Add debug ring support to ExecuteIndirect.

include/vkd3d.h

@@ -86,7 +86,7 @@ extern "C" {
 #define VKD3D_CONFIG_FLAG_FORCE_RAW_VA_CBV (1ull << 28)
 #define VKD3D_CONFIG_FLAG_ZERO_MEMORY_WORKAROUNDS_COMMITTED_BUFFER_UAV (1ull << 29)
 #define VKD3D_CONFIG_FLAG_ALLOW_SBT_COLLECTION (1ull << 30)
-/* Bit 31 is vacant */
+#define VKD3D_CONFIG_FLAG_BREADCRUMBS_TRACE_INDIRECT (1ull << 31)
 #define VKD3D_CONFIG_FLAG_USE_HOST_IMPORT_FALLBACK (1ull << 32)
 #define VKD3D_CONFIG_FLAG_PREALLOCATE_SRV_MIP_CLAMPS (1ull << 33)
 #define VKD3D_CONFIG_FLAG_FORCE_INITIAL_TRANSITION (1ull << 34)

libs/vkd3d/command.c

@@ -6994,6 +6994,68 @@ static bool d3d12_command_list_emit_multi_dispatch_indirect_count(struct d3d12_c
     return true;
 }
 
+static void d3d12_command_list_emit_execute_indirect_debug_ring(struct d3d12_command_list *list,
+        struct d3d12_command_signature *signature,
+        VkDeviceAddress indirect_args, VkDeviceAddress count_arg, uint32_t max_commands)
+{
+    const struct vkd3d_vk_device_procs *vk_procs = &list->device->vk_procs;
+    struct vkd3d_execute_indirect_debug_ring_args args;
+    VkCommandBuffer vk_patch_cmd_buffer;
+    VkMemoryBarrier2 vk_barrier;
+    VkDependencyInfo dep_info;
+
+    memset(&args, 0, sizeof(args));
+    args.api_buffer_va = indirect_args;
+    args.indirect_count_va = count_arg;
+    args.api_buffer_word_stride = signature->desc.ByteStride / sizeof(uint32_t);
+
+    if (vkd3d_config_flags & VKD3D_CONFIG_FLAG_BREADCRUMBS_TRACE_INDIRECT)
+    {
+        args.debug_tag = signature->desc.pArgumentDescs[signature->desc.NumArgumentDescs - 1].Type;
+        args.implicit_instance = vkd3d_atomic_uint32_increment(
+                &list->device->debug_ring.implicit_instance_count, vkd3d_memory_order_relaxed) - 1;
+    }
+
+    /* Allow correlation against breadcrumb log. */
+    VKD3D_BREADCRUMB_TAG("Implicit instance (plain)");
+    VKD3D_BREADCRUMB_AUX32(args.implicit_instance);
+
+    d3d12_command_allocator_allocate_init_post_indirect_command_buffer(list->allocator, list);
+    vk_patch_cmd_buffer = list->cmd.vk_init_commands_post_indirect_barrier;
+
+    if (vk_patch_cmd_buffer == list->cmd.vk_command_buffer)
+        d3d12_command_list_end_current_render_pass(list, true);
+
+    VK_CALL(vkCmdBindPipeline(vk_patch_cmd_buffer, VK_PIPELINE_BIND_POINT_COMPUTE,
+            signature->debug_ring_pipeline.vk_pipeline));
+    VK_CALL(vkCmdPushConstants(vk_patch_cmd_buffer,
+            signature->debug_ring_pipeline.vk_pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT,
+            0, sizeof(args), &args));
+
+    VK_CALL(vkCmdDispatch(vk_patch_cmd_buffer, max_commands, 1, 1));
+
+    if (vk_patch_cmd_buffer == list->cmd.vk_command_buffer)
+    {
+        memset(&dep_info, 0, sizeof(dep_info));
+        dep_info.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO;
+        dep_info.memoryBarrierCount = 1;
+        dep_info.pMemoryBarriers = &vk_barrier;
+
+        memset(&vk_barrier, 0, sizeof(vk_barrier));
+        vk_barrier.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER_2;
+        vk_barrier.srcStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT;
+        vk_barrier.srcAccessMask = 0;
+        vk_barrier.dstStageMask = VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT;
+        vk_barrier.dstAccessMask = 0;
+        VK_CALL(vkCmdPipelineBarrier2(vk_patch_cmd_buffer, &dep_info));
+
+        d3d12_command_list_invalidate_current_pipeline(list, true);
+        d3d12_command_list_invalidate_root_parameters(list, &list->compute_bindings, true, &list->graphics_bindings);
+    }
+    else
+        list->cmd.indirect_meta->need_compute_to_indirect_barrier = true;
+}
+
 static bool d3d12_command_list_emit_multi_dispatch_indirect_count_state(struct d3d12_command_list *list,
         struct d3d12_command_signature *signature,
         VkDeviceAddress indirect_args,
@@ -7045,6 +7107,17 @@ static bool d3d12_command_list_emit_multi_dispatch_indirect_count_state(struct d
     args.stride_words = stride / sizeof(uint32_t);
     args.dispatch_offset_words = signature->state_template.compute.dispatch_offset_words;
 
