From dc0b3a8d450e88d224b2d4b85e775e88f99e7c97 Mon Sep 17 00:00:00 2001 From: Geoffrey McRae Date: Tue, 10 Aug 2021 13:08:54 +1000 Subject: [PATCH] [client] egl: rework post process filters and add AMD FXR --- client/renderers/EGL/CMakeLists.txt | 2 + client/renderers/EGL/desktop.c | 156 ++- client/renderers/EGL/desktop.h | 1 + client/renderers/EGL/egl.c | 25 +- client/renderers/EGL/shader/compat.h | 32 + client/renderers/EGL/shader/desktop.vert | 4 +- client/renderers/EGL/shader/desktop_rgb.frag | 5 +- client/renderers/EGL/shader/ffx_a.h | 1051 ++++++++++++--- client/renderers/EGL/shader/ffx_cas.frag | 19 +- client/renderers/EGL/shader/ffx_fsr1.h | 1199 +++++++++++++++++ .../renderers/EGL/shader/ffx_fsr1_easu.frag | 48 + .../renderers/EGL/shader/ffx_fsr1_rcas.frag | 36 + client/renderers/EGL/texture.c | 88 +- client/renderers/EGL/texture.h | 10 +- 14 files changed, 2441 insertions(+), 235 deletions(-) create mode 100644 client/renderers/EGL/shader/compat.h create mode 100644 client/renderers/EGL/shader/ffx_fsr1.h create mode 100644 client/renderers/EGL/shader/ffx_fsr1_easu.frag create mode 100644 client/renderers/EGL/shader/ffx_fsr1_rcas.frag diff --git a/client/renderers/EGL/CMakeLists.txt b/client/renderers/EGL/CMakeLists.txt index 4d998d31..38836c51 100644 --- a/client/renderers/EGL/CMakeLists.txt +++ b/client/renderers/EGL/CMakeLists.txt @@ -61,6 +61,8 @@ build_shaders( shader/splash_logo.frag shader/basic.vert shader/ffx_cas.frag + shader/ffx_fsr1_easu.frag + shader/ffx_fsr1_rcas.frag ) make_defines( diff --git a/client/renderers/EGL/desktop.c b/client/renderers/EGL/desktop.c index c553f40a..52d7f138 100644 --- a/client/renderers/EGL/desktop.c +++ b/client/renderers/EGL/desktop.c @@ -40,13 +40,15 @@ #include "basic.vert.h" #include "ffx_cas.frag.h" +#include "ffx_fsr1_easu.frag.h" +#include "ffx_fsr1_rcas.frag.h" struct DesktopShader { EGL_Shader * shader; GLint uTransform; GLint uDesktopSize; - GLint uTextureScale; + GLint uTextureSize; GLint uScaleAlgo; GLint uNVGain; GLint uCBMode; @@ -65,6 +67,7 @@ struct EGL_Desktop // internals int width, height; LG_RendererRotate rotate; + bool upscale; // scale algorithm int scaleAlgo; @@ -79,10 +82,15 @@ struct EGL_Desktop bool useDMA; LG_RendererFormat format; + EGL_Shader * ffxFSR1[2]; + bool ffxFSR1Enable; + PostProcessHandle ffxFSR1Handle[2]; + EGL_Uniform ffxFSR1Uniform; + EGL_Shader * ffxCAS; bool ffxCASEnable; PostProcessHandle ffxCASHandle; - EGL_Uniform ffxUniform; + EGL_Uniform ffxCASUniform; }; // forwards @@ -104,20 +112,27 @@ static bool egl_initDesktopShader( return false; } - shader->uTransform = egl_shaderGetUniform(shader->shader, "transform" ); - shader->uDesktopSize = egl_shaderGetUniform(shader->shader, "size" ); - shader->uTextureScale = egl_shaderGetUniform(shader->shader, "textureScale"); - shader->uScaleAlgo = egl_shaderGetUniform(shader->shader, "scaleAlgo" ); - shader->uNVGain = egl_shaderGetUniform(shader->shader, "nvGain" ); - shader->uCBMode = egl_shaderGetUniform(shader->shader, "cbMode" ); + shader->uTransform = egl_shaderGetUniform(shader->shader, "transform" ); + shader->uDesktopSize = egl_shaderGetUniform(shader->shader, "desktopSize"); + shader->uTextureSize = egl_shaderGetUniform(shader->shader, "textureSize"); + shader->uScaleAlgo = egl_shaderGetUniform(shader->shader, "scaleAlgo" ); + shader->uNVGain = egl_shaderGetUniform(shader->shader, "nvGain" ); + shader->uCBMode = egl_shaderGetUniform(shader->shader, "cbMode" ); return true; } static void setupFilters(EGL_Desktop * desktop) { + desktop->ffxFSR1Handle[0] = + egl_textureAddFilter(desktop->texture, desktop->ffxFSR1[0], + desktop->ffxFSR1Enable); + desktop->ffxFSR1Handle[1] = + egl_textureAddFilter(desktop->texture, desktop->ffxFSR1[1], + desktop->ffxFSR1Enable); + desktop->ffxCASHandle = - egl_textureAddFilter(desktop->texture, desktop->ffxCAS, 1.0f, + egl_textureAddFilter(desktop->texture, desktop->ffxCAS, desktop->ffxCASEnable); } @@ -173,18 +188,38 @@ bool egl_desktopInit(EGL * egl, EGL_Desktop ** desktop_, EGLDisplay * display, desktop->scaleAlgo = option_get_int("egl", "scale" ); desktop->useDMA = useDMA; + // AMD FidelidyFX FSR + egl_shaderInit(&desktop->ffxFSR1[0]); + egl_shaderCompile(desktop->ffxFSR1[0], + b_shader_basic_vert , b_shader_basic_vert_size, + b_shader_ffx_fsr1_easu_frag, b_shader_ffx_fsr1_easu_frag_size); + + egl_shaderInit(&desktop->ffxFSR1[1]); + egl_shaderCompile(desktop->ffxFSR1[1], + b_shader_basic_vert , b_shader_basic_vert_size, + b_shader_ffx_fsr1_rcas_frag, b_shader_ffx_fsr1_rcas_frag_size); + + desktop->ffxFSR1Enable = option_get_bool("eglFilter", "ffxFSR"); + desktop->ffxFSR1Uniform.type = EGL_UNIFORM_TYPE_1F; + desktop->ffxFSR1Uniform.location = + egl_shaderGetUniform(desktop->ffxFSR1[1], "uSharpness"); + desktop->ffxFSR1Uniform.f[0] = + option_get_float("eglFilter", "ffxFSRSharpness"); + egl_shaderSetUniforms(desktop->ffxFSR1[1], &desktop->ffxFSR1Uniform, 