]> git.sesse.net Git - movit/blobdiff - effect_chain_test.cpp
Make it easier to turn on debug for the EffectChain test without having it be turned...
[movit] / effect_chain_test.cpp
index ca208da50cac7acaa54742357204e3c6b9fc6ec8..73f13c7957d7c8c4c0b66ea4cd50f4e464b6462b 100644 (file)
@@ -2,12 +2,29 @@
 //
 // Note that this also contains the tests for some of the simpler effects.
 
+#include <locale>
+#include <sstream>
+#include <string>
+
+#include <epoxy/gl.h>
+#include <assert.h>
+
+#include "effect.h"
 #include "effect_chain.h"
 #include "flat_input.h"
 #include "gtest/gtest.h"
+#include "init.h"
+#include "input.h"
 #include "mirror_effect.h"
-#include "opengl.h"
+#include "multiply_effect.h"
+#include "resize_effect.h"
+#include "resource_pool.h"
 #include "test_util.h"
+#include "util.h"
+
+using namespace std;
+
+namespace movit {
 
 TEST(EffectChainTest, EmptyChain) {
        float data[] = {
@@ -25,8 +42,8 @@ TEST(EffectChainTest, EmptyChain) {
 class IdentityEffect : public Effect {
 public:
        IdentityEffect() {}
-       virtual std::string effect_type_id() const { return "IdentityEffect"; }
-       std::string output_fragment_shader() { return read_file("identity.frag"); }
+       string effect_type_id() const override { return "IdentityEffect"; }
+       string output_fragment_shader() override { return read_file("identity.frag"); }
 };
 
 TEST(EffectChainTest, Identity) {
@@ -46,9 +63,10 @@ TEST(EffectChainTest, Identity) {
 class BouncingIdentityEffect : public Effect {
 public:
        BouncingIdentityEffect() {}
-       virtual std::string effect_type_id() const { return "IdentityEffect"; }
-       std::string output_fragment_shader() { return read_file("identity.frag"); }
-       bool needs_texture_bounce() const { return true; }
+       string effect_type_id() const override { return "IdentityEffect"; }
+       string output_fragment_shader() override { return read_file("identity.frag"); }
+       bool needs_texture_bounce() const override { return true; }
+       AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; }
 };
 
 TEST(EffectChainTest, TextureBouncePreservesIdentity) {
@@ -81,33 +99,73 @@ TEST(MirrorTest, BasicTest) {
        expect_equal(expected_data, out_data, 3, 2);
 }
 
+class WithAndWithoutComputeShaderTest : public testing::TestWithParam<string> {
+};
+INSTANTIATE_TEST_CASE_P(WithAndWithoutComputeShaderTest,
+                        WithAndWithoutComputeShaderTest,
+                        testing::Values("fragment", "compute"));
+
+// An effect that does nothing, but as a compute shader.
+class IdentityComputeEffect : public Effect {
+public:
+       IdentityComputeEffect() {}
+       virtual string effect_type_id() const { return "IdentityComputeEffect"; }
+       virtual bool is_compute_shader() const { return true; }
+       string output_fragment_shader() { return read_file("identity.comp"); }
+};
+
+TEST_P(WithAndWithoutComputeShaderTest, TopLeftOrigin) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       // Note that EffectChainTester assumes bottom-left origin, so by setting
+       // top-left, we will get flipped data back.
+       float expected_data[6] = {
+               0.75f, 1.0f, 1.0f,
+               0.0f, 0.25f, 0.3f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       tester.get_chain()->set_output_origin(OUTPUT_ORIGIN_TOP_LEFT);
+       if (GetParam() == "compute") {
+               tester.get_chain()->add_effect(new IdentityComputeEffect());
+       }
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 3, 2);
+}
+
 // A dummy effect that inverts its input.
 class InvertEffect : public Effect {
 public:
        InvertEffect() {}
-       virtual std::string effect_type_id() const { return "InvertEffect"; }
-       std::string output_fragment_shader() { return read_file("invert_effect.frag"); }
+       string effect_type_id() const override { return "InvertEffect"; }
+       string output_fragment_shader() override { return read_file("invert_effect.frag"); }
+
+       // A real invert would actually care about its alpha,
+       // but in this unit test, it only complicates things.
+       AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; }
 };
 
 // Like IdentityEffect, but rewrites itself out of the loop,
-// splicing in a InvertEffect instead. Also stores the new node,
-// so we later can check that there are gamma conversion effects
-// on both sides.
-class RewritingToInvertEffect : public Effect {
+// splicing in a different effect instead. Also stores the new node,
+// so we later can check whatever properties we'd like about the graph.
+template<class T>
+class RewritingEffect : public Effect {
 public:
-       RewritingToInvertEffect() {}
-       virtual std::string effect_type_id() const { return "RewritingToInvertEffect"; }
-       std::string output_fragment_shader() { EXPECT_TRUE(false); return read_file("identity.frag"); }
-       virtual void rewrite_graph(EffectChain *graph, Node *self) {
-               Node *invert_node = graph->add_node(new InvertEffect());
-               graph->replace_receiver(self, invert_node);
-               graph->replace_sender(self, invert_node);
-
+       RewritingEffect() : effect(new T()), replaced_node(nullptr) {}
+       string effect_type_id() const override { return "RewritingEffect[" + effect->effect_type_id() + "]"; }
+       string output_fragment_shader() override { EXPECT_TRUE(false); return read_file("identity.frag"); }
+       void rewrite_graph(EffectChain *graph, Node *self) override {
+               replaced_node = graph->add_node(effect);
+               graph->replace_receiver(self, replaced_node);
+               graph->replace_sender(self, replaced_node);
                self->disabled = true;
-               this->invert_node = invert_node;
        }
 
-       Node *invert_node;      
+       T *effect;
+       Node *replaced_node;
 };
 
 TEST(EffectChainTest, RewritingWorksAndGammaConversionsAreInserted) {
@@ -121,11 +179,11 @@ TEST(EffectChainTest, RewritingWorksAndGammaConversionsAreInserted) {
        };
        float out_data[6];
        EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_sRGB);
-       RewritingToInvertEffect *effect = new RewritingToInvertEffect();
+       RewritingEffect<InvertEffect> *effect = new RewritingEffect<InvertEffect>();
        tester.get_chain()->add_effect(effect);
        tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_sRGB);
 
-       Node *node = effect->invert_node;
+       Node *node = effect->replaced_node;
        ASSERT_EQ(1, node->incoming_links.size());
        ASSERT_EQ(1, node->outgoing_links.size());
        EXPECT_EQ("GammaExpansionEffect", node->incoming_links[0]->effect->effect_type_id());
@@ -134,25 +192,114 @@ TEST(EffectChainTest, RewritingWorksAndGammaConversionsAreInserted) {
        expect_equal(expected_data, out_data, 3, 2);
 }
 
