renderer_vulkan: Wait on present semaphore at queue submit

The present semaphore is being signalled by the call to acquire the
swapchain image. This semaphore is meant to be waited on when rendering
to the swapchain image. Currently it is waited on when presenting, but
moving its usage to be waited on in the command buffer submission allows
for proper usage of this semaphore.

Fixes the device lost when launching titles on the Intel Linux Mesa driver.

src/shader_recompiler/frontend/maxwell/structured_control_flow.cpp

@@ -20,6 +20,7 @@
 #include "shader_recompiler/frontend/maxwell/decode.h"
 #include "shader_recompiler/frontend/maxwell/structured_control_flow.h"
 #include "shader_recompiler/frontend/maxwell/translate/translate.h"
+#include "shader_recompiler/host_translate_info.h"
 #include "shader_recompiler/object_pool.h"
 
 namespace Shader::Maxwell {
@@ -652,7 +653,7 @@ class TranslatePass {
 public:
     TranslatePass(ObjectPool<IR::Inst>& inst_pool_, ObjectPool<IR::Block>& block_pool_,
                   ObjectPool<Statement>& stmt_pool_, Environment& env_, Statement& root_stmt,
-                  IR::AbstractSyntaxList& syntax_list_)
+                  IR::AbstractSyntaxList& syntax_list_, const HostTranslateInfo& host_info)
         : stmt_pool{stmt_pool_}, inst_pool{inst_pool_}, block_pool{block_pool_}, env{env_},
           syntax_list{syntax_list_} {
         Visit(root_stmt, nullptr, nullptr);
@@ -660,6 +661,9 @@ public:
         IR::Block& first_block{*syntax_list.front().data.block};
         IR::IREmitter ir(first_block, first_block.begin());
         ir.Prologue();
+        if (uses_demote_to_helper && host_info.needs_demote_reorder) {
+            DemoteCombinationPass();
+        }
     }
 
 private:
@@ -809,7 +813,14 @@ private:
             }
             case StatementType::Return: {
                 ensure_block();
-                IR::IREmitter{*current_block}.Epilogue();
+                IR::Block* return_block{block_pool.Create(inst_pool)};
+                IR::IREmitter{*return_block}.Epilogue();
+                current_block->AddBranch(return_block);
+
+                auto& merge{syntax_list.emplace_back()};
+                merge.type = IR::AbstractSyntaxNode::Type::Block;
+                merge.data.block = return_block;
+
                 current_block = nullptr;
                 syntax_list.emplace_back().type = IR::AbstractSyntaxNode::Type::Return;
                 break;
@@ -824,6 +835,7 @@ private:
                 auto& merge{syntax_list.emplace_back()};
                 merge.type = IR::AbstractSyntaxNode::Type::Block;
                 merge.data.block = demote_block;
+                uses_demote_to_helper = true;
                 break;
             }
             case StatementType::Unreachable: {
@@ -855,11 +867,117 @@ private:
         return block_pool.Create(inst_pool);
     }
 
