nv2a/vk: Add command buffer debug region markers

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

@@ -81,3 +81,41 @@ void pgraph_vk_insert_debug_marker(PGRAPHVkState *r, VkCommandBuffer cmd,
     vkCmdInsertDebugUtilsLabelEXT(cmd, &label_info);
     free(buf);
 }
+
+void pgraph_vk_begin_debug_marker(PGRAPHVkState *r, VkCommandBuffer cmd,
+                                  float color[4], const char *format, ...)
+{
+    if (!r->debug_utils_extension_enabled) {
+        return;
+    }
+
+    char *buf = NULL;
+
+    va_list args;
+    va_start(args, format);
+    int err = vasprintf(&buf, format, args);
+    assert(err >= 0);
+    va_end(args);
+
+    VkDebugUtilsLabelEXT label_info = {
+        .sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,
+        .pLabelName = buf,
+    };
+    memcpy(label_info.color, color, 4 * sizeof(float));
+    vkCmdBeginDebugUtilsLabelEXT(cmd, &label_info);
+    free(buf);
+
+    r->debug_depth += 1;
+    assert(r->debug_depth < 10 && "Missing pgraph_vk_debug_marker_end?");
+}
+
+void pgraph_vk_end_debug_marker(PGRAPHVkState *r, VkCommandBuffer cmd)
+{
+    if (!r->debug_utils_extension_enabled) {
+        return;
+    }
+
+    vkCmdEndDebugUtilsLabelEXT(cmd);
+    assert(r->debug_depth > 0);
+    r->debug_depth -= 1;
+}

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

@@ -920,6 +920,10 @@ static void render_display(PGRAPHState *pg, SurfaceBinding *surface)
 
     VkCommandBuffer cmd = pgraph_vk_begin_single_time_commands(pg);
 
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_YELLOW,
+        "Display Surface %08"HWADDR_PRIx" %s",
+        surface->vram_addr, disp->pvideo.state.enabled ? "pvideo" : "");
+
     pgraph_vk_transition_image_layout(pg, cmd, surface->image,
                                       surface->host_fmt.vk_format,
                                       VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
@@ -994,6 +998,7 @@ static void render_display(PGRAPHState *pg, SurfaceBinding *surface)
                                       VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
                                       VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
 
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_single_time_commands(pg, cmd);
     nv2a_profile_inc_counter(NV2A_PROF_QUEUE_SUBMIT_5);
 

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

@@ -1263,9 +1263,9 @@ void pgraph_vk_finish(PGRAPHState *pg, FinishReason finish_reason)
     PGRAPHVkState *r = pg->vk_renderer_state;
 
     assert(!r->in_draw);
+    assert(r->debug_depth == 0);
 
     if (r->in_command_buffer) {
-
         nv2a_profile_inc_counter(finish_reason_to_counter_enum[finish_reason]);
 
         if (r->in_render_pass) {
@@ -1713,6 +1713,9 @@ void pgraph_vk_clear_surface(NV2AState *d, uint32_t parameter)
                          write_zeta ? " zeta" : "");
 
     begin_pre_draw(pg);
+    pgraph_vk_begin_debug_marker(r, r->command_buffer,
+        RGBA_BLUE, "Clear %08" HWADDR_PRIx,
+        binding->vram_addr);
     begin_draw(pg);
 
     // FIXME: What does hardware do when min <= max?
@@ -1791,6 +1794,7 @@ void pgraph_vk_clear_surface(NV2AState *d, uint32_t parameter)
                               1, &clear_rect);
     }
     end_draw(pg);
+    pgraph_vk_end_debug_marker(r, r->command_buffer);
 
     pg->clearing = false;
 
@@ -2064,6 +2068,8 @@ void pgraph_vk_flush_draw(NV2AState *d)
         NV2A_VK_DGROUP_BEGIN("Draw Arrays");
         nv2a_profile_inc_counter(NV2A_PROF_DRAW_ARRAYS);
 
