From: Steinar H. Gunderson Date: Fri, 24 Nov 2017 21:11:08 +0000 (+0100) Subject: Support chaining certain effects after compute shaders. X-Git-Tag: 1.6.0~32 X-Git-Url: https://git.sesse.net/?p=movit;a=commitdiff_plain;h=e683d4e49549b2d4c6906d2ebc4184904f4fe9f8 Support chaining certain effects after compute shaders. This rests on the notion of “strong one-to-one” that's very similar to the older “one-to-one“ concept, yet a bit stricter (in particular, PaddingEffect is not strong one-to-one). --- diff --git a/alpha_division_effect.h b/alpha_division_effect.h index 9258171..554a186 100644 --- a/alpha_division_effect.h +++ b/alpha_division_effect.h @@ -14,7 +14,7 @@ public: AlphaDivisionEffect() {} std::string effect_type_id() const override { return "AlphaDivisionEffect"; } std::string output_fragment_shader() override; - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } }; } // namespace movit diff --git a/alpha_multiplication_effect.h b/alpha_multiplication_effect.h index 85f08f8..50333cb 100644 --- a/alpha_multiplication_effect.h +++ b/alpha_multiplication_effect.h @@ -14,7 +14,7 @@ public: AlphaMultiplicationEffect() {} std::string effect_type_id() const override { return "AlphaMultiplicationEffect"; } std::string output_fragment_shader() override; - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } }; } // namespace movit diff --git a/colorspace_conversion_effect.h b/colorspace_conversion_effect.h index d875465..a59d476 100644 --- a/colorspace_conversion_effect.h +++ b/colorspace_conversion_effect.h @@ -28,7 +28,7 @@ public: bool needs_srgb_primaries() const override { return false; } AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } // Get a conversion matrix from the given color space to XYZ. static Eigen::Matrix3d get_xyz_matrix(Colorspace space); diff --git a/diffusion_effect.h b/diffusion_effect.h index 253126e..a8babd0 100644 --- a/diffusion_effect.h +++ b/diffusion_effect.h @@ -57,7 +57,7 @@ public: AlphaHandling alpha_handling() const override { return INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK; } unsigned num_inputs() const override { return 2; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } private: float blurred_mix_amount; diff --git a/dither_effect.h b/dither_effect.h index 71bb6d3..b1bbbff 100644 --- a/dither_effect.h +++ b/dither_effect.h @@ -66,7 +66,7 @@ public: // premultiplied error. However, we need to do dithering in the same // space as quantization, whether that be pre- or postmultiply. AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } void set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) override; diff --git a/effect.h b/effect.h index afac5dc..ba39d6a 100644 --- a/effect.h +++ b/effect.h @@ -196,7 +196,16 @@ public: // set sets_virtual_output_size(), though. // // Does not make a lot of sense together with needs_texture_bounce(). - virtual bool one_to_one_sampling() const { return false; } + // Cannot be set for compute shaders. + virtual bool one_to_one_sampling() const { return strong_one_to_one_sampling(); } + + // Similar in use to one_to_one_sampling(), but even stricter: + // The effect must not use texture coordinate in any way beyond + // giving it unmodified to its (single) input. This allows it to + // also be used after a compute shader, in the same phase. + // + // An effect that it strong one-to-one must also be one-to-one. + virtual bool strong_one_to_one_sampling() const { return false; } // Whether this effect wants to output to a different size than // its input(s) (see inform_input_size(), below). See also diff --git a/effect_chain.cpp b/effect_chain.cpp index 23156cb..c6bac80 100644 --- a/effect_chain.cpp +++ b/effect_chain.cpp @@ -34,12 +34,15 @@ namespace movit { namespace { -// An effect that does nothing. -class IdentityEffect : public Effect { +// An effect whose only purpose is to sit in a phase on its own and take the +// texture output from a compute shader and display it to the normal backbuffer +// (or any FBO). That phase can be skipped when rendering using render_to_textures(). +class