+    if (vkd3d_config_flags & VKD3D_CONFIG_FLAG_BREADCRUMBS_TRACE_INDIRECT)
+    {
+        args.debug_tag = UINT32_MAX;
+        args.implicit_instance = vkd3d_atomic_uint32_increment(
+                &list->device->debug_ring.implicit_instance_count, vkd3d_memory_order_relaxed) - 1;
+    }
+
+    /* Allow correlation against breadcrumb log. */
+    VKD3D_BREADCRUMB_TAG("Implicit instance (compute template)");
+    VKD3D_BREADCRUMB_AUX32(args.implicit_instance);
+
     d3d12_command_allocator_allocate_init_post_indirect_command_buffer(list->allocator, list);
     vk_patch_cmd_buffer = list->cmd.vk_init_commands_post_indirect_barrier;
 
@@ -13584,7 +13657,7 @@ static void d3d12_command_list_execute_indirect_state_template_dgc(
     current_pipeline = list->current_pipeline;
 
     memset(&patch_args, 0, sizeof(patch_args));
-    patch_args.debug_tag = 0; /* Modify to non-zero value as desired when debugging. */
+    patch_args.debug_tag = (vkd3d_config_flags & VKD3D_CONFIG_FLAG_BREADCRUMBS_TRACE_INDIRECT) ? UINT32_MAX : 0;
 
     /* If everything regarding alignment works out, we can just reuse the app indirect buffer instead. */
     require_ibo_update = false;
@@ -13664,9 +13737,12 @@ static void d3d12_command_list_execute_indirect_state_template_dgc(
         if (patch_args.debug_tag != 0)
         {
             /* Makes log easier to understand since a sorted log will appear in-order. */
-            static uint32_t vkd3d_implicit_instance_count;
             patch_args.implicit_instance = vkd3d_atomic_uint32_increment(
-                    &vkd3d_implicit_instance_count, vkd3d_memory_order_relaxed) - 1;
+                    &list->device->debug_ring.implicit_instance_count, vkd3d_memory_order_relaxed) - 1;
+
+            /* Allow correlation against breadcrumb log. */
+            VKD3D_BREADCRUMB_TAG("Implicit instance (template)");
+            VKD3D_BREADCRUMB_AUX32(patch_args.implicit_instance);
         }
 
         d3d12_command_allocator_allocate_init_post_indirect_command_buffer(list->allocator, list);
@@ -14062,6 +14138,16 @@ static void STDMETHODCALLTYPE d3d12_command_list_ExecuteIndirect(d3d12_command_l
         return;
     }
 
+    d3d12_command_list_end_transfer_batch(list);
+
+    if (sig_impl->debug_ring_pipeline.vk_pipeline)
+    {
+        d3d12_command_list_emit_execute_indirect_debug_ring(list, sig_impl,
+                arg_impl->res.va + arg_buffer_offset,
+                count_impl ? count_impl->res.va + count_buffer_offset : 0,
+                max_command_count);
+    }
+
     /* Temporary workaround, since we cannot parse non-draw arguments yet. Point directly
      * to the first argument. Should avoid hard crashes for now. */
     arg_buffer_offset += sig_impl->argument_buffer_offset_for_command;
@@ -14135,7 +14221,6 @@ static void STDMETHODCALLTYPE d3d12_command_list_ExecuteIndirect(d3d12_command_l
             scratch.va = arg_impl->res.va + arg_buffer_offset;
         }
 
