Cycles: add packed_float3 type for storage

Introduce a packed_float3 type for smaller storage that is exactly 3
floats, instead of 4. For computation float3 is still used since it can
use SIMD instructions.

Ref T92212

Differential Revision: https://developer.blender.org/D13243

3 files changed, 63 insertions, 0 deletions

diff --git a/intern/cycles/util/defines.h b/intern/cycles/util/defines.h
index a778bef52b2..edc36b14745 100644
--- a/intern/cycles/util/defines.h
+++ b/intern/cycles/util/defines.h
@@ -44,6 +44,7 @@
 #  if defined(_WIN32) && !defined(FREE_WINDOWS)
 #    define ccl_device_inline static __forceinline
 #    define ccl_device_forceinline static __forceinline
+#    define ccl_device_inline_method __forceinline
 #    define ccl_align(...) __declspec(align(__VA_ARGS__))
 #    ifdef __KERNEL_64_BIT__
 #      define ccl_try_align(...) __declspec(align(__VA_ARGS__))
@@ -58,6 +59,7 @@
 #  else /* _WIN32 && !FREE_WINDOWS */
 #    define ccl_device_inline static inline __attribute__((always_inline))
 #    define ccl_device_forceinline static inline __attribute__((always_inline))
+#    define ccl_device_inline_method __attribute__((always_inline))
 #    define ccl_align(...) __attribute__((aligned(__VA_ARGS__)))
 #    ifndef FREE_WINDOWS64
 #      define __forceinline inline __attribute__((always_inline))
diff --git a/intern/cycles/util/math_float3.h b/intern/cycles/util/math_float3.h
index 81550c5d03c..031aac1b5d4 100644
--- a/intern/cycles/util/math_float3.h
+++ b/intern/cycles/util/math_float3.h
@@ -222,6 +222,32 @@ ccl_device_inline float3 operator/=(float3 &a, float f)
   return a = a * invf;
 }
 
+#if !(defined(__KERNEL_METAL__) || defined(__KERNEL_CUDA__))
+ccl_device_inline packed_float3 operator*=(packed_float3 &a, const float3 &b)
+{
+  a = float3(a) * b;
+  return a;
+}
+
+ccl_device_inline packed_float3 operator*=(packed_float3 &a, float f)
+{
+  a = float3(a) * f;
+  return a;
+}
+
+ccl_device_inline packed_float3 operator/=(packed_float3 &a, const float3 &b)
+{
+  a = float3(a) / b;
+  return a;
+}
+
+ccl_device_inline packed_float3 operator/=(packed_float3 &a, float f)
+{
+  a = float3(a) / f;
+  return a;
+}
+#endif
+
 ccl_device_inline bool operator==(const float3 &a, const float3 &b)
 {
 #ifdef __KERNEL_SSE__
diff --git a/intern/cycles/util/types_float3.h b/intern/cycles/util/types_float3.h
index f990367e7b8..fc0f35fa87f 100644
--- a/intern/cycles/util/types_float3.h
+++ b/intern/cycles/util/types_float3.h
@@ -55,6 +55,41 @@ ccl_device_inline float3 make_float3(float x, float y, float z);
 ccl_device_inline void print_float3(const char *label, const float3 &a);
 #endif /* __KERNEL_GPU__ */
 
+/* Smaller float3 for storage. For math operations this must be converted to float3, so that on the
+ * CPU SIMD instructions can be used. */
+#if defined(__KERNEL_METAL__)
+/* Metal has native packed_float3. */
+#elif defined(__KERNEL_CUDA__)
+/* CUDA float3 is already packed. */
+typedef float3 packed_float3;
+#else
+/* HIP float3 is not packed (https://github.com/ROCm-Developer-Tools/HIP/issues/706). */
+struct packed_float3 {
+  ccl_device_inline_method packed_float3(){};
+
+  ccl_device_inline_method packed_float3(const float3 &a) : x(a.x), y(a.y), z(a.z)
+  {
+  }
+
+  ccl_device_inline_method operator float3() const
+  {
+    return make_float3(x, y, z);
+  }
+
+  ccl_device_inline_method packed_float3 &operator=(const float3 &a)
+  {
+    x = a.x;
+    y = a.y;
+    z = a.z;
+    return *this;
+  }
+
+  float x, y, z;
+};
+#endif
+
+static_assert(sizeof(packed_float3) == 12);
+
 CCL_NAMESPACE_END
 
 #endif /* __UTIL_TYPES_FLOAT3_H__ */