1); + + // AMD FidelidyFX CAS egl_shaderInit(&desktop->ffxCAS); egl_shaderCompile(desktop->ffxCAS, b_shader_basic_vert , b_shader_basic_vert_size, b_shader_ffx_cas_frag, b_shader_ffx_cas_frag_size); desktop->ffxCASEnable = option_get_bool("eglFilter", "ffxCAS"); - desktop->ffxUniform.type = EGL_UNIFORM_TYPE_1F; - desktop->ffxUniform.location = + desktop->ffxCASUniform.type = EGL_UNIFORM_TYPE_1F; + desktop->ffxCASUniform.location = egl_shaderGetUniform(desktop->ffxCAS, "uSharpness"); - desktop->ffxUniform.f[0] = + desktop->ffxCASUniform.f[0] = option_get_float("eglFilter", "ffxCASSharpness"); - egl_shaderSetUniforms(desktop->ffxCAS, &desktop->ffxUniform, 1); + egl_shaderSetUniforms(desktop->ffxCAS, &desktop->ffxCASUniform, 1); setupFilters(desktop); @@ -222,6 +257,9 @@ void egl_desktopFree(EGL_Desktop ** desktop) egl_shaderFree (&(*desktop)->shader.shader); egl_desktopRectsFree(&(*desktop)->mesh ); countedBufferRelease(&(*desktop)->matrix ); + + egl_shaderFree(&(*desktop)->ffxFSR1[0]); + egl_shaderFree(&(*desktop)->ffxFSR1[1]); egl_shaderFree(&(*desktop)->ffxCAS); free(*desktop); @@ -266,25 +304,64 @@ void egl_desktopConfigUI(EGL_Desktop * desktop) igSliderInt("##nvgain", &desktop->nvGain, 0, desktop->nvMax, format, 0); igPopItemWidth(); - bool invalidateCAS = false; + bool invalidateTex = false; + // AMD FidelityFX FSR + bool fsr1 = desktop->ffxFSR1Enable; + igCheckbox("AMD FidelityFX FSR", &fsr1); + if (fsr1 != desktop->ffxFSR1Enable) + { + desktop->ffxFSR1Enable = fsr1; + egl_textureEnableFilter(desktop->ffxFSR1Handle[0], + fsr1 && desktop->upscale); + egl_textureEnableFilter(desktop->ffxFSR1Handle[1], + fsr1 && desktop->upscale); + invalidateTex = true; + } + + float fsr1Sharpness = desktop->ffxFSR1Uniform.f[0]; + igText("Sharpness:"); + igSameLine(0.0f, -1.0f); + igPushItemWidth(igGetWindowWidth() - igGetCursorPosX() - + igGetStyle()->WindowPadding.x); + igSliderFloat("##fsr1Sharpness", &fsr1Sharpness, 0.0f, 1.0f, NULL, 0); + igPopItemWidth(); + + if (fsr1Sharpness != desktop->ffxFSR1Uniform.f[0]) + { + // enable FSR1 if the sharpness was changed + if (!fsr1) + { + fsr1 = true; + desktop->ffxFSR1Enable = fsr1; + egl_textureEnableFilter(desktop->ffxFSR1Handle[0], + fsr1 && desktop->upscale); + egl_textureEnableFilter(desktop->ffxFSR1Handle[1], + fsr1 && desktop->upscale); + } + desktop->ffxFSR1Uniform.f[0] = fsr1Sharpness; + egl_shaderSetUniforms(desktop->ffxFSR1[1], &desktop->ffxFSR1Uniform, 1); + invalidateTex = true; + } + + // AMD FiedlityFX CAS bool cas = desktop->ffxCASEnable; igCheckbox("AMD FidelityFX CAS", &cas); if (cas != desktop->ffxCASEnable) { desktop->ffxCASEnable = cas; egl_textureEnableFilter(desktop->ffxCASHandle, cas); - invalidateCAS = true; + invalidateTex = true; } - float sharpness = desktop->ffxUniform.f[0]; + float casSharpness = desktop->ffxCASUniform.f[0]; igText("Sharpness:"); igSameLine(0.0f, -1.0f); igPushItemWidth(igGetWindowWidth() - igGetCursorPosX() - igGetStyle()->WindowPadding.x); - igSliderFloat("##casSharpness", &sharpness, 0.0f, 1.0f, NULL, 0); + igSliderFloat("##casSharpness", &casSharpness, 0.0f, 1.0f, NULL, 0); igPopItemWidth(); - if (sharpness != desktop->ffxUniform.f[0]) + if (casSharpness != desktop->ffxCASUniform.f[0]) { // enable CAS if the sharpness was changed if (!cas) @@ -293,12 +370,12 @@ void egl_desktopConfigUI(EGL_Desktop * desktop) desktop->ffxCASEnable = cas; egl_textureEnableFilter(desktop->ffxCASHandle, cas); } - desktop->ffxUniform.f[0] = sharpness; - egl_shaderSetUniforms(desktop->ffxCAS, &desktop->ffxUniform, 1); - invalidateCAS = true; + desktop->ffxCASUniform.f[0] = casSharpness; + egl_shaderSetUniforms(desktop->ffxCAS, &desktop->ffxCASUniform, 1); + invalidateTex = true; } - if (invalidateCAS) + if (invalidateTex) { egl_textureInvalidate(desktop->texture); app_invalidateWindow(true); @@ -379,6 +456,29 @@ bool egl_desktop_update(EGL_Desktop * desktop, const FrameBuffer * frame, int dm return egl_textureUpdateFromFrame(desktop->texture, frame, damageRects, damageRectsCount); } +void egl_desktopResize(EGL_Desktop * desktop, int width, int height) +{ + if (width > desktop->width && height > desktop->height) + { + desktop->upscale = true; + if (desktop->ffxFSR1Enable) + { + egl_textureEnableFilter(desktop->ffxFSR1Handle[0], true); + egl_textureEnableFilter(desktop->ffxFSR1Handle[1], true); + } + egl_textureSetFilterRes(desktop->ffxFSR1Handle[0], width, height); + egl_textureSetFilterRes(desktop->ffxFSR1Handle[1], width, height); + egl_textureSetFilterRes(desktop->ffxCASHandle , width, height); + } + else + { + desktop->upscale = false; + egl_textureEnableFilter(desktop->ffxFSR1Handle[0], false); + egl_textureEnableFilter(desktop->ffxFSR1Handle[1], false); + egl_textureSetFilterRes(desktop->ffxCASHandle, 0, 0); + } +} + bool egl_desktopRender(EGL_Desktop * desktop, const float x, const float y, const float scaleX, const float scaleY, enum EGL_DesktopScaleType