+        NV2A_VK_DPRINTF("Primitive Mode: %d", pg->primitive_mode);
+
         assert(pg->inline_elements_length == 0);
         assert(pg->inline_buffer_length == 0);
         assert(pg->inline_array_length == 0);
@@ -2082,6 +2088,8 @@ void pgraph_vk_flush_draw(NV2AState *d)
         copy_remapped_attributes_to_inline_buffer(pg, remap, 0, max_element);
 
         begin_pre_draw(pg);
+        pgraph_vk_begin_debug_marker(r, r->command_buffer, RGBA_BLUE,
+                                     "Draw Arrays");
         begin_draw(pg);
         bind_vertex_buffer(pg, remap.attributes, 0);
         for (int i = 0; i < pg->draw_arrays_length; i++) {
@@ -2091,6 +2099,7 @@ void pgraph_vk_flush_draw(NV2AState *d)
             vkCmdDraw(r->command_buffer, count, 1, start, 0);
         }
         end_draw(pg);
+        pgraph_vk_end_debug_marker(r, r->command_buffer);
 
         NV2A_VK_DGROUP_END();
     } else if (pg->inline_elements_length) {
@@ -2121,6 +2130,8 @@ void pgraph_vk_flush_draw(NV2AState *d)
         begin_pre_draw(pg);
         VkDeviceSize buffer_offset = pgraph_vk_update_index_buffer(
             pg, pg->inline_elements, index_data_size);
+        pgraph_vk_begin_debug_marker(r, r->command_buffer, RGBA_BLUE,
+                                     "Inline Elements");
         begin_draw(pg);
         bind_vertex_buffer(pg, remap.attributes, 0);
         vkCmdBindIndexBuffer(r->command_buffer,
@@ -2129,6 +2140,7 @@ void pgraph_vk_flush_draw(NV2AState *d)
         vkCmdDrawIndexed(r->command_buffer, pg->inline_elements_length, 1, 0, 0,
                          0);
         end_draw(pg);
+        pgraph_vk_end_debug_marker(r, r->command_buffer);
 
         NV2A_VK_DGROUP_END();
     } else if (pg->inline_buffer_length) {
@@ -2159,10 +2171,13 @@ void pgraph_vk_flush_draw(NV2AState *d)
         begin_pre_draw(pg);
         VkDeviceSize buffer_offset = pgraph_vk_update_vertex_inline_buffer(
             pg, data, sizes, r->num_active_vertex_attribute_descriptions);
+        pgraph_vk_begin_debug_marker(r, r->command_buffer, RGBA_BLUE,
+                                     "Inline Buffer");
         begin_draw(pg);
         bind_inline_vertex_buffer(pg, buffer_offset);
         vkCmdDraw(r->command_buffer, pg->inline_buffer_length, 1, 0, 0);
         end_draw(pg);
+        pgraph_vk_end_debug_marker(r, r->command_buffer);
 
         NV2A_VK_DGROUP_END();
     } else if (pg->inline_array_length) {
@@ -2200,10 +2215,13 @@ void pgraph_vk_flush_draw(NV2AState *d)
         void *inline_array_data = pg->inline_array;
         VkDeviceSize buffer_offset = pgraph_vk_update_vertex_inline_buffer(
             pg, &inline_array_data, &inline_array_data_size, 1);
+        pgraph_vk_begin_debug_marker(r, r->command_buffer, RGBA_BLUE,
+                                     "Inline Array");
         begin_draw(pg);
         bind_inline_vertex_buffer(pg, buffer_offset);
         vkCmdDraw(r->command_buffer, index_count, 1, 0, 0);
         end_draw(pg);
+        pgraph_vk_end_debug_marker(r, r->command_buffer);
         NV2A_VK_DGROUP_END();
     } else {
         NV2A_VK_DPRINTF("EMPTY NV097_SET_BEGIN_END");

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

@@ -321,6 +321,7 @@ typedef struct PGRAPHVkState {
     void *window;
     VkInstance instance;
     VkDebugUtilsMessengerEXT debug_messenger;
+    int debug_depth;
 
     bool debug_utils_extension_enabled;
     bool custom_border_color_extension_enabled;
@@ -429,9 +430,19 @@ typedef struct PGRAPHVkState {
 void pgraph_vk_check_memory_budget(PGRAPHState *pg);
 
