revert 7b9798f842

revert Revert "[shader_recompiler/Maxwell] Unstub ISBERD completely (#160)"

This reverts commit 3f12ae1e6e.

Found that this was not the cause of the performance issue on this PR. Reverting so it can be tested once fix for the slowdown can be merged and this PR rebased.

src/core/arm/nce/patcher.h

@@ -16,6 +16,24 @@
 #include "core/hle/kernel/physical_memory.h"
 #include "lru_cache.h"
 #include <utility>
+using ModuleID = std::array<u8, 32>;  // NSO build ID
+struct PatchCacheKey {
+    ModuleID module_id;
+    uintptr_t offset;
+    bool operator==(const PatchCacheKey&) const = default;
+};
+
+template <>
+struct std::hash<PatchCacheKey> {
+    size_t operator()(const PatchCacheKey& key) const {
+        // Simple XOR hash of first few bytes
+        size_t hash = 0;
+        for (size_t i = 0; i < key.module_id.size(); ++i) {
+            hash ^= static_cast<size_t>(key.module_id[i]) << ((i % sizeof(size_t)) * 8);
+        }
+        return hash ^ std::hash<uintptr_t>{}(key.offset);
+    }
+};
 
 namespace Core::NCE {
 
@@ -31,13 +49,15 @@ using EntryTrampolines = std::unordered_map<ModuleTextAddress, PatchTextAddress>
 
 class Patcher {
 public:
+    void SetModuleID(const ModuleID& id) {
+        module_id = id;
+    }
     Patcher(const Patcher&) = delete;
     Patcher& operator=(const Patcher&) = delete;
     Patcher(Patcher&& other) noexcept;
     Patcher& operator=(Patcher&&) noexcept = delete;
     explicit Patcher();
     ~Patcher();
-
     bool PatchText(const Kernel::PhysicalMemory& program_image,
                    const Kernel::CodeSet::Segment& code);
     bool RelocateAndCopy(Common::ProcessAddress load_base, const Kernel::CodeSet::Segment& code,
@@ -50,7 +70,7 @@ public:
 
 private:
     using ModuleDestLabel = uintptr_t;
-
+    ModuleID module_id{};
     struct Trampoline {
         ptrdiff_t patch_offset;
         uintptr_t module_offset;
@@ -68,26 +88,25 @@ private:
 
 private:
     static constexpr size_t CACHE_SIZE = 16384;  // Cache size for patch entries
-    LRUCache<uintptr_t, PatchTextAddress> patch_cache{CACHE_SIZE, Settings::values.lru_cache_enabled.GetValue()};
+    LRUCache<PatchCacheKey, PatchTextAddress> patch_cache{CACHE_SIZE, Settings::values.lru_cache_enabled.GetValue()};
 
     void BranchToPatch(uintptr_t module_dest) {
         if (patch_cache.isEnabled()) {
-            LOG_DEBUG(Core_ARM, "LRU cache lookup for address {:#x}", module_dest);
+            PatchCacheKey key{module_id, module_dest};
+            LOG_DEBUG(Core_ARM, "LRU cache lookup for module={}, offset={:#x}", fmt::ptr(module_id.data()), module_dest);
             // Try to get existing patch entry from cache
-            if (auto* cached_patch = patch_cache.get(module_dest)) {
-                LOG_WARNING(Core_ARM, "LRU cache hit for address {:#x}", module_dest);
+            if (auto* cached_patch = patch_cache.get(key)) {
+                LOG_WARNING(Core_ARM, "LRU cache hit for module offset {:#x}", module_dest);
                 curr_patch->m_branch_to_patch_relocations.push_back({c.offset(), *cached_patch});
                 return;
             }
-            LOG_DEBUG(Core_ARM, "LRU cache miss for address {:#x}, creating new patch", module_dest);
-
-            // If not in cache, create new entry and cache it
+            LOG_DEBUG(Core_ARM, "LRU cache miss for module offset {:#x}, creating new patch", module_dest);
+            // Not in cache: create and store
             const auto patch_addr = c.offset();
             curr_patch->m_branch_to_patch_relocations.push_back({patch_addr, module_dest});
-            patch_cache.put(module_dest, patch_addr);
+            patch_cache.put(key, patch_addr);
         } else {
-            LOG_DEBUG(Core_ARM, "LRU cache disabled - creating direct patch for address {:#x}", module_dest);
-            // LRU disabled - use pre-LRU approach
+            LOG_DEBUG(Core_ARM, "LRU cache disabled - direct patch for offset {:#x}", module_dest);
             curr_patch->m_branch_to_patch_relocations.push_back({c.offset(), module_dest});
         }
     }

