Hack up test for repro.

Useful for test suite since a test can be comprised of several smaller
submissions, and it's easier to debug if we have one trace.

Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>

README.md

@@ -203,6 +203,9 @@ pass `-Denable_renderdoc=true` to Meson.
  vkd3d-proton will automatically make a capture when a specific shader is encountered.
  - `VKD3D_AUTO_CAPTURE_COUNTS` - A comma-separated list of indices. This can be used to control which queue submissions to capture.
  E.g., use `VKD3D_AUTO_CAPTURE_COUNTS=0,4,10` to capture the 0th (first submission), 4th and 10th submissions which are candidates for capturing.
+ If `VKD3D_AUTO_CAPTURE_COUNTS` is `-1`, the entire app runtime can be turned into one big capture.
+ This is only intended to be used when capturing something like the test suite,
+ or tiny applications with a finite runtime to make it easier to debug cross submission work.
 
  If only `VKD3D_AUTO_CAPTURE_COUNTS` is set, any queue submission is considered for capturing.
  If only `VKD3D_AUTO_CAPTURE_SHADER` is set, `VKD3D_AUTO_CAPTURE_COUNTS` is considered to be equal to `"0"`, i.e. a capture is only

libs/vkd3d/command.c

@@ -2438,6 +2438,8 @@ static VkImageLayout vk_separate_stencil_layout(VkImageLayout combined_layout)
             VK_IMAGE_LAYOUT_STENCIL_ATTACHMENT_OPTIMAL : VK_IMAGE_LAYOUT_STENCIL_READ_ONLY_OPTIMAL;
 }
 
+static bool d3d12_resource_may_alias_other_resources(struct d3d12_resource *resource);
+
 static void d3d12_command_list_clear_attachment_pass(struct d3d12_command_list *list, struct d3d12_resource *resource,
         struct vkd3d_view *view, VkImageAspectFlags clear_aspects, const VkClearValue *clear_value, UINT rect_count,
         const D3D12_RECT *rects, bool is_bound)
@@ -2453,6 +2455,8 @@ static void d3d12_command_list_clear_attachment_pass(struct d3d12_command_list *
     VkSubpassEndInfoKHR subpass_end_info;
     VkRenderPassCreateInfo2KHR pass_info;
     VkRenderPassBeginInfo begin_info;
+    bool require_full_src_barrier;
+    bool require_full_dst_barrier;
     VkFramebuffer vk_framebuffer;
     VkRenderPass vk_render_pass;
     VkPipelineStageFlags stages;
@@ -2461,8 +2465,16 @@ static void d3d12_command_list_clear_attachment_pass(struct d3d12_command_list *
     VkAccessFlags access;
     VkExtent3D extent;
     bool clear_op;
+    bool discard;
     VkResult vr;
 
+    /* There is no reason to introduce a full barrier for a clear unless we are forced by
+     * the API to inject a true discard.
+     * For aliased resources in D3D12, a full clear can complete the alias ownership transfer,
+     * and we must be conservative and use UNDEFINED oldLayout. For committed resources (common case),
+     * this can never happen and we rely on the initial transition mechanism instead. */
+    discard = d3d12_resource_may_alias_other_resources(resource);
+
     attachment_desc.sType = VK_STRUCTURE_TYPE_ATTACHMENT_DESCRIPTION_2_KHR;
     attachment_desc.pNext = NULL;
     attachment_desc.flags = 0;
@@ -2562,7 +2574,7 @@ static void d3d12_command_list_clear_attachment_pass(struct d3d12_command_list *
 
         /* Ignore 3D images as re-initializing those may cause us to
          * discard the entire image, not just the layers to clear. */
-        if (resource->desc.Dimension != D3D12_RESOURCE_DIMENSION_TEXTURE3D)
+        if (discard && resource->desc.Dimension != D3D12_RESOURCE_DIMENSION_TEXTURE3D)
         {
             if (separate_ds_layouts)
             {
@@ -2598,28 +2610,60 @@ static void d3d12_command_list_clear_attachment_pass(struct d3d12_command_list *
         subpass_desc.pColorAttachments = &attachment_ref;
     }
 
