12_pipeline_state_is_render_pass_compatible(const struct d3d12_pipeline_ STATIC_ASSERT(sizeof(struct vkd3d_shader_transform_feedback_element) == sizeof(D3D12_SO_DECLARATION_ENTRY));
static HRESULT d3d12_graphics_pipeline_state_create_render_pass( - struct d3d12_graphics_pipeline_state *graphics, struct d3d12_device *device) + struct d3d12_graphics_pipeline_state *graphics, struct d3d12_device *device, + VkFormat dynamic_dsv_format, VkRenderPass *vk_render_pass) { struct vkd3d_render_pass_key key; unsigned int i; @@ -2064,7 +2067,12 @@ static HRESULT d3d12_graphics_pipeline_state_create_render_pass( key.stencil_enable = graphics->ds_desc.stencilTestEnable; key.depth_stencil_write = graphics->ds_desc.depthWriteEnable || graphics->ds_desc.front.writeMask; - key.vk_formats[0] = graphics->dsv_format; + + if (!(key.vk_formats[0] = graphics->dsv_format)) + key.vk_formats[0] = dynamic_dsv_format; + + if (!key.vk_formats[0]) + FIXME("Compiling with DXGI_FORMAT_UNKNOWN.\n"); } else { @@ -2082,7 +2090,7 @@ static HRESULT d3d12_graphics_pipeline_state_create_render_pass( key.padding = 0; key.sample_count = graphics->ms_desc.rasterizationSamples;
- return vkd3d_render_pass_cache_find(&device->render_pass_cache, device, &key, &graphics->render_pass); + return vkd3d_render_pass_cache_find(&device->render_pass_cache, device, &key, vk_render_pass); }
static HRESULT d3d12_pipeline_state_init_graphics(struct d3d12_pipeline_state *state, @@ -2177,26 +2185,29 @@ static HRESULT d3d12_pipeline_state_init_graphics(struct d3d12_pipeline_state *s rt_count = ARRAY_SIZE(graphics->blend_attachments); }
- if (desc->DSVFormat == DXGI_FORMAT_UNKNOWN - && (desc->DepthStencilState.DepthEnable || desc->DepthStencilState.StencilEnable)) - FIXME("DSV format is DXGI_FORMAT_UNKNOWN, disabling depth/stencil tests.\n"); - graphics->rt_idx = 0; graphics->null_attachment_mask = 0; - if (desc->DSVFormat != DXGI_FORMAT_UNKNOWN - && (desc->DepthStencilState.DepthEnable || desc->DepthStencilState.StencilEnable)) + if (desc->DepthStencilState.DepthEnable || desc->DepthStencilState.StencilEnable) { - if (!(format = vkd3d_get_format(device, desc->DSVFormat, true))) + if (desc->DSVFormat == DXGI_FORMAT_UNKNOWN) + { + WARN("DSV format is DXGI_FORMAT_UNKNOWN.\n"); + graphics->dsv_format = VK_FORMAT_UNDEFINED; + } + else if ((format = vkd3d_get_format(device, desc->DSVFormat, true))) + { + if (!(format->vk_aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT))) + FIXME("Format %#x is not depth/stencil format.\n", format->dxgi_format); + + graphics->dsv_format = format->vk_format; + } + else { WARN("Invalid DSV format %#x.\n", desc->DSVFormat); hr = E_INVALIDARG; goto fail; }
- if (!(format->vk_aspect_mask & (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT))) - FIXME("Format %#x is not depth/stencil format.\n", format->dxgi_format); - - graphics->dsv_format = format->vk_format; ++graphics->rt_idx;
if (!desc->PS.pShaderBytecode) @@ -2538,7 +2549,10 @@ static HRESULT d3d12_pipeline_state_init_graphics(struct d3d12_pipeline_state *s
ds_desc_from_d3d12(&graphics->ds_desc, &desc->DepthStencilState);
