util,gallium: put count in pipe_resource & sampler_view on its own cache line

This adds 60 bytes to both structures. It eliminates "False Sharing"
for atomic operations (see wikipedia).

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11618>

src/gallium/auxiliary/util/u_tests.c

@@ -48,7 +48,7 @@ util_create_texture2d(struct pipe_screen *screen, unsigned width,
                       unsigned height, enum pipe_format format,
                       unsigned num_samples)
 {
-   struct pipe_resource templ = {{0}};
+   struct pipe_resource templ = {0};
 
    templ.target = PIPE_TEXTURE_2D;
    templ.width0 = width;
@@ -698,7 +698,7 @@ test_texture_barrier(struct pipe_context *ctx, bool use_fbfetch,
              "ADD OUT[0], TEMP[0], IMM[0]\n"
              "END\n";
    } else {
-      struct pipe_sampler_view templ = {{0}};
+      struct pipe_sampler_view templ = {0};
       templ.format = cb->format;
       templ.target = cb->target;
       templ.swizzle_r = PIPE_SWIZZLE_X;

src/gallium/drivers/d3d12/d3d12_blit.cpp

@@ -469,7 +469,7 @@ create_staging_resource(struct d3d12_context *ctx,
                         unsigned mask)
 
 {
-   struct pipe_resource templ = {{0}};
+   struct pipe_resource templ = {};
    struct pipe_resource *staging_res;
    struct pipe_box copy_src;
 

src/gallium/drivers/d3d12/d3d12_surface.cpp

@@ -320,7 +320,7 @@ d3d12_surface_update_pre_draw(struct d3d12_surface *surface,
 
    if (mode == D3D12_SURFACE_CONVERSION_BGRA_UINT) {
       if (!surface->rgba_texture) {
-         struct pipe_resource templ = {{0}};
+         struct pipe_resource templ = {};
          struct pipe_resource *src = surface->base.texture;
 
          templ.format = PIPE_FORMAT_R8G8B8A8_UNORM;

src/gallium/drivers/r300/r300_context.c

@@ -441,8 +441,8 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
      * dummy texture there. */
     if (!r300->screen->caps.is_r500) {
         struct pipe_resource *tex;
-        struct pipe_resource rtempl = {{0}};
-        struct pipe_sampler_view vtempl = {{0}};
+        struct pipe_resource rtempl = {0};
+        struct pipe_sampler_view vtempl = {0};
 
         rtempl.target = PIPE_TEXTURE_2D;
         rtempl.format = PIPE_FORMAT_I8_UNORM;

src/gallium/include/pipe/p_compiler.h

@@ -155,6 +155,23 @@ typedef unsigned char boolean;
 
 #endif
 
+/**
+ * Declare a variable on its own cache line.
+ *
+ * This helps eliminate "False sharing" to make atomic operations
+ * on pipe_reference::count faster and/or access to adjacent fields faster.
+ *
+ * https://en.wikipedia.org/wiki/False_sharing
+ *
+ * CALLOC_STRUCT_CL or MALLOC_STRUCT_CL and FREE_CL should be used to allocate
+ * structures that contain this.
+ *
+ * NOTE: Don't use PIPE_ALIGN_VAR because it causes the whole structure to be
+ *       aligned, but we only want to align the field.
+ */
+#define EXCLUSIVE_CACHELINE(decl) \
+   union { char __cl_space[CACHE_LINE_SIZE]; \
+           decl; }
 
 #if defined(__GNUC__)
 

src/gallium/include/pipe/p_state.h

@@ -472,7 +472,9 @@ struct pipe_surface
  */
 struct pipe_sampler_view
 {
-   struct pipe_reference reference;
+   /* Put the refcount on its own cache line to prevent "False sharing". */
+   EXCLUSIVE_CACHELINE(struct pipe_reference reference);
+
    enum pipe_format format:15;      /**< typed PIPE_FORMAT_x */
    enum pipe_texture_target target:5; /**< PIPE_TEXTURE_x */
    unsigned swizzle_r:3;         /**< PIPE_SWIZZLE_x for red component */
@@ -543,7 +545,8 @@ struct pipe_box
  */
 struct pipe_resource
 {
-   struct pipe_reference reference;
+   /* Put the refcount on its own cache line to prevent "False sharing". */
+   EXCLUSIVE_CACHELINE(struct pipe_reference reference);
 
    unsigned width0; /**< Used by both buffers and textures. */
    uint16_t height0; /* Textures: The maximum height/depth/array_size is 16k. */

src/util/macros.h

@@ -477,4 +477,7 @@ typedef int lock_cap_t;
 
 #endif
 
+/* TODO: this could be different on non-x86 architectures. */
+#define CACHE_LINE_SIZE 64
+
 #endif /* UTIL_MACROS_H */

src/util/u_memory.h

@@ -90,6 +90,10 @@ mem_dup(const void *src, size_t size)
 #define Offset(TYPE, MEMBER) ((uintptr_t)&(((TYPE *)NULL)->MEMBER))
 
 
+/* Allocate a structure aligned to a cache line. (used to make atomic ops faster) */
+#define MALLOC_STRUCT_CL(T) (struct T *)align_malloc(sizeof(struct T), CACHE_LINE_SIZE)
+#define CALLOC_STRUCT_CL(T) (struct T *)align_calloc(sizeof(struct T), CACHE_LINE_SIZE)
+#define FREE_CL(ptr) align_free(ptr)
 
 #ifdef __cplusplus
 }