[Mesa-dev] [PATCH 14/20] radeonsi: add HUD queries for counting VS/PS/CS partial flushes

Marek Olšák maraeo at gmail.com
Mon Aug 29 15:28:29 UTC 2016


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

---
 src/gallium/drivers/radeon/r600_pipe_common.h |  3 +++
 src/gallium/drivers/radeon/r600_query.c       | 21 +++++++++++++++++++++
 src/gallium/drivers/radeon/r600_query.h       |  3 +++
 src/gallium/drivers/radeonsi/si_state_draw.c  |  8 ++++++++
 4 files changed, 35 insertions(+)

diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index 624dea3..d821eaa 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -551,20 +551,23 @@ struct r600_common_context {
 	unsigned			num_cs_dw_queries_suspend;
 	/* Additional hardware info. */
 	unsigned			backend_mask;
 	unsigned			max_db; /* for OQ */
 	/* Misc stats. */
 	unsigned			num_draw_calls;
 	unsigned			num_spill_draw_calls;
 	unsigned			num_compute_calls;
 	unsigned			num_spill_compute_calls;
 	unsigned			num_dma_calls;
+	unsigned			num_vs_flushes;
+	unsigned			num_ps_flushes;
+	unsigned			num_cs_flushes;
 	uint64_t			num_alloc_tex_transfer_bytes;
 	unsigned			last_tex_ps_draw_ratio; /* for query */
 
 	/* Render condition. */
 	struct r600_atom		render_cond_atom;
 	struct pipe_query		*render_cond;
 	unsigned			render_cond_mode;
 	bool				render_cond_invert;
 	bool				render_cond_force_off; /* for u_blitter */
 
diff --git a/src/gallium/drivers/radeon/r600_query.c b/src/gallium/drivers/radeon/r600_query.c
index 29ad249..2c3d530 100644
--- a/src/gallium/drivers/radeon/r600_query.c
+++ b/src/gallium/drivers/radeon/r600_query.c
@@ -83,20 +83,29 @@ static bool r600_query_sw_begin(struct r600_common_context *rctx,
 		break;
 	case R600_QUERY_COMPUTE_CALLS:
 		query->begin_result = rctx->num_compute_calls;
 		break;
 	case R600_QUERY_SPILL_COMPUTE_CALLS:
 		query->begin_result = rctx->num_spill_compute_calls;
 		break;
 	case R600_QUERY_DMA_CALLS:
 		query->begin_result = rctx->num_dma_calls;
 		break;
+	case R600_QUERY_NUM_VS_FLUSHES:
+		query->begin_result = rctx->num_vs_flushes;
+		break;
+	case R600_QUERY_NUM_PS_FLUSHES:
+		query->begin_result = rctx->num_ps_flushes;
+		break;
+	case R600_QUERY_NUM_CS_FLUSHES:
+		query->begin_result = rctx->num_cs_flushes;
+		break;
 	case R600_QUERY_REQUESTED_VRAM:
 	case R600_QUERY_REQUESTED_GTT:
 	case R600_QUERY_MAPPED_VRAM:
 	case R600_QUERY_MAPPED_GTT:
 	case R600_QUERY_VRAM_USAGE:
 	case R600_QUERY_GTT_USAGE:
 	case R600_QUERY_GPU_TEMPERATURE:
 	case R600_QUERY_CURRENT_GPU_SCLK:
 	case R600_QUERY_CURRENT_GPU_MCLK:
 	case R600_QUERY_BACK_BUFFER_PS_DRAW_RATIO:
@@ -151,20 +160,29 @@ static bool r600_query_sw_end(struct r600_common_context *rctx,
 		break;
 	case R600_QUERY_COMPUTE_CALLS:
 		query->end_result = rctx->num_compute_calls;
 		break;
 	case R600_QUERY_SPILL_COMPUTE_CALLS:
 		query->end_result = rctx->num_spill_compute_calls;
 		break;
 	case R600_QUERY_DMA_CALLS:
 		query->end_result = rctx->num_dma_calls;
 		break;
+	case R600_QUERY_NUM_VS_FLUSHES:
+		query->end_result = rctx->num_vs_flushes;
+		break;
+	case R600_QUERY_NUM_PS_FLUSHES:
+		query->end_result = rctx->num_ps_flushes;
+		break;
+	case R600_QUERY_NUM_CS_FLUSHES:
+		query->end_result = rctx->num_cs_flushes;
+		break;
 	case R600_QUERY_REQUESTED_VRAM:
 	case R600_QUERY_REQUESTED_GTT:
 	case R600_QUERY_MAPPED_VRAM:
 	case R600_QUERY_MAPPED_GTT:
 	case R600_QUERY_VRAM_USAGE:
 	case R600_QUERY_GTT_USAGE:
 	case R600_QUERY_GPU_TEMPERATURE:
 	case R600_QUERY_CURRENT_GPU_SCLK:
 	case R600_QUERY_CURRENT_GPU_MCLK:
 	case R600_QUERY_BUFFER_WAIT_TIME:
@@ -1175,20 +1193,23 @@ err:
 	XFULL(name_, query_type_, type_, result_type_, R600_QUERY_GROUP_##group_)
 