-        d3d12_command_list_end_transfer_batch(list);
         switch (arg_desc->Type)
         {
             case D3D12_INDIRECT_ARGUMENT_TYPE_DRAW:
@@ -20488,6 +20573,13 @@ static HRESULT d3d12_command_signature_init_state_template_dgc_nv(struct d3d12_c
         VKD3D_PATCH_COMMAND_TOKEN_COPY_ROOT_VA_HI,
     };
 
+    static const enum vkd3d_patch_command_token dispatch_types[] =
+    {
+        VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_X,
+        VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_Y,
+        VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_Z,
+    };
+
     static const VkIndexType vk_index_types[] = { VK_INDEX_TYPE_UINT32, VK_INDEX_TYPE_UINT16 };
     static const uint32_t d3d_index_types[] = { DXGI_FORMAT_R32_UINT, DXGI_FORMAT_R16_UINT };
 
@@ -20661,9 +20753,8 @@ static HRESULT d3d12_command_signature_init_state_template_dgc_nv(struct d3d12_c
                 token.offset = stream_stride;
                 stream_stride += sizeof(VkDispatchIndirectCommand);
                 dst_word_offset = token.offset / sizeof(uint32_t);
-                /* TODO: Rebase on top of debug-ring-indirect. */
-                generic_u32_copy_count = 0;
-                generic_u32_copy_types = NULL;
+                generic_u32_copy_count = ARRAY_SIZE(dispatch_types);
+                generic_u32_copy_types = dispatch_types;
                 break;
 
             default:
@@ -21244,6 +21335,15 @@ HRESULT d3d12_command_signature_create(struct d3d12_device *device, struct d3d12
          * for optimal reordering. */
         vkd3d_atomic_uint32_store_explicit(&device->device_has_dgc_templates, 1, vkd3d_memory_order_relaxed);
     }
+    else
+    {
+        if (vkd3d_config_flags & VKD3D_CONFIG_FLAG_BREADCRUMBS_TRACE_INDIRECT)
+        {
+            vkd3d_meta_get_execute_indirect_debug_ring_pipeline(&device->meta_ops,
+                    signature_size / sizeof(uint32_t),
+                    &object->debug_ring_pipeline);
+        }
+    }
 
     object->argument_buffer_offset_for_command = argument_buffer_offset;
     d3d_destruction_notifier_init(&object->destruction_notifier, (IUnknown*)&object->ID3D12CommandSignature_iface);

libs/vkd3d/debug_ring.c

@@ -86,6 +86,9 @@ static const char *vkd3d_patch_command_token_str(enum vkd3d_patch_command_token
         case VKD3D_PATCH_COMMAND_TOKEN_COPY_MESH_TASKS_X: return "Mesh Tasks (X)";
         case VKD3D_PATCH_COMMAND_TOKEN_COPY_MESH_TASKS_Y: return "Mesh Tasks (Y)";
         case VKD3D_PATCH_COMMAND_TOKEN_COPY_MESH_TASKS_Z: return "Mesh Tasks (Z)";
+        case VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_X: return "X";
+        case VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_Y: return "Y";
+        case VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_Z: return "Z";
         default: return "???";
     }
 }
@@ -107,6 +110,29 @@ static bool vkd3d_patch_command_token_is_hex(enum vkd3d_patch_command_token toke
     }
 }
 