ComputeShaderOutputDisplayEffect : public Effect { public: - IdentityEffect() {} - string effect_type_id() const override { return "IdentityEffect"; } + ComputeShaderOutputDisplayEffect() {} + string effect_type_id() const override { return "ComputeShaderOutputDisplayEffect"; } string output_fragment_shader() override { return read_file("identity.frag"); } + bool needs_texture_bounce() const override { return true; } }; } // namespace @@ -162,6 +165,7 @@ Node *EffectChain::add_node(Effect *effect) node->output_alpha_type = ALPHA_INVALID; node->needs_mipmaps = false; node->one_to_one_sampling = false; + node->strong_one_to_one_sampling = false; nodes.push_back(node); node_map[effect] = node; @@ -408,9 +412,17 @@ void EffectChain::compile_glsl_program(Phase *phase) Node *node = phase->effects[i]; const string effect_id = phase->effect_ids[node]; if (node->incoming_links.size() == 1) { - frag_shader += string("#define INPUT ") + phase->effect_ids[node->incoming_links[0]] + "\n"; + Node *input = node->incoming_links[0]; + if (i != 0 && input->effect->is_compute_shader()) { + // First effect after the compute shader reads the value + // that cs_output() wrote to a global variable. + frag_shader += string("#define INPUT(tc) CS_OUTPUT_VAL\n"); + } else { + frag_shader += string("#define INPUT ") + phase->effect_ids[input] + "\n"; + } } else { for (unsigned j = 0; j < node->incoming_links.size(); ++j) { + assert(!node->incoming_links[j]->effect->is_compute_shader()); char buf[256]; sprintf(buf, "#define INPUT%d %s\n", j + 1, phase->effect_ids[node->incoming_links[j]].c_str()); frag_shader += buf; @@ -435,7 +447,17 @@ void EffectChain::compile_glsl_program(Phase *phase) } frag_shader += "\n"; } - frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n"; + if (phase->is_compute_shader) { + frag_shader += string("#define INPUT ") + phase->effect_ids[phase->compute_shader_node] + "\n"; + if (phase->compute_shader_node == phase->effects.back()) { + // No postprocessing. + frag_shader += "#define CS_POSTPROC(tc) CS_OUTPUT_VAL\n"; + } else { + frag_shader += string("#define CS_POSTPROC ") + phase->effect_ids[phase->effects.back()] + "\n"; + } + } else { + frag_shader += string("#define INPUT ") + phase->effect_ids[phase->effects.back()] + "\n"; + } // If we're the last phase, add the right #defines for Y'CbCr multi-output as needed. vector frag_shader_outputs; // In order. @@ -501,8 +523,8 @@ void EffectChain::compile_glsl_program(Phase *phase) if (phase->is_compute_shader) { frag_shader.append(read_file("footer.comp")); - phase->output_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size); - phase->output_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust); + phase->compute_shader_node->effect->register_uniform_vec2("inv_output_size", (float *)&phase->inv_output_size); + phase->compute_shader_node->effect->register_uniform_vec2("output_texcoord_adjust", (float *)&phase->output_texcoord_adjust); } else { frag_shader.append(read_file("footer.frag")); } @@ -597,7 +619,8 @@ Phase *EffectChain::construct_phase(Node *output, map *complete Phase *phase = new Phase; phase->output_node = output; - phase->is_compute_shader = output->effect->is_compute_shader(); + phase->is_compute_shader = false; + phase->compute_shader_node = nullptr; // If the output effect has one-to-one sampling, we try to trace this // status down through the dependency chain. This is important in case @@ -605,6 +628,7 @@ Phase *EffectChain::construct_phase(Node *output, map *complete // output size); if we have one-to-one sampling, we don't have to break // the phase. output->one_to_one_sampling = output->effect->one_to_one_sampling(); + output->strong_one_to_one_sampling = output->effect->strong_one_to_one_sampling(); // Effects that we have yet to calculate, but that we know should // be in the current phase. @@ -615,6 +639,8 @@ Phase *EffectChain::construct_phase(Node *output, map *complete Node *node = effects_todo_this_phase.top(); effects_todo_this_phase.pop(); + assert(node->effect->one_to_one_sampling() >= node->effect->strong_one_to_one_sampling()); + if (node->effect->needs_mipmaps()) { node->needs_mipmaps = true; } @@ -631,6 +657,10 @@ Phase *EffectChain::construct_phase(Node *output, map *complete } phase->effects.push_back(node); + if (node->effect->is_compute_shader()) { + phase->is_compute_shader = true; + phase->compute_shader_node = node; + } // Find all the dependencies of this effect, and add them to the stack. vector deps = node->incoming_links; @@ -644,12 +674,6 @@ Phase *EffectChain::construct_phase(Node *output, map *complete start_new_phase = true; } - // Compute shaders currently always end phases. - // (We might loosen this up in some cases in the future.) - if (deps[i]->effect->is_compute_shader()) { - start_new_phase = true; - } - // Propagate information about needing mipmaps down the chain, // breaking the phase if we notice an incompatibility. // @@ -690,7 +714,15 @@ Phase *EffectChain::construct_phase(Node *output, map *complete } } - if (deps[i]->effect->sets_virtual_output_size()) { + if (deps[i]->effect->is_compute_shader()) { + // Only one compute shader per phase; we should have been stopped + // already due to the fact that compute shaders are not one-to-one. + assert(!phase->is_compute_shader); + + // If all nodes so far are strong one-to-one, we can put them after + // the compute shader (ie., process them on the output). + start_new_phase = !node->strong_one_to_one_sampling; + } else if (deps[i]->effect->sets_virtual_output_size()) { assert(deps[i]->effect->changes_output_size()); // If the next effect sets a virtual size to rely on OpenGL's // bilinear sampling, we'll really need to break the phase here. @@ -709,6 +741,8 @@ Phase *EffectChain::construct_phase(Node *output, map *complete // Propagate the one-to-one status down through the dependency. deps[i]->one_to_one_sampling = node->one_to_one_sampling && deps[i]->effect->one_to_one_sampling(); + deps[i]->strong_one_to_one_sampling = node->strong_one_to_one_sampling && + deps[i]->effect->strong_one_to_one_sampling(); } } } @@ -1722,8 +1756,15 @@ void EffectChain::add_dither_if_needed() void EffectChain::add_dummy_effect_if_needed() { Node *output = find_output_node(); - if (output->effect->is_compute_shader()) { - Node *dummy = add_node(new IdentityEffect()); + + // See if the last effect that's not strong one-to-one is a compute shader. + Node *last_effect = output; + while (last_effect->effect->num_inputs() == 1 && + last_effect->effect->strong_one_to_one_sampling()) { + last_effect = last_effect->incoming_links[0]; + } + if (last_effect->effect->is_compute_shader()) { + Node *dummy = add_node(new ComputeShaderOutputDisplayEffect()); connect_nodes(output, dummy); has_dummy_effect = true; } @@ -1905,7 +1946,7 @@ void EffectChain::render(GLuint dest_fbo, const vector &dest assert(num_phases >= 2); assert(!phases.back()->is_compute_shader); assert(phases.back()->effects.size() == 1); - assert(phases.back()->effects[0]->effect->effect_type_id() == "IdentityEffect"); + assert(phases.back()->effects[0]->effect->effect_type_id() == "ComputeShaderOutputDisplayEffect"); // We are rendering to a set of textures, so we can run the compute shader // directly and skip the dummy phase. diff --git a/effect_chain.h b/effect_chain.h index 716a970..793bf7b 100644 --- a/effect_chain.h +++ b/effect_chain.h @@ -153,6 +153,9 @@ private: // (in the same phase) have one_to_one_sampling() set. bool one_to_one_sampling; + // Same, for strong_one_to_one_sampling(). + bool strong_one_to_one_sampling; + friend class EffectChain; }; @@ -181,6 +184,7 @@ struct Phase { // Whether this phase is compiled as a compute shader, ie., the last effect is // marked as one. bool is_compute_shader; + Node *compute_shader_node; // If , which image unit the output buffer is bound to. // This is used as source for a Uniform below. diff --git a/effect_chain_test.cpp b/effect_chain_test.cpp index 92fe52b..551cdbc 100644 --- a/effect_chain_test.cpp +++ b/effect_chain_test.cpp @@ -1113,16 +1113,31 @@ public: bool sets_virtual_output_size() const override { return false; } }; +class