-// Like RewritingToInvertEffect, but splicing in a MirrorEffect instead,
-// which does not need linear light.
-class RewritingToMirrorEffect : public Effect {
+TEST(EffectChainTest, RewritingWorksAndTexturesAreAskedForsRGB) {
+       unsigned char data[] = {
+                 0,   0,   0, 255,
+                64,  64,  64, 255,
+               128, 128, 128, 255,
+               255, 255, 255, 255,
+       };
+       float expected_data[] = {
+               1.0000f, 1.0000f, 1.0000f, 1.0000f,
+               0.9771f, 0.9771f, 0.9771f, 1.0000f,
+               0.8983f, 0.8983f, 0.8983f, 1.0000f,
+               0.0000f, 0.0000f, 0.0000f, 1.0000f
+       };
+       float out_data[4 * 4];
+       EffectChainTester tester(nullptr, 1, 4);
+       tester.add_input(data, FORMAT_RGBA_POSTMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_sRGB);
+       RewritingEffect<InvertEffect> *effect = new RewritingEffect<InvertEffect>();
+       tester.get_chain()->add_effect(effect);
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_sRGB);
+
+       Node *node = effect->replaced_node;
+       ASSERT_EQ(1, node->incoming_links.size());
+       ASSERT_EQ(1, node->outgoing_links.size());
+       EXPECT_EQ("FlatInput", node->incoming_links[0]->effect->effect_type_id());
+       EXPECT_EQ("GammaCompressionEffect", node->outgoing_links[0]->effect->effect_type_id());
+
+       expect_equal(expected_data, out_data, 4, 4);
+}
+
+TEST(EffectChainTest, RewritingWorksAndColorspaceConversionsAreInserted) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       float expected_data[6] = {
+               1.0f, 0.75f, 0.7f,
+               0.25f, 0.0f, 0.0f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_REC_601_525, GAMMA_LINEAR);
+       RewritingEffect<InvertEffect> *effect = new RewritingEffect<InvertEffect>();
+       tester.get_chain()->add_effect(effect);
+       tester.run(out_data, GL_RED, COLORSPACE_REC_601_525, GAMMA_LINEAR);
+
+       Node *node = effect->replaced_node;
+       ASSERT_EQ(1, node->incoming_links.size());
+       ASSERT_EQ(1, node->outgoing_links.size());
+       EXPECT_EQ("ColorspaceConversionEffect", node->incoming_links[0]->effect->effect_type_id());
+       EXPECT_EQ("ColorspaceConversionEffect", node->outgoing_links[0]->effect->effect_type_id());
+
+       expect_equal(expected_data, out_data, 3, 2);
+}
+
+// A fake input that can change its output colorspace and gamma between instantiation
+// and finalize.
+class UnknownColorspaceInput : public FlatInput {
 public:
-       RewritingToMirrorEffect() {}
-       virtual std::string effect_type_id() const { return "RewritingToMirrorEffect"; }
-       std::string output_fragment_shader() { EXPECT_TRUE(false); return read_file("identity.frag"); }
-       virtual void rewrite_graph(EffectChain *graph, Node *self) {
-               Node *mirror_node = graph->add_node(new MirrorEffect());
-               graph->replace_receiver(self, mirror_node);
-               graph->replace_sender(self, mirror_node);
+       UnknownColorspaceInput(ImageFormat format, MovitPixelFormat pixel_format, GLenum type, unsigned width, unsigned height)
+           : FlatInput(format, pixel_format, type, width, height),
+             overridden_color_space(format.color_space),
+             overridden_gamma_curve(format.gamma_curve) {}
+       string effect_type_id() const override { return "UnknownColorspaceInput"; }
 
-               self->disabled = true;
-               this->mirror_node = mirror_node;
+       void set_color_space(Colorspace colorspace) {
+               overridden_color_space = colorspace;
        }
+       void set_gamma_curve(GammaCurve gamma_curve) {
+               overridden_gamma_curve = gamma_curve;
+       }
+       Colorspace get_color_space() const override { return overridden_color_space; }
+       GammaCurve get_gamma_curve() const override { return overridden_gamma_curve; }
 
-       Node *mirror_node;
+private:
+       Colorspace overridden_color_space;
+       GammaCurve overridden_gamma_curve;
 };
 
+TEST(EffectChainTest, HandlesInputChangingColorspace) {
+       const int size = 4;
+
+       float data[size] = {
+               0.0,
+               0.5,
+               0.7,
+               1.0,
+       };
+       float out_data[size];
+
+       EffectChainTester tester(nullptr, 4, 1, FORMAT_GRAYSCALE);
+
+       // First say that we have sRGB, linear input.
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+
+       UnknownColorspaceInput *input = new UnknownColorspaceInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 4, 1);
+       input->set_pixel_data(data);
+       tester.get_chain()->add_input(input);
+
+       // Now we change to Rec. 601 input.
+       input->set_color_space(COLORSPACE_REC_601_625);
+       input->set_gamma_curve(GAMMA_REC_601);
+
+       // Now ask for Rec. 601 output. Thus, our chain should now be a no-op.
+       tester.run(out_data, GL_RED, COLORSPACE_REC_601_625, GAMMA_REC_601);
+       expect_equal(data, out_data, 4, 1);
+}
+
 TEST(EffectChainTest, NoGammaConversionsWhenLinearLightNotNeeded) {
        float data[] = {
                0.0f, 0.25f, 0.3f,
@@ -164,11 +311,34 @@ TEST(EffectChainTest, NoGammaConversionsWhenLinearLightNotNeeded) {
        };
        float out_data[6];
        EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_sRGB);
-       RewritingToMirrorEffect *effect = new RewritingToMirrorEffect();
+       RewritingEffect<MirrorEffect> *effect = new RewritingEffect<MirrorEffect>();
        tester.get_chain()->add_effect(effect);
        tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_sRGB);
 
-       Node *node = effect->mirror_node;
+       Node *node = effect->replaced_node;
+       ASSERT_EQ(1, node->incoming_links.size());
+       EXPECT_EQ(0, node->outgoing_links.size());
+       EXPECT_EQ("FlatInput", node->incoming_links[0]->effect->effect_type_id());
+
+       expect_equal(expected_data, out_data, 3, 2);
+}
+
+TEST(EffectChainTest, NoColorspaceConversionsWhensRGBPrimariesNotNeeded) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       float expected_data[6] = {
+               0.3f, 0.25f, 0.0f,
+               1.0f, 1.0f, 0.75f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_REC_601_525, GAMMA_LINEAR);
+       RewritingEffect<MirrorEffect> *effect = new RewritingEffect<MirrorEffect>();
+       tester.get_chain()->add_effect(effect);
+       tester.run(out_data, GL_RED, COLORSPACE_REC_601_525, GAMMA_LINEAR);
+
+       Node *node = effect->replaced_node;
        ASSERT_EQ(1, node->incoming_links.size());
        EXPECT_EQ(0, node->outgoing_links.size());
        EXPECT_EQ("FlatInput", node->incoming_links[0]->effect->effect_type_id());
@@ -191,6 +361,23 @@ TEST(EffectChainTest, IdentityThroughsRGBConversions) {
        expect_equal(data, out_data, 256, 1);
 }
 
+// Same, but uses the forward sRGB table from the GPU.
+TEST(EffectChainTest, IdentityThroughGPUsRGBConversions) {
+       unsigned char data[256];
+       float expected_data[256];
+       for (unsigned i = 0; i < 256; ++i) {
+               data[i] = i;
+               expected_data[i] = i / 255.0;
+       };
+       float out_data[256];
+       EffectChainTester tester(nullptr, 256, 1);
+       tester.add_input(data, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_sRGB);
+       tester.get_chain()->add_effect(new IdentityEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_sRGB);
+
+       expect_equal(expected_data, out_data, 256, 1);
+}
+
 // Same, for the Rec. 601/709 gamma curve.
 TEST(EffectChainTest, IdentityThroughRec709) {
        float data[256];
@@ -205,23 +392,201 @@ TEST(EffectChainTest, IdentityThroughRec709) {
        expect_equal(data, out_data, 256, 1);
 }
 
