radeonsi/gfx9: keep reusing the same buffer/address for the gfx9 flush fence

instead of using a monotonic suballocator

v2: initialize the memory at context creation

Reviewed-by: Nicolai Hähnle <[email protected]>

3 files changed, 28 insertions, 8 deletions

diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index a4f3c2d6149..74ac0ec0389 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -64,6 +64,7 @@ static void si_destroy_context(struct pipe_context *context)
 	free(sctx->border_color_table);
 	r600_resource_reference(&sctx->scratch_buffer, NULL);
 	r600_resource_reference(&sctx->compute_scratch_buffer, NULL);
+	r600_resource_reference(&sctx->wait_mem_scratch, NULL);
 
 	si_pm4_free_state(sctx, sctx->init_config, ~0);
 	if (sctx->init_config_gs_rings)
@@ -269,6 +270,23 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
 	/* these must be last */
 	si_begin_new_cs(sctx);
 
+	if (sctx->b.chip_class >= GFX9) {
+		sctx->wait_mem_scratch = (struct r600_resource*)
+			pipe_buffer_create(screen, 0, PIPE_USAGE_DEFAULT, 4);
+		if (!sctx->wait_mem_scratch)
+			goto fail;
+
+		/* Initialize the memory. */
+		struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
+		radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
+		radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
+			    S_370_WR_CONFIRM(1) |
+			    S_370_ENGINE_SEL(V_370_ME));
+		radeon_emit(cs, sctx->wait_mem_scratch->gpu_address);
+		radeon_emit(cs, sctx->wait_mem_scratch->gpu_address >> 32);
+		radeon_emit(cs, sctx->wait_mem_number);
+	}
+
 	/* CIK cannot unbind a constant buffer (S_BUFFER_LOAD doesn't skip loads
 	 * if NUM_RECORDS == 0). We need to use a dummy buffer instead. */
 	if (sctx->b.chip_class == CIK) {
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 1c174083819..bd724e80a06 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -263,6 +263,8 @@ struct si_context {
 	struct si_screen		*screen;
 	LLVMTargetMachineRef		tm; /* only non-threaded compilation */
 	struct si_shader_ctx_state	fixed_func_tcs_shader;
+	struct r600_resource		*wait_mem_scratch;
+	unsigned			wait_mem_number;
 
 	struct radeon_winsys_cs		*ce_ib;
 	struct radeon_winsys_cs		*ce_preamble_ib;
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 2b000e74696..85ceacad80f 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -954,9 +954,8 @@ void si_emit_cache_flush(struct si_context *sctx)
 	 * wait for idle on GFX9. We have to use a TS event.
 	 */
 	if (sctx->b.chip_class >= GFX9 && flush_cb_db) {
-		struct r600_resource *rbuf = NULL;
 		uint64_t va;
-		unsigned offset = 0, tc_flags, cb_db_event;
+		unsigned tc_flags, cb_db_event;
 
 		/* Set the CB/DB flush event. */
 		switch (flush_cb_db) {
@@ -997,14 +996,15 @@ void si_emit_cache_flush(struct si_context *sctx)
 			sctx->b.num_L2_invalidates++;
 		}
 
-		/* Allocate memory for the fence. */
-		u_suballocator_alloc(rctx->allocator_zeroed_memory, 4, 4,
-				     &offset, (struct pipe_resource**)&rbuf);
-		va = rbuf->gpu_address + offset;
+		/* Do the flush (enqueue the event and wait for it). */
+		va = sctx->wait_mem_scratch->gpu_address;
+		sctx->wait_mem_number++;
 
 		r600_gfx_write_event_eop(rctx, cb_db_event, tc_flags, 1,
-					 rbuf, va, 0, 1);
-		r600_gfx_wait_fence(rctx, va, 1, 0xffffffff);
+					 sctx->wait_mem_scratch, va,
+					 sctx->wait_mem_number - 1,
+					 sctx->wait_mem_number);
+		r600_gfx_wait_fence(rctx, va, sctx->wait_mem_number, 0xffffffff);
 	}
 
 	/* Make sure ME is idle (it executes most packets) before continuing.