scaleType, LG_RendererRotate rotate, const struct DamageRects * rects) @@ -412,12 +512,14 @@ bool egl_desktopRender(EGL_Desktop * desktop, const float x, const float y, scaleAlgo = desktop->scaleAlgo; } + struct Rect finalSize; + egl_textureBind(desktop->texture); + egl_textureGetFinalSize(desktop->texture, &finalSize); + egl_desktopRectsMatrix((float *)desktop->matrix->data, desktop->width, desktop->height, x, y, scaleX, scaleY, rotate); egl_desktopRectsUpdate(desktop->mesh, rects, desktop->width, desktop->height); - egl_textureBind(desktop->texture); - const struct DesktopShader * shader = &desktop->shader; EGL_Uniform uniforms[] = { @@ -432,9 +534,9 @@ bool egl_desktopRender(EGL_Desktop * desktop, const float x, const float y, .f = { desktop->width, desktop->height }, }, { - .type = EGL_UNIFORM_TYPE_1F, - .location = shader->uTextureScale, - .f = { egl_textureGetScale(desktop->texture) }, + .type = EGL_UNIFORM_TYPE_2I, + .location = shader->uTextureSize, + .i = { finalSize.x, finalSize.y }, }, { .type = EGL_UNIFORM_TYPE_M3x2FV, diff --git a/client/renderers/EGL/desktop.h b/client/renderers/EGL/desktop.h index 484afd67..7c70b3bc 100644 --- a/client/renderers/EGL/desktop.h +++ b/client/renderers/EGL/desktop.h @@ -45,6 +45,7 @@ void egl_desktopConfigUI(EGL_Desktop * desktop); bool egl_desktopSetup (EGL_Desktop * desktop, const LG_RendererFormat format); bool egl_desktop_update(EGL_Desktop * desktop, const FrameBuffer * frame, int dmaFd, const FrameDamageRect * damageRects, int damageRectsCount); +void egl_desktopResize(EGL_Desktop * desktop, int width, int height); bool egl_desktopRender(EGL_Desktop * desktop, const float x, const float y, const float scaleX, const float scaleY, enum EGL_DesktopScaleType scaleType, LG_RendererRotate rotate, const struct DamageRects * rects); diff --git a/client/renderers/EGL/egl.c b/client/renderers/EGL/egl.c index 34e8b1a0..be9a0305 100644 --- a/client/renderers/EGL/egl.c +++ b/client/renderers/EGL/egl.c @@ -183,11 +183,25 @@ static struct Option egl_options[] = }, { - .module = "eglFilter", - .name = "ffxCAS", - .description = "AMD FidelityFX CAS", - .type = OPTION_TYPE_BOOL, - .value.x_bool = false + .module = "eglFilter", + .name = "ffxFSR", + .description = "AMD FidelityFX FSR", + .type = OPTION_TYPE_BOOL, + .value.x_bool = false + }, + { + .module = "eglFilter", + .name = "ffxFSRSharpness", + .description = "AMD FidelityFX FSR Sharpness", + .type = OPTION_TYPE_FLOAT, + .value.x_float = 1.0f + }, + { + .module = "eglFilter", + .name = "ffxCAS", + .description = "AMD FidelityFX CAS", + .type = OPTION_TYPE_BOOL, + .value.x_bool = false }, { .module = "eglFilter", @@ -466,6 +480,7 @@ static void egl_onResize(LG_Renderer * renderer, const int width, const int heig ImGui_ImplOpenGL3_NewFrame(); egl_damageResize(this->damage, this->translateX, this->translateY, this->scaleX, this->scaleY); + egl_desktopResize(this->desktop, this->width, this->height); } static bool egl_onMouseShape(LG_Renderer * renderer, const LG_RendererCursor cursor, diff --git a/client/renderers/EGL/shader/compat.h b/client/renderers/EGL/shader/compat.h new file mode 100644 index 00000000..49958876 --- /dev/null +++ b/client/renderers/EGL/shader/compat.h @@ -0,0 +1,32 @@ +#if __VERSION__ == 300 + vec4 textureGather(sampler2D tex, vec2 uv, int comp) + { + vec4 c0 = textureOffset(tex, uv, ivec2(0,1)); + vec4 c1 = textureOffset(tex, uv, ivec2(1,1)); + vec4 c2 = textureOffset(tex, uv, ivec2(1,0)); + vec4 c3 = textureOffset(tex, uv, ivec2(0,0)); + return vec4(c0[comp], c1[comp], c2[comp],c3[comp]); + } +#elif __VERSION__ < 300 + vec4 textureGather(sampler2D tex, vec2 uv, int comp) + { + vec4 c3 = texture2D(tex, uv); + return vec4(c3[comp], c3[comp], c3[comp],c3[comp]); + } +#endif + +#if __VERSION__ < 310 + uint bitfieldExtract(uint val, int off, int size) + { + uint mask = uint((1 << size) - 1); + return uint(val >> off) & mask; + } + + uint bitfieldInsert(uint a, uint b, int c, int d) + { + uint mask = ~(0xffffffffu << d) << c; + mask = ~mask; + a &= mask; + return a | (b << c); + } +#endif diff --git a/client/renderers/EGL/shader/desktop.vert b/client/renderers/EGL/shader/desktop.vert index 5849b46f..8049940c 100644 --- a/client/renderers/EGL/shader/desktop.vert +++ b/client/renderers/EGL/shader/desktop.vert @@ -3,11 +3,11 @@ layout(location = 0) in vec2 vertex; out highp vec2 uv; -uniform highp vec2 size; +uniform highp vec2 desktopSize; uniform mat3x2 transform; void main() { gl_Position = vec4(transform * vec3(vertex, 1.0), 0.0, 1.0); - uv = vertex / size; + uv = vertex / desktopSize; } diff --git a/client/renderers/EGL/shader/desktop_rgb.frag b/client/renderers/EGL/shader/desktop_rgb.frag index 91629db3..74f26719 100644 --- a/client/renderers/EGL/shader/desktop_rgb.frag +++ b/client/renderers/EGL/shader/desktop_rgb.frag @@ -13,8 +13,7 @@ out highp vec4 color; uniform sampler2D sampler1; uniform int scaleAlgo; -uniform highp vec2 