+// The identity effect needs premultiplied alpha, and thus will get conversions on both sides.
+TEST(EffectChainTest, IdentityThroughAlphaConversions) {
+       const int size = 3;
+       float data[4 * size] = {
+               0.8f, 0.0f, 0.0f, 0.5f,
+               0.0f, 0.2f, 0.2f, 0.3f,
+               0.1f, 0.0f, 1.0f, 1.0f,
+       };
+       float out_data[4 * size];
+       EffectChainTester tester(data, size, 1, FORMAT_RGBA_POSTMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
+       tester.get_chain()->add_effect(new IdentityEffect());
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, 4, size);
+}
+
+TEST(EffectChainTest, NoAlphaConversionsWhenPremultipliedAlphaNotNeeded) {
+       const int size = 3;
+       float data[4 * size] = {
+               0.8f, 0.0f, 0.0f, 0.5f,
+               0.0f, 0.2f, 0.2f, 0.3f,
+               0.1f, 0.0f, 1.0f, 1.0f,
+       };
+       float expected_data[4 * size] = {
+               0.1f, 0.0f, 1.0f, 1.0f,
+               0.0f, 0.2f, 0.2f, 0.3f,
+               0.8f, 0.0f, 0.0f, 0.5f,
+       };
+       float out_data[4 * size];
+       EffectChainTester tester(data, size, 1, FORMAT_RGBA_POSTMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
+       RewritingEffect<MirrorEffect> *effect = new RewritingEffect<MirrorEffect>();
+       tester.get_chain()->add_effect(effect);
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       Node *node = effect->replaced_node;
+       ASSERT_EQ(1, node->incoming_links.size());
+       EXPECT_EQ(0, node->outgoing_links.size());
+       EXPECT_EQ("FlatInput", node->incoming_links[0]->effect->effect_type_id());
+
+       expect_equal(expected_data, out_data, 4, size);
+}
+
+// An input that outputs only blue, which has blank alpha.
+class BlueInput : public Input {
+public:
+       BlueInput() { register_int("needs_mipmaps", &needs_mipmaps); }
+       string effect_type_id() const override { return "IdentityEffect"; }
+       string output_fragment_shader() override { return read_file("blue.frag"); }
+       AlphaHandling alpha_handling() const override { return OUTPUT_BLANK_ALPHA; }
+       bool can_output_linear_gamma() const override { return true; }
+       unsigned get_width() const override { return 1; }
+       unsigned get_height() const override { return 1; }
+       Colorspace get_color_space() const override { return COLORSPACE_sRGB; }
+       GammaCurve get_gamma_curve() const override { return GAMMA_LINEAR; }
+
+private:
+       int needs_mipmaps;
+};
+
+// Like RewritingEffect<InvertEffect>, but splicing in a BlueInput instead,
+// which outputs blank alpha.
+class RewritingToBlueInput : public Input {
+public:
+       RewritingToBlueInput() : blue_node(nullptr) { register_int("needs_mipmaps", &needs_mipmaps); }
+       string effect_type_id() const override { return "RewritingToBlueInput"; }
+       string output_fragment_shader() override { EXPECT_TRUE(false); return read_file("identity.frag"); }
+       void rewrite_graph(EffectChain *graph, Node *self) override {
+               Node *blue_node = graph->add_node(new BlueInput());
+               graph->replace_receiver(self, blue_node);
+               graph->replace_sender(self, blue_node);
+
+               self->disabled = true;
+               this->blue_node = blue_node;
+       }
+
+       // Dummy values that we need to implement because we inherit from Input.
+       // Same as BlueInput.
+       AlphaHandling alpha_handling() const override { return OUTPUT_BLANK_ALPHA; }
+       bool can_output_linear_gamma() const override { return true; }
+       unsigned get_width() const override { return 1; }
+       unsigned get_height() const override { return 1; }
+       Colorspace get_color_space() const override { return COLORSPACE_sRGB; }
+       GammaCurve get_gamma_curve() const override { return GAMMA_LINEAR; }
+
+       Node *blue_node;
+
+private:
+       int needs_mipmaps;
+};
+
+TEST(EffectChainTest, NoAlphaConversionsWithBlankAlpha) {
+       const int size = 3;
+       float data[4 * size] = {
+               0.0f, 0.0f, 1.0f, 1.0f,
+               0.0f, 0.0f, 1.0f, 1.0f,
+               0.0f, 0.0f, 1.0f, 1.0f,
+       };
+       float out_data[4 * size];
+       EffectChainTester tester(nullptr, size, 1);
+       RewritingToBlueInput *input = new RewritingToBlueInput();
+       tester.get_chain()->add_input(input);
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_PREMULTIPLIED);
+
+       Node *node = input->blue_node;
+       EXPECT_EQ(0, node->incoming_links.size());
+       EXPECT_EQ(0, node->outgoing_links.size());
+
+       expect_equal(data, out_data, 4, size);
+}
+
+// An effect that does nothing, and specifies that it preserves blank alpha.
+class BlankAlphaPreservingEffect : public Effect {
+public:
+       BlankAlphaPreservingEffect() {}
+       string effect_type_id() const override { return "BlankAlphaPreservingEffect"; }
+       string output_fragment_shader() override { return read_file("identity.frag"); }
+       AlphaHandling alpha_handling() const override { return INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK; }
+};
+
+TEST(EffectChainTest, NoAlphaConversionsWithBlankAlphaPreservingEffect) {
+       const int size = 3;
+       float data[4 * size] = {
+               0.0f, 0.0f, 1.0f, 1.0f,
+               0.0f, 0.0f, 1.0f, 1.0f,
+               0.0f, 0.0f, 1.0f, 1.0f,
+       };
+       float out_data[4 * size];
+       EffectChainTester tester(nullptr, size, 1);
+       tester.get_chain()->add_input(new BlueInput());
+       tester.get_chain()->add_effect(new BlankAlphaPreservingEffect());
+       RewritingEffect<MirrorEffect> *effect = new RewritingEffect<MirrorEffect>();
+       tester.get_chain()->add_effect(effect);
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
+
+       Node *node = effect->replaced_node;
+       EXPECT_EQ(1, node->incoming_links.size());
+       EXPECT_EQ(0, node->outgoing_links.size());
+
+       expect_equal(data, out_data, 4, size);
+}
+
+// This is the counter-test to NoAlphaConversionsWithBlankAlphaPreservingEffect;
+// just to be sure that with a normal INPUT_AND_OUTPUT_PREMULTIPLIED_ALPHA effect,
+// an alpha conversion _should_ be inserted at the very end. (There is some overlap
+// with other tests.)
+TEST(EffectChainTest, AlphaConversionsWithNonBlankAlphaPreservingEffect) {
+       const int size = 3;
+       float data[4 * size] = {
+               0.0f, 0.0f, 1.0f, 1.0f,
+               0.0f, 0.0f, 1.0f, 1.0f,
+               0.0f, 0.0f, 1.0f, 1.0f,
+       };
+       float out_data[4 * size];
+       EffectChainTester tester(nullptr, size, 1);
+       tester.get_chain()->add_input(new BlueInput());
+       tester.get_chain()->add_effect(new IdentityEffect());  // Not BlankAlphaPreservingEffect.
+       RewritingEffect<MirrorEffect> *effect = new RewritingEffect<MirrorEffect>();
+       tester.get_chain()->add_effect(effect);
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
+
+       Node *node = effect->replaced_node;
+       EXPECT_EQ(1, node->incoming_links.size());
+       EXPECT_EQ(1, node->outgoing_links.size());
+       EXPECT_EQ("AlphaDivisionEffect", node->outgoing_links[0]->effect->effect_type_id());
+
+       expect_equal(data, out_data, 4, size);
+}
+
 // Effectively scales down its input linearly by 4x (and repeating it),
 // which is not attainable without mipmaps.
 class MipmapNeedingEffect : public Effect {
 public:
        MipmapNeedingEffect() {}
-       virtual bool needs_mipmaps() const { return true; }
-       virtual std::string effect_type_id() const { return "MipmapNeedingEffect"; }
-       std::string output_fragment_shader() { return read_file("mipmap_needing_effect.frag"); }
-       void set_gl_state(GLuint glsl_program_num, const std::string& prefix, unsigned *sampler_num)
+       bool needs_mipmaps() const override { return true; }
+
+       // To be allowed to mess with the sampler state.
+       bool needs_texture_bounce() const override { return true; }
+
+       string effect_type_id() const override { return "MipmapNeedingEffect"; }
+       string output_fragment_shader() override { return read_file("mipmap_needing_effect.frag"); }
+       void inform_added(EffectChain *chain) override { this->chain = chain; }
+
+       void set_gl_state(GLuint glsl_program_num, const string& prefix, unsigned *sampler_num) override
        {
-               glActiveTexture(GL_TEXTURE0);
+               Node *self = chain->find_node_for_effect(this);
+               glActiveTexture(chain->get_input_sampler(self, 0));
                check_error();
                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
                check_error();
                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
                check_error();
        }
+
+private:
+       EffectChain *chain;
 };
 