- if (FAILED(hr = d3d12_graphics_pipeline_state_create_render_pass(graphics, device))) + if (graphics->dsv_format == VK_FORMAT_UNDEFINED) + graphics->render_pass = VK_NULL_HANDLE; + else if (FAILED(hr = d3d12_graphics_pipeline_state_create_render_pass(graphics, + device, 0, &graphics->render_pass))) goto fail;
graphics->root_signature = root_signature; @@ -2640,7 +2654,7 @@ static enum VkPrimitiveTopology vk_topology_from_d3d12_topology(D3D12_PRIMITIVE_ }
static VkPipeline d3d12_pipeline_state_find_compiled_pipeline(const struct d3d12_pipeline_state *state, - const struct vkd3d_pipeline_key *key) + const struct vkd3d_pipeline_key *key, VkRenderPass *vk_render_pass) { const struct d3d12_graphics_pipeline_state *graphics = &state->u.graphics; struct d3d12_device *device = state->device; @@ -2648,6 +2662,8 @@ static VkPipeline d3d12_pipeline_state_find_compiled_pipeline(const struct d3d12 struct vkd3d_compiled_pipeline *current; int rc;
+ *vk_render_pass = VK_NULL_HANDLE; + if (!(rc = pthread_mutex_lock(&device->mutex))) { LIST_FOR_EACH_ENTRY(current, &graphics->compiled_pipelines, struct vkd3d_compiled_pipeline, entry) @@ -2655,6 +2671,7 @@ static VkPipeline d3d12_pipeline_state_find_compiled_pipeline(const struct d3d12 if (!memcmp(¤t->key, key, sizeof(*key))) { vk_pipeline = current->vk_pipeline; + *vk_render_pass = current->vk_render_pass; break; } } @@ -2669,7 +2686,7 @@ static VkPipeline d3d12_pipeline_state_find_compiled_pipeline(const struct d3d12 }
static bool d3d12_pipeline_state_put_pipeline_to_cache(struct d3d12_pipeline_state *state, - const struct vkd3d_pipeline_key *key, VkPipeline vk_pipeline) + const struct vkd3d_pipeline_key *key, VkPipeline vk_pipeline, VkRenderPass vk_render_pass) { struct d3d12_graphics_pipeline_state *graphics = &state->u.graphics; struct vkd3d_compiled_pipeline *compiled_pipeline, *current; @@ -2681,6 +2698,7 @@ static bool d3d12_pipeline_state_put_pipeline_to_cache(struct d3d12_pipeline_sta
compiled_pipeline->key = *key; compiled_pipeline->vk_pipeline = vk_pipeline; + compiled_pipeline->vk_render_pass = vk_render_pass;
if ((rc = pthread_mutex_lock(&device->mutex))) { @@ -2707,7 +2725,8 @@ static bool d3d12_pipeline_state_put_pipeline_to_cache(struct d3d12_pipeline_sta }
VkPipeline d3d12_pipeline_state_get_or_create_pipeline(struct d3d12_pipeline_state *state, - D3D12_PRIMITIVE_TOPOLOGY topology, const uint32_t *strides) + D3D12_PRIMITIVE_TOPOLOGY topology, const uint32_t *strides, VkFormat dsv_format, + VkRenderPass *vk_render_pass) { VkVertexInputBindingDescription bindings[D3D12_IA_VERTEX_INPUT_RESOURCE_SLOT_COUNT]; const struct vkd3d_vk_device_procs *vk_procs = &state->device->vk_procs; @@ -2725,6 +2744,7 @@ VkPipeline d3d12_pipeline_state_get_or_create_pipeline(struct d3d12_pipeline_sta unsigned int i; uint32_t mask; VkResult vr; + HRESULT hr;
static const VkPipelineViewportStateCreateInfo vp_desc = { @@ -2754,6 +2774,8 @@ VkPipeline d3d12_pipeline_state_get_or_create_pipeline(struct d3d12_pipeline_sta
assert(d3d12_pipeline_state_is_graphics(state));
+ *vk_render_pass = VK_NULL_HANDLE; + memset(&pipeline_key, 0, sizeof(pipeline_key)); pipeline_key.topology = topology;
@@ -2783,7 +2805,9 @@ VkPipeline d3d12_pipeline_state_get_or_create_pipeline(struct d3d12_pipeline_sta ++binding_count; }