 // debug.c
+#define RGBA_RED     (float[4]){1,0,0,1}
+#define RGBA_YELLOW  (float[4]){1,1,0,1}
+#define RGBA_GREEN   (float[4]){0,1,0,1}
+#define RGBA_BLUE    (float[4]){0,0,1,1}
+#define RGBA_PINK    (float[4]){1,0,1,1}
+#define RGBA_DEFAULT (float[4]){0,0,0,0}
+
 void pgraph_vk_debug_init(void);
 void pgraph_vk_insert_debug_marker(PGRAPHVkState *r, VkCommandBuffer cmd,
                                    float color[4], const char *format, ...) __attribute__ ((format (printf, 4, 5)));
+void pgraph_vk_begin_debug_marker(PGRAPHVkState *r, VkCommandBuffer cmd,
+                                  float color[4], const char *format, ...) __attribute__ ((format (printf, 4, 5)));
+void pgraph_vk_end_debug_marker(PGRAPHVkState *r, VkCommandBuffer cmd);
 
 // instance.c
 void pgraph_vk_init_instance(PGRAPHState *pg, Error **errp);

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

@@ -434,6 +434,7 @@ void pgraph_vk_pack_depth_stencil(PGRAPHState *pg, SurfaceBinding *surface,
 
     // FIXME: Smarter workgroup scaling
 
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_PINK, "%s", __func__);
     vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline->pipeline);
     vkCmdBindDescriptorSets(
         cmd, VK_PIPELINE_BIND_POINT_COMPUTE, r->compute.pipeline_layout, 0, 1,
@@ -449,6 +450,7 @@ void pgraph_vk_pack_depth_stencil(PGRAPHState *pg, SurfaceBinding *surface,
     // FIXME: Check max group count
 
     vkCmdDispatch(cmd, group_count, 1, 1);
+    pgraph_vk_end_debug_marker(r, cmd);
 }
 
 void pgraph_vk_unpack_depth_stencil(PGRAPHState *pg, SurfaceBinding *surface,
@@ -505,6 +507,7 @@ void pgraph_vk_unpack_depth_stencil(PGRAPHState *pg, SurfaceBinding *surface,
 
     // FIXME: Smarter workgroup scaling
 
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_PINK, "%s", __func__);
     vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline->pipeline);
     vkCmdBindDescriptorSets(
         cmd, VK_PIPELINE_BIND_POINT_COMPUTE, r->compute.pipeline_layout, 0, 1,
@@ -518,6 +521,7 @@ void pgraph_vk_unpack_depth_stencil(PGRAPHState *pg, SurfaceBinding *surface,
                        VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(push_constants),
                        push_constants);
     vkCmdDispatch(cmd, group_count, 1, 1);
+    pgraph_vk_end_debug_marker(r, cmd);
 }
 
 static void pipeline_cache_entry_init(Lru *lru, LruNode *node, void *state)

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

@@ -191,6 +191,7 @@ static void download_surface_to_buffer(NV2AState *d, SurfaceBinding *surface,
     pgraph_apply_scaling_factor(pg, &scaled_width, &scaled_height);
 
     VkCommandBuffer cmd = pgraph_vk_begin_single_time_commands(pg);
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_RED, "%s", __func__);
 
     pgraph_vk_transition_image_layout(
         pg, cmd, surface->image, surface->host_fmt.vk_format,
@@ -432,6 +433,7 @@ static void download_surface_to_buffer(NV2AState *d, SurfaceBinding *surface,
                          &post_copy_dst_barrier, 0, NULL);
 
     nv2a_profile_inc_counter(NV2A_PROF_QUEUE_SUBMIT_1);
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_single_time_commands(pg, cmd);
 
     void *mapped_memory_ptr = NULL;
@@ -782,13 +784,16 @@ static void create_surface_image(PGRAPHState *pg, SurfaceBinding *surface)
 
     // FIXME: Go right into main command buffer
     VkCommandBuffer cmd = pgraph_vk_begin_single_time_commands(pg);
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_RED, "%s", __func__);
+
     pgraph_vk_transition_image_layout(
         pg, cmd, surface->image, surface->host_fmt.vk_format,
         VK_IMAGE_LAYOUT_UNDEFINED,
         surface->color ? VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL :
                          VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL);
 
     nv2a_profile_inc_counter(NV2A_PROF_QUEUE_SUBMIT_3);
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_single_time_commands(pg, cmd);
     nv2a_profile_inc_counter(NV2A_PROF_SURF_CREATE);
 }
@@ -977,6 +982,7 @@ void pgraph_vk_upload_surface_data(NV2AState *d, SurfaceBinding *surface,
     vmaUnmapMemory(r->allocator, copy_buffer->allocation);
 