 TEST(EffectChainTest, MipmapGenerationWorks) {
@@ -267,10 +632,1007 @@ TEST(EffectChainTest, MipmapGenerationWorks) {
                1.0f,     1.0f,     1.0f,     1.0f,
                0.25f,    0.25f,    0.25f,    0.25f,
        };
-       float out_data[16 * 4];
+       float out_data[4 * 16];
        EffectChainTester tester(data, 4, 16, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
        tester.get_chain()->add_effect(new MipmapNeedingEffect());
        tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
 
        expect_equal(expected_data, out_data, 4, 16);
 }
+
+class NonMipmapCapableInput : public FlatInput {
+public:
+       NonMipmapCapableInput(ImageFormat format, MovitPixelFormat pixel_format, GLenum type, unsigned width, unsigned height)
+               : FlatInput(format, pixel_format, type, width, height) {}
+
+       bool can_supply_mipmaps() const override { return false; }
+       bool set_int(const std::string& key, int value) override {
+               if (key == "needs_mipmaps") {
+                       assert(value == 0);
+               }
+               return FlatInput::set_int(key, value);
+       }
+};
+
+// The same test as MipmapGenerationWorks, but with an input that refuses
+// to supply mipmaps.
+TEST(EffectChainTest, MipmapsWithNonMipmapCapableInput) {
+       float data[] = {  // In 4x4 blocks.
+               1.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 1.0f,
+
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.5f, 0.0f, 0.0f,
+               0.0f, 0.0f, 1.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+
+               1.0f, 1.0f, 1.0f, 1.0f,
+               1.0f, 1.0f, 1.0f, 1.0f,
+               1.0f, 1.0f, 1.0f, 1.0f,
+               1.0f, 1.0f, 1.0f, 1.0f,
+
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 1.0f, 1.0f, 0.0f,
+               0.0f, 1.0f, 1.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+       };
+       float expected_data[] = {  // Repeated four times each way.
+               0.125f,   0.125f,   0.125f,   0.125f,
+               0.09375f, 0.09375f, 0.09375f, 0.09375f,
+               1.0f,     1.0f,     1.0f,     1.0f,
+               0.25f,    0.25f,    0.25f,    0.25f,
+
+               0.125f,   0.125f,   0.125f,   0.125f,
+               0.09375f, 0.09375f, 0.09375f, 0.09375f,
+               1.0f,     1.0f,     1.0f,     1.0f,
+               0.25f,    0.25f,    0.25f,    0.25f,
+
+               0.125f,   0.125f,   0.125f,   0.125f,
+               0.09375f, 0.09375f, 0.09375f, 0.09375f,
+               1.0f,     1.0f,     1.0f,     1.0f,
+               0.25f,    0.25f,    0.25f,    0.25f,
+
+               0.125f,   0.125f,   0.125f,   0.125f,
+               0.09375f, 0.09375f, 0.09375f, 0.09375f,
+               1.0f,     1.0f,     1.0f,     1.0f,
+               0.25f,    0.25f,    0.25f,    0.25f,
+       };
+       float out_data[4 * 16];
+       EffectChainTester tester(nullptr, 4, 16, FORMAT_GRAYSCALE);
+
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+
+       NonMipmapCapableInput *input = new NonMipmapCapableInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 4, 16);
+       input->set_pixel_data(data);
+       tester.get_chain()->add_input(input);
+       tester.get_chain()->add_effect(new MipmapNeedingEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 4, 16);
+}
+
+TEST(EffectChainTest, ResizeDownByFourThenUpByFour) {
+       float data[] = {  // In 4x4 blocks.
+               1.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 1.0f,
+
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.5f, 0.0f, 0.0f,
+               0.0f, 0.0f, 1.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+
+               1.0f, 1.0f, 1.0f, 1.0f,
+               1.0f, 1.0f, 1.0f, 1.0f,
+               1.0f, 1.0f, 1.0f, 1.0f,
+               1.0f, 1.0f, 1.0f, 1.0f,
+
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 1.0f, 1.0f, 0.0f,
+               0.0f, 1.0f, 1.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+       };
+       float expected_data[] = {  // Repeated four times horizontaly, interpolated vertically.
+               0.1250f, 0.1250f, 0.1250f, 0.1250f,
+               0.1250f, 0.1250f, 0.1250f, 0.1250f,
+               0.1211f, 0.1211f, 0.1211f, 0.1211f,
+               0.1133f, 0.1133f, 0.1133f, 0.1133f,
+               0.1055f, 0.1055f, 0.1055f, 0.1055f,
+               0.0977f, 0.0977f, 0.0977f, 0.0977f,
+               0.2070f, 0.2070f, 0.2070f, 0.2070f,
+               0.4336f, 0.4336f, 0.4336f, 0.4336f,
+               0.6602f, 0.6602f, 0.6602f, 0.6602f,
+               0.8867f, 0.8867f, 0.8867f, 0.8867f,
+               0.9062f, 0.9062f, 0.9062f, 0.9062f,
+               0.7188f, 0.7188f, 0.7188f, 0.7188f,
+               0.5312f, 0.5312f, 0.5312f, 0.5312f,
+               0.3438f, 0.3438f, 0.3438f, 0.3438f,
+               0.2500f, 0.2500f, 0.2500f, 0.2500f,
+               0.2500f, 0.2500f, 0.2500f, 0.2500f,
+       };
+       float out_data[4 * 16];
+
+       ResizeEffect *downscale = new ResizeEffect();
+       ASSERT_TRUE(downscale->set_int("width", 1));
+       ASSERT_TRUE(downscale->set_int("height", 4));
+
+       ResizeEffect *upscale = new ResizeEffect();
+       ASSERT_TRUE(upscale->set_int("width", 4));
+       ASSERT_TRUE(upscale->set_int("height", 16));
+
+       EffectChainTester tester(data, 4, 16, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       tester.get_chain()->add_effect(downscale);
+       tester.get_chain()->add_effect(upscale);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 4, 16);
+}
+
+// An effect that adds its two inputs together. Used below.
+class AddEffect : public Effect {
+public:
+       AddEffect() {}
+       string effect_type_id() const override { return "AddEffect"; }
+       string output_fragment_shader() override { return read_file("add.frag"); }
+       unsigned num_inputs() const override { return 2; }
+       AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; }
+};
+
+// Constructs the graph
+//
+//             FlatInput               |
+//            /         \              |
+//  MultiplyEffect  MultiplyEffect     |
+//            \         /              |
+//             AddEffect               |
+//
+// and verifies that it gives the correct output.
+TEST(EffectChainTest, DiamondGraph) {
+       float data[] = {
+               1.0f, 1.0f,
+               1.0f, 0.0f,
+       };
+       float expected_data[] = {
+               2.5f, 2.5f,
+               2.5f, 0.0f,
+       };
+       float out_data[2 * 2];
+
+       const float half[] = { 0.5f, 0.5f, 0.5f, 0.5f };
+       const float two[] = { 2.0f, 2.0f, 2.0f, 0.5f };
+
+       MultiplyEffect *mul_half = new MultiplyEffect();
+       ASSERT_TRUE(mul_half->set_vec4("factor", half));
+       
+       MultiplyEffect *mul_two = new MultiplyEffect();
+       ASSERT_TRUE(mul_two->set_vec4("factor", two));
+
+       EffectChainTester tester(nullptr, 2, 2);
+
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+
+       FlatInput *input = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 2, 2);
+       input->set_pixel_data(data);
+
+       tester.get_chain()->add_input(input);
+       tester.get_chain()->add_effect(mul_half, input);
+       tester.get_chain()->add_effect(mul_two, input);
+       tester.get_chain()->add_effect(new AddEffect(), mul_half, mul_two);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 2, 2);
+}
+
+// Constructs the graph
+//
+//             FlatInput                     |
+//            /         \                    |
+//  MultiplyEffect  MultiplyEffect           |
+//         \             |                   |
+//          \    BouncingIdentityEffect      |  
+//            \         /                    |
+//             AddEffect                     |
+//
+// and verifies that it gives the correct output.