 static struct pipe_driver_query_info r600_driver_query_list[] = {
 	X("num-compilations",		NUM_COMPILATIONS,	UINT64, CUMULATIVE),
 	X("num-shaders-created",	NUM_SHADERS_CREATED,	UINT64, CUMULATIVE),
 	X("draw-calls",			DRAW_CALLS,		UINT64, AVERAGE),
 	X("spill-draw-calls",		SPILL_DRAW_CALLS,	UINT64, AVERAGE),
 	X("compute-calls",		COMPUTE_CALLS,		UINT64, AVERAGE),
 	X("spill-compute-calls",	SPILL_COMPUTE_CALLS,	UINT64, AVERAGE),
 	X("dma-calls",			DMA_CALLS,		UINT64, AVERAGE),
+	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("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-ctx-flushes",		NUM_CTX_FLUSHES,	UINT64, AVERAGE),
 	X("num-bytes-moved",		NUM_BYTES_MOVED,	BYTES, CUMULATIVE),
 	X("num-evictions",		NUM_EVICTIONS,		UINT64, CUMULATIVE),
 	X("VRAM-usage",			VRAM_USAGE,		BYTES, AVERAGE),
 	X("GTT-usage",			GTT_USAGE,		BYTES, AVERAGE),
diff --git a/src/gallium/drivers/radeon/r600_query.h b/src/gallium/drivers/radeon/r600_query.h
index 89f5516..0cd1a02 100644
--- a/src/gallium/drivers/radeon/r600_query.h
+++ b/src/gallium/drivers/radeon/r600_query.h
@@ -39,20 +39,23 @@ struct r600_common_screen;
 struct r600_query;
 struct r600_query_hw;
 struct r600_resource;
 
 enum {
 	R600_QUERY_DRAW_CALLS = PIPE_QUERY_DRIVER_SPECIFIC,
 	R600_QUERY_SPILL_DRAW_CALLS,
 	R600_QUERY_COMPUTE_CALLS,
 	R600_QUERY_SPILL_COMPUTE_CALLS,
 	R600_QUERY_DMA_CALLS,
+	R600_QUERY_NUM_VS_FLUSHES,
+	R600_QUERY_NUM_PS_FLUSHES,
+	R600_QUERY_NUM_CS_FLUSHES,
 	R600_QUERY_REQUESTED_VRAM,
 	R600_QUERY_REQUESTED_GTT,
 	R600_QUERY_MAPPED_VRAM,
 	R600_QUERY_MAPPED_GTT,
 	R600_QUERY_BUFFER_WAIT_TIME,
 	R600_QUERY_NUM_CTX_FLUSHES,
 	R600_QUERY_NUM_BYTES_MOVED,
 	R600_QUERY_NUM_EVICTIONS,
 	R600_QUERY_VRAM_USAGE,
 	R600_QUERY_GTT_USAGE,
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 60cc3f0..9e50bb2 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -766,28 +766,36 @@ void si_emit_cache_flush(struct si_context *si_ctx, struct r600_atom *atom)
 
 	/* Wait for shader engines to go idle.
 	 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
 	 * for everything including CB/DB cache flushes.
 	 */
 	if (!(sctx->flags & (SI_CONTEXT_FLUSH_AND_INV_CB |
 			     SI_CONTEXT_FLUSH_AND_INV_DB))) {
 		if (sctx->flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
 			radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
 			radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
+			/* Only count explicit shader flushes, not implicit ones
+			 * done by SURFACE_SYNC.
+			 */
+			sctx->num_vs_flushes++;
+			sctx->num_ps_flushes++;
 		} else if (sctx->flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
 			radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
 			radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
+			sctx->num_vs_flushes++;
 		}
 	}
+
 	if (sctx->flags & SI_CONTEXT_CS_PARTIAL_FLUSH) {
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
 		radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
+		sctx->num_cs_flushes++;
 	}
 
 	/* VGT state synchronization. */
 	if (sctx->flags & SI_CONTEXT_VGT_FLUSH) {
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
 		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
 	}
 	if (sctx->flags & SI_CONTEXT_VGT_STREAMOUT_SYNC) {
 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
 		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
-- 
2.7.4



More information about the mesa-dev mailing list