     VkCommandBuffer cmd = pgraph_vk_begin_single_time_commands(pg);
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_RED, "%s", __func__);
 
     VkBufferMemoryBarrier host_barrier = {
         .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
@@ -1226,6 +1232,7 @@ void pgraph_vk_upload_surface_data(NV2AState *d, SurfaceBinding *surface,
                          VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL);
 
     nv2a_profile_inc_counter(NV2A_PROF_QUEUE_SUBMIT_2);
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_single_time_commands(pg, cmd);
 
     surface->initialized = true;

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

@@ -556,6 +556,7 @@ static void upload_texture_image(PGRAPHState *pg, int texture_idx,
 
     // FIXME: Use nondraw. Need to fill and copy tex buffer at once
     VkCommandBuffer cmd = pgraph_vk_begin_single_time_commands(pg);
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_GREEN, "%s", __func__);
 
     VkBufferMemoryBarrier host_barrier = {
         .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
@@ -585,6 +586,7 @@ static void upload_texture_image(PGRAPHState *pg, int texture_idx,
     binding->current_layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 
     nv2a_profile_inc_counter(NV2A_PROF_QUEUE_SUBMIT_4);
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_single_time_commands(pg, cmd);
 
     // Release decoded texture data
@@ -621,6 +623,7 @@ static void copy_zeta_surface_to_texture(PGRAPHState *pg, SurfaceBinding *surfac
         surface->vram_addr, surface->width, surface->height);
 
     VkCommandBuffer cmd = pgraph_vk_begin_nondraw_commands(pg);
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_GREEN, "%s", __func__);
 
     unsigned int scaled_width = surface->width,
                  scaled_height = surface->height;
@@ -781,6 +784,7 @@ static void copy_zeta_surface_to_texture(PGRAPHState *pg, SurfaceBinding *surfac
                                       VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
     texture->current_layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_nondraw_commands(pg, cmd);
 
     texture->draw_time = surface->draw_time;
@@ -795,6 +799,7 @@ static void copy_surface_to_texture(PGRAPHState *pg, SurfaceBinding *surface,
         return;
     }
 
+    PGRAPHVkState *r = pg->vk_renderer_state;
     TextureShape *state = &texture->key.state;
     VkColorFormatInfo vkf = kelvin_color_format_vk_map[state->color_format];
 
@@ -804,6 +809,7 @@ static void copy_surface_to_texture(PGRAPHState *pg, SurfaceBinding *surface,
         surface->vram_addr, surface->width, surface->height);
 
     VkCommandBuffer cmd = pgraph_vk_begin_nondraw_commands(pg);
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_GREEN, "%s", __func__);
 
     pgraph_vk_transition_image_layout(
         pg, cmd, surface->image, surface->host_fmt.vk_format,
@@ -842,6 +848,7 @@ static void copy_surface_to_texture(PGRAPHState *pg, SurfaceBinding *surface,
                                       VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
     texture->current_layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_nondraw_commands(pg, cmd);
 
     texture->draw_time = surface->draw_time;
@@ -1001,6 +1008,7 @@ static void create_dummy_texture(PGRAPHState *pg)
                    r->storage_buffers[BUFFER_STAGING_SRC].allocation);
 
     VkCommandBuffer cmd = pgraph_vk_begin_single_time_commands(pg);
+    pgraph_vk_begin_debug_marker(r, cmd, RGBA_GREEN, "%s", __func__);
 
     pgraph_vk_transition_image_layout(
         pg, cmd, texture_image, VK_FORMAT_R8_UNORM, VK_IMAGE_LAYOUT_UNDEFINED,
@@ -1026,6 +1034,8 @@ static void create_dummy_texture(PGRAPHState *pg)
                                       VK_FORMAT_R8_UNORM,
                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
                                       VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+
+    pgraph_vk_end_debug_marker(r, cmd);
     pgraph_vk_end_single_time_commands(pg, cmd);
 
     r->dummy_texture = (TextureBinding){