Merge pull request #8218 from huffman/feat/load-improvements

Speed up gamma correction in cube map processing

libraries/gpu/src/gpu/Texture.cpp

@@ -788,12 +788,9 @@ bool sphericalHarmonicsFromTexture(const gpu::Texture& cubeTexture, std::vector<
                 uint pixOffsetIndex = (x + y * width) * numComponents;
 
                 // get color from texture and map to range [0, 1]
-                glm::vec3 clr(float(data[pixOffsetIndex]) * UCHAR_TO_FLOAT,
-                            float(data[pixOffsetIndex+1]) * UCHAR_TO_FLOAT,
-                            float(data[pixOffsetIndex+2]) * UCHAR_TO_FLOAT);
-
-                // Gamma correct
-                clr = ColorUtils::sRGBToLinearVec3(clr);
+                glm::vec3 clr(ColorUtils::sRGB8ToLinearFloat(data[pixOffsetIndex]),
+                              ColorUtils::sRGB8ToLinearFloat(data[pixOffsetIndex + 1]),
+                              ColorUtils::sRGB8ToLinearFloat(data[pixOffsetIndex + 2]));
 
                 // scale color and add to previously accumulated coefficients
                 sphericalHarmonicsScale(shBuffB.data(), order,

libraries/shared/src/ColorUtils.cpp

@@ -0,0 +1,59 @@
+//
+//  ColorUtils.cpp
+//  libraries/shared/src
+//
+//  Created by Ryan Huffman on 8/8/16.
+//  Copyright 2016 High Fidelity, Inc.
+//
+//  Distributed under the Apache License, Version 2.0.
+//  See the accompanying file LICENSE or http://www.apache.org/licenses/LICENSE-2.0.html
+//
+
+#include "ColorUtils.h"
+
+// Generated from python script in repository at `tools/srgb_gen.py`
+const float srgbToLinearLookupTable[256] = {
+    0.0f, 0.000303526983549f, 0.000607053967098f, 0.000910580950647f, 0.0012141079342f, 0.00151763491774f, 0.00182116190129f,
+    0.00212468888484f, 0.00242821586839f, 0.00273174285194f, 0.00303526983549f, 0.0033465357639f, 0.00367650732405f,
+    0.0040247170185f, 0.00439144203741f, 0.00477695348069f, 0.00518151670234f, 0.0056053916242f, 0.00604883302286f,
+    0.00651209079259f, 0.00699541018727f, 0.00749903204323f, 0.00802319298538f, 0.00856812561807f, 0.00913405870222f,
+    0.00972121732024f, 0.0103298230296f, 0.0109600940065f, 0.0116122451797f, 0.0122864883569f, 0.0129830323422f,
+    0.0137020830473f, 0.0144438435961f, 0.0152085144229f, 0.0159962933655f, 0.0168073757529f, 0.0176419544884f,
+    0.0185002201284f, 0.0193823609569f, 0.0202885630567f, 0.021219010376f, 0.0221738847934f, 0.0231533661781f,
+    0.0241576324485f, 0.0251868596274f, 0.0262412218948f, 0.0273208916391f, 0.0284260395044f, 0.0295568344378f,
+    0.030713443733f, 0.031896033073f, 0.0331047665709f, 0.0343398068087f, 0.035601314875f, 0.0368894504011f,
+    0.0382043715953f, 0.0395462352767f, 0.0409151969069f, 0.0423114106208f, 0.043735029257f, 0.0451862043857f,
+    0.0466650863369f, 0.0481718242269f, 0.0497065659841f, 0.051269458374f, 0.0528606470232f, 0.0544802764424f,
+    0.0561284900496f, 0.0578054301911f, 0.059511238163f, 0.0612460542316f, 0.0630100176532f, 0.0648032666929f,
+    0.0666259386438f, 0.0684781698444f, 0.0703600956966f, 0.0722718506823f, 0.0742135683801f, 0.0761853814813f,
+    0.0781874218052f, 0.0802198203145f, 0.0822827071298f, 0.0843762115441f, 0.0865004620365f, 0.0886555862858f,
+    0.0908417111834f, 0.0930589628467f, 0.095307466631f, 0.0975873471419f, 0.0998987282471f, 0.102241733088f,
+    0.104616484091f, 0.107023102978f, 0.109461710778f, 0.111932427837f, 0.114435373827f, 0.116970667759f,
+    0.119538427988f, 0.12213877223f, 0.124771817561f, 0.127437680436f, 0.13013647669f, 0.132868321554f,
+    0.135633329655f, 0.138431615032f, 0.14126329114f, 0.144128470858f, 0.147027266498f, 0.149959789811f,
+    0.152926151996f, 0.155926463708f, 0.158960835061f, 0.162029375639f, 0.165132194502f, 0.16826940019f,
+    0.171441100733f, 0.174647403656f, 0.177888415984f, 0.18116424425f, 0.1844749945f, 0.187820772301f,
+    0.191201682741f, 0.194617830442f, 0.19806931956f, 0.201556253794f, 0.20507873639f, 0.208636870145f,
+    0.212230757414f, 0.215860500114f, 0.219526199729f, 0.223227957317f, 0.22696587351f, 0.230740048524f,
+    0.234550582161f, 0.238397573812f, 0.242281122466f, 0.246201326708f, 0.25015828473f, 0.254152094331f,
+    0.258182852922f, 0.26225065753f, 0.266355604803f, 0.270497791013f, 0.27467731206f, 0.278894263477f,
+    0.28314874043f, 0.287440837727f, 0.291770649818f, 0.296138270798f, 0.300543794416f, 0.30498731407f,
+    0.309468922818f, 0.313988713376f, 0.318546778125f, 0.323143209113f, 0.327778098057f, 0.332451536346f,
+    0.337163615048f, 0.341914424909f, 0.346704056355f, 0.3515325995f, 0.356400144146f, 0.361306779784f,
+    0.366252595599f, 0.371237680474f, 0.376262122991f, 0.381326011433f, 0.386429433787f, 0.39157247775f,
+    0.396755230726f, 0.401977779832f, 0.407240211902f, 0.412542613484f, 0.417885070848f, 0.423267669986f,
+    0.428690496614f, 0.434153636175f, 0.439657173841f, 0.445201194516f, 0.450785782838f, 0.45641102318f,
+    0.462076999654f, 0.467783796112f, 0.473531496148f, 0.479320183101f, 0.485149940056f, 0.491020849848f,
+    0.496932995061f, 0.502886458033f, 0.508881320855f, 0.514917665377f, 0.520995573204f, 0.527115125706f,
+    0.533276404011f, 0.539479489012f, 0.54572446137f, 0.552011401512f, 0.558340389634f, 0.564711505705f,
+    0.571124829465f, 0.57758044043f, 0.584078417891f, 0.590618840919f, 0.597201788364f, 0.603827338855f,
+    0.610495570808f, 0.61720656242f, 0.623960391675f, 0.630757136346f, 0.637596873994f, 0.644479681971f,
+    0.65140563742f, 0.658374817279f, 0.665387298282f, 0.672443156958f, 0.679542469633f, 0.686685312435f,
+    0.693871761292f, 0.701101891933f, 0.708375779892f, 0.715693500506f, 0.723055128922f, 0.73046074009f,
+    0.737910408773f, 0.74540420954f, 0.752942216776f, 0.760524504675f, 0.768151147248f, 0.775822218317f,
+    0.783537791526f, 0.791297940333f, 0.799102738014f, 0.806952257669f, 0.814846572216f, 0.822785754396f,
+    0.830769876775f, 0.838799011741f, 0.84687323151f, 0.854992608124f, 0.863157213454f, 0.871367119199f,
+    0.879622396888f, 0.887923117882f, 0.896269353374f, 0.904661174391f, 0.913098651793f, 0.921581856277f,
+    0.930110858375f, 0.938685728458f, 0.947306536733f, 0.955973353249f, 0.964686247894f, 0.973445290398f,
+    0.982250550333f, 0.991102097114f, 1.0f
+};