+    void DemoteCombinationPass() {
+        using Type = IR::AbstractSyntaxNode::Type;
+        std::vector<IR::Block*> demote_blocks;
+        std::vector<IR::U1> demote_conds;
+        u32 num_epilogues{};
+        u32 branch_depth{};
+        for (const IR::AbstractSyntaxNode& node : syntax_list) {
+            if (node.type == Type::If) {
+                ++branch_depth;
+            }
+            if (node.type == Type::EndIf) {
+                --branch_depth;
+            }
+            if (node.type != Type::Block) {
+                continue;
+            }
+            if (branch_depth > 1) {
+                // Skip reordering nested demote branches.
+                continue;
+            }
+            for (const IR::Inst& inst : node.data.block->Instructions()) {
+                const IR::Opcode op{inst.GetOpcode()};
+                if (op == IR::Opcode::DemoteToHelperInvocation) {
+                    demote_blocks.push_back(node.data.block);
+                    break;
+                }
+                if (op == IR::Opcode::Epilogue) {
+                    ++num_epilogues;
+                }
+            }
+        }
+        if (demote_blocks.size() == 0) {
+            return;
+        }
+        if (num_epilogues > 1) {
+            LOG_DEBUG(Shader, "Combining demotes with more than one return is not implemented.");
+            return;
+        }
+        s64 last_iterator_offset{};
+        auto& asl{syntax_list};
+        for (const IR::Block* demote_block : demote_blocks) {
+            const auto start_it{asl.begin() + last_iterator_offset};
+            auto asl_it{std::find_if(start_it, asl.end(), [&](const IR::AbstractSyntaxNode& asn) {
+                return asn.type == Type::If && asn.data.if_node.body == demote_block;
+            })};
+            if (asl_it == asl.end()) {
+                // Demote without a conditional branch.
+                // No need to proceed since all fragment instances will be demoted regardless.
+                return;
+            }
+            const IR::Block* const end_if = asl_it->data.if_node.merge;
+            demote_conds.push_back(asl_it->data.if_node.cond);
+            last_iterator_offset = std::distance(asl.begin(), asl_it);
+
+            asl_it = asl.erase(asl_it);
+            asl_it = std::find_if(asl_it, asl.end(), [&](const IR::AbstractSyntaxNode& asn) {
+                return asn.type == Type::Block && asn.data.block == demote_block;
+            });
+
+            asl_it = asl.erase(asl_it);
+            asl_it = std::find_if(asl_it, asl.end(), [&](const IR::AbstractSyntaxNode& asn) {
+                return asn.type == Type::EndIf && asn.data.end_if.merge == end_if;
+            });
+            asl_it = asl.erase(asl_it);
+        }
+        const auto epilogue_func{[](const IR::AbstractSyntaxNode& asn) {
+            if (asn.type != Type::Block) {
+                return false;
+            }
+            for (const auto& inst : asn.data.block->Instructions()) {
+                if (inst.GetOpcode() == IR::Opcode::Epilogue) {
+                    return true;
+                }
+            }
+            return false;
+        }};
+        const auto reverse_it{std::find_if(asl.rbegin(), asl.rend(), epilogue_func)};
+        const auto return_block_it{(reverse_it + 1).base()};
+
+        IR::IREmitter ir{*(return_block_it - 1)->data.block};
+        IR::U1 cond(IR::Value(false));
+        for (const auto& demote_cond : demote_conds) {
+            cond = ir.LogicalOr(cond, demote_cond);
+        }
+        cond.Inst()->DestructiveAddUsage(1);
+
+        IR::AbstractSyntaxNode demote_if_node{};
+        demote_if_node.type = Type::If;
+        demote_if_node.data.if_node.cond = cond;
+        demote_if_node.data.if_node.body = demote_blocks[0];
+        demote_if_node.data.if_node.merge = return_block_it->data.block;
+
+        IR::AbstractSyntaxNode demote_node{};
+        demote_node.type = Type::Block;
+        demote_node.data.block = demote_blocks[0];
+
+        IR::AbstractSyntaxNode demote_endif_node{};
+        demote_endif_node.type = Type::EndIf;
+        demote_endif_node.data.end_if.merge = return_block_it->data.block;
+
+        asl.insert(return_block_it, demote_endif_node);
+        asl.insert(return_block_it, demote_node);
+        asl.insert(return_block_it, demote_if_node);
+    }
+
     ObjectPool<Statement>& stmt_pool;
     ObjectPool<IR::Inst>& inst_pool;
     ObjectPool<IR::Block>& block_pool;
     Environment& env;
     IR::AbstractSyntaxList& syntax_list;
+    bool uses_demote_to_helper{};
 
 // TODO: C++20 Remove this when all compilers support constexpr std::vector
 #if __cpp_lib_constexpr_vector >= 201907
@@ -871,12 +989,13 @@ private:
 } // Anonymous namespace
 
 IR::AbstractSyntaxList BuildASL(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Block>& block_pool,
-                                Environment& env, Flow::CFG& cfg) {
+                                Environment& env, Flow::CFG& cfg,
+                                const HostTranslateInfo& host_info) {
     ObjectPool<Statement> stmt_pool{64};
     GotoPass goto_pass{cfg, stmt_pool};
     Statement& root{goto_pass.RootStatement()};
     IR::AbstractSyntaxList syntax_list;
-    TranslatePass{inst_pool, block_pool, stmt_pool, env, root, syntax_list};
+    TranslatePass{inst_pool, block_pool, stmt_pool, env, root, syntax_list, host_info};
     return syntax_list;
 }
 

src/shader_recompiler/frontend/maxwell/structured_control_flow.h

@@ -11,10 +11,13 @@
 #include "shader_recompiler/frontend/maxwell/control_flow.h"
 #include "shader_recompiler/object_pool.h"
 
