vc4: Use runtime CPU detection for whether NEON is available.

[mesa.git] / src / gallium / drivers / vc4 / vc4_tiling.h

diff --git a/src/gallium/drivers/vc4/vc4_tiling.h b/src/gallium/drivers/vc4/vc4_tiling.h
index b90bba7020036c4356afe5496e1a30786c461095..3168ec20a606e9f0d4cf22e7d920c95ca701f442 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_tiling.h
+++ b/src/gallium/drivers/vc4/vc4_tiling.h
@@ -24,11 +24,57 @@
 #ifndef VC4_TILING_H
 #define VC4_TILING_H
 
-uint32_t vc4_utile_width(int cpp) ATTRIBUTE_CONST;
-uint32_t vc4_utile_height(int cpp) ATTRIBUTE_CONST;
+#include <stdbool.h>
+#include <stdint.h>
+#include "util/macros.h"
+#include "util/u_cpu_detect.h"
+
+/** Return the width in pixels of a 64-byte microtile. */
+static inline uint32_t
+vc4_utile_width(int cpp)
+{
+        switch (cpp) {
+        case 1:
+        case 2:
+                return 8;
+        case 4:
+                return 4;
+        case 8:
+                return 2;
+        default:
+                unreachable("unknown cpp");
+        }
+}
+
+/** Return the height in pixels of a 64-byte microtile. */
+static inline uint32_t
+vc4_utile_height(int cpp)
+{
+        switch (cpp) {
+        case 1:
+                return 8;
+        case 2:
+        case 4:
+        case 8:
+                return 4;
+        default:
+                unreachable("unknown cpp");
+        }
+}
+
 bool vc4_size_is_lt(uint32_t width, uint32_t height, int cpp) ATTRIBUTE_CONST;
-void vc4_load_utile(void *dst, void *src, uint32_t dst_stride, uint32_t cpp);
-void vc4_store_utile(void *dst, void *src, uint32_t src_stride, uint32_t cpp);
+void vc4_load_lt_image_base(void *dst, uint32_t dst_stride,
+                            void *src, uint32_t src_stride,
+                            int cpp, const struct pipe_box *box);
+void vc4_store_lt_image_base(void *dst, uint32_t dst_stride,
+                             void *src, uint32_t src_stride,
+                             int cpp, const struct pipe_box *box);
+void vc4_load_lt_image_neon(void *dst, uint32_t dst_stride,
+                            void *src, uint32_t src_stride,
+                            int cpp, const struct pipe_box *box);
+void vc4_store_lt_image_neon(void *dst, uint32_t dst_stride,
+                             void *src, uint32_t src_stride,
+                             int cpp, const struct pipe_box *box);
 void vc4_load_tiled_image(void *dst, uint32_t dst_stride,
                           void *src, uint32_t src_stride,
                           uint8_t tiling_format, int cpp,
@@ -38,4 +84,32 @@ void vc4_store_tiled_image(void *dst, uint32_t dst_stride,
                            uint8_t tiling_format, int cpp,
                            const struct pipe_box *box);
 
+static inline void
+vc4_load_lt_image(void *dst, uint32_t dst_stride,
+                  void *src, uint32_t src_stride,
+                  int cpp, const struct pipe_box *box)
+{
+        if (util_cpu_caps.has_neon) {
+                vc4_load_lt_image_neon(dst, dst_stride, src, src_stride,
+                                       cpp, box);
+        } else {
+                vc4_load_lt_image_base(dst, dst_stride, src, src_stride,
+                                       cpp, box);
+        }
+}
+
+static inline void
+vc4_store_lt_image(void *dst, uint32_t dst_stride,
+                   void *src, uint32_t src_stride,
+                   int cpp, const struct pipe_box *box)
+{
+        if (util_cpu_caps.has_neon) {
+                vc4_store_lt_image_neon(dst, dst_stride, src, src_stride,
+                                        cpp, box);
+        } else {
+                vc4_store_lt_image_base(dst, dst_stride, src, src_stride,
+                                        cpp, box);
+        }
+}
+
 #endif /* VC4_TILING_H */