vk_renderpass_cache: Pack renderpass cache key to 12 bytes

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -288,7 +288,7 @@ vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpa
     depth_stencil_ci.maxDepthBounds = 0.0f;
 
     std::array<VkPipelineColorBlendAttachmentState, Maxwell::NumRenderTargets> cb_attachments;
-    const std::size_t num_attachments = renderpass_params.color_attachments.size();
+    const auto num_attachments = static_cast<std::size_t>(renderpass_params.num_color_attachments);
     for (std::size_t index = 0; index < num_attachments; ++index) {
         static constexpr std::array COMPONENT_TABLE = {
             VK_COLOR_COMPONENT_R_BIT, VK_COLOR_COMPONENT_G_BIT, VK_COLOR_COMPONENT_B_BIT,

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -1245,28 +1245,15 @@ std::size_t RasterizerVulkan::CalculateConstBufferSize(
 }
 
 RenderPassParams RasterizerVulkan::GetRenderPassParams(Texceptions texceptions) const {
-    using namespace VideoCore::Surface;
-
     const auto& regs = system.GPU().Maxwell3D().regs;
-    RenderPassParams renderpass_params;
-
-    for (std::size_t rt = 0; rt < static_cast<std::size_t>(regs.rt_control.count); ++rt) {
-        const auto& rendertarget = regs.rt[rt];
-        if (rendertarget.Address() == 0 || rendertarget.format == Tegra::RenderTargetFormat::NONE) {
-            continue;
-        }
-        renderpass_params.color_attachments.push_back(RenderPassParams::ColorAttachment{
-            static_cast<u32>(rt), PixelFormatFromRenderTargetFormat(rendertarget.format),
-            texceptions[rt]});
-    }
-
-    renderpass_params.has_zeta = regs.zeta_enable;
-    if (renderpass_params.has_zeta) {
-        renderpass_params.zeta_pixel_format = PixelFormatFromDepthFormat(regs.zeta.format);
-        renderpass_params.zeta_texception = texceptions[ZETA_TEXCEPTION_INDEX];
-    }
-
-    return renderpass_params;
+    RenderPassParams params;
+    params.num_color_attachments = static_cast<u8>(regs.rt_control.count);
+    std::transform(regs.rt.begin(), regs.rt.end(), params.color_formats.begin(),
+                   [](const auto& rt) { return static_cast<u8>(rt.format); });
+    params.texceptions = static_cast<u8>(texceptions.to_ullong());
+    params.zeta_format = regs.zeta_enable ? static_cast<u8>(regs.zeta.format) : 0;
+    params.zeta_texception = texceptions[ZETA_TEXCEPTION_INDEX];
+    return params;
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_renderpass_cache.cpp

@@ -2,9 +2,11 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <cstring>
 #include <memory>
 #include <vector>
 
+#include "common/cityhash.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_vulkan/maxwell_to_vk.h"
 #include "video_core/renderer_vulkan/vk_device.h"
@@ -13,6 +15,15 @@
 
 namespace Vulkan {
 
+std::size_t RenderPassParams::Hash() const noexcept {
+    const u64 hash = Common::CityHash64(reinterpret_cast<const char*>(this), sizeof *this);
+    return static_cast<std::size_t>(hash);
+}
+
+bool RenderPassParams::operator==(const RenderPassParams& rhs) const noexcept {
+    return std::memcmp(&rhs, this, sizeof *this) == 0;
+}
+
 VKRenderPassCache::VKRenderPassCache(const VKDevice& device) : device{device} {}
 
 VKRenderPassCache::~VKRenderPassCache() = default;
@@ -27,18 +38,20 @@ VkRenderPass VKRenderPassCache::GetRenderPass(const RenderPassParams& params) {
 }
 
 vk::RenderPass VKRenderPassCache::CreateRenderPass(const RenderPassParams& params) const {
+    using namespace VideoCore::Surface;
     std::vector<VkAttachmentDescription> descriptors;
     std::vector<VkAttachmentReference> color_references;
 
-    for (std::size_t rt = 0; rt < params.color_attachments.size(); ++rt) {
-        const auto attachment = params.color_attachments[rt];
-        const auto format =
-            MaxwellToVK::SurfaceFormat(device, FormatType::Optimal, attachment.pixel_format);
+    const std::size_t num_attachments = static_cast<std::size_t>(params.num_color_attachments);
+    for (std::size_t rt = 0; rt < num_attachments; ++rt) {
+        const auto guest_format = static_cast<Tegra::RenderTargetFormat>(params.color_formats[rt]);
+        const PixelFormat pixel_format = PixelFormatFromRenderTargetFormat(guest_format);
+        const auto format = MaxwellToVK::SurfaceFormat(device, FormatType::Optimal, pixel_format);
         ASSERT_MSG(format.attachable, "Trying to attach a non-attachable format with format={}",
-                   static_cast<u32>(attachment.pixel_format));
+                   static_cast<int>(pixel_format));
 
-        // TODO(Rodrigo): Add eMayAlias when it's needed.
-        const auto color_layout = attachment.is_texception
+        // TODO(Rodrigo): Add MAY_ALIAS_BIT when it's needed.
+        const VkImageLayout color_layout = ((params.texceptions >> rt) & 1) != 0
                                                ? VK_IMAGE_LAYOUT_GENERAL
                                                : VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
         VkAttachmentDescription& descriptor = descriptors.emplace_back();
@@ -58,13 +71,15 @@ vk::RenderPass VKRenderPassCache::CreateRenderPass(const RenderPassParams& param
     }
 