+TEST(EffectChainTest, DiamondGraphWithOneInputUsedInTwoPhases) {
+       float data[] = {
+               1.0f, 1.0f,
+               1.0f, 0.0f,
+       };
+       float expected_data[] = {
+               2.5f, 2.5f,
+               2.5f, 0.0f,
+       };
+       float out_data[2 * 2];
+
+       const float half[] = { 0.5f, 0.5f, 0.5f, 0.5f };
+       const float two[] = { 2.0f, 2.0f, 2.0f, 0.5f };
+
+       MultiplyEffect *mul_half = new MultiplyEffect();
+       ASSERT_TRUE(mul_half->set_vec4("factor", half));
+       
+       MultiplyEffect *mul_two = new MultiplyEffect();
+       ASSERT_TRUE(mul_two->set_vec4("factor", two));
+       
+       BouncingIdentityEffect *bounce = new BouncingIdentityEffect();
+
+       EffectChainTester tester(nullptr, 2, 2);
+
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+
+       FlatInput *input = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 2, 2);
+       input->set_pixel_data(data);
+
+       tester.get_chain()->add_input(input);
+       tester.get_chain()->add_effect(mul_half, input);
+       tester.get_chain()->add_effect(mul_two, input);
+       tester.get_chain()->add_effect(bounce, mul_two);
+       tester.get_chain()->add_effect(new AddEffect(), mul_half, bounce);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 2, 2);
+}
+
+TEST(EffectChainTest, EffectUsedTwiceOnlyGetsOneGammaConversion) {
+       float data[] = {
+               0.735f, 0.0f,
+               0.735f, 0.0f,
+       };
+       float expected_data[] = {
+               0.0f, 0.5f,  // 0.5 and not 1.0, since AddEffect doesn't clamp alpha properly.
+               0.0f, 0.5f,
+       };
+       float out_data[2 * 2];
+       
+       EffectChainTester tester(nullptr, 2, 2);
+       tester.add_input(data, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_sRGB);
+
+       // MirrorEffect does not get linear light, so the conversions will be
+       // inserted after it, not before.
+       RewritingEffect<MirrorEffect> *effect = new RewritingEffect<MirrorEffect>();
+       tester.get_chain()->add_effect(effect);
+
+       Effect *identity1 = tester.get_chain()->add_effect(new IdentityEffect(), effect);
+       Effect *identity2 = tester.get_chain()->add_effect(new IdentityEffect(), effect);
+       tester.get_chain()->add_effect(new AddEffect(), identity1, identity2);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 2, 2);
+
+       Node *node = effect->replaced_node;
+       ASSERT_EQ(1, node->incoming_links.size());
+       ASSERT_EQ(1, node->outgoing_links.size());
+       EXPECT_EQ("FlatInput", node->incoming_links[0]->effect->effect_type_id());
+       EXPECT_EQ("GammaExpansionEffect", node->outgoing_links[0]->effect->effect_type_id());
+}
+
+TEST(EffectChainTest, EffectUsedTwiceOnlyGetsOneColorspaceConversion) {
+       float data[] = {
+               0.5f, 0.0f,
+               0.5f, 0.0f,
+       };
+       float expected_data[] = {
+               0.0f, 0.5f,  // 0.5 and not 1.0, since AddEffect doesn't clamp alpha properly.
+               0.0f, 0.5f,
+       };
+       float out_data[2 * 2];
+       
+       EffectChainTester tester(nullptr, 2, 2);
+       tester.add_input(data, FORMAT_GRAYSCALE, COLORSPACE_REC_601_625, GAMMA_LINEAR);
+
+       // MirrorEffect does not get linear light, so the conversions will be
+       // inserted after it, not before.
+       RewritingEffect<MirrorEffect> *effect = new RewritingEffect<MirrorEffect>();
+       tester.get_chain()->add_effect(effect);
+
+       Effect *identity1 = tester.get_chain()->add_effect(new IdentityEffect(), effect);
+       Effect *identity2 = tester.get_chain()->add_effect(new IdentityEffect(), effect);
+       tester.get_chain()->add_effect(new AddEffect(), identity1, identity2);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 2, 2);
+
+       Node *node = effect->replaced_node;
+       ASSERT_EQ(1, node->incoming_links.size());
+       ASSERT_EQ(1, node->outgoing_links.size());
+       EXPECT_EQ("FlatInput", node->incoming_links[0]->effect->effect_type_id());
+       EXPECT_EQ("ColorspaceConversionEffect", node->outgoing_links[0]->effect->effect_type_id());
+}
+
+// An effect that does nothing, but requests texture bounce and stores
+// its input size.
+class SizeStoringEffect : public BouncingIdentityEffect {
+public:
+       SizeStoringEffect() : input_width(-1), input_height(-1) {}
+       void inform_input_size(unsigned input_num, unsigned width, unsigned height) override {
+               assert(input_num == 0);
+               input_width = width;
+               input_height = height;
+       }
+       string effect_type_id() const override { return "SizeStoringEffect"; }
+
+       int input_width, input_height;
+};
+
+TEST(EffectChainTest, SameInputsGiveSameOutputs) {
+       float data[2 * 2] = {
+               0.0f, 0.0f,
+               0.0f, 0.0f,
+       };
+       float out_data[4 * 3];
+       
+       EffectChainTester tester(nullptr, 4, 3);  // Note non-square aspect.
+
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+
+       FlatInput *input1 = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 2, 2);
+       input1->set_pixel_data(data);
+       
+       FlatInput *input2 = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 2, 2);
+       input2->set_pixel_data(data);
+
+       SizeStoringEffect *input_store = new SizeStoringEffect();
+
+       tester.get_chain()->add_input(input1);
+       tester.get_chain()->add_input(input2);
+       tester.get_chain()->add_effect(new AddEffect(), input1, input2);
+       tester.get_chain()->add_effect(input_store);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       EXPECT_EQ(2, input_store->input_width);
+       EXPECT_EQ(2, input_store->input_height);
+}
+
+TEST(EffectChainTest, AspectRatioConversion) {
+       float data1[4 * 3] = {
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f,
+       };
+       float data2[7 * 7] = {
+               0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+               0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f,
+       };
+
+       // The right conversion here is that the 7x7 image decides the size,
+       // since it is the biggest, so everything is scaled up to 9x7
+       // (keep the height, round the width 9.333 to 9). 
+       float out_data[9 * 7];
+       
+       EffectChainTester tester(nullptr, 4, 3);
+
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+
+       FlatInput *input1 = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 4, 3);
+       input1->set_pixel_data(data1);
+       
+       FlatInput *input2 = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 7, 7);
+       input2->set_pixel_data(data2);
+
+       SizeStoringEffect *input_store = new SizeStoringEffect();
+
+       tester.get_chain()->add_input(input1);
+       tester.get_chain()->add_input(input2);
+       tester.get_chain()->add_effect(new AddEffect(), input1, input2);
+       tester.get_chain()->add_effect(input_store);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       EXPECT_EQ(9, input_store->input_width);
+       EXPECT_EQ(7, input_store->input_height);
+}
+
+// Tests that putting a BlueInput (constant color) into its own pass,
+// which creates a phase that doesn't need texture coordinates,
+// doesn't mess up a second phase that actually does.
+TEST(EffectChainTest, FirstPhaseWithNoTextureCoordinates) {
+       const int size = 2;
+       float data[] = {
+               1.0f,
+               0.0f,
+       };
+       float expected_data[] = {
+               1.0f, 1.0f, 2.0f, 2.0f,
+               0.0f, 0.0f, 1.0f, 2.0f,
+       };
+       float out_data[size * 4];
+       // First say that we have sRGB, linear input.
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+       FlatInput *input = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, 1, size);