libraries/shared/src/ColorUtils.h

@@ -13,10 +13,13 @@
 #define hifi_ColorUtils_h
 
 #include <glm/glm.hpp>
-#include <SharedUtil.h>
+
+#include "SharedUtil.h"
 
 #include "DependencyManager.h"
 
+extern const float srgbToLinearLookupTable[256];
+
 class ColorUtils {
 public:
     inline static glm::vec3 toVec3(const xColor& color);
@@ -33,6 +36,7 @@ public:
     inline static glm::vec4 tosRGBVec4(const glm::vec4& srgb);
     
     inline static float sRGBToLinearFloat(const float& srgb);
+    inline static float sRGB8ToLinearFloat(const uint8_t srgb);
     inline static float tosRGBFloat(const float& linear);
 };
 
@@ -82,6 +86,9 @@ inline float ColorUtils::sRGBToLinearFloat(const float &srgb) {
     
     return linearValue;
 }
+inline float ColorUtils::sRGB8ToLinearFloat(const uint8_t srgb) {
+    return srgbToLinearLookupTable[srgb];
+}
 
 // This is based upon the conversions found in section 17.3.9 of the OpenGL 4.4 specification.
 // glm::pow(color, 1.0f/2.2f) is approximate, and will cause subtle differences when used with sRGB framebuffers.

tools/srgb_gen.py

@@ -0,0 +1,34 @@
+#
+#  srgb_gen.py
+#  tools/
+#
+#  Created by Ryan Huffman on 8/8/2016.
+#  Copyright 2016 High Fidelity, Inc.
+#
+#  Distributed under the Apache License, Version 2.0.
+#  See the accompanying file LICENSE or http://www.apache.org/licenses/LICENSE-2.0.html
+
+# Generates a lookup table for SRGB to Linear color transformations
+
+NUM_VALUES = 256
+srgb_to_linear = []
+
+# Calculate srgb to linear
+for i in range(NUM_VALUES):
+    s = float(i) / 255
+    if s < 0.04045:
+        l = s / 12.92
+    else:
+        l = ((s + 0.055) / 1.055) ** 2.4
+    srgb_to_linear.append(l)
+
+# Format and print
+data = "{\n    "
+for i, v in enumerate(srgb_to_linear):
+    data += str(v) + "f"
+    if i < NUM_VALUES - 1:
+        data += ", "
+    if i > 0 and i % 6 == 0:
+        data += "\n    "
+data += "\n}"
+print(data)