[Mesa-dev] [PATCH 2/2] radeonsi: decrease the number of compiler threads

[Marek Olšák]

This will also go to stable.

Marek

On Tue, Jul 25, 2017 at 5:36 PM, Marek Olšák <maraeo at gmail.com> wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> ---
>  src/gallium/drivers/radeonsi/si_pipe.c | 2 +-
>  src/gallium/drivers/radeonsi/si_pipe.h | 9 +++++++--
>  2 files changed, 8 insertions(+), 3 deletions(-)
>
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
> index 0bc3002..234469f 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.c
> +++ b/src/gallium/drivers/radeonsi/si_pipe.c
> @@ -963,21 +963,21 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws,
>         if (!util_queue_init(&sscreen->shader_compiler_queue, "si_shader",
>                              32, num_compiler_threads,
>                              UTIL_QUEUE_INIT_RESIZE_IF_FULL)) {
>                 si_destroy_shader_cache(sscreen);
>                 FREE(sscreen);
>                 return NULL;
>         }
>
>         if (!util_queue_init(&sscreen->shader_compiler_queue_low_priority,
>                              "si_shader_low",
> -                            32, num_compiler_threads,
> +                            32, num_compiler_threads_lowprio,
>                              UTIL_QUEUE_INIT_RESIZE_IF_FULL |
>                              UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY)) {
>                si_destroy_shader_cache(sscreen);
>                FREE(sscreen);
>                return NULL;
>         }
>
>         si_handle_env_var_force_family(sscreen);
>
>         if (!debug_get_bool_option("RADEON_DISABLE_PERFCOUNTERS", false))
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
> index c028aba..d25705b 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.h
> +++ b/src/gallium/drivers/radeonsi/si_pipe.h
> @@ -106,24 +106,29 @@ struct si_screen {
>          * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
>          *   variants of VS and TES are cached, so LS and ES aren't.
>          * - GS and CS aren't cached, but it's certainly possible to cache
>          *   those as well.
>          */
>         mtx_t                   shader_cache_mutex;
>         struct hash_table               *shader_cache;
>
>         /* Shader compiler queue for multithreaded compilation. */
>         struct util_queue               shader_compiler_queue;
> -       LLVMTargetMachineRef            tm[4]; /* used by the queue only */
> +       /* Use at most 3 normal compiler threads on quadcore and better.
> +        * Hyperthreaded CPUs report the number of threads, but we want
> +        * the number of cores. */
> +       LLVMTargetMachineRef            tm[3]; /* used by the queue only */
>
>         struct util_queue               shader_compiler_queue_low_priority;
> -       LLVMTargetMachineRef            tm_low_priority[4];
> +       /* Use at most 2 low priority threads on quadcore and better.
> +        * We want to minimize the impact on multithreaded Mesa. */
> +       LLVMTargetMachineRef            tm_low_priority[2]; /* at most 2 threads */
>  };
>
>  struct si_blend_color {
>         struct r600_atom                atom;
>         struct pipe_blend_color         state;
>  };
>
>  struct si_sampler_view {
>         struct pipe_sampler_view        base;
>          /* [0..7] = image descriptor
> --
> 2.7.4
>