ACES 2.0 Output Transform performance optimisation

src/OpenColorIO/CMakeLists.txt

@@ -220,6 +220,16 @@ if(OCIO_USE_SIMD AND (OCIO_ARCH_X86 OR OCIO_USE_SSE2NEON))
     set_property(SOURCE ops/lut3d/Lut3DOpCPU_AVX.cpp APPEND PROPERTY COMPILE_OPTIONS ${OCIO_AVX_ARGS})
     set_property(SOURCE ops/lut3d/Lut3DOpCPU_AVX2.cpp APPEND PROPERTY COMPILE_OPTIONS ${OCIO_AVX2_ARGS})
     set_property(SOURCE ops/lut3d/Lut3DOpCPU_AVX512.cpp APPEND PROPERTY COMPILE_OPTIONS ${OCIO_AVX512_ARGS})
+    set_property(SOURCE ops/fixedfunction/ACES2/Transform.cpp APPEND PROPERTY COMPILE_OPTIONS  ${OCIO_SSE2_ARGS} ${OCIO_AVX_ARGS})
+endif()
+
+# TODO: What to do for other compilers?
+if(USE_GCC)
+    set_property(SOURCE ops/fixedfunction/ACES2/Transform.cpp APPEND PROPERTY COMPILE_OPTIONS
+                -fno-math-errno -fno-signed-zeros -fno-trapping-math -fno-signaling-nans -ffinite-math-only -freciprocal-math -ftree-vectorize)
+endif()
+if(MSVC)
+    set_property(SOURCE ops/fixedfunction/ACES2/Transform.cpp APPEND PROPERTY COMPILE_OPTIONS /Qvec)
 endif()
 
 configure_file(CPUInfoConfig.h.in CPUInfoConfig.h)

src/OpenColorIO/ops/fixedfunction/ACES2/ColorLib.h

@@ -7,6 +7,7 @@
 #include "transforms/builtins/ColorMatrixHelpers.h"
 #include "MatrixLib.h"
 
+#include <cmath>
 
 namespace OCIO_NAMESPACE
 {

src/OpenColorIO/ops/fixedfunction/ACES2/Common.h

@@ -7,44 +7,98 @@
 #include "MatrixLib.h"
 #include "ColorLib.h"
 
+#include <cmath>
 
 namespace OCIO_NAMESPACE
 {
 
 namespace ACES2
 {
+constexpr float PI = 3.14159265358979f;
 
-constexpr int TABLE_SIZE = 360;
-constexpr int TABLE_ADDITION_ENTRIES = 2;
-constexpr int TABLE_TOTAL_SIZE = TABLE_SIZE + TABLE_ADDITION_ENTRIES;
-constexpr int GAMUT_TABLE_BASE_INDEX = 1;
+constexpr float hue_limit = 360.0f;
+//constexpr float hue_limit = 2.0f * PI;
+inline float _wrap_to_hue_limit(float y)
+{
+    if ( y < 0.f)
+    {
+        y = y + hue_limit;
+    }
+    return y;
+}
+
+inline float wrap_to_hue_limit(float hue)
+{
+    float y = std::fmod(hue, hue_limit);
+    return _wrap_to_hue_limit(y);
+}
+inline constexpr float to_degrees(const float v) { return v; }
+inline float from_degrees(const float v) { return wrap_to_hue_limit(v); }
+inline constexpr float to_radians(const float v) { return PI * v / 180.0f; };
+inline float _from_radians(const float v) { return _wrap_to_hue_limit(180.0f * v / PI); }; // v needs to be wrapped already
+inline float from_radians(const float v) { return wrap_to_hue_limit(180.0f * v / PI); };
+/*
+inline constexpr float to_degrees(const float v) { return 180.0f * v / PI; }
+inline float from_degrees(const float v) { return wrap_to_hue_limit(PI * v / 180.0f); }
+inline constexpr float to_radians(const float v) { return v; }
+inline float _from_radians(const float v) { return _wrap_to_hue_limit(v); };
+inline float from_radians(const float v) { return wrap_to_hue_limit(v); };
+*/
+
+struct TableBase
+{
+    static constexpr unsigned int _TABLE_ADDITION_ENTRIES = 2;
+    static constexpr unsigned int base_index = 1;
+    static constexpr unsigned int nominal_size = 360;
+    static constexpr unsigned int total_size = nominal_size + _TABLE_ADDITION_ENTRIES;
+
+    static constexpr unsigned int lower_wrap_index = 0;
+    static constexpr unsigned int upper_wrap_index = base_index + nominal_size;
+    static constexpr unsigned int first_nominal_index = base_index;
+    static constexpr unsigned int last_nominal_index = upper_wrap_index - 1;
+
+    inline float base_hue_for_position(unsigned int i_lo) const
+    {
+        if (hue_limit == float(nominal_size)) // TODO C++ 17 if constexpr
+            return float(i_lo);
+
+        const float result = i_lo * hue_limit / nominal_size;
+        return result;
+    }
+
+    inline unsigned int hue_position_in_uniform_table(float wrapped_hue) const 
+    {
+        if (hue_limit == float(nominal_size)) // TODO C++ 17 if constexpr
+            return static_cast<unsigned int>(wrapped_hue);
+        else
+            return static_cast<unsigned int>(wrapped_hue / hue_limit * float(nominal_size)); // TODO: can we use the 'lost' fraction for the lerps?
+    }
+
+   inline unsigned int nominal_hue_position_in_uniform_table(float wrapped_hue) const 
+    {
+        return first_nominal_index + hue_position_in_uniform_table(wrapped_hue);
+    }
+};
 