- if ((vk_pipeline = d3d12_pipeline_state_find_compiled_pipeline(state, &pipeline_key))) + pipeline_key.dsv_format = dsv_format; + + if ((vk_pipeline = d3d12_pipeline_state_find_compiled_pipeline(state, &pipeline_key, vk_render_pass))) return vk_pipeline;
input_desc.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO; @@ -2846,6 +2870,18 @@ VkPipeline d3d12_pipeline_state_get_or_create_pipeline(struct d3d12_pipeline_sta pipeline_desc.subpass = 0; pipeline_desc.basePipelineHandle = VK_NULL_HANDLE; pipeline_desc.basePipelineIndex = -1; + + /* Create a render pass for pipelines with DXGI_FORMAT_UNKNOWN. */ + if (!pipeline_desc.renderPass) + { + TRACE("Compiling %p with DSV format %#x.\n", state, dsv_format); + if (FAILED(hr = d3d12_graphics_pipeline_state_create_render_pass(graphics, device, dsv_format, + &pipeline_desc.renderPass))) + return VK_NULL_HANDLE; + + *vk_render_pass = pipeline_desc.renderPass; + } + if ((vr = VK_CALL(vkCreateGraphicsPipelines(device->vk_device, device->vk_pipeline_cache, 1, &pipeline_desc, NULL, &vk_pipeline))) < 0) { @@ -2853,12 +2889,12 @@ VkPipeline d3d12_pipeline_state_get_or_create_pipeline(struct d3d12_pipeline_sta return VK_NULL_HANDLE; }
- if (d3d12_pipeline_state_put_pipeline_to_cache(state, &pipeline_key, vk_pipeline)) + if (d3d12_pipeline_state_put_pipeline_to_cache(state, &pipeline_key, vk_pipeline, pipeline_desc.renderPass)) return vk_pipeline;
/* Other thread compiled the pipeline before us. */ VK_CALL(vkDestroyPipeline(device->vk_device, vk_pipeline, NULL)); - vk_pipeline = d3d12_pipeline_state_find_compiled_pipeline(state, &pipeline_key); + vk_pipeline = d3d12_pipeline_state_find_compiled_pipeline(state, &pipeline_key, vk_render_pass); if (!vk_pipeline) ERR("Could not get the pipeline compiled by other thread from the cache.\n"); return vk_pipeline; diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h index 4ae6afe128db..3fd1eacbfc77 100644 --- a/libs/vkd3d/vkd3d_private.h +++ b/libs/vkd3d/vkd3d_private.h @@ -741,7 +741,8 @@ HRESULT d3d12_pipeline_state_create_compute(struct d3d12_device *device, HRESULT d3d12_pipeline_state_create_graphics(struct d3d12_device *device, const D3D12_GRAPHICS_PIPELINE_STATE_DESC *desc, struct d3d12_pipeline_state **state) DECLSPEC_HIDDEN; VkPipeline d3d12_pipeline_state_get_or_create_pipeline(struct d3d12_pipeline_state *state, - D3D12_PRIMITIVE_TOPOLOGY topology, const uint32_t *strides) DECLSPEC_HIDDEN; + D3D12_PRIMITIVE_TOPOLOGY topology, const uint32_t *strides, VkFormat dsv_format, + VkRenderPass *vk_render_pass) DECLSPEC_HIDDEN; bool d3d12_pipeline_state_is_render_pass_compatible(const struct d3d12_pipeline_state *state_a, const struct d3d12_pipeline_state *state_b) DECLSPEC_HIDDEN; struct d3d12_pipeline_state *unsafe_impl_from_ID3D12PipelineState(ID3D12PipelineState *iface) DECLSPEC_HIDDEN; @@ -861,11 +862,13 @@ struct d3d12_command_list unsigned int fb_width; unsigned int fb_height; unsigned int fb_layer_count; + VkFormat dsv_format;
bool xfb_enabled;
VkFramebuffer current_framebuffer; VkPipeline current_pipeline; + VkRenderPass pso_render_pass; VkRenderPass current_render_pass; struct vkd3d_pipeline_bindings pipeline_bindings[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
-
Henri Verbeet -
Józef Kucia