turnip: Add an override to uncached memory type

When uncached memory type is used under emulation then most games have a
significant performance penalty due to accessing the buffer atomically.

Instead when this option is set, it will override uncached buffer
allocations to instead be cached+coherent if the host supports it. This
allows the atomic accesses to still be done but not have abysmal
performance.

src/freedreno/vulkan/tu_cmd_buffer.cc

@@ -43,7 +43,10 @@ tu_cmd_buffer_setup_status_tracking(struct tu_device *device)
       device, NULL, &status_bo, sizeof(enum tu_cmd_buffer_status), 0,
       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
          VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-         VK_MEMORY_PROPERTY_HOST_COHERENT_BIT,
+         VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
+         (device->physical_device->preferred_uncached_as_cached_index >= 0 ?
+          VK_MEMORY_PROPERTY_HOST_CACHED_BIT : 0)
+         ,
       TU_BO_ALLOC_INTERNAL_RESOURCE, NULL, "cmd_buffer_status");
    if (result != VK_SUCCESS)
       return NULL;

src/freedreno/vulkan/tu_device.cc

@@ -1687,6 +1687,7 @@ tu_physical_device_init(struct tu_physical_device *device,
    device->level1_dcache_size = util_cache_granularity();
    device->has_cached_non_coherent_memory =
       device->level1_dcache_size > 0 && !DETECT_ARCH_ARM;
+   device->preferred_uncached_as_cached_index = -1;
 
    device->memory.type_count = 1;
    device->memory.types[0] =
@@ -1700,6 +1701,11 @@ tu_physical_device_init(struct tu_physical_device *device,
          VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
          VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
          VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
+
+      if (instance->override_uncached_as_cache_coherent) {
+          /* Retain this memory type index to override later. */
+          device->preferred_uncached_as_cached_index = device->memory.type_count;
+      }
       device->memory.type_count++;
    }
 
@@ -1851,6 +1857,7 @@ static const driOptionDescription tu_dri_options[] = {
       DRI_CONF_TU_ENABLE_SOFTFLOAT32(false)
       DRI_CONF_TU_EMULATE_ALPHA_TO_COVERAGE(false)
       DRI_CONF_TU_AUTOTUNE_ALGORITHM()
+      DRI_CONF_TU_OVERRIDE_UNCACHED_AS_CACHE_COHERENT(false)
    DRI_CONF_SECTION_END
 };
 
@@ -1885,6 +1892,8 @@ tu_init_dri_options(struct tu_instance *instance)
          driQueryOptionb(&instance->dri_options, "tu_emulate_alpha_to_coverage");
    instance->autotune_algo =
          driQueryOptionstr(&instance->dri_options, "tu_autotune_algorithm");
+   instance->override_uncached_as_cache_coherent =
+         driQueryOptionb(&instance->dri_options, "tu_override_uncached_as_cache_coherent");
 }
 
 static uint32_t instance_count = 0;

src/freedreno/vulkan/tu_device.h

@@ -134,6 +134,10 @@ struct tu_physical_device
 
    bool has_cached_coherent_memory;
    bool has_cached_non_coherent_memory;
+   /* Index for device local, host-coherent, host-cached memory.
+    * Only set to positive index when overwriting device-local, uncached memory
+    */
+   int32_t preferred_uncached_as_cached_index;
    uintptr_t level1_dcache_size;
 
    struct fdl_ubwc_config ubwc_config;
@@ -240,6 +244,11 @@ struct tu_instance
 
    /* Configuration option to use a specific autotune algorithm by default. */
    const char *autotune_algo;
+
+   /* When enabled, replaces uncached+host_visible allocations
+    * with cached+coherent+host_visible when the hardware supports it.
+    */
+   bool override_uncached_as_cache_coherent;
 };
 VK_DEFINE_HANDLE_CASTS(tu_instance, vk.base, VkInstance,
                        VK_OBJECT_TYPE_INSTANCE)

src/freedreno/vulkan/tu_knl.cc

@@ -43,6 +43,26 @@ tu_bo_init_new_explicit_iova(struct tu_device *dev,
 
    size = align64(size, os_page_size);
 
+   const VkMemoryPropertyFlags replace_flags_mask =
+      VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+      VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+      VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
+      VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
+
+   const VkMemoryPropertyFlags replace_flags_match =
+      VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
+      VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+      VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
+
+   if (dev->physical_device->preferred_uncached_as_cached_index >= 0 &&
+      (mem_property & replace_flags_mask) == replace_flags_match) {
+      /* Override the memory type if the requested type was uncached, only replacing
+       * if the device supports cached-coherent memory type.
+       */
+      mem_property =
+          dev->physical_device->memory.types[dev->physical_device->preferred_uncached_as_cached_index];
+   }
+
    VkResult result =
       dev->instance->knl->bo_init(dev, base, out_bo, size, client_iova,
                                   mem_property, flags, lazy_vma, name);

src/util/driconf.h

@@ -668,6 +668,9 @@
    DRI_CONF_OPT_S_NODEF(tu_autotune_algorithm, \
                         "Set the preferred autotune algorithm")
 
+#define DRI_CONF_TU_OVERRIDE_UNCACHED_AS_CACHE_COHERENT(def) \
+   DRI_CONF_OPT_B(tu_override_uncached_as_cache_coherent, def, \
+                  "Replaces uncached-host allocations with cached-coherent-host when possible. Only useful under x86 emulation where memory accesses tend to be atomic")
 /**
  * \brief Honeykrisp specific configuration options
  */