1 files changed, 23 insertions, 30 deletions

diff --git a/intern/cycles/kernel/kernel_write_passes.h b/intern/cycles/kernel/kernel_write_passes.h
index 410218d91d4..9d379495629 100644
--- a/intern/cycles/kernel/kernel_write_passes.h
+++ b/intern/cycles/kernel/kernel_write_passes.h
@@ -14,23 +14,25 @@
  * limitations under the License.
  */
 
-#if defined(__SPLIT_KERNEL__) || defined(__KERNEL_CUDA__)
+#pragma once
+
+#ifdef __KERNEL_GPU__
 #  define __ATOMIC_PASS_WRITE__
 #endif
 
 CCL_NAMESPACE_BEGIN
 
-ccl_device_inline void kernel_write_pass_float(ccl_global float *buffer, float value)
+ccl_device_inline void kernel_write_pass_float(ccl_global float *ccl_restrict buffer, float value)
 {
-  ccl_global float *buf = buffer;
 #ifdef __ATOMIC_PASS_WRITE__
-  atomic_add_and_fetch_float(buf, value);
+  atomic_add_and_fetch_float(buffer, value);
 #else
-  *buf += value;
+  *buffer += value;
 #endif
 }
 
-ccl_device_inline void kernel_write_pass_float3(ccl_global float *buffer, float3 value)
+ccl_device_inline void kernel_write_pass_float3(ccl_global float *ccl_restrict buffer,
+                                                float3 value)
 {
 #ifdef __ATOMIC_PASS_WRITE__
   ccl_global float *buf_x = buffer + 0;
@@ -41,12 +43,14 @@ ccl_device_inline void kernel_write_pass_float3(ccl_global float *buffer, float3
   atomic_add_and_fetch_float(buf_y, value.y);
   atomic_add_and_fetch_float(buf_z, value.z);
 #else
-  ccl_global float3 *buf = (ccl_global float3 *)buffer;
-  *buf += value;
+  buffer[0] += value.x;
+  buffer[1] += value.y;
+  buffer[2] += value.z;
 #endif
 }
 
-ccl_device_inline void kernel_write_pass_float4(ccl_global float *buffer, float4 value)
+ccl_device_inline void kernel_write_pass_float4(ccl_global float *ccl_restrict buffer,
+                                                float4 value)
 {
 #ifdef __ATOMIC_PASS_WRITE__
   ccl_global float *buf_x = buffer + 0;
@@ -59,37 +63,26 @@ ccl_device_inline void kernel_write_pass_float4(ccl_global float *buffer, float4
   atomic_add_and_fetch_float(buf_z, value.z);
   atomic_add_and_fetch_float(buf_w, value.w);
 #else
-  ccl_global float4 *buf = (ccl_global float4 *)buffer;
-  *buf += value;
+  buffer[0] += value.x;
+  buffer[1] += value.y;
+  buffer[2] += value.z;
+  buffer[3] += value.w;
 #endif
 }
 
-#ifdef __DENOISING_FEATURES__
-ccl_device_inline void kernel_write_pass_float_variance(ccl_global float *buffer, float value)
+ccl_device_inline float kernel_read_pass_float(ccl_global float *ccl_restrict buffer)
 {
-  kernel_write_pass_float(buffer, value);
-
-  /* The online one-pass variance update that's used for the megakernel can't easily be implemented
-   * with atomics, so for the split kernel the E[x^2] - 1/N * (E[x])^2 fallback is used. */
-  kernel_write_pass_float(buffer + 1, value * value);
+  return *buffer;
 }
 
-#  ifdef __ATOMIC_PASS_WRITE__
-#    define kernel_write_pass_float3_unaligned kernel_write_pass_float3
-#  else
-ccl_device_inline void kernel_write_pass_float3_unaligned(ccl_global float *buffer, float3 value)
+ccl_device_inline float3 kernel_read_pass_float3(ccl_global float *ccl_restrict buffer)
 {
-  buffer[0] += value.x;
-  buffer[1] += value.y;
-  buffer[2] += value.z;
+  return make_float3(buffer[0], buffer[1], buffer[2]);
 }
-#  endif
 
-ccl_device_inline void kernel_write_pass_float3_variance(ccl_global float *buffer, float3 value)
+ccl_device_inline float4 kernel_read_pass_float4(ccl_global float *ccl_restrict buffer)
 {
-  kernel_write_pass_float3_unaligned(buffer, value);
-  kernel_write_pass_float3_unaligned(buffer + 3, value * value);
+  return make_float4(buffer[0], buffer[1], buffer[2], buffer[3]);
 }
-#endif /* __DENOISING_FEATURES__ */
 
 CCL_NAMESPACE_END