buffer_cache: Skip cache on small uploads on Vulkan

Ports from OpenGL the optimization to skip small 3D uniform buffer
uploads. This will take advantage of the previously introduced stream
buffer.

Fixes instances where the staging buffer offset was being ignored.

src/video_core/buffer_cache/buffer_cache.h

@@ -660,24 +660,25 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     const VAddr cpu_addr = binding.cpu_addr;
     const u32 size = binding.size;
     Buffer& buffer = slot_buffers[binding.buffer_id];
-    if constexpr (IS_OPENGL) {
+    if (size <= SKIP_CACHE_SIZE && !buffer.IsRegionGpuModified(cpu_addr, size)) {
-        if (size <= SKIP_CACHE_SIZE && !buffer.IsRegionGpuModified(cpu_addr, size)) {
+        if constexpr (IS_OPENGL) {
             if (runtime.HasFastBufferSubData()) {
                 // Fast path for Nvidia
                 if (!HasFastUniformBufferBound(stage, binding_index)) {
                     // We only have to bind when the currently bound buffer is not the fast version
-                    fast_bound_uniform_buffers[stage] |= 1U << binding_index;
                     runtime.BindFastUniformBuffer(stage, binding_index, size);
                 }
                 const auto span = ImmediateBufferWithData(cpu_addr, size);
                 runtime.PushFastUniformBuffer(stage, binding_index, span);
-            } else {
+                return;
-                // Stream buffer path to avoid stalling on non-Nvidia drivers
-                const auto span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
-                cpu_memory.ReadBlockUnsafe(cpu_addr, span.data(), size);
             }
-            return;
         }
+        fast_bound_uniform_buffers[stage] |= 1U << binding_index;
+
+        // Stream buffer path to avoid stalling on non-Nvidia drivers or Vulkan
+        const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
+        cpu_memory.ReadBlockUnsafe(cpu_addr, span.data(), size);
+        return;
     }
     // Classic cached path
     SynchronizeBuffer(buffer, cpu_addr, size);

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -69,6 +69,13 @@ public:
 
     void BindTransformFeedbackBuffer(u32 index, VkBuffer buffer, u32 offset, u32 size);
 
+    std::span<u8> BindMappedUniformBuffer([[maybe_unused]] size_t stage,
+                                          [[maybe_unused]] u32 binding_index, u32 size) {
+        const StagingBufferRef ref = staging_pool.Request(size, MemoryUsage::Upload);
+        BindBuffer(ref.buffer, static_cast<u32>(ref.offset), size);
+        return ref.mapped_span;
+    }
+
     void BindUniformBuffer(VkBuffer buffer, u32 offset, u32 size) {
         BindBuffer(buffer, offset, size);
     }

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -86,7 +86,8 @@ StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& mem
         .pNext = nullptr,
         .flags = 0,
         .size = STREAM_BUFFER_SIZE,
-        .usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT,
+        .usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
+                 VK_BUFFER_USAGE_INDEX_BUFFER_BIT,
         .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
         .queueFamilyIndexCount = 0,
         .pQueueFamilyIndices = nullptr,