+    require_full_src_barrier = (attachment_desc.initialLayout != attachment_ref.layout) ||
+            (separate_ds_layouts && (stencil_attachment_desc.stencilInitialLayout != stencil_attachment_ref.stencilLayout));
+    require_full_dst_barrier = (attachment_desc.finalLayout != attachment_ref.layout) ||
+            (separate_ds_layouts && (stencil_attachment_desc.stencilFinalLayout != stencil_attachment_ref.stencilLayout));
+
     dependencies[0].sType = VK_STRUCTURE_TYPE_SUBPASS_DEPENDENCY_2_KHR;
     dependencies[0].pNext = NULL;
     dependencies[0].srcSubpass = VK_SUBPASS_EXTERNAL;
     dependencies[0].dstSubpass = 0;
+
     dependencies[0].srcStageMask = stages;
     dependencies[0].dstStageMask = stages;
-    dependencies[0].srcAccessMask = clear_op ? access : 0;
-    dependencies[0].dstAccessMask = access;
-    dependencies[0].dependencyFlags = 0;
     dependencies[0].viewOffset = 0;
 
+    if (require_full_src_barrier)
+    {
+        dependencies[0].srcAccessMask = clear_op ? access : 0;
+        dependencies[0].dstAccessMask = access;
+        dependencies[0].dependencyFlags = 0;
+    }
+    else
+    {
+        /* Same setup as normal render passes.
+         * The 0 access mask is technically a hack.
+         * See vkd3d_render_pass_cache_create_pass_locked(). */
+        dependencies[0].srcAccessMask = 0;
+        dependencies[0].dstAccessMask = 0;
+        dependencies[0].dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT;
+    }
+
     dependencies[1].sType = VK_STRUCTURE_TYPE_SUBPASS_DEPENDENCY_2_KHR;
     dependencies[1].pNext = NULL;
     dependencies[1].srcSubpass = 0;
     dependencies[1].dstSubpass = VK_SUBPASS_EXTERNAL;
     dependencies[1].srcStageMask = stages;
     dependencies[1].dstStageMask = stages;
-    dependencies[1].srcAccessMask = access;
-    dependencies[1].dstAccessMask = 0;
-    dependencies[1].dependencyFlags = 0;
     dependencies[1].viewOffset = 0;
 
+    if (require_full_dst_barrier)
+    {
+        dependencies[1].srcAccessMask = access;
+        dependencies[1].dstAccessMask = 0;
+        dependencies[1].dependencyFlags = 0;
+    }
+    else
+    {
+        /* Same setup as normal render passes.
+        * The 0 access mask is technically a hack.
+        * See vkd3d_render_pass_cache_create_pass_locked(). */
+        dependencies[1].srcAccessMask = 0;
+        dependencies[1].dstAccessMask = 0;
+        dependencies[1].dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT;
+    }
+
     pass_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO_2_KHR;
     pass_info.pNext = NULL;
     pass_info.flags = 0;
@@ -8239,6 +8283,13 @@ static void STDMETHODCALLTYPE d3d12_command_list_DiscardResource(d3d12_command_l
         return;
     }
 
+    /* Only care about resources which could alias memory with something else.
+     * We only need to transition committed resources once,
+     * and render targets tend to be that.
+     * There are no particular performance benefits in using DiscardResource as far as I know ... */
+    if (!d3d12_resource_may_alias_other_resources(texture))
+        return;
+
     /* Assume that pRegion == NULL means that we should discard
      * the entire resource. This does not seem to be documented. */
     resource_subresource_count = d3d12_resource_get_sub_resource_count(texture);

libs/vkd3d/device.c

@@ -2538,6 +2538,12 @@ static void d3d12_device_destroy(struct d3d12_device *device)
     vkd3d_memory_allocator_cleanup(&device->memory_allocator, device);
     /* Tear down descriptor global info late, so we catch last minute faults after we drain the queues. */
     vkd3d_descriptor_debug_free_global_info(device->descriptor_qa_global_info, device);
+
+#ifdef VKD3D_ENABLE_RENDERDOC
+    if (vkd3d_renderdoc_active() && vkd3d_renderdoc_global_capture_enabled())
+        vkd3d_renderdoc_end_capture(device->vkd3d_instance->vk_instance);
+#endif
+
     VK_CALL(vkDestroyDevice(device->vk_device, NULL));
     pthread_mutex_destroy(&device->mutex);
     if (device->parent)
@@ -5242,6 +5248,12 @@ static HRESULT d3d12_device_init(struct d3d12_device *device,
         IUnknown_AddRef(device->parent);
 