size; -uniform highp float textureScale; +uniform highp ivec2 textureSize; uniform highp float nvGain; uniform int cbMode; @@ -24,7 +23,7 @@ void main() switch (scaleAlgo) { case EGL_SCALE_NEAREST: - color = texelFetch(sampler1, ivec2(uv * size * textureScale), 0); + color = texelFetch(sampler1, ivec2(uv * vec2(textureSize)), 0); break; case EGL_SCALE_LINEAR: diff --git a/client/renderers/EGL/shader/ffx_a.h b/client/renderers/EGL/shader/ffx_a.h index 7da869f5..d04bff55 100644 --- a/client/renderers/EGL/shader/ffx_a.h +++ b/client/renderers/EGL/shader/ffx_a.h @@ -1,22 +1,39 @@ -//_____________________________________________________________/\_______________________________________________________________ //============================================================================================================================== // -// [A] SHADER PORTABILITY 1.20190530 +// [A] SHADER PORTABILITY 1.20210629 // //============================================================================================================================== -// LICENSE -// ======= -// Copyright (c) 2017-2019 Advanced Micro Devices, Inc. All rights reserved. -// Copyright (c) <2014> -// ------- +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +// MIT LICENSE +// =========== +// Copyright (c) 2014 Michal Drobot (for concepts used in "FLOAT APPROXIMATIONS"). +// ----------- // Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation // files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, // modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the // Software is furnished to do so, subject to the following conditions: -// ------- +// ----------- // The above copyright notice and this permission notice shall be included in all copies or substantial portions of the // Software. -// ------- +// ----------- // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE // WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR // COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, @@ -32,6 +49,8 @@ // A_GPU ..... Include the GPU related code. // A_GLSL .... Using GLSL. // A_HLSL .... Using HLSL. +// A_HLSL_6_2 Using HLSL 6.2 with new 'uint16_t' and related types (requires '-enable-16bit-types'). +// A_NO_16_BIT_CAST Don't use instructions that are not availabe in SPIR-V (needed for running A_HLSL_6_2 on Vulkan) // A_GCC ..... Using a GCC compatible compiler (else assume MSVC compatible compiler by default). // ======= // A_BYTE .... Support 8-bit integer. @@ -60,26 +79,11 @@ // TODO // ==== // - Make sure 'ALerp*(a,b,m)' does 'b*m+(-a*m+a)' (2 ops). -// - Add subgroup ops. //------------------------------------------------------------------------------------------------------------------------------ // CHANGE LOG // ========== -// 20190531 - Fixed changed to llabs() because long is int on Windows. -// 20190530 - Updated for new CPU/GPU portability. -// 20190528 - Fix AU1_AH2_x() on HLSL (had incorrectly swapped x and y), fixed asuint() cases. -// 20190527 - Added min3/max3 for low precision for HLSL. -// 20190526 - Updated with half approximations, added ARsq*(), and ASat*() for CPU. -// 20190519 - Added more approximations. -// 20190514 - Added long conversions. -// 20190513 - Added the real BFI moved the other one to ABfiM(). -// 20190507 - Added extra remap useful for 2D reductions. -// 20190507 - Started adding wave ops, add parabolic sin/cos. -// 20190505 - Added ASigned*() and friends, setup more auto-typecast, GLSL extensions, etc. -// 20190504 - Added min3/max3 for 32-bit integers. -// 20190503 - Added type reinterpretation for half. -// 20190416 - Added min3/max3 for half. -// 20190405 - Misc bug fixing. -// 20190404 - Cleaned up color conversion code. Switched "splat" to shorter naming "type_". Misc bug fixing. +// 20200914 - Expanded wave ops and prx code. +// 20200713 - Added [ZOL] section, fixed serious bugs in sRGB and Rec.709 color conversion code, etc. //============================================================================================================================== //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -102,13 +106,6 @@ // // //============================================================================================================================== -// Requires standard C types: stdint.h -// Requires a collection of standard math intrinsics. -// - Requires VS2013 when not using GCC to get exp2() and log2(). -// - https://blogs.msdn.microsoft.com/vcblog/2013/07/19/c99-library-support-in-visual-studio-2013/ -//------------------------------------------------------------------------------------------------------------------------------ -// This provides a minimum subset of functionality compared to the GPU parts. -//============================================================================================================================== #ifdef A_CPU // Supporting user defined overrides. #ifndef