Port hakuX Phase 3 conservative draw reordering

hw/xbox/nv2a/pgraph/vk/renderer.c

@@ -192,6 +192,7 @@ static void pgraph_vk_flush(NV2AState *d)
 {
     PGRAPHState *pg = &d->pgraph;
 
+    pgraph_vk_flush_reorder_window(d);
     pgraph_vk_finish(pg, VK_FINISH_REASON_FLUSH);
     pgraph_vk_surface_flush(d);
     pgraph_vk_mark_textures_possibly_dirty(d, 0, memory_region_size(d->vram));

hw/xbox/nv2a/pgraph/vk/renderer.h

@@ -47,6 +47,8 @@
 #define MAX_FRAMEBUFFERS 256
 #define FB_CACHE_MAX 32
 #define OPT_BINDLESS_TEXTURES 1
+#define OPT_REORDER_SAFE_WINDOWS 1
+#define REORDER_WINDOW_MAX 64
 #define MAX_BINDLESS_TEXTURES 1024
 #define BINDLESS_STAGE_SLOT_BASE (MAX_BINDLESS_TEXTURES - NV2A_MAX_TEXTURES)
 
@@ -256,6 +258,64 @@ typedef struct TextureBinding {
 #endif
 } TextureBinding;
 
+#if OPT_REORDER_SAFE_WINDOWS
+typedef enum ReorderDrawMode {
+    RW_DRAW_INDEXED,
+    RW_DRAW_INDIRECT,
+    RW_DRAW_DIRECT,
+} ReorderDrawMode;
+
+typedef struct ReorderWindowEntry {
+    PipelineBinding *pipeline_binding;
+    VkDescriptorSet descriptor_set;
+
+    VkBuffer vertex_buffers[NV2A_VERTEXSHADER_ATTRIBUTES];
+    VkDeviceSize vertex_offsets[NV2A_VERTEXSHADER_ATTRIBUTES];
+    int num_vertex_bindings;
+
+    ReorderDrawMode draw_mode;
+    VkDeviceSize index_indirect_offset;
+    uint32_t draw_count;
+    uint32_t vertex_count;
+    int32_t first_vertex;
+
+    bool has_dynamic_line_width;
+    float line_width;
+
+    VkViewport viewport;
+    VkRect2D scissor;
+
+    float push_values[NV2A_VERTEXSHADER_ATTRIBUTES * 4];
+    int num_push_values;
+    bool use_push_constants;
+    bool use_bindless_textures;
+#if OPT_BINDLESS_TEXTURES
+    uint32_t tex_bindless_indices[NV2A_MAX_TEXTURES];
+#endif
+    VkPipelineLayout layout;
+
+    int sequence_number;
+    int group_order;
+
+    bool color_write;
+    bool depth_test;
+    bool stencil_test;
+} ReorderWindowEntry;
+
+#define REORDER_MAX_PIPELINES 32
+
+typedef struct ReorderWindow {
+    ReorderWindowEntry entries[REORDER_WINDOW_MAX];
+    int count;
+    bool active;
+
+    PipelineBinding *seen_pipelines[REORDER_MAX_PIPELINES];
+    int seen_pipeline_group[REORDER_MAX_PIPELINES];
+    int num_seen_pipelines;
+    int next_group;
+} ReorderWindow;
+#endif
+
 typedef struct QueryReport {
     QSIMPLEQ_ENTRY(QueryReport) entry;
     bool clear;
@@ -491,6 +551,10 @@ typedef struct PGRAPHVkState {
     bool shader_bindings_changed;
     bool use_push_constants_for_uniform_attrs;
 
+#if OPT_REORDER_SAFE_WINDOWS
+    ReorderWindow reorder_window;
+#endif
+
     Lru shader_module_cache;
     ShaderModuleCacheEntry *shader_module_cache_entries;
     size_t shader_module_cache_target;
@@ -685,6 +749,7 @@ void pgraph_vk_draw_begin(NV2AState *d);
 void pgraph_vk_draw_end(NV2AState *d);
 void pgraph_vk_finish(PGRAPHState *pg, FinishReason why);
 void pgraph_vk_flush_draw(NV2AState *d);
+void pgraph_vk_flush_reorder_window(NV2AState *d);
 void pgraph_vk_begin_command_buffer(PGRAPHState *pg);
 void pgraph_vk_ensure_command_buffer(PGRAPHState *pg);
 void pgraph_vk_ensure_not_in_render_pass(PGRAPHState *pg);

hw/xbox/nv2a/pgraph/vk/surface.c

@@ -2029,6 +2029,9 @@ void pgraph_vk_surface_update(NV2AState *d, bool upload, bool color_write,
             update_surface_part(d, true, false);
         }
     } else {
+#if OPT_REORDER_SAFE_WINDOWS
+        pgraph_vk_flush_reorder_window(d);
+#endif
         if ((color_write || pg->surface_color.write_enabled_cache)
             && pg->surface_color.draw_dirty) {
             update_surface_part(d, false, true);

hw/xbox/nv2a/pgraph/vk/texture.c

@@ -689,6 +689,13 @@ static void copy_zeta_surface_to_texture(PGRAPHState *pg, SurfaceBinding *surfac
     TextureShape *state = &texture->key.state;
     VkColorFormatInfo vkf = kelvin_color_format_vk_map[state->color_format];
 
+#if OPT_REORDER_SAFE_WINDOWS
+    if (r->reorder_window.count > 0) {
+        NV2AState *d = container_of(pg, NV2AState, pgraph);
+        pgraph_vk_flush_reorder_window(d);
+    }
+#endif
+
     bool use_compute_to_convert_depth_stencil =
         surface->host_fmt.vk_format == VK_FORMAT_D24_UNORM_S8_UINT ||
         surface->host_fmt.vk_format == VK_FORMAT_D32_SFLOAT_S8_UINT;
@@ -1216,6 +1223,13 @@ static void copy_surface_to_texture(PGRAPHState *pg, SurfaceBinding *surface,
     TextureShape *state = &texture->key.state;
     VkColorFormatInfo vkf = kelvin_color_format_vk_map[state->color_format];
 
+#if OPT_REORDER_SAFE_WINDOWS
+    if (r->reorder_window.count > 0) {
+        NV2AState *d = container_of(pg, NV2AState, pgraph);
+        pgraph_vk_flush_reorder_window(d);
+    }
+#endif
+
     nv2a_profile_inc_counter(NV2A_PROF_SURF_TO_TEX);
 
     trace_nv2a_pgraph_surface_render_to_texture(