     d3d12_device_caps_init(device);
+
+#ifdef VKD3D_ENABLE_RENDERDOC
+    if (vkd3d_renderdoc_active() && vkd3d_renderdoc_global_capture_enabled())
+        vkd3d_renderdoc_begin_capture(device->vkd3d_instance->vk_instance);
+#endif
+
     return S_OK;
 
 out_cleanup_global_pipeline_cache:

libs/vkd3d/renderdoc.c

@@ -42,6 +42,7 @@ static vkd3d_shader_hash_t renderdoc_capture_shader_hash;
 static uint32_t *renderdoc_capture_counts;
 static size_t renderdoc_capture_counts_count;
 static bool vkd3d_renderdoc_is_active;
+static bool vkd3d_renderdoc_global_capture;
 
 static void vkd3d_renderdoc_init_capture_count_list(const char *env)
 {
@@ -49,6 +50,13 @@ static void vkd3d_renderdoc_init_capture_count_list(const char *env)
     uint32_t count;
     char *endp;
 
+    if (strcmp(env, "-1") == 0)
+    {
+        INFO("Doing one big capture of the entire lifetime of a device.\n");
+        vkd3d_renderdoc_global_capture = true;
+        return;
+    }
+
     while (*env != '\0')
     {
         errno = 0;
@@ -180,6 +188,11 @@ bool vkd3d_renderdoc_active(void)
     return vkd3d_renderdoc_is_active;
 }
 
+bool vkd3d_renderdoc_global_capture_enabled(void)
+{
+    return vkd3d_renderdoc_global_capture;
+}
+
 bool vkd3d_renderdoc_should_capture_shader_hash(vkd3d_shader_hash_t hash)
 {
     return (renderdoc_capture_shader_hash == hash) || (renderdoc_capture_shader_hash == 0);
@@ -190,9 +203,12 @@ bool vkd3d_renderdoc_begin_capture(void *instance)
     static uint32_t overall_counter;
     uint32_t counter;
 
-    counter = vkd3d_atomic_uint32_increment(&overall_counter, vkd3d_memory_order_relaxed) - 1;
-    if (!vkd3d_renderdoc_enable_submit_counter(counter))
-        return false;
+    if (!vkd3d_renderdoc_global_capture)
+    {
+        counter = vkd3d_atomic_uint32_increment(&overall_counter, vkd3d_memory_order_relaxed) - 1;
+        if (!vkd3d_renderdoc_enable_submit_counter(counter))
+            return false;
+    }
 
     if (renderdoc_api)
         renderdoc_api->StartFrameCapture(RENDERDOC_DEVICEPOINTER_FROM_VKINSTANCE(instance), NULL);
@@ -215,6 +231,9 @@ void vkd3d_renderdoc_command_list_check_capture(struct d3d12_command_list *list,
 {
     unsigned int i;
 
+    if (vkd3d_renderdoc_global_capture_enabled())
+        return;
+
     if (vkd3d_renderdoc_active() && state)
     {
         if (state->vk_bind_point == VK_PIPELINE_BIND_POINT_COMPUTE)
@@ -246,6 +265,9 @@ bool vkd3d_renderdoc_command_queue_begin_capture(struct d3d12_command_queue *com
     VkDebugUtilsLabelEXT capture_label;
     bool debug_capture;
 
+    if (vkd3d_renderdoc_global_capture_enabled())
+        return false;
+
     debug_capture = vkd3d_renderdoc_begin_capture(command_queue->device->vkd3d_instance->vk_instance);
     if (debug_capture && !vkd3d_renderdoc_loaded_api())
     {
@@ -273,6 +295,9 @@ void vkd3d_renderdoc_command_queue_end_capture(struct d3d12_command_queue *comma
     const struct vkd3d_vk_device_procs *vk_procs = &command_queue->device->vk_procs;
     VkDebugUtilsLabelEXT capture_label;
 
+    if (vkd3d_renderdoc_global_capture_enabled())
+        return;
+
     if (!vkd3d_renderdoc_loaded_api())
     {
         /* Magic fallback which lets us bridge the Wine barrier over to Linux RenderDoc. */

libs/vkd3d/vkd3d_renderdoc.h

@@ -26,6 +26,7 @@
 bool vkd3d_renderdoc_active(void);
 bool vkd3d_renderdoc_loaded_api(void);
 bool vkd3d_renderdoc_should_capture_shader_hash(vkd3d_shader_hash_t hash);
+bool vkd3d_renderdoc_global_capture_enabled(void);
 
 bool vkd3d_renderdoc_begin_capture(void *instance);
 void vkd3d_renderdoc_end_capture(void *instance);

tests/d3d12_sync.c

@@ -87,7 +87,7 @@ void test_queue_wait(void)
     context.pipeline_state = create_pipeline_state(context.device,
             context.root_signature, context.render_target_desc.Format, NULL, &ps, NULL);
 