A_RESTRICT @@ -155,8 +152,7 @@ // CPU/GPU PORTING // //------------------------------------------------------------------------------------------------------------------------------ -// Hackary to get CPU and GPU to share all setup code, without duplicate code paths. -// Unfortunately this is the level of "ugly" that is required since the languages are very different. +// Get CPU and GPU to share all setup code, without duplicate code paths. // This uses a lower-case prefix for special vector constructs. // - In C restrict pointers are used. // - In the shading language, in/inout/out arguments are used. @@ -253,18 +249,18 @@ //------------------------------------------------------------------------------------------------------------------------------ // TODO // ==== -// - Replace transcendentals with manual versions. +// - Replace transcendentals with manual versions. //============================================================================================================================== #ifdef A_GCC A_STATIC AD1 AAbsD1(AD1 a){return __builtin_fabs(a);} A_STATIC AF1 AAbsF1(AF1 a){return __builtin_fabsf(a);} A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(__builtin_abs(ASU1_(a)));} - A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(__builtin_labs(ASL1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(__builtin_llabs(ASL1_(a)));} #else A_STATIC AD1 AAbsD1(AD1 a){return fabs(a);} A_STATIC AF1 AAbsF1(AF1 a){return fabsf(a);} A_STATIC AU1 AAbsSU1(AU1 a){return AU1_(abs(ASU1_(a)));} - A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(llabs(ASL1_(a)));} + A_STATIC AL1 AAbsSL1(AL1 a){return AL1_(labs((long)ASL1_(a)));} #endif //------------------------------------------------------------------------------------------------------------------------------ #ifdef A_GCC @@ -314,7 +310,7 @@ A_STATIC AL1 AMaxL1(AL1 a,AL1 b){return a>b?a:b;} A_STATIC AU1 AMaxU1(AU1 a,AU1 b){return a>b?a:b;} //------------------------------------------------------------------------------------------------------------------------------ - // These follow the convention that A integer types don't have signage, until they are operated on. + // These follow the convention that A integer types don't have signage, until they are operated on. A_STATIC AL1 AMaxSL1(AL1 a,AL1 b){return (ASL1_(a)>ASL1_(b))?a:b;} A_STATIC AU1 AMaxSU1(AU1 a,AU1 b){return (ASU1_(a)>ASU1_(b))?a:b;} //------------------------------------------------------------------------------------------------------------------------------ @@ -353,6 +349,9 @@ //============================================================================================================================== // SCALAR RETURN OPS - DEPENDENT //============================================================================================================================== + A_STATIC AD1 AClampD1(AD1 x,AD1 n,AD1 m){return AMaxD1(n,AMinD1(x,m));} + A_STATIC AF1 AClampF1(AF1 x,AF1 n,AF1 m){return AMaxF1(n,AMinF1(x,m));} +//------------------------------------------------------------------------------------------------------------------------------ A_STATIC AD1 AFractD1(AD1 a){return a-AFloorD1(a);} A_STATIC AF1 AFractF1(AF1 a){return a-AFloorF1(a);} //------------------------------------------------------------------------------------------------------------------------------ @@ -388,6 +387,14 @@ A_STATIC retAF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];return d;} A_STATIC retAF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];return d;} A_STATIC retAF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d[0]=a[0]+b[0];d[1]=a[1]+b[1];d[2]=a[2]+b[2];d[3]=a[3]+b[3];return d;} +//============================================================================================================================== + A_STATIC retAD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + A_STATIC retAF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;return d;} + A_STATIC retAF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;return d;} + A_STATIC retAF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d[0]=a[0]+b;d[1]=a[1]+b;d[2]=a[2]+b;d[3]=a[3]+b;return d;} //============================================================================================================================== A_STATIC retAD2 opACpyD2(outAD2 d,inAD2 a){d[0]=a[0];d[1]=a[1];return d;} A_STATIC retAD3 opACpyD3(outAD3 d,inAD3 a){d[0]=a[0];d[1]=a[1];d[2]=a[2];return d;} @@ -562,12 +569,11 @@ #ifndef A_SKIP_EXT #ifdef A_HALF #extension GL_EXT_shader_16bit_storage:require - #extension GL_EXT_shader_explicit_arithmetic_types:require + #extension GL_EXT_shader_explicit_arithmetic_types:require #endif //------------------------------------------------------------------------------------------------------------------------------ #ifdef A_LONG #extension GL_ARB_gpu_shader_int64:require - // TODO: Fixme to more portable extension!