[Mesa-dev] [PATCH 1/2] gallium/radeon: add a query for monitoring Gallium thread load

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

---
 src/gallium/drivers/radeon/r600_query.c | 12 ++++++++++++
 src/gallium/drivers/radeon/r600_query.h |  1 +
 2 files changed, 13 insertions(+)

diff --git a/src/gallium/drivers/radeon/r600_query.c b/src/gallium/drivers/radeon/r600_query.c
index 61fd662..0d960bc 100644
--- a/src/gallium/drivers/radeon/r600_query.c
+++ b/src/gallium/drivers/radeon/r600_query.c
@@ -163,20 +163,25 @@ static bool r600_query_sw_begin(struct r600_common_context *rctx,
 	case R600_QUERY_NUM_EVICTIONS: {
 		enum radeon_value_id ws_id = winsys_id_from_type(query->b.type);
 		query->begin_result = rctx->ws->query_value(rctx->ws, ws_id);
 		break;
 	}
 	case R600_QUERY_CS_THREAD_BUSY:
 		ws_id = winsys_id_from_type(query->b.type);
 		query->begin_result = rctx->ws->query_value(rctx->ws, ws_id);
 		query->begin_time = os_time_get_nano();
 		break;
+	case R600_QUERY_GALLIUM_THREAD_BUSY:
+		query->begin_result =
+			rctx->tc ? util_queue_get_thread_time_nano(&rctx->tc->queue, 0) : 0;
+		query->begin_time = os_time_get_nano();
+		break;
 	case R600_QUERY_GPU_LOAD:
 	case R600_QUERY_GPU_SHADERS_BUSY:
 	case R600_QUERY_GPU_TA_BUSY:
 	case R600_QUERY_GPU_GDS_BUSY:
 	case R600_QUERY_GPU_VGT_BUSY:
 	case R600_QUERY_GPU_IA_BUSY:
 	case R600_QUERY_GPU_SX_BUSY:
 	case R600_QUERY_GPU_WD_BUSY:
 	case R600_QUERY_GPU_BCI_BUSY:
 	case R600_QUERY_GPU_SC_BUSY:
@@ -296,20 +301,25 @@ static bool r600_query_sw_end(struct r600_common_context *rctx,
 	case R600_QUERY_NUM_EVICTIONS: {
 		enum radeon_value_id ws_id = winsys_id_from_type(query->b.type);
 		query->end_result = rctx->ws->query_value(rctx->ws, ws_id);
 		break;
 	}
 	case R600_QUERY_CS_THREAD_BUSY:
 		ws_id = winsys_id_from_type(query->b.type);
 		query->end_result = rctx->ws->query_value(rctx->ws, ws_id);
 		query->end_time = os_time_get_nano();
 		break;
+	case R600_QUERY_GALLIUM_THREAD_BUSY:
+		query->end_result =
+			rctx->tc ? util_queue_get_thread_time_nano(&rctx->tc->queue, 0) : 0;
+		query->end_time = os_time_get_nano();
+		break;
 	case R600_QUERY_GPU_LOAD:
 	case R600_QUERY_GPU_SHADERS_BUSY:
 	case R600_QUERY_GPU_TA_BUSY:
 	case R600_QUERY_GPU_GDS_BUSY:
 	case R600_QUERY_GPU_VGT_BUSY:
 	case R600_QUERY_GPU_IA_BUSY:
 	case R600_QUERY_GPU_SX_BUSY:
 	case R600_QUERY_GPU_WD_BUSY:
 	case R600_QUERY_GPU_BCI_BUSY:
 	case R600_QUERY_GPU_SC_BUSY:
@@ -373,20 +383,21 @@ static bool r600_query_sw_get_result(struct r600_common_context *rctx,
 	case PIPE_QUERY_GPU_FINISHED: {
 		struct pipe_screen *screen = rctx->b.screen;
 		struct pipe_context *ctx = rquery->b.flushed ? NULL : &rctx->b;
 
 		result->b = screen->fence_finish(screen, ctx, query->fence,
 						 wait ? PIPE_TIMEOUT_INFINITE : 0);
 		return result->b;
 	}
 