+
+       input->set_pixel_data(data);
+       EffectChainTester tester(nullptr, 1, size);
+       tester.get_chain()->add_input(new BlueInput());
+       Effect *phase1_end = tester.get_chain()->add_effect(new BouncingIdentityEffect());
+       tester.get_chain()->add_input(input);
+       tester.get_chain()->add_effect(new AddEffect(), phase1_end, input);
+
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
+
+       expect_equal(expected_data, out_data, 4, size);
+}
+
+// An effect that does nothing except changing its output sizes.
+class VirtualResizeEffect : public Effect {
+public:
+       VirtualResizeEffect(int width, int height, int virtual_width, int virtual_height)
+               : width(width),
+                 height(height),
+                 virtual_width(virtual_width),
+                 virtual_height(virtual_height) {}
+       string effect_type_id() const override { return "VirtualResizeEffect"; }
+       string output_fragment_shader() override { return read_file("identity.frag"); }
+
+       bool changes_output_size() const override { return true; }
+
+       void get_output_size(unsigned *width, unsigned *height,
+                            unsigned *virtual_width, unsigned *virtual_height) const override {
+               *width = this->width;
+               *height = this->height;
+               *virtual_width = this->virtual_width;
+               *virtual_height = this->virtual_height;
+       }
+
+private:
+       int width, height, virtual_width, virtual_height;
+};
+
+TEST(EffectChainTest, VirtualSizeIsSentOnToInputs) {
+       const int size = 2, bigger_size = 3;
+       float data[size * size] = {
+               1.0f, 0.0f,
+               0.0f, 1.0f,
+       };
+       float out_data[size * size];
+       
+       EffectChainTester tester(data, size, size, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       SizeStoringEffect *size_store = new SizeStoringEffect();
+
+       tester.get_chain()->add_effect(new VirtualResizeEffect(size, size, bigger_size, bigger_size));
+       tester.get_chain()->add_effect(size_store);
+       tester.get_chain()->add_effect(new VirtualResizeEffect(size, size, size, size));
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       EXPECT_EQ(bigger_size, size_store->input_width);
+       EXPECT_EQ(bigger_size, size_store->input_height);
+
+       // If the resize is implemented as non-virtual, we'll fail here,
+       // since bilinear scaling from 2x2 → 3x3 → 2x2 is not very exact.
+       expect_equal(data, out_data, size, size);
+}
+
+// An effect that is like VirtualResizeEffect, but always has virtual and real
+// sizes the same (and promises this).
+class NonVirtualResizeEffect : public VirtualResizeEffect {
+public:
+       NonVirtualResizeEffect(int width, int height)
+               : VirtualResizeEffect(width, height, width, height) {}
+       string effect_type_id() const override { return "NonVirtualResizeEffect"; }
+       bool sets_virtual_output_size() const override { return false; }
+};
+
+// An effect that promises one-to-one sampling (unlike IdentityEffect).
+class OneToOneEffect : public Effect {
+public:
+       OneToOneEffect() {}
+       string effect_type_id() const override { return "OneToOneEffect"; }
+       string output_fragment_shader() override { return read_file("identity.frag"); }
+       bool strong_one_to_one_sampling() const override { return true; }
+};
+
+TEST_P(WithAndWithoutComputeShaderTest, NoBounceWithOneToOneSampling) {
+       const int size = 2;
+       float data[size * size] = {
+               1.0f, 0.0f,
+               0.0f, 1.0f,
+       };
+       float out_data[size * size];
+
+       EffectChainTester tester(data, size, size, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       RewritingEffect<OneToOneEffect> *effect1 = new RewritingEffect<OneToOneEffect>();
+       RewritingEffect<OneToOneEffect> *effect2 = new RewritingEffect<OneToOneEffect>();
+
+       if (GetParam() == "compute") {
+               tester.get_chain()->add_effect(new IdentityComputeEffect());
+       } else {
+               tester.get_chain()->add_effect(new NonVirtualResizeEffect(size, size));
+       }
+       tester.get_chain()->add_effect(effect1);
+       tester.get_chain()->add_effect(effect2);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, size, size);
+
+       // The first OneToOneEffect should be in the same phase as its input.
+       ASSERT_EQ(1, effect1->replaced_node->incoming_links.size());
+       EXPECT_EQ(effect1->replaced_node->incoming_links[0]->containing_phase,
+                 effect1->replaced_node->containing_phase);
+
+       // The second OneToOneEffect, too.
+       EXPECT_EQ(effect1->replaced_node->containing_phase,
+                 effect2->replaced_node->containing_phase);
+}
+
+TEST(EffectChainTest, BounceWhenOneToOneIsBroken) {
+       const int size = 2;
+       float data[size * size] = {
+               1.0f, 0.0f,
+               0.0f, 1.0f,
+       };
+       float out_data[size * size];
+
+       EffectChainTester tester(data, size, size, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       RewritingEffect<OneToOneEffect> *effect1 = new RewritingEffect<OneToOneEffect>();
+       RewritingEffect<OneToOneEffect> *effect2 = new RewritingEffect<OneToOneEffect>();
+       RewritingEffect<IdentityEffect> *effect3 = new RewritingEffect<IdentityEffect>();
+       RewritingEffect<OneToOneEffect> *effect4 = new RewritingEffect<OneToOneEffect>();
+
+       tester.get_chain()->add_effect(new NonVirtualResizeEffect(size, size));
+       tester.get_chain()->add_effect(effect1);
+       tester.get_chain()->add_effect(effect2);
+       tester.get_chain()->add_effect(effect3);
+       tester.get_chain()->add_effect(effect4);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, size, size);
+
+       // The NonVirtualResizeEffect should be in a different phase from
+       // the IdentityEffect (since the latter is not one-to-one),
+       // ie., the chain should be broken somewhere between them, but exactly
+       // where doesn't matter.
+       ASSERT_EQ(1, effect1->replaced_node->incoming_links.size());
+       EXPECT_NE(effect1->replaced_node->incoming_links[0]->containing_phase,
+                 effect3->replaced_node->containing_phase);
+
+       // The last OneToOneEffect should also be in the same phase as the
+       // IdentityEffect (the phase was already broken).
+       EXPECT_EQ(effect3->replaced_node->containing_phase,
+                 effect4->replaced_node->containing_phase);
+}
+
+// Does not use EffectChainTest, so that it can construct an EffectChain without
+// a shared ResourcePool (which is also properly destroyed afterwards).
+// Also turns on debugging to test that code path.
+TEST(EffectChainTest, IdentityWithOwnPool) {
+       const int width = 3, height = 2;
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       const float expected_data[] = {
+               0.75f, 1.0f, 1.0f,
+               0.0f, 0.25f, 0.3f,
+       };
+       float out_data[6], temp[6 * 4];
+
+       EffectChain chain(width, height);
+       MovitDebugLevel old_movit_debug_level = movit_debug_level;
+       movit_debug_level = MOVIT_DEBUG_ON;
+
+       ImageFormat format;
+       format.color_space = COLORSPACE_sRGB;
+       format.gamma_curve = GAMMA_LINEAR;
+
+       FlatInput *input = new FlatInput(format, FORMAT_GRAYSCALE, GL_FLOAT, width, height);
+       input->set_pixel_data(data);
+       chain.add_input(input);
+       chain.add_output(format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
+
+       GLuint texnum, fbo;
+       glGenTextures(1, &texnum);
+       check_error();
+       glBindTexture(GL_TEXTURE_2D, texnum);
+       check_error();
+       glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA32F, width, height, 0, GL_RGBA, GL_FLOAT, nullptr);