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! #extension GL_NV_shader_atomic_int64:require #endif //------------------------------------------------------------------------------------------------------------------------------ @@ -608,6 +614,9 @@ #define AU2_AF2(x) floatBitsToUint(AF2(x)) #define AU3_AF3(x) floatBitsToUint(AF3(x)) #define AU4_AF4(x) floatBitsToUint(AF4(x)) +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AU1_AH1_AF1_x(AF1 a){return packHalf2x16(AF2(a,0.0));} + #define AU1_AH1_AF1(a) AU1_AH1_AF1_x(AF1(a)) //------------------------------------------------------------------------------------------------------------------------------ #define AU1_AH2_AF2 packHalf2x16 #define AU1_AW2Unorm_AF2 packUnorm2x16 @@ -644,6 +653,12 @@ AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} // Proxy for V_BFI_B32 where the 'mask' is set as 'bits', 'mask=(1<>16));} @@ -1062,9 +1119,14 @@ AU3 AAbsSU3(AU3 a){return AU3(abs(ASU3(a)));} AU4 AAbsSU4(AU4 a){return AU4(abs(ASU4(a)));} //------------------------------------------------------------------------------------------------------------------------------ - AU1 ABfe(AU1 src,AU1 off,AU1 bits){AU1 mask=(1<>off)&mask;} + AU1 ABfe(AU1 src,AU1 off,AU1 bits){AU1 mask=(1u<>off)&mask;} AU1 ABfi(AU1 src,AU1 ins,AU1 mask){return (ins&mask)|(src&(~mask));} - AU1 ABfiM(AU1 src,AU1 ins,AU1 bits){AU1 mask=(1<> 8)&0xffu)|((i.y<< 8)&0xff0000u);} + AU1 APerm0G0C(AU2 i){return((i.x>>16)&0xffu)|((i.y )&0xff0000u);} + AU1 APerm0H0D(AU2 i){return((i.x>>24)&0xffu)|((i.y>> 8)&0xff0000u);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermHGFA(AU2 i){return((i.x )&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGFC(AU2 i){return((i.x>>16)&0x000000ffu)|(i.y&0xffffff00u);} + AU1 APermHGAE(AU2 i){return((i.x<< 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHGCE(AU2 i){return((i.x>> 8)&0x0000ff00u)|(i.y&0xffff00ffu);} + AU1 APermHAFE(AU2 i){return((i.x<<16)&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermHCFE(AU2 i){return((i.x )&0x00ff0000u)|(i.y&0xff00ffffu);} + AU1 APermAGFE(AU2 i){return((i.x<<24)&0xff000000u)|(i.y&0x00ffffffu);} + AU1 APermCGFE(AU2 i){return((i.x<< 8)&0xff000000u)|(i.y&0x00ffffffu);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 APermGCEA(AU2 i){return((i.x)&0x00ff00ffu)|((i.y<<8)&0xff00ff00u);} + AU1 APermGECA(AU2 i){return(((i.x)&0xffu)|((i.x>>8)&0xff00u)|((i.y<<16)&0xff0000u)|((i.y<<8)&0xff000000u));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BUC] BYTE UNSIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Designed to use the optimal conversion, enables the scaling to possibly be factored into other computation. +// Works on a range of {0 to A_BUC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// OPCODE NOTES +// ============ +// GCN does not do UNORM or SNORM for bytes in opcodes. +// - V_CVT_F32_UBYTE{0,1,2,3} - Unsigned byte to float. +// - V_CVT_PKACC_U8_F32 - Float to unsigned byte (does bit-field insert into 32-bit integer). +// V_PERM_B32 does byte packing with ability to zero fill bytes as well. +// - Can pull out byte values from two sources, and zero fill upper 8-bits of packed hi and lo. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U1() - Designed for V_CVT_F32_UBYTE* and V_CVT_PKACCUM_U8_F32 ops. +// ==== ===== +// 0 : 0 +// 1 : 1 +// ... +// 255 : 255 +// : 256 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABuc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : 0 +// 1 : 1/512 +// 2 : 1/256 +// ... +// 64 : 1/8 +// 128 : 1/4 +// 255 : 255/512 +// : 1/2 (just outside the encoding range) +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMAL IMPLEMENTATIONS ON AMD ARCHITECTURES +// ============================================ +// r=ABuc0FromU1(i) +// V_CVT_F32_UBYTE0 r,i +// -------------------------------------------- +// r=ABuc0ToU1(d,i) +// V_CVT_PKACCUM_U8_F32 r,i,0,d +// -------------------------------------------- +// d=ABuc0FromU2(i) +// Where 'k0' is an SGPR with 0x0E0A +// Where 'k1' is an SGPR with {32768.0} packed into the lower 16-bits +// V_PERM_B32 d,i.x,i.y,k0 +// V_PK_FMA_F16 d,d,k1.x,0 +// -------------------------------------------- +// r=ABuc0ToU2(d,i) +// Where 'k0' is an SGPR with {1.0/32768.0} packed into the lower 16-bits +// Where 'k1' is an SGPR with 0x???? +// Where 'k2' is an SGPR with 0x???? +// V_PK_FMA_F16 i,i,k0.x,0 +// V_PERM_B32 r.x,i,i,k1 +// V_PERM_B32 r.y,i,i,k2 +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BUC_32 (255.0) + #define A_BUC_16 (255.0/512.0) +//============================================================================================================================== + #if 1 + // Designed to be one V_CVT_PKACCUM_U8_F32. + // The extra min is required to pattern match to V_CVT_PKACCUM_U8_F32. + AU1 ABuc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i),255u) )&(0x000000ffu));} + AU1 ABuc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i),255u)<< 