From: Marek Olšák <marek.ol...@amd.com>
This is used when glthread is disabled.
Mesa pretty much chases the app thread on the CPU.
The performance is the same as pinning the app thread.
---
src/mesa/state_tracker/st_context.h | 2 ++
src/mesa/state_tracker/st_draw.c | 32 +++++++++++++++++++++++++++++
2 files changed, 34 insertions(+)
diff --git a/src/mesa/state_tracker/st_context.h
b/src/mesa/state_tracker/st_context.h
index 14b9b018809..95133c7020f 100644
--- a/src/mesa/state_tracker/st_context.h
+++ b/src/mesa/state_tracker/st_context.h
@@ -186,20 +186,22 @@ struct st_context
uint8_t sample_locations[
PIPE_MAX_SAMPLE_LOCATION_GRID_SIZE *
PIPE_MAX_SAMPLE_LOCATION_GRID_SIZE * 32];
} state;
uint64_t dirty; /**< dirty states */
/** This masks out unused shader resources. Only valid in draw calls. */
uint64_t active_states;
+ unsigned pin_thread_counter; /* for L3 thread pinning on AMD Zen */
+
/* If true, further analysis of states is required to know if something
* has changed. Used mainly for shaders.
*/
bool gfx_shaders_may_be_dirty;
bool compute_shader_may_be_dirty;
GLboolean vertdata_edgeflags;
GLboolean edgeflag_culls_prims;
struct st_vertex_program *vp; /**< Currently bound vertex program */
diff --git a/src/mesa/state_tracker/st_draw.c b/src/mesa/state_tracker/st_draw.c
index eb52d9505b6..5910ffa5bda 100644
--- a/src/mesa/state_tracker/st_draw.c
+++ b/src/mesa/state_tracker/st_draw.c
@@ -51,28 +51,36 @@
#include "st_atom.h"
#include "st_cb_bitmap.h"
#include "st_cb_bufferobjects.h"
#include "st_cb_xformfb.h"
#include "st_debug.h"
#include "st_draw.h"
#include "st_program.h"
#include "pipe/p_context.h"
#include "pipe/p_defines.h"
+#include "util/u_cpu_detect.h"
#include "util/u_inlines.h"
#include "util/u_format.h"
#include "util/u_prim.h"
#include "util/u_draw.h"
#include "util/u_upload_mgr.h"
#include "draw/draw_context.h"
#include "cso_cache/cso_context.h"
+#ifdef PIPE_OS_LINUX
+#include <sched.h>
+#define HAVE_SCHED_GETCPU 1
+#else
+#define sched_getcpu() 0
+#define HAVE_SCHED_GETCPU 0
+#endif
/**
* Set the restart index.
*/
static void
setup_primitive_restart(struct gl_context *ctx, struct pipe_draw_info *info)
{
if (ctx->Array._PrimitiveRestart) {
unsigned index_size = info->index_size;
@@ -115,20 +123,44 @@ prepare_draw(struct st_context *st, struct gl_context
*ctx)
if (unlikely(!st->bitmap.cache.empty))
st_flush_bitmap_cache(st);
st_invalidate_readpix_cache(st);
/* Validate state. */
if ((st->dirty | ctx->NewDriverState) & ST_PIPELINE_RENDER_STATE_MASK ||
st->gfx_shaders_may_be_dirty) {
st_validate_state(st, ST_PIPELINE_RENDER);
}
+
+ struct pipe_context *pipe = st->pipe;
+
+ /* Pin threads regularly to the same Zen CCX that the main thread is
+ * running on. The main thread can move between CCXs.
+ */
+ if (unlikely(HAVE_SCHED_GETCPU && /* Linux */
+ /* AMD Zen */
+ util_cpu_caps.nr_cpus != util_cpu_caps.cores_per_L3 &&
+ /* no glthread */
+ ctx->CurrentClientDispatch != ctx->MarshalExec &&
+ /* driver support */
+ pipe->set_context_param &&
+ /* do it occasionally */
+ ++st->pin_thread_counter % 512 == 0)) {
+ int cpu = sched_getcpu();
+ if (cpu >= 0) {
+ unsigned L3_cache = cpu / util_cpu_caps.cores_per_L3;
+
+ pipe->set_context_param(pipe,
+ PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE,
+ L3_cache);
+ }
+ }
}
/**
* This function gets plugged into the VBO module and is called when
* we have something to render.
* Basically, translate the information into the format expected by gallium.
*/
static void
st_draw_vbo(struct gl_context *ctx,
const struct _mesa_prim *prims,
--
2.17.1
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