-namespace Shader::Maxwell {
+namespace Shader {
+struct HostTranslateInfo;
+namespace Maxwell {
 
 [[nodiscard]] IR::AbstractSyntaxList BuildASL(ObjectPool<IR::Inst>& inst_pool,
                                               ObjectPool<IR::Block>& block_pool, Environment& env,
-                                              Flow::CFG& cfg);
+                                              Flow::CFG& cfg, const HostTranslateInfo& host_info);
 
-} // namespace Shader::Maxwell
+} // namespace Maxwell
+} // namespace Shader

src/shader_recompiler/frontend/maxwell/translate_program.cpp

@@ -130,7 +130,7 @@ void AddNVNStorageBuffers(IR::Program& program) {
 IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Block>& block_pool,
                              Environment& env, Flow::CFG& cfg, const HostTranslateInfo& host_info) {
     IR::Program program;
-    program.syntax_list = BuildASL(inst_pool, block_pool, env, cfg);
+    program.syntax_list = BuildASL(inst_pool, block_pool, env, cfg, host_info);
     program.blocks = GenerateBlocks(program.syntax_list);
     program.post_order_blocks = PostOrder(program.syntax_list.front());
     program.stage = env.ShaderStage();

src/shader_recompiler/host_translate_info.h

@@ -11,8 +11,9 @@ namespace Shader {
 
 /// Misc information about the host
 struct HostTranslateInfo {
-    bool support_float16{}; ///< True when the device supports 16-bit floats
-    bool support_int64{};   ///< True when the device supports 64-bit integers
+    bool support_float16{};      ///< True when the device supports 16-bit floats
+    bool support_int64{};        ///< True when the device supports 64-bit integers
+    bool needs_demote_reorder{}; ///< True when the device needs DemoteToHelperInvocation reordered
 };
 
 } // namespace Shader

src/video_core/renderer_opengl/gl_device.h

@@ -156,6 +156,10 @@ public:
         return shader_backend;
     }
 
+    bool IsAmd() const {
+        return vendor_name == "ATI Technologies Inc.";
+    }
+
 private:
     static bool TestVariableAoffi();
     static bool TestPreciseBug();

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -219,6 +219,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo
       host_info{
           .support_float16 = false,
           .support_int64 = device.HasShaderInt64(),
+          .needs_demote_reorder = device.IsAmd(),
       } {
     if (use_asynchronous_shaders) {
         workers = CreateWorkers();

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -164,7 +164,8 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
         blit_screen.Recreate();
     }
     const VkSemaphore render_semaphore = blit_screen.DrawToSwapchain(*framebuffer, use_accelerated);
-    scheduler.Flush(render_semaphore);
+    const VkSemaphore present_semaphore = swapchain.CurrentPresentSemaphore();
+    scheduler.Flush(render_semaphore, present_semaphore);
     scheduler.WaitWorker();
     swapchain.Present(render_semaphore);
 

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -325,6 +325,8 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw
     host_info = Shader::HostTranslateInfo{
         .support_float16 = device.IsFloat16Supported(),
         .support_int64 = device.IsShaderInt64Supported(),
+        .needs_demote_reorder = driver_id == VK_DRIVER_ID_AMD_PROPRIETARY_KHR ||
+                                driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE_KHR,
     };
 }
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -55,14 +55,14 @@ VKScheduler::~VKScheduler() {
     worker_thread.join();
 }
 
-void VKScheduler::Flush(VkSemaphore semaphore) {
-    SubmitExecution(semaphore);
+void VKScheduler::Flush(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
+    SubmitExecution(signal_semaphore, wait_semaphore);
     AllocateNewContext();
 }
 
-void VKScheduler::Finish(VkSemaphore semaphore) {
+void VKScheduler::Finish(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
     const u64 presubmit_tick = CurrentTick();
-    SubmitExecution(semaphore);
+    SubmitExecution(signal_semaphore, wait_semaphore);
     WaitWorker();
     Wait(presubmit_tick);
     AllocateNewContext();
@@ -171,37 +171,41 @@ void VKScheduler::AllocateWorkerCommandBuffer() {
     });
 }
 
