mesa/st: rework thread scheduler handling + add dispatch tracking

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36715>

src/mesa/main/compute.c

@@ -296,6 +296,7 @@ prepare_compute(struct gl_context *ctx)
       _mesa_update_state(ctx);
 
    st_validate_state(st, ST_PIPELINE_COMPUTE_STATE_MASK);
+   st_context_add_work(st);
 }
 
 static ALWAYS_INLINE void

src/mesa/main/glthread.c

@@ -299,7 +299,7 @@ void _mesa_glthread_enable(struct gl_context *ctx)
    ctx->GLApi = ctx->MarshalExec;
 
    /* glthread takes over all thread scheduling. */
-   ctx->st->pin_thread_counter = ST_THREAD_SCHEDULER_DISABLED;
+   ctx->st->thread_scheduler_disabled = true;
 
    /* Update the dispatch only if the dispatch is current. */
    if (GET_DISPATCH() == ctx->Dispatch.Current) {
@@ -319,7 +319,7 @@ void _mesa_glthread_disable(struct gl_context *ctx)
 
    /* Re-enable thread scheduling in st/mesa when glthread is disabled. */
    if (ctx->pipe->set_context_param && util_thread_scheduler_enabled())
-      ctx->st->pin_thread_counter = 0;
+      ctx->st->thread_scheduler_disabled = false;
 
    /* Update the dispatch only if the dispatch is current. */
    if (GET_DISPATCH() == ctx->MarshalExec) {

src/mesa/state_tracker/st_context.c

@@ -717,7 +717,7 @@ st_create_context_priv(struct gl_context *ctx, struct pipe_context *pipe,
    st->shader_has_one_variant[MESA_SHADER_COMPUTE] = st->has_shareable_shaders;
 
    if (!st->pipe->set_context_param || !util_thread_scheduler_enabled())
-      st->pin_thread_counter = ST_THREAD_SCHEDULER_DISABLED;
+      st->thread_scheduler_disabled = true;
 
    st->bitmap.cache.empty = true;
 

src/mesa/state_tracker/st_context.h

@@ -39,7 +39,7 @@
 #include "vbo/vbo.h"
 #include "util/list.h"
 #include "cso_cache/cso_context.h"
-
+#include "util/u_cpu_detect.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -53,8 +53,6 @@ struct st_context;
 struct st_program;
 struct u_upload_mgr;
 
-#define ST_THREAD_SCHEDULER_DISABLED 0xffffffff
-
 struct st_bitmap_cache
 {
    /** Window pos to render the cached image */
@@ -141,10 +139,11 @@ struct st_context
    struct draw_stage *selection_stage;  /**< For GL_SELECT rendermode */
    struct draw_stage *rastpos_stage;  /**< For glRasterPos */
 
-   unsigned pin_thread_counter; /* for L3 thread pinning on AMD Zen */
+   unsigned work_counter; /* for L3 thread pinning on AMD Zen and resource pruning */
 
    GLboolean clamp_frag_color_in_shader;
    GLboolean clamp_vert_color_in_shader;
+   bool thread_scheduler_disabled;
    bool has_stencil_export; /**< can do shader stencil export? */
    bool has_time_elapsed;
    bool has_etc1;
@@ -517,6 +516,37 @@ st_api_destroy_drawable(struct pipe_frontend_drawable *drawable);
 void
 st_screen_destroy(struct pipe_frontend_screen *fscreen);
 
+static inline void
+st_context_apply_scheduler_policy(struct st_context *st)
+{
+   int cpu = util_get_current_cpu();
+   if (cpu >= 0) {
+      struct pipe_context *pipe = st->pipe;
+      uint16_t L3_cache = util_get_cpu_caps()->cpu_to_L3[cpu];
+
+      if (L3_cache != U_CPU_INVALID_L3) {
+         pipe->set_context_param(pipe,
+                                 PIPE_CONTEXT_PARAM_UPDATE_THREAD_SCHEDULING,
+                                 cpu);
+      }
+   }
+}
+
+static inline void
+st_context_add_work(struct st_context *st)
+{
+
+   /* Apply our thread scheduling policy for better multithreading
+    * performance.
+    */
+   if (unlikely(++st->work_counter % 512 == 0)) {
+      st->work_counter = 0;
+      if (!st->thread_scheduler_disabled)
+         st_context_apply_scheduler_policy(st);
+   }
+}
+
+
 #ifdef __cplusplus
 }
 #endif

src/mesa/state_tracker/st_draw.c

@@ -55,7 +55,6 @@
 
 #include "pipe/p_context.h"
 #include "pipe/p_defines.h"
-#include "util/u_cpu_detect.h"
 #include "util/u_inlines.h"
 #include "util/format/u_format.h"
 #include "util/u_prim.h"
@@ -87,27 +86,7 @@ st_prepare_draw(struct gl_context *ctx, uint64_t state_mask)
 
    /* Validate state. */
    st_validate_state(st, state_mask);
-
-   /* Apply our thread scheduling policy for better multithreading
-    * performance.
-    */
-   if (unlikely(st->pin_thread_counter != ST_THREAD_SCHEDULER_DISABLED &&
-                /* do it occasionally */
-                ++st->pin_thread_counter % 512 == 0)) {
-      st->pin_thread_counter = 0;
-
-      int cpu = util_get_current_cpu();
-      if (cpu >= 0) {
-         struct pipe_context *pipe = st->pipe;
-         uint16_t L3_cache = util_get_cpu_caps()->cpu_to_L3[cpu];
-
-         if (L3_cache != U_CPU_INVALID_L3) {
-            pipe->set_context_param(pipe,
-                                    PIPE_CONTEXT_PARAM_UPDATE_THREAD_SCHEDULING,
-                                    cpu);
-         }
-      }
-   }
+   st_context_add_work(st);
 }
 
 void