-struct Table3D
+struct Table3D : public TableBase, std::array<float[3], TableBase::total_size>
 {
-    static constexpr int base_index = GAMUT_TABLE_BASE_INDEX;
-    static constexpr int size = TABLE_SIZE;
-    static constexpr int total_size = TABLE_TOTAL_SIZE;
-    float table[TABLE_TOTAL_SIZE][3];
 };
 
-struct Table1D
+struct Table1D : public TableBase, std::array<float, TableBase::total_size>
 {
-    static constexpr int base_index = GAMUT_TABLE_BASE_INDEX;
-    static constexpr int size = TABLE_SIZE;
-    static constexpr int total_size = TABLE_TOTAL_SIZE;
-    float table[TABLE_TOTAL_SIZE];
 };
 
 struct JMhParams
 {
-    float F_L;
-    float z;
-    float A_w;
+    m33f MATRIX_RGB_to_CAM16_c;
+    m33f MATRIX_CAM16_c_to_RGB;
+    m33f MATRIX_cone_response_to_Aab;
+    m33f MATRIX_Aab_to_cone_response;
+    float F_L_n;    // F_L normalised
+    float cz;
+    float inv_cz;   // 1/cz
     float A_w_J;
-    f3 XYZ_w;
-    f3 D_RGB;
-    m33f MATRIX_RGB_to_CAM16;
-    m33f MATRIX_CAM16_to_RGB;
+    float inv_A_w_J; // 1/A_w_J
 };
 
 struct ToneScaleParams
@@ -57,41 +111,63 @@ struct ToneScaleParams
     float s_2;
     float u_2;
     float m_2;
+    float forward_limit;
+    float inverse_limit;
+    float log_peak;
 };
 
-struct ChromaCompressParams
+struct SharedCompressionParameters
 {
     float limit_J_max;
-    float model_gamma;
+    float model_gamma_inv;
+    Table1D reach_m_table;
+};
+
+struct ResolvedSharedCompressionParameters
+{
+    float limit_J_max;
+    float model_gamma_inv;
+    float reachMaxM;
+};
+
+struct ChromaCompressParams
+{
     float sat;
     float sat_thr;
     float compr;
-    Table1D reach_m_table;
     float chroma_compress_scale;
     static constexpr float cusp_mid_blend = 1.3f;
 };
 
+struct HueDependantGamutParams
+{
+    float gamma_bottom_inv;
+    f2 JMcusp;
+    float gamma_top_inv;
+    float focusJ;
+    float analytical_threshold;
+};
 struct GamutCompressParams
 {
-    float limit_J_max;
     float mid_J;
-    float model_gamma;
     float focus_dist;
-    float lower_hull_gamma;
-    Table1D reach_m_table;
+    float lower_hull_gamma_inv;
+    std::array<int, 2> hue_linearity_search_range;
+    Table1D hue_table;;
     Table3D gamut_cusp_table;
-    Table1D upper_hull_gamma_table;
 };
 
 // CAM
 constexpr float reference_luminance = 100.f;
 constexpr float L_A = 100.f;
 constexpr float Y_b = 20.f;
-constexpr float ac_resp = 1.f;
-constexpr float ra = 2.f * ac_resp;
-constexpr float ba = 0.05f + (2.f - ra);
 constexpr f3 surround = {0.9f, 0.59f, 0.9f}; // Dim surround
 
+constexpr float J_scale = 100.0f;
+constexpr float cam_nl_Y_reference = 100.0f;
+constexpr float cam_nl_offset = 0.2713f * cam_nl_Y_reference;
+constexpr float cam_nl_scale = 4.0f * cam_nl_Y_reference;
+
 // Chroma compression
 constexpr float chroma_compress = 2.4f;
 constexpr float chroma_compress_fact = 3.3f;
@@ -100,11 +176,11 @@ constexpr float chroma_expand_fact = 0.69f;
 constexpr float chroma_expand_thr = 0.5f;
 
 // Gamut compression
-constexpr float smooth_cusps = 0.12f;
+constexpr float smooth_cusps =  0.12f; // C++ 14 required for constexpr std::max(0.000001f, 0.12f);
 constexpr float smooth_m = 0.27f;
 constexpr float cusp_mid_blend = 1.3f;
 constexpr float focus_gain_blend = 0.3f;
-constexpr float focus_adjust_gain = 0.55f;
+constexpr float focus_adjust_gain_inv = 1.0f / 0.55f;
 constexpr float focus_distance = 1.35f;
 constexpr float focus_distance_scaling = 1.75f;
 constexpr float compression_threshold = 0.75f;
@@ -125,6 +201,11 @@ constexpr float gammaMaximum = 5.0f;
 constexpr float gammaSearchStep = 0.4f;
 constexpr float gammaAccuracy = 1e-5f;
 
+constexpr int cuspCornerCount = 6;
+constexpr int totalCornerCount = cuspCornerCount + 2;
+constexpr int max_sorted_corners = 2 * cuspCornerCount;
+constexpr float reach_cusp_tolerance = 1e-3f;
+constexpr float display_cusp_tolerance = 1e-7f;
 
 } // namespace ACES2
 

src/OpenColorIO/ops/fixedfunction/ACES2/MatrixLib.h

@@ -6,6 +6,7 @@
 
 #include "ops/matrix/MatrixOpData.h"
 
+#include <array>
 
 namespace OCIO_NAMESPACE
 {