-    cb = create_upload_buffer(device, sizeof(color), NULL);
+    cb = create_upload_buffer(device, 512, NULL);
 
     resource_desc = ID3D12Resource_GetDesc(context.render_target);
 
@@ -95,20 +95,6 @@ void test_queue_wait(void)
     buffer_size = row_pitch * resource_desc.Height;
     readback_buffer = create_readback_buffer(device, buffer_size);
 
-    ID3D12GraphicsCommandList_ClearRenderTargetView(command_list, context.rtv, white, 0, NULL);
-    ID3D12GraphicsCommandList_OMSetRenderTargets(command_list, 1, &context.rtv, false, NULL);
-    ID3D12GraphicsCommandList_SetGraphicsRootSignature(command_list, context.root_signature);
-    ID3D12GraphicsCommandList_SetPipelineState(command_list, context.pipeline_state);
-    ID3D12GraphicsCommandList_IASetPrimitiveTopology(command_list, D3D_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
-    ID3D12GraphicsCommandList_RSSetViewports(command_list, 1, &context.viewport);
-    ID3D12GraphicsCommandList_RSSetScissorRects(command_list, 1, &context.scissor_rect);
-    ID3D12GraphicsCommandList_SetGraphicsRootConstantBufferView(command_list, 0,
-            ID3D12Resource_GetGPUVirtualAddress(cb));
-    ID3D12GraphicsCommandList_DrawInstanced(command_list, 3, 1, 0, 0);
-
-    transition_resource_state(command_list, context.render_target,
-            D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE);
-
     dst_location.pResource = readback_buffer;
     dst_location.Type = D3D12_TEXTURE_COPY_TYPE_PLACED_FOOTPRINT;
     dst_location.PlacedFootprint.Offset = 0;
@@ -120,14 +106,47 @@ void test_queue_wait(void)
     src_location.pResource = context.render_target;
     src_location.Type = D3D12_TEXTURE_COPY_TYPE_SUBRESOURCE_INDEX;
     src_location.SubresourceIndex = 0;
-    ID3D12GraphicsCommandList_CopyTextureRegion(command_list, &dst_location, 0, 0, 0, &src_location, NULL);
 
+    ID3D12GraphicsCommandList_ClearRenderTargetView(command_list, context.rtv, white, 0, NULL);
+    ID3D12GraphicsCommandList_OMSetRenderTargets(command_list, 1, &context.rtv, false, NULL);
+    ID3D12GraphicsCommandList_SetGraphicsRootSignature(command_list, context.root_signature);
+    ID3D12GraphicsCommandList_SetPipelineState(command_list, context.pipeline_state);
+    ID3D12GraphicsCommandList_IASetPrimitiveTopology(command_list, D3D_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
+    ID3D12GraphicsCommandList_RSSetViewports(command_list, 1, &context.viewport);
+    ID3D12GraphicsCommandList_RSSetScissorRects(command_list, 1, &context.scissor_rect);
+    ID3D12GraphicsCommandList_SetGraphicsRootConstantBufferView(command_list, 0,
+            ID3D12Resource_GetGPUVirtualAddress(cb));
+    ID3D12GraphicsCommandList_DrawInstanced(command_list, 3, 1, 0, 0);
+    transition_resource_state(command_list, context.render_target,
+            D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE);
+    ID3D12GraphicsCommandList_CopyTextureRegion(command_list, &dst_location, 0, 0, 0, &src_location, NULL);
     transition_resource_state(command_list, context.render_target,
             D3D12_RESOURCE_STATE_COPY_SOURCE, D3D12_RESOURCE_STATE_RENDER_TARGET);
-
     hr = ID3D12GraphicsCommandList_Close(command_list);
     ok(hr == S_OK, "Failed to close command list, hr %#x.\n", hr);
 