     VkAttachmentReference zeta_attachment_ref;
-    if (params.has_zeta) {
-        const auto format =
-            MaxwellToVK::SurfaceFormat(device, FormatType::Optimal, params.zeta_pixel_format);
+    const bool has_zeta = params.zeta_format != 0;
+    if (has_zeta) {
+        const auto guest_format = static_cast<Tegra::DepthFormat>(params.zeta_format);
+        const PixelFormat pixel_format = PixelFormatFromDepthFormat(guest_format);
+        const auto format = MaxwellToVK::SurfaceFormat(device, FormatType::Optimal, pixel_format);
         ASSERT_MSG(format.attachable, "Trying to attach a non-attachable format with format={}",
-                   static_cast<u32>(params.zeta_pixel_format));
+                   static_cast<int>(pixel_format));
 
-        const auto zeta_layout = params.zeta_texception
+        const VkImageLayout zeta_layout = params.zeta_texception != 0
                                               ? VK_IMAGE_LAYOUT_GENERAL
                                               : VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
         VkAttachmentDescription& descriptor = descriptors.emplace_back();
@@ -78,7 +93,7 @@ vk::RenderPass VKRenderPassCache::CreateRenderPass(const RenderPassParams& param
         descriptor.initialLayout = zeta_layout;
         descriptor.finalLayout = zeta_layout;
 
-        zeta_attachment_ref.attachment = static_cast<u32>(params.color_attachments.size());
+        zeta_attachment_ref.attachment = static_cast<u32>(num_attachments);
         zeta_attachment_ref.layout = zeta_layout;
     }
 
@@ -90,7 +105,7 @@ vk::RenderPass VKRenderPassCache::CreateRenderPass(const RenderPassParams& param
     subpass_description.colorAttachmentCount = static_cast<u32>(color_references.size());
     subpass_description.pColorAttachments = color_references.data();
     subpass_description.pResolveAttachments = nullptr;
-    subpass_description.pDepthStencilAttachment = params.has_zeta ? &zeta_attachment_ref : nullptr;
+    subpass_description.pDepthStencilAttachment = has_zeta ? &zeta_attachment_ref : nullptr;
     subpass_description.preserveAttachmentCount = 0;
     subpass_description.pPreserveAttachments = nullptr;
 
@@ -101,7 +116,7 @@ vk::RenderPass VKRenderPassCache::CreateRenderPass(const RenderPassParams& param
         stage |= VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
     }
 
-    if (params.has_zeta) {
+    if (has_zeta) {
         access |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
                   VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
         stage |= VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;

src/video_core/renderer_vulkan/vk_renderpass_cache.h

@@ -4,8 +4,7 @@
 
 #pragma once
 
-#include <memory>
-#include <tuple>
+#include <type_traits>
 #include <unordered_map>
 
 #include <boost/container/static_vector.hpp>
@@ -19,51 +18,25 @@ namespace Vulkan {
 
 class VKDevice;
 
-// TODO(Rodrigo): Optimize this structure for faster hashing
-
 struct RenderPassParams {
-    struct ColorAttachment {
-        u32 index = 0;
-        VideoCore::Surface::PixelFormat pixel_format = VideoCore::Surface::PixelFormat::Invalid;
-        bool is_texception = false;
-
-        std::size_t Hash() const noexcept {
-            return static_cast<std::size_t>(pixel_format) |
-                   static_cast<std::size_t>(is_texception) << 6 |
-                   static_cast<std::size_t>(index) << 7;
-        }
+    u8 num_color_attachments;
+    std::array<u8, Tegra::Engines::Maxwell3D::Regs::NumRenderTargets> color_formats;
+    u8 texceptions;
 
-        bool operator==(const ColorAttachment& rhs) const noexcept {
-            return std::tie(index, pixel_format, is_texception) ==
-                   std::tie(rhs.index, rhs.pixel_format, rhs.is_texception);
-        }
-    };
-
-    boost::container::static_vector<ColorAttachment,
-                                    Tegra::Engines::Maxwell3D::Regs::NumRenderTargets>
-        color_attachments{};
-    // TODO(Rodrigo): Unify has_zeta into zeta_pixel_format and zeta_component_type.
-    VideoCore::Surface::PixelFormat zeta_pixel_format = VideoCore::Surface::PixelFormat::Invalid;
-    bool has_zeta = false;
-    bool zeta_texception = false;
-
-    std::size_t Hash() const noexcept {
-        std::size_t hash = 0;
-        for (const auto& rt : color_attachments) {
-            boost::hash_combine(hash, rt.Hash());
-        }
-        boost::hash_combine(hash, zeta_pixel_format);
-        boost::hash_combine(hash, has_zeta);
-        boost::hash_combine(hash, zeta_texception);
-        return hash;
-    }
+    u8 zeta_format;
+    u8 zeta_texception;
+
+    std::size_t Hash() const noexcept;
+
+    bool operator==(const RenderPassParams& rhs) const noexcept;
 
-    bool operator==(const RenderPassParams& rhs) const {
-        return std::tie(color_attachments, zeta_pixel_format, has_zeta, zeta_texception) ==
-               std::tie(rhs.color_attachments, rhs.zeta_pixel_format, rhs.has_zeta,
-                        rhs.zeta_texception);
+    bool operator!=(const RenderPassParams& rhs) const noexcept {
+        return !operator==(rhs);
     }
 };
+static_assert(std::has_unique_object_representations_v<RenderPassParams>);
+static_assert(std::is_trivially_copyable_v<RenderPassParams>);
+static_assert(std::is_trivially_constructible_v<RenderPassParams>);
 
 } // namespace Vulkan