8)&(0x0000ff00u));} + AU1 ABuc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i),255u)<<16)&(0x00ff0000u));} + AU1 ABuc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed to be one V_CVT_F32_UBYTE*. + AF1 ABuc0FromU1(AU1 i){return AF1((i )&255u);} + AF1 ABuc1FromU1(AU1 i){return AF1((i>> 8)&255u);} + AF1 ABuc2FromU1(AU1 i){return AF1((i>>16)&255u);} + AF1 ABuc3FromU1(AU1 i){return AF1((i>>24)&255u);} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABuc01ToW2(AH2 x,AH2 y){x*=AH2_(1.0/32768.0);y*=AH2_(1.0/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 3 ops to do SOA to AOS and conversion. + AU2 ABuc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABuc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABuc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABuc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + // Designed for 2 ops to do both AOS to SOA, and conversion. + AH2 ABuc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0);} + AH2 ABuc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0);} + AH2 ABuc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0);} + AH2 ABuc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0);} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [BSC] BYTE SIGNED CONVERSION +//------------------------------------------------------------------------------------------------------------------------------ +// Similar to [BUC]. +// Works on a range of {-/+ A_BSC_<32,16>}, for <32-bit, and 16-bit> respectively. +//------------------------------------------------------------------------------------------------------------------------------ +// ENCODING (without zero-based encoding) +// ======== +// 0 = unused (can be used to mean something else) +// 1 = lowest value +// 128 = exact zero center (zero based encoding +// 255 = highest value +//------------------------------------------------------------------------------------------------------------------------------ +// Zero-based [Zb] flips the MSB bit of the byte (making 128 "exact zero" actually zero). +// This is useful if there is a desire for cleared values to decode as zero. +//------------------------------------------------------------------------------------------------------------------------------ +// BYTE : FLOAT - ABsc{0,1,2,3}{To,From}U2() - Designed for 16-bit denormal tricks and V_PERM_B32. +// ==== ===== +// 0 : -127/512 (unused) +// 1 : -126/512 +// 2 : -125/512 +// ... +// 128 : 0 +// ... +// 255 : 127/512 +// : 1/4 (just outside the encoding range) +//============================================================================================================================== + // Peak range for 32-bit and 16-bit operations. + #define A_BSC_32 (127.0) + #define A_BSC_16 (127.0/512.0) +//============================================================================================================================== + #if 1 + AU1 ABsc0ToU1(AU1 d,AF1 i){return (d&0xffffff00u)|((min(AU1(i+128.0),255u) )&(0x000000ffu));} + AU1 ABsc1ToU1(AU1 d,AF1 i){return (d&0xffff00ffu)|((min(AU1(i+128.0),255u)<< 8)&(0x0000ff00u));} + AU1 ABsc2ToU1(AU1 d,AF1 i){return (d&0xff00ffffu)|((min(AU1(i+128.0),255u)<<16)&(0x00ff0000u));} + AU1 ABsc3ToU1(AU1 d,AF1 i){return (d&0x00ffffffu)|((min(AU1(i+128.0),255u)<<24)&(0xff000000u));} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 ABsc0ToZbU1(AU1 d,AF1 i){return ((d&0xffffff00u)|((min(AU1(trunc(i)+128.0),255u) )&(0x000000ffu)))^0x00000080u;} + AU1 ABsc1ToZbU1(AU1 d,AF1 i){return ((d&0xffff00ffu)|((min(AU1(trunc(i)+128.0),255u)<< 8)&(0x0000ff00u)))^0x00008000u;} + AU1 ABsc2ToZbU1(AU1 d,AF1 i){return ((d&0xff00ffffu)|((min(AU1(trunc(i)+128.0),255u)<<16)&(0x00ff0000u)))^0x00800000u;} + AU1 ABsc3ToZbU1(AU1 d,AF1 i){return ((d&0x00ffffffu)|((min(AU1(trunc(i)+128.0),255u)<<24)&(0xff000000u)))^0x80000000u;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromU1(AU1 i){return AF1((i )&255u)-128.0;} + AF1 ABsc1FromU1(AU1 i){return AF1((i>> 8)&255u)-128.0;} + AF1 ABsc2FromU1(AU1 i){return AF1((i>>16)&255u)-128.0;} + AF1 ABsc3FromU1(AU1 i){return AF1((i>>24)&255u)-128.0;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ABsc0FromZbU1(AU1 i){return AF1(((i )&255u)^0x80u)-128.0;} + AF1 ABsc1FromZbU1(AU1 i){return AF1(((i>> 8)&255u)^0x80u)-128.0;} + AF1 ABsc2FromZbU1(AU1 i){return AF1(((i>>16)&255u)^0x80u)-128.0;} + AF1 ABsc3FromZbU1(AU1 i){return AF1(((i>>24)&255u)^0x80u)-128.0;} + #endif +//============================================================================================================================== + #ifdef A_HALF + // Takes {x0,x1} and {y0,y1} and builds {{x0,y0},{x1,y1}}. + AW2 ABsc01ToW2(AH2 x,AH2 y){x=x*AH2_(1.0/32768.0)+AH2_(0.25/32768.0);y=y*AH2_(1.0/32768.0)+AH2_(0.25/32768.0); + return AW2_AU1(APermGCEA(AU2(AU1_AW2(AW2_AH2(x)),AU1_AW2(AW2_AH2(y)))));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0))); + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AU2 