+    ID3D12GraphicsCommandList *command_list2;
+
+    ID3D12Device_CreateCommandList(device, 0, D3D12_COMMAND_LIST_TYPE_DIRECT,
+            context.allocator, NULL, &IID_ID3D12GraphicsCommandList, (void**)&command_list2);
+    ID3D12GraphicsCommandList_ClearRenderTargetView(command_list2, context.rtv, white, 0, NULL);
+    ID3D12GraphicsCommandList_OMSetRenderTargets(command_list2, 1, &context.rtv, false, NULL);
+    ID3D12GraphicsCommandList_SetGraphicsRootSignature(command_list2, context.root_signature);
+    ID3D12GraphicsCommandList_SetPipelineState(command_list2, context.pipeline_state);
+    ID3D12GraphicsCommandList_IASetPrimitiveTopology(command_list2, D3D_PRIMITIVE_TOPOLOGY_TRIANGLELIST);
+    ID3D12GraphicsCommandList_RSSetViewports(command_list2, 1, &context.viewport);
+    ID3D12GraphicsCommandList_RSSetScissorRects(command_list2, 1, &context.scissor_rect);
+    ID3D12GraphicsCommandList_SetGraphicsRootConstantBufferView(command_list2, 0,
+            ID3D12Resource_GetGPUVirtualAddress(cb) + 0);
+    ID3D12GraphicsCommandList_DrawInstanced(command_list2, 3, 1, 0, 0);
+    transition_resource_state(command_list2, context.render_target,
+            D3D12_RESOURCE_STATE_RENDER_TARGET, D3D12_RESOURCE_STATE_COPY_SOURCE);
+    ID3D12GraphicsCommandList_CopyTextureRegion(command_list2, &dst_location, 0, 0, 0, &src_location, NULL);
+    transition_resource_state(command_list2, context.render_target,
+            D3D12_RESOURCE_STATE_COPY_SOURCE, D3D12_RESOURCE_STATE_RENDER_TARGET);
+    hr = ID3D12GraphicsCommandList_Close(command_list2);
+    ok(hr == S_OK, "Failed to close command list, hr %#x.\n", hr);
+
     /* Wait() with signaled fence */
     update_buffer_data(cb, 0, sizeof(green), &green);
     queue_wait(queue, fence, 1);
@@ -137,28 +156,23 @@ void test_queue_wait(void)
     check_readback_data_uint(&rb, NULL, 0xff00ff00, 0);
     release_resource_readback(&rb);
 
-    if (!vkd3d_test_platform_is_windows())
-    {
-        skip("Wait() is not implemented yet.\n"); /* FIXME */
-        goto skip_tests;
-    }
-
     /* Wait() before CPU signal */
     update_buffer_data(cb, 0, sizeof(blue), &blue);
-    queue_wait(queue, fence, 2);
+    queue_wait(queue, fence, 1000);
     exec_command_list(queue, command_list);
     queue_signal(queue, fence2, 1);
     hr = ID3D12Fence_SetEventOnCompletion(fence2, 1, event);
     ok(hr == S_OK, "Failed to set event on completion, hr %#x.\n", hr);
-    ret = wait_event(event, 0);
-    ok(ret == WAIT_TIMEOUT, "Got unexpected return value %#x.\n", ret);
-    init_readback(&rb, readback_buffer, buffer_size, resource_desc.Width, resource_desc.Height, 1, row_pitch);
-    check_readback_data_uint(&rb, NULL, 0xff00ff00, 0);
-    release_resource_readback(&rb);
-    value = ID3D12Fence_GetCompletedValue(fence2);
-    ok(value == 0, "Got unexpected value %"PRIu64".\n", value);
 
-    hr = ID3D12Fence_Signal(fence, 2);
+    //ret = wait_event(event, 0);
+    //ok(ret == WAIT_TIMEOUT, "Got unexpected return value %#x.\n", ret);
+    //init_readback(&rb, readback_buffer, buffer_size, resource_desc.Width, resource_desc.Height, 1, row_pitch);
+    //check_readback_data_uint(&rb, NULL, 0xff00ff00, 0);
+    //release_resource_readback(&rb);
+    //value = ID3D12Fence_GetCompletedValue(fence2);
+    //ok(value == 0, "Got unexpected value %"PRIu64".\n", value);
+
+    hr = ID3D12Fence_Signal(fence, 1000);
     ok(hr == S_OK, "Failed to signal fence, hr %#x.\n", hr);
     ret = wait_event(event, INFINITE);
     ok(ret == WAIT_OBJECT_0, "Got unexpected return value %#x.\n", ret);
@@ -170,6 +184,7 @@ void test_queue_wait(void)
     value = ID3D12Fence_GetCompletedValue(fence2);
     ok(value == 1, "Got unexpected value %"PRIu64".\n", value);
 
