[Mesa-dev] [PATCH 3/3] radeonsi: move descriptor logs to after corresponding draw/compute packet

[Samuel Pitoiset]

For the series:

Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>

On 09/26/2017 04:39 PM, Nicolai Hähnle wrote:
> From: Nicolai Hähnle <nicolai.haehnle at amd.com>
> 
> It has to happen after descriptor uploads since otherwise we'll print out
> the wrong GPU list / incorrectly claim descriptor corruption.
> ---
>   src/gallium/drivers/radeonsi/si_compute.c    | 7 +++----
>   src/gallium/drivers/radeonsi/si_state_draw.c | 7 +++----
>   2 files changed, 6 insertions(+), 8 deletions(-)
> 
> diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
> index ca334949d77..2346e2e95bc 100644
> --- a/src/gallium/drivers/radeonsi/si_compute.c
> +++ b/src/gallium/drivers/radeonsi/si_compute.c
> @@ -799,23 +799,20 @@ static void si_launch_grid(
>   		/* Indirect buffers use TC L2 on GFX9, but not older hw. */
>   		if (sctx->b.chip_class <= VI &&
>   		    r600_resource(info->indirect)->TC_L2_dirty) {
>   			sctx->b.flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
>   			r600_resource(info->indirect)->TC_L2_dirty = false;
>   		}
>   	}
>   
>   	si_need_cs_space(sctx);
>   
> -	if (sctx->b.log)
> -		si_log_compute_state(sctx, sctx->b.log);
> -
>   	if (!sctx->cs_shader_state.initialized)
>   		si_initialize_compute(sctx);
>   
>   	if (sctx->b.flags)
>   		si_emit_cache_flush(sctx);
>   
>   	if (!si_switch_compute_shader(sctx, program, &program->shader,
>   					code_object, info->pc))
>   		return;
>   
> @@ -844,22 +841,24 @@ static void si_launch_grid(
>   		radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, buffer,
>   					  RADEON_USAGE_READWRITE,
>   					  RADEON_PRIO_COMPUTE_GLOBAL);
>   	}
>   
>   	if (program->ir_type == PIPE_SHADER_IR_TGSI)
>   		si_setup_tgsi_grid(sctx, info);
>   
>   	si_emit_dispatch_packets(sctx, info);
>   
> -	if (unlikely(sctx->current_saved_cs))
> +	if (unlikely(sctx->current_saved_cs)) {
>   		si_trace_emit(sctx);
> +		si_log_compute_state(sctx, sctx->b.log);
> +	}
>   
>   	sctx->compute_is_busy = true;
>   	sctx->b.num_compute_calls++;
>   	if (sctx->cs_shader_state.uses_scratch)
>   		sctx->b.num_spill_compute_calls++;
>   
>   	if (cs_regalloc_hang)
>   		sctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
>   }
>   
> diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
> index 05ed85475bf..e29d716580a 100644
> --- a/src/gallium/drivers/radeonsi/si_state_draw.c
> +++ b/src/gallium/drivers/radeonsi/si_state_draw.c
> @@ -1369,23 +1369,20 @@ void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
>   			if (indirect->indirect_draw_count &&
>   			    r600_resource(indirect->indirect_draw_count)->TC_L2_dirty) {
>   				sctx->b.flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
>   				r600_resource(indirect->indirect_draw_count)->TC_L2_dirty = false;
>   			}
>   		}
>   	}
>   
>   	si_need_cs_space(sctx);
>   
> -	if (unlikely(sctx->b.log))
> -		si_log_draw_state(sctx, sctx->b.log);
> -
>   	/* Since we've called r600_context_add_resource_size for vertex buffers,
>   	 * this must be called after si_need_cs_space, because we must let
>   	 * need_cs_space flush before we add buffers to the buffer list.
>   	 */
>   	if (!si_upload_vertex_buffer_descriptors(sctx))
>   		return;
>   
>   	/* GFX9 scissor bug workaround. This must be done before VPORT scissor
>   	 * registers are changed. There is also a more efficient but more
>   	 * involved alternative workaround.
> @@ -1447,22 +1444,24 @@ void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
>   		if (sctx->b.chip_class >= CIK && sctx->prefetch_L2_mask)
>   			cik_emit_prefetch_L2(sctx);
>   
>   		if (!si_upload_graphics_shader_descriptors(sctx))
>   			return;
>   
>   		si_emit_all_states(sctx, info, 0);
>   		si_emit_draw_packets(sctx, info, indexbuf, index_size, index_offset);
>   	}
>   
> -	if (unlikely(sctx->current_saved_cs))
> +	if (unlikely(sctx->current_saved_cs)) {
>   		si_trace_emit(sctx);
> +		si_log_draw_state(sctx, sctx->b.log);
> +	}
>   
>   	/* Workaround for a VGT hang when streamout is enabled.
>   	 * It must be done after drawing. */
>   	if ((sctx->b.family == CHIP_HAWAII ||
>   	     sctx->b.family == CHIP_TONGA ||
>   	     sctx->b.family == CHIP_FIJI) &&
>   	    r600_get_strmout_en(&sctx->b)) {
>   		sctx->b.flags |= SI_CONTEXT_VGT_STREAMOUT_SYNC;
>   	}
>   
>