src/core/loader/nso.cpp

@@ -166,6 +166,8 @@ std::optional<VAddr> AppLoader_NSO::LoadModule(Kernel::KProcess& process, Core::
     const auto& code = codeset.CodeSegment();
     auto* patch = patches ? &patches->operator[](patch_index) : nullptr;
     if (patch && !load_into_process) {
+        //Set module ID using build_id from the NSO header
+        patch->SetModuleID(nso_header.build_id);
         // Patch SVCs and MRS calls in the guest code
         while (!patch->PatchText(program_image, code)) {
             patch = &patches->emplace_back();

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -507,58 +507,84 @@ TransformBufferCopies(std::span<const VideoCommon::BufferCopy> copies, size_t bu
         return value;
     }
 }
+struct RangedBarrierRange {
+    u32 min_mip = std::numeric_limits<u32>::max();
+    u32 max_mip = std::numeric_limits<u32>::min();
+    u32 min_layer = std::numeric_limits<u32>::max();
+    u32 max_layer = std::numeric_limits<u32>::min();
 
+    void AddLayers(const VkImageSubresourceLayers& layers) {
+        min_mip = std::min(min_mip, layers.mipLevel);
+        max_mip = std::max(max_mip, layers.mipLevel + 1);
+        min_layer = std::min(min_layer, layers.baseArrayLayer);
+        max_layer = std::max(max_layer, layers.baseArrayLayer + layers.layerCount);
+    }
+
+    VkImageSubresourceRange SubresourceRange(VkImageAspectFlags aspect_mask) const noexcept {
+        return VkImageSubresourceRange{
+            .aspectMask = aspect_mask,
+            .baseMipLevel = min_mip,
+            .levelCount = max_mip - min_mip,
+            .baseArrayLayer = min_layer,
+            .layerCount = max_layer - min_layer,
+        };
+    }
+};
 void CopyBufferToImage(vk::CommandBuffer cmdbuf, VkBuffer src_buffer, VkImage image,
                        VkImageAspectFlags aspect_mask, bool is_initialized,
                        std::span<const VkBufferImageCopy> copies) {
     static constexpr VkAccessFlags WRITE_ACCESS_FLAGS =
-        VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
-        VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
+                                           VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
+                                           VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
     static constexpr VkAccessFlags READ_ACCESS_FLAGS = VK_ACCESS_SHADER_READ_BIT |
                                                        VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
                                                        VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT;
+
+    //  Compute exact mip/layer range being written to
+    RangedBarrierRange range;
+    for (const auto& region : copies) {
+        range.AddLayers(region.imageSubresource);
+    }
+    const VkImageSubresourceRange subresource_range = range.SubresourceRange(aspect_mask);
+
     const VkImageMemoryBarrier read_barrier{
-        .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
-        .pNext = nullptr,
-        .srcAccessMask = WRITE_ACCESS_FLAGS,
-        .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-        .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
-        .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
-        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .image = image,
-        .subresourceRange{
-            .aspectMask = aspect_mask,
-            .baseMipLevel = 0,
-            .levelCount = VK_REMAINING_MIP_LEVELS,
-            .baseArrayLayer = 0,
-            .layerCount = VK_REMAINING_ARRAY_LAYERS,
-        },
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = WRITE_ACCESS_FLAGS,
+            .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+            .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
+            .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = image,
+            .subresourceRange = subresource_range,
     };
+
     const VkImageMemoryBarrier write_barrier{
-        .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
-        .pNext = nullptr,
-        .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-        .dstAccessMask = WRITE_ACCESS_FLAGS | READ_ACCESS_FLAGS,
-        .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
-        .newLayout = VK_IMAGE_LAYOUT_GENERAL,
-        .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-        .image = image,
-        .subresourceRange{
-            .aspectMask = aspect_mask,
-            .baseMipLevel = 0,
-            .levelCount = VK_REMAINING_MIP_LEVELS,
-            .baseArrayLayer = 0,
-            .layerCount = VK_REMAINING_ARRAY_LAYERS,
-        },
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+            .dstAccessMask = WRITE_ACCESS_FLAGS | READ_ACCESS_FLAGS,
+            .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = image,
+            .subresourceRange = subresource_range,
     };
-    cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
+
+    cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
+                           VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT |
+                           VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
                            read_barrier);
     cmdbuf.CopyBufferToImage(src_buffer, image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, copies);
     // TODO: Move this to another API
-    cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0,
-                           write_barrier);
+    cmdbuf.PipelineBarrier(
+            VK_PIPELINE_STAGE_TRANSFER_BIT,
+            VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
+            VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
+            VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+            0, nullptr, nullptr, write_barrier);
 }
 