+#if 0
     /* Wait() before GPU signal */
     update_buffer_data(cb, 0, sizeof(green), &green);
     queue_wait(queue, fence, 3);
@@ -218,7 +233,6 @@ void test_queue_wait(void)
     check_readback_data_uint(&rb, NULL, 0xff00ff00, 0);
     release_resource_readback(&rb);
 
-skip_tests:
     /* Signal() and Wait() in the same command queue */
     update_buffer_data(cb, 0, sizeof(blue), &blue);
     queue_signal(queue, fence, 7);
@@ -231,6 +245,7 @@ skip_tests:
 
     value = ID3D12Fence_GetCompletedValue(fence);
     ok(value == 7, "Got unexpected value %"PRIu64".\n", value);
+#endif
 
     destroy_event(event);
     ID3D12Fence_Release(fence);
@@ -238,6 +253,7 @@ skip_tests:
     ID3D12Resource_Release(cb);
     ID3D12CommandQueue_Release(queue2);
     ID3D12Resource_Release(readback_buffer);
+    ID3D12GraphicsCommandList_Release(command_list2);
     destroy_test_context(&context);
 }
 

tests/meson.build

@@ -61,3 +61,11 @@ executable('descriptor-performance', 'descriptor_performance.c',
   c_args              : vkd3d_test_flags,
   override_options    : [ 'c_std='+vkd3d_c_std ],
   link_with           : [ d3d12_test_utils_lib ])
+
+executable('renderpass-performance', 'renderpass_performance.c',
+  dependencies        : vkd3d_test_deps,
+  include_directories : vkd3d_private_includes,
+  install             : false,
+  c_args              : vkd3d_test_flags,
+  override_options    : [ 'c_std='+vkd3d_c_std ],
+  link_with           : [ d3d12_test_utils_lib ])

