#include "st_debug.h"
#include "st_draw.h"
#include "st_program.h"
+#include "st_util.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 "draw/draw_context.h"
#include "cso_cache/cso_context.h"
+#if defined(PIPE_OS_LINUX) && !defined(ANDROID)
+#include <sched.h>
+#define HAVE_SCHED_GETCPU 1
+#else
+#define sched_getcpu() 0
+#define HAVE_SCHED_GETCPU 0
+#endif
/**
* Set the restart index.
_mesa_primitive_restart_index(ctx, index_size);
/* Enable primitive restart only when the restart index can have an
- * effect. This is required for correctness in radeonsi VI support.
+ * effect. This is required for correctness in radeonsi GFX8 support.
* Other hardware may also benefit from taking a faster, non-restart path
* when possible.
*/
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.
+ *
+ * Try to keep this logic in sync with st_feedback_draw_vbo.
*/
static void
st_draw_vbo(struct gl_context *ctx,
unsigned i;
unsigned start = 0;
- /* The initial pushdown of the inputs array into the drivers */
- _mesa_set_drawing_arrays(ctx, st->draw_arrays.inputs);
- _vbo_update_inputs(ctx, &st->draw_arrays);
-
prepare_draw(st, ctx);
- if (st->vertex_array_out_of_memory)
- return;
-
/* Initialize pipe_draw_info. */
info.primitive_restart = false;
info.vertices_per_patch = ctx->TessCtrlProgram.patch_vertices;
info.indirect = NULL;
info.count_from_stream_output = NULL;
+ info.restart_index = 0;
if (ib) {
struct gl_buffer_object *bufobj = ib->obj;
struct pipe_draw_info info;
struct pipe_draw_indirect_info indirect;
- /* The initial pushdown of the inputs array into the drivers */
- _mesa_set_drawing_arrays(ctx, st->draw_arrays.inputs);
- _vbo_update_inputs(ctx, &st->draw_arrays);
-
assert(stride);
prepare_draw(st, ctx);
- if (st->vertex_array_out_of_memory)
- return;
-
memset(&indirect, 0, sizeof(indirect));
util_draw_init_info(&info);
info.start = 0; /* index offset / index size */
+ info.max_index = ~0u; /* so that u_vbuf can tell that it's unknown */
if (ib) {
struct gl_buffer_object *bufobj = ib->obj;
u_upload_unmap(st->pipe->stream_uploader);
- /* At the time of writing, cso_get_aux_vertex_buffer_slot() always returns
- * zero. If that ever changes we need to audit the calls to that function
- * and make sure the slot number is used consistently everywhere.
- */
- assert(cso_get_aux_vertex_buffer_slot(st->cso_context) == 0);
-
- cso_set_vertex_buffers(st->cso_context,
- cso_get_aux_vertex_buffer_slot(st->cso_context),
- 1, &vb);
+ cso_set_vertex_buffers(st->cso_context, 0, 1, &vb);
if (num_instances > 1) {
cso_draw_arrays_instanced(st->cso_context, PIPE_PRIM_TRIANGLE_FAN, 0, 4,