[Mesa-dev] [PATCH 13/13] radeonsi: enable threaded_context

[Dieter Nützel]

For the series:

Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>

on radeonsi / RX580, 8 GB

Unigine_Heaven-4.0
Unigine_Valley-1.0
Unigine_Superposition-1.0

running wine-2.7_gallium_nine+staging

Steam: PES2015 + TS2017
LS2015 + LS2017 (!!! ;-)

Except: currently _only_ Linux 4.11
(have to compile 4.12-rcx or Alex's tree next night)

Dieter

Am 11.05.2017 00:45, schrieb Marek Olšák:
> From: Marek Olšák <marek.olsak at amd.com>
> 
> ---
>  src/gallium/drivers/radeonsi/si_pipe.c | 37 
> +++++++++++++++++++++++++++++++---
>  1 file changed, 34 insertions(+), 3 deletions(-)
> 
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.c
> b/src/gallium/drivers/radeonsi/si_pipe.c
> index 53738a5..c7ca928 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.c
> +++ b/src/gallium/drivers/radeonsi/si_pipe.c
> @@ -131,21 +131,21 @@ si_create_llvm_target_machine(struct si_screen 
> *sscreen)
>  				       r600_get_llvm_processor_name(sscreen->b.family),
>  				       sscreen->b.debug_flags & DBG_SI_SCHED ?
>  					       SI_LLVM_DEFAULT_FEATURES ",+si-scheduler" :
>  					       SI_LLVM_DEFAULT_FEATURES,
>  				       LLVMCodeGenLevelDefault,
>  				       LLVMRelocDefault,
>  				       LLVMCodeModelDefault);
>  }
> 
>  static struct pipe_context *si_create_context(struct pipe_screen 
> *screen,
> -                                              void *priv, unsigned 
> flags)
> +                                              unsigned flags)
>  {
>  	struct si_context *sctx = CALLOC_STRUCT(si_context);
>  	struct si_screen* sscreen = (struct si_screen *)screen;
>  	struct radeon_winsys *ws = sscreen->b.ws;
>  	int shader, i;
> 
>  	if (!sctx)
>  		return NULL;
> 
>  	if (sscreen->b.debug_flags & DBG_CHECK_VM)
> @@ -313,20 +313,51 @@ static struct pipe_context
> *si_create_context(struct pipe_screen *screen,
> 
>  	sctx->tm = si_create_llvm_target_machine(sscreen);
> 
>  	return &sctx->b.b;
>  fail:
>  	fprintf(stderr, "radeonsi: Failed to create a context.\n");
>  	si_destroy_context(&sctx->b.b);
>  	return NULL;
>  }
> 
> +static struct pipe_context *si_pipe_create_context(struct pipe_screen 
> *screen,
> +						   void *priv, unsigned flags)
> +{
> +	struct si_screen *sscreen = (struct si_screen *)screen;
> +	struct pipe_context *ctx = si_create_context(screen, flags);
> +
> +	if (!(flags & PIPE_CONTEXT_PREFER_THREADED))
> +		return ctx;
> +
> +	/* Clover (compute-only) is unsupported.
> +	 *
> +	 * Since the threaded context creates shader states from the 
> non-driver
> +	 * thread, asynchronous compilation is required for create_{shader}_-
> +	 * state not to use pipe_context. Debug contexts (ddebug) disable
> +	 * asynchronous compilation, so don't use the threaded context with
> +	 * those.
> +	 */
> +	if (flags & (PIPE_CONTEXT_COMPUTE_ONLY | PIPE_CONTEXT_DEBUG))
> +		return ctx;
> +
> +	/* When shaders are logged to stderr, asynchronous compilation is
> +	 * disabled too. */
> +	if (sscreen->b.debug_flags & (DBG_VS | DBG_TCS | DBG_TES | DBG_GS |
> +				      DBG_PS | DBG_CS))
> +		return ctx;
> +
> +	return threaded_context_create(ctx, &sscreen->b.pool_transfers,
> +				       r600_replace_buffer_storage,
> +				       &((struct si_context*)ctx)->b.tc);
> +}
> +
>  /*
>   * pipe_screen
>   */
>  static bool si_have_tgsi_compute(struct si_screen *sscreen)
>  {
>  	/* Old kernels disallowed some register writes for SI
>  	 * that are used for indirect dispatches. */
>  	return (sscreen->b.chip_class >= CIK ||
>  		sscreen->b.info.drm_major == 3 ||
>  		(sscreen->b.info.drm_major == 2 &&
> @@ -823,21 +854,21 @@ static void si_test_vmfault(struct si_screen 
> *sscreen)
>  struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
>  {
>  	struct si_screen *sscreen = CALLOC_STRUCT(si_screen);
>  	unsigned num_cpus, num_compiler_threads, i;
> 
>  	if (!sscreen) {
>  		return NULL;
>  	}
> 
>  	/* Set functions first. */
> -	sscreen->b.b.context_create = si_create_context;
> +	sscreen->b.b.context_create = si_pipe_create_context;
>  	sscreen->b.b.destroy = si_destroy_screen;
>  	sscreen->b.b.get_param = si_get_param;
>  	sscreen->b.b.get_shader_param = si_get_shader_param;
>  	sscreen->b.b.resource_create = r600_resource_create_common;
> 
>  	si_init_screen_state_functions(sscreen);
> 
>  	if (!r600_common_screen_init(&sscreen->b, ws) ||
>  	    !si_init_gs_info(sscreen) ||
>  	    !si_init_shader_cache(sscreen)) {
> @@ -912,21 +943,21 @@ struct pipe_screen
> *radeonsi_screen_create(struct radeon_winsys *ws)
>  					    SI_CONTEXT_INV_GLOBAL_L2;
>  	sscreen->b.barrier_flags.compute_to_L2 = SI_CONTEXT_CS_PARTIAL_FLUSH;
> 
>  	if (debug_get_bool_option("RADEON_DUMP_SHADERS", false))
>  		sscreen->b.debug_flags |= DBG_FS | DBG_VS | DBG_GS | DBG_PS | 
> DBG_CS;
> 
>  	for (i = 0; i < num_compiler_threads; i++)
>  		sscreen->tm[i] = si_create_llvm_target_machine(sscreen);
> 
>  	/* Create the auxiliary context. This must be done last. */
> -	sscreen->b.aux_context = sscreen->b.b.context_create(&sscreen->b.b, 
> NULL, 0);
> +	sscreen->b.aux_context = si_create_context(&sscreen->b.b, 0);
> 
>  	if (sscreen->b.debug_flags & DBG_TEST_DMA)
>  		r600_test_dma(&sscreen->b);
> 
>  	if (sscreen->b.debug_flags & (DBG_TEST_VMFAULT_CP |
>  				      DBG_TEST_VMFAULT_SDMA |
>  				      DBG_TEST_VMFAULT_SHADER))
>  		si_test_vmfault(sscreen);
> 
>  	return &sscreen->b.b;