tests/renderpass_performance.c

@@ -0,0 +1,231 @@
+/*
+ * Copyright 2021 Hans-Kristian Arntzen for Valve Corporation
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
+ */
+
+#define VKD3D_DBG_CHANNEL VKD3D_DBG_CHANNEL_API
+
+#define INITGUID
+#define VKD3D_TEST_DECLARE_MAIN
+#include "d3d12_crosstest.h"
+
+struct context
+{
+    ID3D12Device *device;
+    ID3D12CommandQueue *queue;
+    ID3D12CommandAllocator *allocator;
+    ID3D12GraphicsCommandList *list;
+    ID3D12Resource *resources[64];
+    ID3D12DescriptorHeap *heap;
+};
+
+static void setup(int argc, char **argv)
+{
+    pfn_D3D12CreateDevice = get_d3d12_pfn(D3D12CreateDevice);
+    pfn_D3D12GetDebugInterface = get_d3d12_pfn(D3D12GetDebugInterface);
+
+    parse_args(argc, argv);
+    enable_d3d12_debug_layer(argc, argv);
+    init_adapter_info();
+
+    pfn_D3D12CreateVersionedRootSignatureDeserializer = get_d3d12_pfn(D3D12CreateVersionedRootSignatureDeserializer);
+    pfn_D3D12SerializeVersionedRootSignature = get_d3d12_pfn(D3D12SerializeVersionedRootSignature);
+}
+
+static double get_time(void)
+{
+#ifdef _WIN32
+    LARGE_INTEGER lc, lf;
+    QueryPerformanceCounter(&lc);
+    QueryPerformanceFrequency(&lf);
+    return (double)lc.QuadPart / (double)lf.QuadPart;
+#else
+    struct timespec ts;
+    clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
+    return ts.tv_sec + 1e-9 * ts.tv_nsec;
+#endif
+}
+
+static bool init_context(struct context *ctx)
+{
+    unsigned int i;
+    memset(ctx, 0, sizeof(*ctx));
+
+    if (!(ctx->device = create_device()))
+        return false;
+
+    ctx->heap = create_cpu_descriptor_heap(ctx->device, D3D12_DESCRIPTOR_HEAP_TYPE_RTV, 64);
+    for (i = 0; i < ARRAY_SIZE(ctx->resources); i++)
+    {
+        ctx->resources[i] = create_default_texture2d(ctx->device, 512, 512, 1, 1,
+                DXGI_FORMAT_R8G8B8A8_UNORM, D3D12_RESOURCE_FLAG_ALLOW_RENDER_TARGET,
+                D3D12_RESOURCE_STATE_RENDER_TARGET);
+    }
+    ID3D12Device_CreateCommandAllocator(ctx->device, D3D12_COMMAND_LIST_TYPE_DIRECT,
+            &IID_ID3D12CommandAllocator, (void**)&ctx->allocator);
+    ctx->queue = create_command_queue(ctx->device, D3D12_COMMAND_LIST_TYPE_DIRECT, 0);
+    ID3D12Device_CreateCommandList(ctx->device, 0, D3D12_COMMAND_LIST_TYPE_DIRECT, ctx->allocator,
+            NULL, &IID_ID3D12GraphicsCommandList, (void**)&ctx->list);
+    ID3D12GraphicsCommandList_Close(ctx->list);
+
+    for (i = 0; i < 64; i++)
+    {
+        D3D12_CPU_DESCRIPTOR_HANDLE h;
+        h = ID3D12DescriptorHeap_GetCPUDescriptorHandleForHeapStart(ctx->heap);
+        h.ptr += i * ID3D12Device_GetDescriptorHandleIncrementSize(ctx->device, D3D12_DESCRIPTOR_HEAP_TYPE_RTV);
+        ID3D12Device_CreateRenderTargetView(ctx->device, ctx->resources[i], NULL, h);
+    }
+
+    return true;
+}
+
+static void destroy_context(struct context *ctx)
+{
+    unsigned int i;
+    ID3D12DescriptorHeap_Release(ctx->heap);
+    for (i = 0; i < ARRAY_SIZE(ctx->resources); i++)
+        ID3D12Resource_Release(ctx->resources[i]);
+    ID3D12GraphicsCommandList_Release(ctx->list);
+    ID3D12CommandAllocator_Release(ctx->allocator);
+    ID3D12CommandQueue_Release(ctx->queue);
+    ID3D12Device_Release(ctx->device);
+}
+
+enum clear_mode
+{
+    CLEAR_MODE_NULL,
+    CLEAR_MODE_FULL_RECT,
+    CLEAR_MODE_PARTIAL
+};
+
+static double do_benchmark_run(struct context *ctx, unsigned int clear_iterations,
+        enum clear_mode mode, bool pre_discard, unsigned int num_resources)
+{
+    const D3D12_RECT partial_rect = { 257, 0, 509, 259 };
+    const D3D12_RECT full_rect = { 0, 0, 512, 512 };
+    D3D12_CPU_DESCRIPTOR_HANDLE base_h, h;
+    D3D12_DISCARD_REGION discard_region;
+    double start_time, end_time;
+    unsigned int resource_index;
+    UINT descriptor_size;
+    FLOAT clear_color[4];
+    ID3D12Fence *fence;
+    unsigned int i;
+
+    base_h = ID3D12DescriptorHeap_GetCPUDescriptorHandleForHeapStart(ctx->heap);
+    ID3D12GraphicsCommandList_Reset(ctx->list, ctx->allocator, NULL);
+    descriptor_size = ID3D12Device_GetDescriptorHandleIncrementSize(ctx->device, D3D12_DESCRIPTOR_HEAP_TYPE_RTV);
+
+    discard_region.FirstSubresource = 0;
+    discard_region.NumSubresources = 1;
+    discard_region.NumRects = 0;
+    discard_region.pRects = NULL;
+
+    /* Spam clear on different RTVs and see what happens to perf. */