+static const char *vkd3d_debug_tag_to_str(uint32_t value)
+{
+    switch (value)
+    {
+        case D3D12_INDIRECT_ARGUMENT_TYPE_DRAW:
+            return "Draw";
+        case D3D12_INDIRECT_ARGUMENT_TYPE_DRAW_INDEXED:
+            return "DrawIndexed";
+        case D3D12_INDIRECT_ARGUMENT_TYPE_DISPATCH:
+            return "Dispatch";
+        case D3D12_INDIRECT_ARGUMENT_TYPE_DISPATCH_MESH:
+            return "Mesh";
+        case D3D12_INDIRECT_ARGUMENT_TYPE_DISPATCH_RAYS:
+            return "RayGen";
+        case UINT32_MAX:
+            return "Template";
+        default:
+            break;
+    }
+
+    return "???";
+}
+
 static bool vkd3d_shader_debug_ring_print_message(struct vkd3d_shader_debug_ring *ring,
         uint32_t word_offset, uint32_t message_word_count)
 {
@@ -136,8 +162,8 @@ static bool vkd3d_shader_debug_ring_print_message(struct vkd3d_shader_debug_ring
          * Make sure the log is sortable for easier debug.
          * TODO: Might consider a callback system that listeners from different subsystems can listen to and print their own messages,
          * but that is overengineering at this time ... */
-        snprintf(message_buffer, sizeof(message_buffer), "ExecuteIndirect: GlobalCommandIndex %010u, Debug tag %010u, DrawID %04u (ThreadID %04u): ",
-                debug_instance, debug_thread_id[0], debug_thread_id[1], debug_thread_id[2]);
+        snprintf(message_buffer, sizeof(message_buffer), "ExecuteIndirect: GlobalCommandIndex %010u, %s, DrawID %04u (ThreadID %04u): ",
+                debug_instance, vkd3d_debug_tag_to_str(debug_thread_id[0]), debug_thread_id[1], debug_thread_id[2]);
 
         if (message_word_count == 2)
         {
@@ -147,6 +173,76 @@ static bool vkd3d_shader_debug_ring_print_message(struct vkd3d_shader_debug_ring
                     READ_RING_WORD(word_offset + 0),
                     READ_RING_WORD(word_offset + 1));
         }
+        else if (message_word_count == 3)
+        {
+            static const enum vkd3d_patch_command_token draw_types[] =
+            {
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_VERTEX_COUNT,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_INSTANCE_COUNT,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_FIRST_VERTEX,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_FIRST_INSTANCE,
+            };
+
+            static const enum vkd3d_patch_command_token draw_indexed_types[] =
+            {
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_INDEX_COUNT,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_INSTANCE_COUNT,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_FIRST_INDEX,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_VERTEX_OFFSET,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_FIRST_INSTANCE,
+            };
+
+            static const enum vkd3d_patch_command_token dispatch_types[] =
+            {
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_X,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_Y,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_DISPATCH_Z,
+            };
+
+            static const enum vkd3d_patch_command_token mesh_types[] =
+            {
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_MESH_TASKS_X,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_MESH_TASKS_Y,
+                VKD3D_PATCH_COMMAND_TOKEN_COPY_MESH_TASKS_Z,
+            };
+
+            const char *tag_str = "?";
+            uint32_t value, index;
+
+            len = strlen(message_buffer);
+            avail = sizeof(message_buffer) - len;
+            /* word 0 is a dummy value. */
+            index = READ_RING_WORD(word_offset + 1);
+            value = READ_RING_WORD(word_offset + 2);
+
+            switch (debug_thread_id[0])
+            {
+                case D3D12_INDIRECT_ARGUMENT_TYPE_DRAW:
+                    if (index < ARRAY_SIZE(draw_types))
+                        tag_str = vkd3d_patch_command_token_str(draw_types[index]);
+                    break;
+
+                case D3D12_INDIRECT_ARGUMENT_TYPE_DRAW_INDEXED:
+                    if (index < ARRAY_SIZE(draw_indexed_types))
+                        tag_str = vkd3d_patch_command_token_str(draw_indexed_types[index]);
+                    break;
+
+                case D3D12_INDIRECT_ARGUMENT_TYPE_DISPATCH:
+                    if (index < ARRAY_SIZE(dispatch_types))
+                        tag_str = vkd3d_patch_command_token_str(dispatch_types[index]);
+                    break;
+
+                case D3D12_INDIRECT_ARGUMENT_TYPE_DISPATCH_MESH:
+                    if (index < ARRAY_SIZE(mesh_types))
+                        tag_str = vkd3d_patch_command_token_str(mesh_types[index]);
+                    break;
+
+                default:
+                    break;
+            }
+
+            snprintf(message_buffer + len, avail, "%s <- %u", tag_str, value);
+        }
         else if (message_word_count == 4)
         {
             union { uint32_t u32; float f32; int32_t s32; } value;