 	case R600_QUERY_CS_THREAD_BUSY:
+	case R600_QUERY_GALLIUM_THREAD_BUSY:
 		result->u64 = (query->end_result - query->begin_result) * 100 /
 			      (query->end_time - query->begin_time);
 		return true;
 	case R600_QUERY_GPIN_ASIC_ID:
 		result->u32 = 0;
 		return true;
 	case R600_QUERY_GPIN_NUM_SIMD:
 		result->u32 = rctx->screen->info.num_good_compute_units;
 		return true;
 	case R600_QUERY_GPIN_NUM_RB:
@@ -1817,20 +1828,21 @@ static struct pipe_driver_query_info r600_driver_query_list[] = {
 	X("num-vs-flushes",		NUM_VS_FLUSHES,		UINT64, AVERAGE),
 	X("num-ps-flushes",		NUM_PS_FLUSHES,		UINT64, AVERAGE),
 	X("num-cs-flushes",		NUM_CS_FLUSHES,		UINT64, AVERAGE),
 	X("num-fb-cache-flushes",	NUM_FB_CACHE_FLUSHES,	UINT64, AVERAGE),
 	X("num-L2-invalidates",		NUM_L2_INVALIDATES,	UINT64, AVERAGE),
 	X("num-L2-writebacks",		NUM_L2_WRITEBACKS,	UINT64, AVERAGE),
 	X("tc-offloaded-slots",		TC_OFFLOADED_SLOTS,     UINT64, AVERAGE),
 	X("tc-direct-slots",		TC_DIRECT_SLOTS,	UINT64, AVERAGE),
 	X("tc-num-syncs",		TC_NUM_SYNCS,		UINT64, AVERAGE),
 	X("CS-thread-busy",		CS_THREAD_BUSY,		UINT64, AVERAGE),
+	X("gallium-thread-busy",	GALLIUM_THREAD_BUSY,	UINT64, AVERAGE),
 	X("requested-VRAM",		REQUESTED_VRAM,		BYTES, AVERAGE),
 	X("requested-GTT",		REQUESTED_GTT,		BYTES, AVERAGE),
 	X("mapped-VRAM",		MAPPED_VRAM,		BYTES, AVERAGE),
 	X("mapped-GTT",			MAPPED_GTT,		BYTES, AVERAGE),
 	X("buffer-wait-time",		BUFFER_WAIT_TIME,	MICROSECONDS, CUMULATIVE),
 	X("num-mapped-buffers",		NUM_MAPPED_BUFFERS,	UINT64, AVERAGE),
 	X("num-GFX-IBs",		NUM_GFX_IBS,		UINT64, AVERAGE),
 	X("num-SDMA-IBs",		NUM_SDMA_IBS,		UINT64, AVERAGE),
 	X("num-bytes-moved",		NUM_BYTES_MOVED,	BYTES, CUMULATIVE),
 	X("num-evictions",		NUM_EVICTIONS,		UINT64, CUMULATIVE),
diff --git a/src/gallium/drivers/radeon/r600_query.h b/src/gallium/drivers/radeon/r600_query.h
index cf99d3a..9017fdf 100644
--- a/src/gallium/drivers/radeon/r600_query.h
+++ b/src/gallium/drivers/radeon/r600_query.h
@@ -51,20 +51,21 @@ enum {
 	R600_QUERY_NUM_VS_FLUSHES,
 	R600_QUERY_NUM_PS_FLUSHES,
 	R600_QUERY_NUM_CS_FLUSHES,
 	R600_QUERY_NUM_FB_CACHE_FLUSHES,
 	R600_QUERY_NUM_L2_INVALIDATES,
 	R600_QUERY_NUM_L2_WRITEBACKS,
 	R600_QUERY_TC_OFFLOADED_SLOTS,
 	R600_QUERY_TC_DIRECT_SLOTS,
 	R600_QUERY_TC_NUM_SYNCS,
 	R600_QUERY_CS_THREAD_BUSY,
+	R600_QUERY_GALLIUM_THREAD_BUSY,
 	R600_QUERY_REQUESTED_VRAM,
 	R600_QUERY_REQUESTED_GTT,
 	R600_QUERY_MAPPED_VRAM,
 	R600_QUERY_MAPPED_GTT,
 	R600_QUERY_BUFFER_WAIT_TIME,
 	R600_QUERY_NUM_MAPPED_BUFFERS,
 	R600_QUERY_NUM_GFX_IBS,
 	R600_QUERY_NUM_SDMA_IBS,
 	R600_QUERY_NUM_BYTES_MOVED,
 	R600_QUERY_NUM_EVICTIONS,
-- 
2.7.4