 [[nodiscard]] VkImageBlit MakeImageBlit(const Region2D& dst_region, const Region2D& src_region,
@@ -651,29 +677,7 @@ void TryTransformSwizzleIfNeeded(PixelFormat format, std::array<SwizzleSource, 4
     }
 }
 
-struct RangedBarrierRange {
-    u32 min_mip = std::numeric_limits<u32>::max();
-    u32 max_mip = std::numeric_limits<u32>::min();
-    u32 min_layer = std::numeric_limits<u32>::max();
-    u32 max_layer = std::numeric_limits<u32>::min();
 
-    void AddLayers(const VkImageSubresourceLayers& layers) {
-        min_mip = std::min(min_mip, layers.mipLevel);
-        max_mip = std::max(max_mip, layers.mipLevel + 1);
-        min_layer = std::min(min_layer, layers.baseArrayLayer);
-        max_layer = std::max(max_layer, layers.baseArrayLayer + layers.layerCount);
-    }
-
-    VkImageSubresourceRange SubresourceRange(VkImageAspectFlags aspect_mask) const noexcept {
-        return VkImageSubresourceRange{
-            .aspectMask = aspect_mask,
-            .baseMipLevel = min_mip,
-            .levelCount = max_mip - min_mip,
-            .baseArrayLayer = min_layer,
-            .layerCount = max_layer - min_layer,
-        };
-    }
-};
 
 [[nodiscard]] VkFormat Format(Shader::ImageFormat format) {
     switch (format) {
@@ -1457,12 +1461,18 @@ void TextureCacheRuntime::CopyImage(Image& dst, Image& src,
                 .subresourceRange = dst_range.SubresourceRange(aspect_mask),
             },
         };
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-                               0, {}, {}, pre_barriers);
+        cmdbuf.PipelineBarrier(
+                VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
+                VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT,
+                VK_PIPELINE_STAGE_TRANSFER_BIT,
+                0, nullptr, nullptr, pre_barriers);
         cmdbuf.CopyImage(src_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, dst_image,
                          VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, vk_copies);
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
-                               0, {}, {}, post_barriers);
+        cmdbuf.PipelineBarrier(
+                VK_PIPELINE_STAGE_TRANSFER_BIT,
+                VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
+                VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT,
+                0, nullptr, nullptr, post_barriers);
     });
 }
 
@@ -1546,7 +1556,7 @@ void Image::UploadMemory(VkBuffer buffer, VkDeviceSize offset,
     }
 
     // Handle MSAA upload if necessary
-    /* WARNING, TODO: This code uses some hacks, besides being fundamentally ugly 
+    /* WARNING, TODO: This code uses some hacks, besides being fundamentally ugly
     since tropic didn't want to touch it for a long time, so it needs a rewrite from someone better than me at vulkan.*/
     if (info.num_samples > 1 && runtime->CanUploadMSAA()) {
         // Only use MSAA copy pass for color formats
@@ -2352,7 +2362,7 @@ void TextureCacheRuntime::TransitionImageLayout(Image& image) {
         };
         scheduler.RequestOutsideRenderPassOperationContext();
         scheduler.Record([barrier](vk::CommandBuffer cmdbuf) {
-            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
                                    VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, 0, barrier);
         });
     }