radeonsi: allocate GDS only once per process

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16885>

3 files changed, 23 insertions, 14 deletions

diff --git a/src/gallium/drivers/radeonsi/si_gfx_cs.c b/src/gallium/drivers/radeonsi/si_gfx_cs.c
index d19225b0fc9..56fb36f7cb7 100644
--- a/src/gallium/drivers/radeonsi/si_gfx_cs.c
+++ b/src/gallium/drivers/radeonsi/si_gfx_cs.c
@@ -211,10 +211,10 @@ static void si_begin_gfx_cs_debug(struct si_context *ctx)
 
 static void si_add_gds_to_buffer_list(struct si_context *sctx)
 {
-   if (sctx->gds) {
-      sctx->ws->cs_add_buffer(&sctx->gfx_cs, sctx->gds, RADEON_USAGE_READWRITE, 0);
-      if (sctx->gds_oa) {
-         sctx->ws->cs_add_buffer(&sctx->gfx_cs, sctx->gds_oa, RADEON_USAGE_READWRITE, 0);
+   if (sctx->screen->gds) {
+      sctx->ws->cs_add_buffer(&sctx->gfx_cs, sctx->screen->gds, RADEON_USAGE_READWRITE, 0);
+      if (sctx->screen->gds_oa) {
+         sctx->ws->cs_add_buffer(&sctx->gfx_cs, sctx->screen->gds_oa, RADEON_USAGE_READWRITE, 0);
       }
    }
 }
@@ -223,7 +223,7 @@ void si_allocate_gds(struct si_context *sctx)
 {
    struct radeon_winsys *ws = sctx->ws;
 
-   if (sctx->gds)
+   if (sctx->screen->gds && sctx->screen->gds_oa)
       return;
 
    assert(sctx->screen->use_ngg_streamout);
@@ -231,10 +231,15 @@ void si_allocate_gds(struct si_context *sctx)
    /* 4 streamout GDS counters.
     * We need 256B (64 dw) of GDS, otherwise streamout hangs.
     */
-   sctx->gds = ws->buffer_create(ws, 256, 4, RADEON_DOMAIN_GDS, RADEON_FLAG_DRIVER_INTERNAL);
-   sctx->gds_oa = ws->buffer_create(ws, 4, 1, RADEON_DOMAIN_OA, RADEON_FLAG_DRIVER_INTERNAL);
+   simple_mtx_lock(&sctx->screen->gds_mutex);
+   if (!sctx->screen->gds)
+      sctx->screen->gds = ws->buffer_create(ws, 256, 4, RADEON_DOMAIN_GDS, RADEON_FLAG_DRIVER_INTERNAL);
+   if (!sctx->screen->gds_oa)
+      sctx->screen->gds_oa = ws->buffer_create(ws, 4, 1, RADEON_DOMAIN_OA, RADEON_FLAG_DRIVER_INTERNAL);
+   simple_mtx_unlock(&sctx->screen->gds_mutex);
+
+   assert(sctx->screen->gds && sctx->screen->gds_oa);
 
-   assert(sctx->gds && sctx->gds_oa);
    si_add_gds_to_buffer_list(sctx);
 }
 
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 8122b61919e..98d93af749e 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -334,8 +334,6 @@ static void si_destroy_context(struct pipe_context *context)
    si_resource_reference(&sctx->eop_bug_scratch, NULL);
    si_resource_reference(&sctx->eop_bug_scratch_tmz, NULL);
    si_resource_reference(&sctx->shadowed_regs, NULL);
-   radeon_bo_reference(sctx->screen->ws, &sctx->gds, NULL);
-   radeon_bo_reference(sctx->screen->ws, &sctx->gds_oa, NULL);
 
    si_destroy_compiler(&sctx->compiler);
 
@@ -969,6 +967,10 @@ static void si_destroy_screen(struct pipe_screen *pscreen)
    si_gpu_load_kill_thread(sscreen);
 
    simple_mtx_destroy(&sscreen->gpu_load_mutex);
+   simple_mtx_destroy(&sscreen->gds_mutex);
+
+   radeon_bo_reference(sscreen->ws, &sscreen->gds, NULL);
+   radeon_bo_reference(sscreen->ws, &sscreen->gds_oa, NULL);
 
    slab_destroy_parent(&sscreen->pool_transfers);
 
@@ -1188,6 +1190,7 @@ static struct pipe_screen *radeonsi_screen_create_impl(struct radeon_winsys *ws,
    (void)simple_mtx_init(&sscreen->aux_context_lock, mtx_plain);
    (void)simple_mtx_init(&sscreen->async_compute_context_lock, mtx_plain);
    (void)simple_mtx_init(&sscreen->gpu_load_mutex, mtx_plain);
+   (void)simple_mtx_init(&sscreen->gds_mutex, mtx_plain);
 
    si_init_gs_info(sscreen);
    if (!si_init_shader_cache(sscreen)) {
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 658ee5d608e..092f1f74ee7 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -710,6 +710,11 @@ struct si_screen {
    struct util_vertex_state_cache vertex_state_cache;
 
    struct si_resource *attribute_ring;
+
+   /* NGG streamout. */
+   simple_mtx_t gds_mutex;
+   struct pb_buffer *gds;
+   struct pb_buffer *gds_oa;
 };
 
 struct si_sampler_view {
@@ -1020,10 +1025,6 @@ struct si_context {
    /* Current unaccounted memory usage. */
    uint32_t memory_usage_kb;
 
-   /* NGG streamout. */
-   struct pb_buffer *gds;
-   struct pb_buffer *gds_oa;
-
    /* Atoms (direct states). */
    union si_state_atoms atoms;
    unsigned dirty_atoms; /* mask */