-void VKScheduler::SubmitExecution(VkSemaphore semaphore) {
+void VKScheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
     EndPendingOperations();
     InvalidateState();
 
     const u64 signal_value = master_semaphore->NextTick();
-    Record([semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
+    Record([signal_semaphore, wait_semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
         cmdbuf.End();
-
-        const u32 num_signal_semaphores = semaphore ? 2U : 1U;
-
-        const u64 wait_value = signal_value - 1;
-        const VkPipelineStageFlags wait_stage_mask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
-
         const VkSemaphore timeline_semaphore = master_semaphore->Handle();
+
+        const u32 num_signal_semaphores = signal_semaphore ? 2U : 1U;
         const std::array signal_values{signal_value, u64(0)};
-        const std::array signal_semaphores{timeline_semaphore, semaphore};
+        const std::array signal_semaphores{timeline_semaphore, signal_semaphore};
+
+        const u32 num_wait_semaphores = wait_semaphore ? 2U : 1U;
+        const std::array wait_values{signal_value - 1, u64(1)};
+        const std::array wait_semaphores{timeline_semaphore, wait_semaphore};
+        static constexpr std::array<VkPipelineStageFlags, 2> wait_stage_masks{
+            VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+            VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+        };
 
         const VkTimelineSemaphoreSubmitInfoKHR timeline_si{
             .sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO_KHR,
             .pNext = nullptr,
-            .waitSemaphoreValueCount = 1,
-            .pWaitSemaphoreValues = &wait_value,
+            .waitSemaphoreValueCount = num_wait_semaphores,
+            .pWaitSemaphoreValues = wait_values.data(),
             .signalSemaphoreValueCount = num_signal_semaphores,
             .pSignalSemaphoreValues = signal_values.data(),
         };
         const VkSubmitInfo submit_info{
             .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
             .pNext = &timeline_si,
-            .waitSemaphoreCount = 1,
-            .pWaitSemaphores = &timeline_semaphore,
-            .pWaitDstStageMask = &wait_stage_mask,
+            .waitSemaphoreCount = num_wait_semaphores,
+            .pWaitSemaphores = wait_semaphores.data(),
+            .pWaitDstStageMask = wait_stage_masks.data(),
             .commandBufferCount = 1,
             .pCommandBuffers = cmdbuf.address(),
             .signalSemaphoreCount = num_signal_semaphores,

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -34,10 +34,10 @@ public:
     ~VKScheduler();
 
     /// Sends the current execution context to the GPU.
-    void Flush(VkSemaphore semaphore = nullptr);
+    void Flush(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
 
     /// Sends the current execution context to the GPU and waits for it to complete.
-    void Finish(VkSemaphore semaphore = nullptr);
+    void Finish(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
 
     /// Waits for the worker thread to finish executing everything. After this function returns it's
     /// safe to touch worker resources.
@@ -191,7 +191,7 @@ private:
 
     void AllocateWorkerCommandBuffer();
 
-    void SubmitExecution(VkSemaphore semaphore);
+    void SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore);
 
     void AllocateNewContext();
 

src/video_core/renderer_vulkan/vk_swapchain.cpp

@@ -107,14 +107,12 @@ void VKSwapchain::AcquireNextImage() {
 }
 
 void VKSwapchain::Present(VkSemaphore render_semaphore) {
-    const VkSemaphore present_semaphore{*present_semaphores[frame_index]};
-    const std::array<VkSemaphore, 2> semaphores{present_semaphore, render_semaphore};
     const auto present_queue{device.GetPresentQueue()};
     const VkPresentInfoKHR present_info{
         .sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
         .pNext = nullptr,
-        .waitSemaphoreCount = render_semaphore ? 2U : 1U,
-        .pWaitSemaphores = semaphores.data(),
+        .waitSemaphoreCount = render_semaphore ? 1U : 0U,
+        .pWaitSemaphores = &render_semaphore,
         .swapchainCount = 1,
         .pSwapchains = swapchain.address(),
         .pImageIndices = &image_index,

src/video_core/renderer_vulkan/vk_swapchain.h

@@ -72,6 +72,10 @@ public:
         return image_format;
     }
 
+    VkSemaphore CurrentPresentSemaphore() const {
+        return *present_semaphores[frame_index];
+    }
+
 private:
     void CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities, u32 width, u32 height,
                          bool srgb);