radeonsi: add a HUD query showing the number of compiler invocations

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

src/gallium/drivers/radeon/r600_pipe_common.c

@@ -689,6 +689,8 @@ static int r600_get_driver_query_info(struct pipe_screen *screen,
 {
 	struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
 	struct pipe_driver_query_info list[] = {
+		{"num-compilations", R600_QUERY_NUM_COMPILATIONS, {0}, PIPE_DRIVER_QUERY_TYPE_UINT64,
+		 PIPE_DRIVER_QUERY_RESULT_TYPE_CUMULATIVE},
 		{"draw-calls", R600_QUERY_DRAW_CALLS, {0}},
 		{"requested-VRAM", R600_QUERY_REQUESTED_VRAM, {rscreen->info.vram_size}, PIPE_DRIVER_QUERY_TYPE_BYTES},
 		{"requested-GTT", R600_QUERY_REQUESTED_GTT, {rscreen->info.gart_size}, PIPE_DRIVER_QUERY_TYPE_BYTES},
@@ -709,7 +711,7 @@ static int r600_get_driver_query_info(struct pipe_screen *screen,
 	if (rscreen->info.drm_major == 2 && rscreen->info.drm_minor >= 42)
 		num_queries = Elements(list);
 	else
-		num_queries = 8;
+		num_queries = 9;
 
 	if (!info)
 		return num_queries;

src/gallium/drivers/radeon/r600_pipe_common.h

@@ -59,6 +59,7 @@
 #define R600_QUERY_CURRENT_GPU_SCLK	(PIPE_QUERY_DRIVER_SPECIFIC + 9)
 #define R600_QUERY_CURRENT_GPU_MCLK	(PIPE_QUERY_DRIVER_SPECIFIC + 10)
 #define R600_QUERY_GPU_LOAD		(PIPE_QUERY_DRIVER_SPECIFIC + 11)
+#define R600_QUERY_NUM_COMPILATIONS	(PIPE_QUERY_DRIVER_SPECIFIC + 12)
 
 #define R600_CONTEXT_STREAMOUT_FLUSH		(1u << 0)
 #define R600_CONTEXT_PRIVATE_FLAG		(1u << 1)
@@ -288,6 +289,11 @@ struct r600_common_screen {
 	uint32_t			*trace_ptr;
 	unsigned			cs_count;
 
+	/* This must be in the screen, because UE4 uses one context for
+	 * compilation and another one for rendering.
+	 */
+	unsigned			num_compilations;
+
 	/* GPU load thread. */
 	pipe_mutex			gpu_load_mutex;
 	pipe_thread			gpu_load_thread;

src/gallium/drivers/radeon/r600_query.c

@@ -92,6 +92,7 @@ static struct r600_resource *r600_new_query_buffer(struct r600_common_context *c
 	case R600_QUERY_CURRENT_GPU_SCLK:
 	case R600_QUERY_CURRENT_GPU_MCLK:
 	case R600_QUERY_GPU_LOAD:
+	case R600_QUERY_NUM_COMPILATIONS:
 		return NULL;
 	}
 
@@ -408,6 +409,7 @@ static struct pipe_query *r600_create_query(struct pipe_context *ctx, unsigned q
 	case R600_QUERY_CURRENT_GPU_SCLK:
 	case R600_QUERY_CURRENT_GPU_MCLK:
 	case R600_QUERY_GPU_LOAD:
+	case R600_QUERY_NUM_COMPILATIONS:
 		skip_allocation = true;
 		break;
 	default:
@@ -483,6 +485,9 @@ static boolean r600_begin_query(struct pipe_context *ctx,
 	case R600_QUERY_GPU_LOAD:
 		rquery->begin_result = r600_gpu_load_begin(rctx->screen);
 		return true;
+	case R600_QUERY_NUM_COMPILATIONS:
+		rquery->begin_result = p_atomic_read(&rctx->screen->num_compilations);
+		return true;
 	}
 
 	/* Discard the old query buffers. */
@@ -560,6 +565,9 @@ static void r600_end_query(struct pipe_context *ctx, struct pipe_query *query)
 	case R600_QUERY_GPU_LOAD:
 		rquery->end_result = r600_gpu_load_end(rctx->screen, rquery->begin_result);
 		return;
+	case R600_QUERY_NUM_COMPILATIONS:
+		rquery->end_result = p_atomic_read(&rctx->screen->num_compilations);
+		return;
 	}
 
 	r600_emit_query_end(rctx, rquery);
@@ -619,6 +627,7 @@ static boolean r600_get_query_buffer_result(struct r600_common_context *ctx,
 	case R600_QUERY_GPU_TEMPERATURE:
 	case R600_QUERY_CURRENT_GPU_SCLK:
 	case R600_QUERY_CURRENT_GPU_MCLK:
+	case R600_QUERY_NUM_COMPILATIONS:
 		result->u64 = query->end_result - query->begin_result;
 		return TRUE;
 	case R600_QUERY_GPU_LOAD:

src/gallium/drivers/radeonsi/si_state_shaders.c

@@ -635,6 +635,7 @@ static int si_shader_select(struct pipe_context *ctx,
 		}
 		si_shader_init_pm4_state(shader);
 		sel->num_shaders++;
+		p_atomic_inc(&sctx->screen->b.num_compilations);
 	}
 
 	return 0;