+       check_error();
+
+       glGenFramebuffers(1, &fbo);
+       check_error();
+       glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+       check_error();
+       glFramebufferTexture2D(
+               GL_FRAMEBUFFER,
+               GL_COLOR_ATTACHMENT0,
+               GL_TEXTURE_2D,
+               texnum,
+               0);
+       check_error();
+       glBindFramebuffer(GL_FRAMEBUFFER, 0);
+       check_error();
+
+       chain.finalize();
+
+       chain.render_to_fbo(fbo, width, height);
+
+       glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+       check_error();
+       glReadPixels(0, 0, width, height, GL_RGBA, GL_FLOAT, temp);
+       check_error();
+       for (unsigned i = 0; i < 6; ++i) {
+               out_data[i] = temp[i * 4];
+       }
+
+       expect_equal(expected_data, out_data, width, height);
+
+       // Reset the debug status again.
+       movit_debug_level = old_movit_debug_level;
+}
+
+// A dummy effect whose only purpose is to test sprintf decimal behavior.
+class PrintfingBlueEffect : public Effect {
+public:
+       PrintfingBlueEffect() {}
+       string effect_type_id() const override { return "PrintfingBlueEffect"; }
+       string output_fragment_shader() override {
+               stringstream ss;
+               ss.imbue(locale("C"));
+               ss.precision(8);
+               ss << "vec4 FUNCNAME(vec2 tc) { return vec4("
+                  << 0.0f << ", " << 0.0f << ", "
+                  << 0.5f << ", " << 1.0f << "); }\n";
+               return ss.str();
+       }
+};
+
+TEST(EffectChainTest, StringStreamLocalesWork) {
+       // An example of a locale with comma instead of period as decimal separator.
+       // Obviously, if you run on a machine without this locale available,
+       // the test will always succeed. Note that the OpenGL driver might call
+       // setlocale() behind-the-scenes, and that might corrupt the returned
+       // pointer, so we need to take our own copy of it here.
+       char *saved_locale = setlocale(LC_ALL, "nb_NO.UTF_8");
+       if (saved_locale == nullptr) {
+               // The locale wasn't available.
+               return;
+       }
+       saved_locale = strdup(saved_locale);
+       float data[] = {
+               0.0f, 0.0f, 0.0f, 0.0f,
+       };
+       float expected_data[] = {
+               0.0f, 0.0f, 0.5f, 1.0f,
+       };
+       float out_data[4];
+       EffectChainTester tester(data, 1, 1, FORMAT_RGBA_PREMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
+       tester.get_chain()->add_effect(new PrintfingBlueEffect());
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 4, 1);
+
+       setlocale(LC_ALL, saved_locale);
+       free(saved_locale);
+}
+
+TEST(EffectChainTest, sRGBIntermediate) {
+       float data[] = {
+               0.0f, 0.5f, 0.0f, 1.0f,
+       };
+       float out_data[4];
+       EffectChainTester tester(data, 1, 1, FORMAT_RGBA_PREMULTIPLIED_ALPHA, COLORSPACE_sRGB, GAMMA_LINEAR);
+       tester.get_chain()->set_intermediate_format(GL_SRGB8);
+       tester.get_chain()->add_effect(new IdentityEffect());
+       tester.get_chain()->add_effect(new BouncingIdentityEffect());
+       tester.run(out_data, GL_RGBA, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       EXPECT_GE(fabs(out_data[1] - data[1]), 1e-3)
+           << "Expected sRGB not to be able to represent 0.5 exactly (got " << out_data[1] << ")";
+       EXPECT_LT(fabs(out_data[1] - data[1]), 0.1f)
+           << "Expected sRGB to be able to represent 0.5 approximately (got " << out_data[1] << ")";
+
+       // This state should have been preserved.
+       EXPECT_FALSE(glIsEnabled(GL_FRAMEBUFFER_SRGB));
+}
+
+// An effect that is like IdentityEffect, but also does not require linear light.
+class PassThroughEffect : public IdentityEffect {
+public:
+       PassThroughEffect() {}
+       string effect_type_id() const override { return "PassThroughEffect"; }
+       bool needs_linear_light() const override { return false; }
+       AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; }
+};
+
+// Same, just also bouncing.
+class BouncingPassThroughEffect : public BouncingIdentityEffect {
+public:
+       BouncingPassThroughEffect() {}
+       string effect_type_id() const override { return "BouncingPassThroughEffect"; }
+       bool needs_linear_light() const override { return false; }
+       bool needs_texture_bounce() const override { return true; }
+       AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; }
+};
+
+TEST(EffectChainTest, Linear10bitIntermediateAccuracy) {
+       // Note that we do the comparison in sRGB space, which is what we
+       // typically would want; however, we do the sRGB conversion ourself
+       // to avoid compounding errors from shader conversions into the
+       // analysis.
+       const int size = 4096;  // 12-bit.
+       float linear_data[size], data[size], out_data[size];
+
+       for (int i = 0; i < size; ++i) {
+               linear_data[i] = i / double(size - 1);
+               data[i] = srgb_to_linear(linear_data[i]);
+       }
+
+       EffectChainTester tester(data, size, 1, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR, GL_RGBA32F);
+       tester.get_chain()->set_intermediate_format(GL_RGB10_A2);
+       tester.get_chain()->add_effect(new IdentityEffect());
+       tester.get_chain()->add_effect(new BouncingIdentityEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       for (int i = 0; i < size; ++i) {
+               out_data[i] = linear_to_srgb(out_data[i]);
+       }
+
+       // This maximum error is pretty bad; about 6.5 levels of a 10-bit sRGB
+       // framebuffer. (Slightly more on NVIDIA cards.)
+       expect_equal(linear_data, out_data, size, 1, 7.5e-3, 2e-5);
+}
+
+TEST_P(WithAndWithoutComputeShaderTest, SquareRoot10bitIntermediateAccuracy) {
+       // Note that we do the comparison in sRGB space, which is what we
+       // typically would want; however, we do the sRGB conversion ourself
+       // to avoid compounding errors from shader conversions into the
+       // analysis.
+       const int size = 4096;  // 12-bit.
+       float linear_data[size], data[size], out_data[size];
+
+       for (int i = 0; i < size; ++i) {
+               linear_data[i] = i / double(size - 1);
+               data[i] = srgb_to_linear(linear_data[i]);
+       }
+
+       EffectChainTester tester(data, size, 1, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR, GL_RGBA32F);
+       tester.get_chain()->set_intermediate_format(GL_RGB10_A2, SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION);
+       if (GetParam() == "compute") {
+               tester.get_chain()->add_effect(new IdentityComputeEffect());
+       } else {
+               tester.get_chain()->add_effect(new IdentityEffect());
+       }
+       tester.get_chain()->add_effect(new BouncingIdentityEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       for (int i = 0; i < size; ++i) {
+               out_data[i] = linear_to_srgb(out_data[i]);
+       }
+
+       // This maximum error is much better; about 0.7 levels of a 10-bit sRGB
+       // framebuffer (ideal would be 0.5). That is an order of magnitude better
+       // than in the linear test above. The RMS error is much better, too.
+       expect_equal(linear_data, out_data, size, 1, 7.5e-4, 5e-6);
+}
+
+TEST(EffectChainTest, SquareRootIntermediateIsTurnedOffForNonLinearData) {
+       const int size = 256;  // 8-bit.
+       float data[size], out_data[size];
+
+       for (int i = 0; i < size; ++i) {
+               data[i] = i / double(size - 1);
+       }
+
+       EffectChainTester tester(data, size, 1, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_REC_601, GL_RGBA32F);