WithAndWithoutComputeShaderTest : public testing::TestWithParam { +}; +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"); } +}; + // 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 one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } }; -TEST(EffectChainTest, NoBounceWithOneToOneSampling) { +TEST_P(WithAndWithoutComputeShaderTest, NoBounceWithOneToOneSampling) { const int size = 2; float data[size * size] = { 1.0f, 0.0f, @@ -1135,7 +1150,11 @@ TEST(EffectChainTest, NoBounceWithOneToOneSampling) { RewritingEffect *effect1 = new RewritingEffect(); RewritingEffect *effect2 = new RewritingEffect(); - tester.get_chain()->add_effect(new NonVirtualResizeEffect(size, size)); + 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); @@ -1302,21 +1321,6 @@ TEST(EffectChainTest, StringStreamLocalesWork) { free(saved_locale); } -// An effect that does nothing, but as a compute shader. -class IdentityComputeEffect : public Effect { -public: - IdentityComputeEffect() {} - string effect_type_id() const override { return "IdentityComputeEffect"; } - bool is_compute_shader() const override { return true; } - string output_fragment_shader() override { return read_file("identity.comp"); } -}; - -class WithAndWithoutComputeShaderTest : public testing::TestWithParam { -}; -INSTANTIATE_TEST_CASE_P(WithAndWithoutComputeShaderTest, - WithAndWithoutComputeShaderTest, - testing::Values("fragment", "compute")); - TEST(EffectChainTest, sRGBIntermediate) { float data[] = { 0.0f, 0.5f, 0.0f, 1.0f, diff --git a/footer.comp b/footer.comp index 16197a6..a28014d 100644 --- a/footer.comp +++ b/footer.comp @@ -22,6 +22,10 @@ void cs_output(uvec2 coord, vec4 val) void cs_output(ivec2 coord, vec4 val) { + // Run the value through any preprocessing steps we might have. + CS_OUTPUT_VAL = val; + val = CS_POSTPROC(vec2(0.0, 0.0)); + #if SQUARE_ROOT_TRANSFORMATION // Make sure we don't give negative values to sqrt. val.rgb = sqrt(max(val.rgb, 0.0)); diff --git a/gamma_compression_effect.h b/gamma_compression_effect.h index f9c3297..9094f19 100644 --- a/gamma_compression_effect.h +++ b/gamma_compression_effect.h @@ -29,7 +29,7 @@ public: void set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) override; bool needs_srgb_primaries() const override { return false; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } // Actually needs postmultiplied input as well as outputting it. // EffectChain will take care of that. diff --git a/gamma_expansion_effect.h b/gamma_expansion_effect.h index c4968b2..9565a3b 100644 --- a/gamma_expansion_effect.h +++ b/gamma_expansion_effect.h @@ -30,7 +30,7 @@ public: bool needs_linear_light() const override { return false; } bool needs_srgb_primaries() const override { return false; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } // Actually processes its input in a nonlinear fashion, // but does not touch alpha, and we are a special case anyway. diff --git a/glow_effect.h b/glow_effect.h index e159999..340cf3d 100644 --- a/glow_effect.h +++ b/glow_effect.h @@ -52,7 +52,7 @@ public: std::string output_fragment_shader() override; AlphaHandling alpha_handling() const override { return INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } private: float cutoff; diff --git a/header.comp b/header.comp index d922e33..3c07007 100644 --- a/header.comp +++ b/header.comp @@ -10,4 +10,8 @@ vec4 tex2D(sampler2D s, vec2 coord); void cs_output(uvec2 coord, vec4 val); void cs_output(ivec2 coord, vec4 val); +// Used if there are any steps used to postprocess compute shader output. +// Initialized due to https://bugs.freedesktop.org/show_bug.cgi?id=103895. +vec4 CS_OUTPUT_VAL = vec4(0.0); + #define OUTPUT(tc, val) cs_output(tc, val) diff --git a/lift_gamma_gain_effect.h b/lift_gamma_gain_effect.h index 19d9b20..6e2ea19 100644 --- a/lift_gamma_gain_effect.h +++ b/lift_gamma_gain_effect.h @@ -32,7 +32,7 @@ public: LiftGammaGainEffect(); std::string effect_type_id() const override { return "LiftGammaGainEffect"; } AlphaHandling alpha_handling() const override { return INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } std::string output_fragment_shader() override; void set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) override; diff --git a/luma_mix_effect.h b/luma_mix_effect.h index fec402f..1272731 100644 --- a/luma_mix_effect.h +++ b/luma_mix_effect.h @@ -26,7 +26,7 @@ public: bool needs_srgb_primaries() const override { return false; } unsigned num_inputs() const override { return 3; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } AlphaHandling alpha_handling() const override { return INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK; } private: diff --git a/mix_effect.h b/mix_effect.h index 1305433..850baae 100644 --- a/mix_effect.h +++ b/mix_effect.h @@ -18,7 +18,7 @@ public: bool needs_srgb_primaries() const override { return false; } unsigned num_inputs() const override { return 2; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } // TODO: In the common case where a+b=1, it would be useful to be able to set // alpha_handling() to INPUT_PREMULTIPLIED_ALPHA_KEEP_BLANK. However, right now diff --git a/multiply_effect.h b/multiply_effect.h index 9e3542d..40a7a72 100644 --- a/multiply_effect.h +++ b/multiply_effect.h @@ -18,7 +18,7 @@ public: MultiplyEffect(); std::string effect_type_id() const override { return "MultiplyEffect"; } std::string output_fragment_shader() override; - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } private: RGBATuple factor; diff --git a/overlay_effect.h b/overlay_effect.h index 2f88d12..29c0941 100644 --- a/overlay_effect.h +++ b/overlay_effect.h @@ -24,7 +24,7 @@ public: bool needs_srgb_primaries() const override { return false; } unsigned num_inputs() const override { return 2; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } // Actually, if _either_ image has blank alpha, our output will have // blank alpha, too (this only tells the framework that having _both_ diff --git a/saturation_effect.h b/saturation_effect.h index 7e108e8..8a3aece 100644 --- a/saturation_effect.h +++ b/saturation_effect.h @@ -18,7 +18,7 @@ public: SaturationEffect(); std::string effect_type_id() const override { return "SaturationEffect"; } AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } std::string output_fragment_shader() override; private: diff --git a/version.h b/version.h index e73f188..da4b1ee 100644 --- a/version.h +++ b/version.h @@ -5,6 +5,6 @@ // changes, even within git versions. There is no specific version // documentation outside the regular changelogs, though. -#define MOVIT_VERSION 33 +#define MOVIT_VERSION 34 #endif // !defined(_MOVIT_VERSION_H) diff --git a/vignette_effect.h b/vignette_effect.h index 8a2229c..2239f96 100644 --- a/vignette_effect.h +++ b/vignette_effect.h @@ -19,7 +19,7 @@ public: bool needs_srgb_primaries() const override { return false; } AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } void inform_input_size(unsigned input_num, unsigned width, unsigned height) override; void set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) override; diff --git a/white_balance_effect.h b/white_balance_effect.h index cf2c58b..01dcc7e 100644 --- a/white_balance_effect.h +++ b/white_balance_effect.h @@ -16,7 +16,7 @@ public: WhiteBalanceEffect(); std::string effect_type_id() const override { return "WhiteBalanceEffect"; } AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } std::string output_fragment_shader() override; void set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) override; diff --git a/ycbcr_conversion_effect.h b/ycbcr_conversion_effect.h index ce456ce..752eb43 100644 --- a/ycbcr_conversion_effect.h +++ b/ycbcr_conversion_effect.h @@ -26,7 +26,7 @@ public: std::string output_fragment_shader() override; void set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) override; AlphaHandling alpha_handling() const override { return DONT_CARE_ALPHA_TYPE; } - bool one_to_one_sampling() const override { return true; } + bool strong_one_to_one_sampling() const override { return true; } // Should not be called by end users; call // EffectChain::change_ycbcr_output_format() instead.