Cycles: perform CPU film reading in the kernel, to use AVX2 half conversion

Adds a bunch of CPU kernel function to process on row of pixels, and use those
instead of calling unoptimized implementations.

Fixes T92598

3 files changed, 112 insertions, 1 deletions

diff --git a/intern/cycles/kernel/device/cpu/kernel.h b/intern/cycles/kernel/device/cpu/kernel.h
index c49d7ca445a..6af8094b1ea 100644
--- a/intern/cycles/kernel/device/cpu/kernel.h
+++ b/intern/cycles/kernel/device/cpu/kernel.h
@@ -18,6 +18,7 @@
 
 /* CPU Kernel Interface */
 
+#include "util/half.h"
 #include "util/types.h"
 
 #include "kernel/types.h"
diff --git a/intern/cycles/kernel/device/cpu/kernel_arch.h b/intern/cycles/kernel/device/cpu/kernel_arch.h
index 432ac5e15a9..2f9a3f7c59d 100644
--- a/intern/cycles/kernel/device/cpu/kernel_arch.h
+++ b/intern/cycles/kernel/device/cpu/kernel_arch.h
@@ -52,6 +52,37 @@ KERNEL_INTEGRATOR_SHADE_FUNCTION(megakernel);
 #undef KERNEL_INTEGRATOR_INIT_FUNCTION
 #undef KERNEL_INTEGRATOR_SHADE_FUNCTION
 
+#define KERNEL_FILM_CONVERT_FUNCTION(name) \
+  void KERNEL_FUNCTION_FULL_NAME(film_convert_##name)(const KernelFilmConvert *kfilm_convert, \
+                                                      const float *buffer, \
+                                                      float *pixel, \
+                                                      const int width, \
+                                                      const int buffer_stride, \
+                                                      const int pixel_stride); \
+  void KERNEL_FUNCTION_FULL_NAME(film_convert_half_rgba_##name)( \
+      const KernelFilmConvert *kfilm_convert, \
+      const float *buffer, \
+      half4 *pixel, \
+      const int width, \
+      const int buffer_stride);
+
+KERNEL_FILM_CONVERT_FUNCTION(depth)
+KERNEL_FILM_CONVERT_FUNCTION(mist)
+KERNEL_FILM_CONVERT_FUNCTION(sample_count)
+KERNEL_FILM_CONVERT_FUNCTION(float)
+
+KERNEL_FILM_CONVERT_FUNCTION(light_path)
+KERNEL_FILM_CONVERT_FUNCTION(float3)
+
+KERNEL_FILM_CONVERT_FUNCTION(motion)
+KERNEL_FILM_CONVERT_FUNCTION(cryptomatte)
+KERNEL_FILM_CONVERT_FUNCTION(shadow_catcher)
+KERNEL_FILM_CONVERT_FUNCTION(shadow_catcher_matte_with_shadow)
+KERNEL_FILM_CONVERT_FUNCTION(combined)
+KERNEL_FILM_CONVERT_FUNCTION(float4)
+
+#undef KERNEL_FILM_CONVERT_FUNCTION
+
 /* --------------------------------------------------------------------
  * Shader evaluation.
  */
diff --git a/intern/cycles/kernel/device/cpu/kernel_arch_impl.h b/intern/cycles/kernel/device/cpu/kernel_arch_impl.h
index 6df5d7787fc..1ea5002e300 100644
--- a/intern/cycles/kernel/device/cpu/kernel_arch_impl.h
+++ b/intern/cycles/kernel/device/cpu/kernel_arch_impl.h
@@ -47,8 +47,8 @@
 #    include "kernel/integrator/megakernel.h"
 
 #    include "kernel/film/adaptive_sampling.h"
-#    include "kernel/film/read.h"
 #    include "kernel/film/id_passes.h"
+#    include "kernel/film/read.h"
 
 #    include "kernel/bake/bake.h"
 
@@ -232,6 +232,85 @@ void KERNEL_FUNCTION_FULL_NAME(cryptomatte_postprocess)(const KernelGlobalsCPU *
 #endif
 }
 