+       tester.get_chain()->set_intermediate_format(GL_RGB8, SQUARE_ROOT_FRAMEBUFFER_TRANSFORMATION);
+       tester.get_chain()->add_effect(new PassThroughEffect());
+       tester.get_chain()->add_effect(new BouncingPassThroughEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_REC_601);
+
+       // The data should be passed through nearly exactly, since there is no effect
+       // on the path that requires linear light. (Actually, it _is_ exact modulo
+       // fp32 errors, but the error bounds is strictly _less than_, not zero.)
+       expect_equal(data, out_data, size, 1, 1e-6, 1e-6);
+}
+
+// An effect that stores which program number was last run under.
+class RecordingIdentityEffect : public Effect {
+public:
+       RecordingIdentityEffect() {}
+       string effect_type_id() const override { return "RecordingIdentityEffect"; }
+       string output_fragment_shader() override { return read_file("identity.frag"); }
+
+       GLuint last_glsl_program_num;
+       void set_gl_state(GLuint glsl_program_num, const std::string& prefix, unsigned *sampler_num) override
+       {
+               last_glsl_program_num = glsl_program_num;
+       }
+};
+
+TEST(EffectChainTest, ProgramsAreClonedForMultipleThreads) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       RecordingIdentityEffect *effect = new RecordingIdentityEffect();
+       tester.get_chain()->add_effect(effect);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, 3, 2);
+
+       ASSERT_NE(0, effect->last_glsl_program_num);
+
+       // Now pretend some other effect is using this program number;
+       // ResourcePool will then need to clone it.
+       ResourcePool *resource_pool = tester.get_chain()->get_resource_pool();
+       GLuint master_program_num = resource_pool->use_glsl_program(effect->last_glsl_program_num);
+       EXPECT_EQ(effect->last_glsl_program_num, master_program_num);
+
+       // Re-run should still give the correct data, but it should have run
+       // with a different program.
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+       expect_equal(data, out_data, 3, 2);
+       EXPECT_NE(effect->last_glsl_program_num, master_program_num);
+
+       // Release the program, and check one final time.
+       resource_pool->unuse_glsl_program(master_program_num);
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+       expect_equal(data, out_data, 3, 2);
+}
+
+TEST(ComputeShaderTest, Identity) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       if (!movit_compute_shaders_supported) {
+               fprintf(stderr, "Skipping test; no support for compile shaders.\n");
+               return;
+       }
+       tester.get_chain()->add_effect(new IdentityComputeEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, 3, 2);
+}
+
+// Like IdentityComputeEffect, but due to the alpha handling, this will be
+// the very last effect in the chain, which means we can't output it directly
+// to the screen.
+class IdentityAlphaComputeEffect : public IdentityComputeEffect {
+       AlphaHandling alpha_handling() const { return DONT_CARE_ALPHA_TYPE; }
+};
+
+TEST(ComputeShaderTest, LastEffectInChain) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f,
+               0.75f, 1.0f, 1.0f,
+       };
+       float out_data[6];
+       EffectChainTester tester(data, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       if (!movit_compute_shaders_supported) {
+               fprintf(stderr, "Skipping test; no support for compile shaders.\n");
+               return;
+       }
+       tester.get_chain()->add_effect(new IdentityAlphaComputeEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, 3, 2);
+}
+
+TEST(ComputeShaderTest, Render8BitTo8Bit) {
+       uint8_t data[] = {
+               14, 200, 80,
+               90, 100, 110,
+       };
+       uint8_t out_data[6];
+       EffectChainTester tester(nullptr, 3, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR, GL_RGBA8);
+       if (!movit_compute_shaders_supported) {
+               fprintf(stderr, "Skipping test; no support for compile shaders.\n");
+               return;
+       }
+       tester.add_input(data, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR, 3, 2);
+       tester.get_chain()->add_effect(new IdentityAlphaComputeEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(data, out_data, 3, 2);
+}
+
+// A compute shader to mirror the inputs, in 2x2 blocks.
+class MirrorComputeEffect : public Effect {
+public:
+       MirrorComputeEffect() {}
+       string effect_type_id() const override { return "MirrorComputeEffect"; }
+       bool is_compute_shader() const override { return true; }
+       string output_fragment_shader() override { return read_file("mirror.comp"); }
+       void get_compute_dimensions(unsigned output_width, unsigned output_height,
+                                   unsigned *x, unsigned *y, unsigned *z) const override {
+               *x = output_width / 2;
+               *y = output_height / 2;
+               *z = 1;
+       }
+};
+
+TEST(ComputeShaderTest, ComputeThenOneToOne) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f, 0.8f,
+               0.75f, 1.0f, 1.0f, 0.2f,
+       };
+       float expected_data[] = {
+               0.8f, 0.3f, 0.25f, 0.0f,
+               0.2f, 1.0f, 1.0f, 0.75f,
+       };
+       float out_data[8];
+       EffectChainTester tester(data, 4, 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR);
+       tester.get_chain()->add_effect(new MirrorComputeEffect());
+       tester.get_chain()->add_effect(new OneToOneEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 4, 2);
+}
+
+// A compute shader that also resizes its input, taking the upper-left pixel
+// of every 2x2 group. (The shader is hard-coded to 4x2 input for simplicity.)
+class Downscale2xComputeEffect : public Effect {
+public:
+       Downscale2xComputeEffect() {}
+       string effect_type_id() const override { return "Downscale2xComputeEffect"; }
+       bool is_compute_shader() const override { return true; }
+       string output_fragment_shader() override { return read_file("downscale2x.comp"); }
+       bool changes_output_size() const override { return true; }
+       void inform_input_size(unsigned input_num, unsigned width, unsigned height) override
+       {
+               this->width = width;
+               this->height = height;
+       }
+       void get_output_size(unsigned *width, unsigned *height,
+                            unsigned *virtual_width, unsigned *virtual_height) const override {
+                *width = *virtual_width = this->width / 2;
+                *height = *virtual_height = this->height / 2;
+        }
+
+private:
+       unsigned width, height;
+};
+
+// Even if the compute shader is not the last effect, it's the one that should decide
+// the output size of the phase.
+TEST(ComputeShaderTest, ResizingComputeThenOneToOne) {
+       float data[] = {
+               0.0f, 0.25f, 0.3f, 0.8f,
+               0.75f, 1.0f, 1.0f, 0.2f,
+       };
+       float expected_data[] = {
+               0.0f, 0.3f,
+       };
+       float out_data[2];
+       EffectChainTester tester(nullptr, 2, 1);
+       tester.add_input(data, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR, 4, 2);
+
+       RewritingEffect<Downscale2xComputeEffect> *downscale_effect = new RewritingEffect<Downscale2xComputeEffect>();
+       tester.get_chain()->add_effect(downscale_effect);
+       tester.get_chain()->add_effect(new OneToOneEffect());
+       tester.get_chain()->add_effect(new BouncingIdentityEffect());
+       tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR);
+
+       expect_equal(expected_data, out_data, 2, 1);
+
+       Phase *phase = downscale_effect->replaced_node->containing_phase;
+       EXPECT_EQ(2, phase->output_width);
+       EXPECT_EQ(1, phase->output_height);
+}
+
+}  // namespace movit