+    for (i = 0; i < clear_iterations; i++)
+    {
+        resource_index = i & (num_resources - 1);
+        h = base_h;
+        h.ptr += resource_index * descriptor_size;
+        clear_color[0] = (float)(i & 255) / 255.0f;
+        clear_color[1] = (float)((i + 1) & 255) / 255.0f;
+        clear_color[2] = (float)((i + 2) & 255) / 255.0f;
+        clear_color[3] = (float)((i + 3) & 255) / 255.0f;
+
+        if (pre_discard)
+            ID3D12GraphicsCommandList_DiscardResource(ctx->list, ctx->resources[resource_index], &discard_region);
+
+        switch (mode)
+        {
+            default:
+                ID3D12GraphicsCommandList_ClearRenderTargetView(ctx->list, h, clear_color, 0, NULL);
+                break;
+
+            case CLEAR_MODE_FULL_RECT:
+                ID3D12GraphicsCommandList_ClearRenderTargetView(ctx->list, h, clear_color, 1, &full_rect);
+                break;
+
+            case CLEAR_MODE_PARTIAL:
+                ID3D12GraphicsCommandList_ClearRenderTargetView(ctx->list, h, clear_color, 1, &partial_rect);
+                break;
+        }
+    }
+
+    ID3D12GraphicsCommandList_Close(ctx->list);
+    ID3D12Device_CreateFence(ctx->device, 0, D3D12_FENCE_FLAG_NONE, &IID_ID3D12Fence, (void**)&fence);
+    start_time = get_time();
+    ID3D12CommandQueue_ExecuteCommandLists(ctx->queue, 1, (ID3D12CommandList * const *)&ctx->list);
+    ID3D12CommandQueue_Signal(ctx->queue, fence, 1);
+    ID3D12Fence_SetEventOnCompletion(fence, 1, NULL);
+    end_time = get_time();
+    ID3D12Fence_Release(fence);
+    ID3D12CommandAllocator_Reset(ctx->allocator);
+    return end_time - start_time;
+}
+
+START_TEST(renderpass_performance)
+{
+    struct test
+    {
+        unsigned int iteration_count;
+        const char *desc;
+        enum clear_mode mode;
+        bool pre_discard;
+        unsigned int num_resources;
+    };
+    static const struct test tests[] =
+    {
+        { 8 * 1024, "8k clear, NULL rect", CLEAR_MODE_NULL, false, 64 },
+        { 32 * 1024, "32k clear, NULL rect", CLEAR_MODE_NULL, false, 64 },
+        { 128 * 1024, "128k clear, NULL rect", CLEAR_MODE_NULL, false, 64 },
+        { 8 * 1024, "8k clear, full rect", CLEAR_MODE_FULL_RECT, false, 64 },
+        { 32 * 1024, "32k clear, full rect", CLEAR_MODE_FULL_RECT, false, 64 },
+        { 128 * 1024, "128k clear, full rect", CLEAR_MODE_FULL_RECT, false, 64 },
+        { 8 * 1024, "8k clear, partial rects", CLEAR_MODE_PARTIAL, false, 64 },
+        { 32 * 1024, "32k clear, partial rects", CLEAR_MODE_PARTIAL, false, 64 },
+        { 128 * 1024, "128k clear, partial rects", CLEAR_MODE_PARTIAL, false, 64 },
+        { 128 * 1024, "128k clear, NULL rect, discard", CLEAR_MODE_NULL, true, 64 },
+        { 128 * 1024, "128k clear, full rect, discard", CLEAR_MODE_FULL_RECT, true, 64 },
+        { 128 * 1024, "128k clear, partial rect, discard", CLEAR_MODE_PARTIAL, true, 64 },
+        { 64 * 1024, "64k clear, batch size 1", CLEAR_MODE_NULL, false, 1 },
+        { 64 * 1024, "64k clear, batch size 2", CLEAR_MODE_NULL, false, 2 },
+        { 64 * 1024, "64k clear, batch size 4", CLEAR_MODE_NULL, false, 4 },
+        { 64 * 1024, "64k clear, batch size 8", CLEAR_MODE_NULL, false, 8 },
+        { 64 * 1024, "64k clear, batch size 16", CLEAR_MODE_NULL, false, 16 },
+        { 64 * 1024, "64k clear, batch size 32", CLEAR_MODE_NULL, false, 32 },
+        { 128 * 1024, "128k clear, batch size 1", CLEAR_MODE_NULL, false, 1 },
+        { 128 * 1024, "128k clear, batch size 2", CLEAR_MODE_NULL, false, 2 },
+        { 128 * 1024, "128k clear, batch size 4", CLEAR_MODE_NULL, false, 4 },
+        { 128 * 1024, "128k clear, batch size 8", CLEAR_MODE_NULL, false, 8 },
+        { 128 * 1024, "128k clear, batch size 16", CLEAR_MODE_NULL, false, 16 },
+        { 128 * 1024, "128k clear, batch size 32", CLEAR_MODE_NULL, false, 32 },
+    };
+    struct context ctx;
+    unsigned int i;
+    double t;
+
+    setup(argc, argv);
+    if (!init_context(&ctx))
+        return;
+
+    for (i = 0; i < ARRAY_SIZE(tests); i++)
+    {
+        t = do_benchmark_run(&ctx, tests[i].iteration_count, tests[i].mode, tests[i].pre_discard, tests[i].num_resources);
+        printf("[%40s] => %8.3f ms.\n", tests[i].desc, 1e3 * t);
+    }
+
+    destroy_context(&ctx);
+}