+/* --------------------------------------------------------------------
+ * Film Convert.
+ */
+
+#ifdef KERNEL_STUB
+
+#  define KERNEL_FILM_CONVERT_FUNCTION(name, is_float) \
+    void KERNEL_FUNCTION_FULL_NAME(film_convert_##name)(const KernelFilmConvert *kfilm_convert, \
+                                                        const float *buffer, \
+                                                        float *pixel, \
+                                                        const int width, \
+                                                        const int buffer_stride, \
+                                                        const int pixel_stride) \
+    { \
+      STUB_ASSERT(KERNEL_ARCH, film_convert_##name); \
+    } \
+    void KERNEL_FUNCTION_FULL_NAME(film_convert_half_rgba_##name)( \
+        const KernelFilmConvert *kfilm_convert, \
+        const float *buffer, \
+        half4 *pixel, \
+        const int width, \
+        const int buffer_stride) \
+    { \
+      STUB_ASSERT(KERNEL_ARCH, film_convert_##name); \
+    }
+
+#else
+
+#  define KERNEL_FILM_CONVERT_FUNCTION(name, is_float) \
+    void KERNEL_FUNCTION_FULL_NAME(film_convert_##name)(const KernelFilmConvert *kfilm_convert, \
+                                                        const float *buffer, \
+                                                        float *pixel, \
+                                                        const int width, \
+                                                        const int buffer_stride, \
+                                                        const int pixel_stride) \
+    { \
+      for (int i = 0; i < width; i++, buffer += buffer_stride, pixel += pixel_stride) { \
+        film_get_pass_pixel_##name(kfilm_convert, buffer, pixel); \
+      } \
+    } \
+    void KERNEL_FUNCTION_FULL_NAME(film_convert_half_rgba_##name)( \
+        const KernelFilmConvert *kfilm_convert, \
+        const float *buffer, \
+        half4 *pixel, \
+        const int width, \
+        const int buffer_stride) \
+    { \
+      for (int i = 0; i < width; i++, buffer += buffer_stride, pixel++) { \
+        float pixel_rgba[4] = {0.0f, 0.0f, 0.0f, 1.0f}; \
+        film_get_pass_pixel_##name(kfilm_convert, buffer, pixel_rgba); \
+        if (is_float) { \
+          pixel_rgba[1] = pixel_rgba[0]; \
+          pixel_rgba[2] = pixel_rgba[0]; \
+        } \
+        film_apply_pass_pixel_overlays_rgba(kfilm_convert, buffer, pixel_rgba); \
+        *pixel = float4_to_half4_display( \
+            make_float4(pixel_rgba[0], pixel_rgba[1], pixel_rgba[2], pixel_rgba[3])); \
+      } \
+    }
+
+#endif
+
+KERNEL_FILM_CONVERT_FUNCTION(depth, true)
+KERNEL_FILM_CONVERT_FUNCTION(mist, true)
+KERNEL_FILM_CONVERT_FUNCTION(sample_count, true)
+KERNEL_FILM_CONVERT_FUNCTION(float, true)
+
+KERNEL_FILM_CONVERT_FUNCTION(light_path, false)
+KERNEL_FILM_CONVERT_FUNCTION(float3, false)
+
+KERNEL_FILM_CONVERT_FUNCTION(motion, false)
+KERNEL_FILM_CONVERT_FUNCTION(cryptomatte, false)
+KERNEL_FILM_CONVERT_FUNCTION(shadow_catcher, false)
+KERNEL_FILM_CONVERT_FUNCTION(shadow_catcher_matte_with_shadow, false)
+KERNEL_FILM_CONVERT_FUNCTION(combined, false)
+KERNEL_FILM_CONVERT_FUNCTION(float4, false)
+
+#undef KERNEL_FILM_CONVERT_FUNCTION
+
 #undef KERNEL_INVOKE
 #undef DEFINE_INTEGRATOR_KERNEL
 #undef DEFINE_INTEGRATOR_SHADE_KERNEL