amdgpu_bo: make cache_entry a extensible array

Improves performance in SPECviewperf13 snx.
e.g.: test10 fps evolution: 270 -> 280.

"pahole radeonsi_dri.so -C amdgpu_winsys_bo" after:

struct amdgpu_winsys_bo {
	struct pb_buffer           base;                 /*     0    32 */
	union {
		struct {
			amdgpu_va_handle va_handle;      /*    32     8 */
			uint32_t   kms_handle;           /*    40     4 */
			int        map_count;            /*    44     4 */
		} real;                                  /*    32    16 */
		[...]
	} u;                                             /*    32    40 */
	/* --- cacheline 1 boundary (64 bytes) was 8 bytes ago --- */
	[...]
	struct pb_cache_entry      cache_entry[];        /*   144     0 */

	/* size: 144, cachelines: 3, members: 17 */
};

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7532>

2 files changed, 15 insertions, 11 deletions

diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
index d17606eef73..1d037b96874 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.c
@@ -230,8 +230,8 @@ static void amdgpu_bo_destroy_or_cache(struct pb_buffer *_buf)
 
    assert(bo->bo); /* slab buffers have a separate vtbl */
 
-   if (bo->u.real.use_reusable_pool)
-      pb_cache_add_buffer(&bo->u.real.cache_entry);
+   if (bo->use_reusable_pool)
+      pb_cache_add_buffer(bo->cache_entry);
    else
       amdgpu_bo_destroy(_buf);
 }
@@ -476,6 +476,7 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws,
    struct amdgpu_winsys_bo *bo;
    amdgpu_va_handle va_handle = NULL;
    int r;
+   bool init_pb_cache;
 
    /* VRAM or GTT must be specified, but not both at the same time. */
    assert(util_bitcount(initial_domain & (RADEON_DOMAIN_VRAM_GTT |
@@ -484,13 +485,17 @@ static struct amdgpu_winsys_bo *amdgpu_create_bo(struct amdgpu_winsys *ws,
 
    alignment = amdgpu_get_optimal_alignment(ws, size, alignment);
 
-   bo = CALLOC_STRUCT(amdgpu_winsys_bo);
+   init_pb_cache = heap >= 0 && (flags & RADEON_FLAG_NO_INTERPROCESS_SHARING);
+
+   bo = CALLOC(1, sizeof(struct amdgpu_winsys_bo) +
+                  init_pb_cache * sizeof(struct pb_cache_entry));
    if (!bo) {
       return NULL;
    }
 
-   if (heap >= 0) {
-      pb_cache_init_entry(&ws->bo_cache, &bo->u.real.cache_entry, &bo->base,
+   if (init_pb_cache) {
+      bo->use_reusable_pool = true;
+      pb_cache_init_entry(&ws->bo_cache, bo->cache_entry, &bo->base,
                           heap);
    }
    request.alloc_size = size;
@@ -1378,7 +1383,6 @@ no_slab:
          return NULL;
    }
 
-   bo->u.real.use_reusable_pool = use_reusable_pool;
    return &bo->base;
 }
 
@@ -1533,7 +1537,7 @@ static bool amdgpu_bo_get_handle(struct radeon_winsys *rws,
    if (!bo->bo)
       return false;
 
-   bo->u.real.use_reusable_pool = false;
+   bo->use_reusable_pool = false;
 
    switch (whandle->type) {
    case WINSYS_HANDLE_TYPE_SHARED:
diff --git a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h
index 2190794102c..d33c141e90f 100644
--- a/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h
+++ b/src/gallium/winsys/amdgpu/drm/amdgpu_bo.h
@@ -59,15 +59,12 @@ struct amdgpu_winsys_bo {
    struct pb_buffer base;
    union {
       struct {
-         struct pb_cache_entry cache_entry;
-
          amdgpu_va_handle va_handle;
-         int map_count;
-         bool use_reusable_pool;
 #if DEBUG
          struct list_head global_list_item;
 #endif
          uint32_t kms_handle;
+         int map_count;
       } real;
       struct {
          struct pb_slab_entry entry;
@@ -91,6 +88,7 @@ struct amdgpu_winsys_bo {
 
    amdgpu_bo_handle bo; /* NULL for slab entries and sparse buffers */
    bool is_user_ptr;
+   bool use_reusable_pool;
    uint32_t unique_id;
    uint64_t va;
    simple_mtx_t lock;
@@ -111,6 +109,8 @@ struct amdgpu_winsys_bo {
    unsigned num_fences;
    unsigned max_fences;
    struct pipe_fence_handle **fences;
+
+   struct pb_cache_entry cache_entry[];
 };
 
 struct amdgpu_slab {