libs/vkd3d/device.c

@@ -992,6 +992,7 @@ static const struct vkd3d_debug_option vkd3d_config_options[] =
     {"preallocate_srv_mip_clamps", VKD3D_CONFIG_FLAG_PREALLOCATE_SRV_MIP_CLAMPS},
     {"force_initial_transition", VKD3D_CONFIG_FLAG_FORCE_INITIAL_TRANSITION},
     {"breadcrumbs_trace", VKD3D_CONFIG_FLAG_BREADCRUMBS | VKD3D_CONFIG_FLAG_BREADCRUMBS_TRACE},
+    {"breadcrumbs_trace_indirect", VKD3D_CONFIG_FLAG_BREADCRUMBS | VKD3D_CONFIG_FLAG_BREADCRUMBS_TRACE_INDIRECT},
     {"requires_compute_indirect_templates", VKD3D_CONFIG_FLAG_REQUIRES_COMPUTE_INDIRECT_TEMPLATES},
     {"skip_driver_workarounds", VKD3D_CONFIG_FLAG_SKIP_DRIVER_WORKAROUNDS},
     {"enable_experimental_features", VKD3D_CONFIG_FLAG_ENABLE_EXPERIMENTAL_FEATURES},
@@ -9138,14 +9139,14 @@ static HRESULT d3d12_device_init(struct d3d12_device *device,
     if (FAILED(hr = d3d12_device_create_sparse_init_timeline(device)))
         goto out_cleanup_sampler_state;
 
-    if (FAILED(hr = vkd3d_meta_ops_init(&device->meta_ops, device)))
+    if (FAILED(hr = vkd3d_shader_debug_ring_init(&device->debug_ring, device)))
         goto out_cleanup_sparse_timeline;
 
-    if (FAILED(hr = vkd3d_shader_debug_ring_init(&device->debug_ring, device)))
-        goto out_cleanup_meta_ops;
+    if (FAILED(hr = vkd3d_meta_ops_init(&device->meta_ops, device)))
+        goto out_cleanup_debug_ring;
 
     if (FAILED(hr = vkd3d_queue_timeline_trace_init(&device->queue_timeline_trace, device)))
-        goto out_cleanup_debug_ring;
+        goto out_cleanup_meta_ops;
 
     if (FAILED(hr = vkd3d_address_binding_tracker_init(&device->address_binding_tracker, device)))
         goto out_cleanup_queue_timeline_trace;
@@ -9215,10 +9216,10 @@ static HRESULT d3d12_device_init(struct d3d12_device *device,
     vkd3d_address_binding_tracker_cleanup(&device->address_binding_tracker, device);
 out_cleanup_queue_timeline_trace:
     vkd3d_queue_timeline_trace_cleanup(&device->queue_timeline_trace);
-out_cleanup_debug_ring:
-    vkd3d_shader_debug_ring_cleanup(&device->debug_ring, device);
 out_cleanup_meta_ops:
     vkd3d_meta_ops_cleanup(&device->meta_ops, device);
+out_cleanup_debug_ring:
+    vkd3d_shader_debug_ring_cleanup(&device->debug_ring, device);
 out_cleanup_sparse_timeline:
     vk_procs = &device->vk_procs;
     VK_CALL(vkDestroySemaphore(device->vk_device, device->sparse_init_timeline, NULL));

libs/vkd3d/meson.build

@@ -34,7 +34,6 @@ vkd3d_shaders =[
   'shaders/vs_swapchain_fullscreen.vert',
   'shaders/fs_swapchain_fullscreen.frag',
   'shaders/cs_execute_indirect_patch.comp',
-  'shaders/cs_execute_indirect_patch_debug_ring.comp',
   'shaders/cs_execute_indirect_multi_dispatch.comp',
   'shaders/cs_execute_indirect_multi_dispatch_state.comp',
 
@@ -60,6 +59,10 @@ vkd3d_shaders =[
   'shaders/cs_workgraph_distribute_payload_offsets.comp',
   'shaders/cs_workgraph_complete_compaction.comp',
   'shaders/cs_workgraph_setup_gpu_input.comp',
+
+  'shaders/cs_execute_indirect_debug_ring.comp',
+  'shaders/cs_execute_indirect_patch_debug_ring.comp',
+  'shaders/cs_execute_indirect_multi_dispatch_state_debug_ring.comp',
 ]
 
 vkd3d_src = [