ABsc0ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGFA(AU2(d.x,b)),APermHGFC(AU2(d.y,b)));} + AU2 ABsc1ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHGAE(AU2(d.x,b)),APermHGCE(AU2(d.y,b)));} + AU2 ABsc2ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermHAFE(AU2(d.x,b)),APermHCFE(AU2(d.y,b)));} + AU2 ABsc3ToZbU2(AU2 d,AH2 i){AU1 b=AU1_AW2(AW2_AH2(i*AH2_(1.0/32768.0)+AH2_(0.25/32768.0)))^0x00800080u; + return AU2(APermAGFE(AU2(d.x,b)),APermCGFE(AU2(d.y,b)));} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)))*AH2_(32768.0)-AH2_(0.25);} +//------------------------------------------------------------------------------------------------------------------------------ + AH2 ABsc0FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0E0A(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc1FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0F0B(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc2FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0G0C(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} + AH2 ABsc3FromZbU2(AU2 i){return AH2_AW2(AW2_AU1(APerm0H0D(i)^0x00800080u))*AH2_(32768.0)-AH2_(0.25);} #endif //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -1409,31 +1794,33 @@ //------------------------------------------------------------------------------------------------------------------------------ // The one Newton Raphson iteration form of rsq() was skipped (requires 6 ops total). // Same with sqrt(), as this could be x*rsq() (7 ops). -//------------------------------------------------------------------------------------------------------------------------------ -// IDEAS -// ===== -// - Polaris hardware has 16-bit support, but non-double rate. -// Could be possible still get part double rate for some of this logic, -// by clearing out the lower half's sign when necessary and using 32-bit ops... //============================================================================================================================== #ifdef A_HALF // Minimize squared error across full positive range, 2 ops. // The 0x1de2 based approximation maps {0 to 1} input maps to < 1 output. AH1 APrxLoSqrtH1(AH1 a){return AH1_AW1((AW1_AH1(a)>>AW1_(1))+AW1_(0x1de2));} AH2 APrxLoSqrtH2(AH2 a){return AH2_AW2((AW2_AH2(a)>>AW2_(1))+AW2_(0x1de2));} + AH3 APrxLoSqrtH3(AH3 a){return AH3_AW3((AW3_AH3(a)>>AW3_(1))+AW3_(0x1de2));} + AH4 APrxLoSqrtH4(AH4 a){return AH4_AW4((AW4_AH4(a)>>AW4_(1))+AW4_(0x1de2));} //------------------------------------------------------------------------------------------------------------------------------ // Lower precision estimation, 1 op. // Minimize squared error across {smallest normal to 16384.0}. AH1 APrxLoRcpH1(AH1 a){return AH1_AW1(AW1_(0x7784)-AW1_AH1(a));} AH2 APrxLoRcpH2(AH2 a){return AH2_AW2(AW2_(0x7784)-AW2_AH2(a));} + AH3 APrxLoRcpH3(AH3 a){return AH3_AW3(AW3_(0x7784)-AW3_AH3(a));} + AH4 APrxLoRcpH4(AH4 a){return AH4_AW4(AW4_(0x7784)-AW4_AH4(a));} //------------------------------------------------------------------------------------------------------------------------------ // Medium precision estimation, one Newton Raphson iteration, 3 ops. AH1 APrxMedRcpH1(AH1 a){AH1 b=AH1_AW1(AW1_(0x778d)-AW1_AH1(a));return b*(-b*a+AH1_(2.0));} AH2 APrxMedRcpH2(AH2 a){AH2 b=AH2_AW2(AW2_(0x778d)-AW2_AH2(a));return b*(-b*a+AH2_(2.0));} + AH3 APrxMedRcpH3(AH3 a){AH3 b=AH3_AW3(AW3_(0x778d)-AW3_AH3(a));return b*(-b*a+AH3_(2.0));} + AH4 APrxMedRcpH4(AH4 a){AH4 b=AH4_AW4(AW4_(0x778d)-AW4_AH4(a));return b*(-b*a+AH4_(2.0));} //------------------------------------------------------------------------------------------------------------------------------ // Minimize squared error across {smallest normal to 16384.0}, 2 ops. AH1 APrxLoRsqH1(AH1 a){return AH1_AW1(AW1_(0x59a3)-(AW1_AH1(a)>>AW1_(1)));} AH2 APrxLoRsqH2(AH2 a){return AH2_AW2(AW2_(0x59a3)-(AW2_AH2(a)>>AW2_(1)));} + AH3 APrxLoRsqH3(AH3 a){return AH3_AW3(AW3_(0x59a3)-(AW3_AH3(a)>>AW3_(1)));} + AH4 APrxLoRsqH4(AH4 a){return AH4_AW4(AW4_(0x59a3)-(AW4_AH4(a)>>AW4_(1)));} #endif //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -1456,6 +1843,75 @@ AF1 APrxLoRcpF1(AF1 a){return AF1_AU1(AU1_(0x7ef07ebb)-AU1_AF1(a));} AF1 APrxMedRcpF1(AF1 a){AF1 b=AF1_AU1(AU1_(0x7ef19fff)-AU1_AF1(a));return b*(-b*a+AF1_(2.0));} AF1 APrxLoRsqF1(AF1 a){return AF1_AU1(AU1_(0x5f347d74)-(AU1_AF1(a)>>AU1_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxLoSqrtF2(AF2 a){return AF2_AU2((AU2_AF2(a)>>AU2_(1))+AU2_(0x1fbc4639));} + AF2 APrxLoRcpF2(AF2 a){return AF2_AU2(AU2_(0x7ef07ebb)-AU2_AF2(a));} + AF2 APrxMedRcpF2(AF2 a){AF2 b=AF2_AU2(AU2_(0x7ef19fff)-AU2_AF2(a));return b*(-b*a+AF2_(2.0));} + AF2 APrxLoRsqF2(AF2 a){return AF2_AU2(AU2_(0x5f347d74)-(AU2_AF2(a)>>AU2_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxLoSqrtF3(AF3 a){return AF3_AU3((AU3_AF3(a)>>AU3_(1))+AU3_(0x1fbc4639));} + AF3 APrxLoRcpF3(AF3 a){return AF3_AU3(AU3_(0x7ef07ebb)-AU3_AF3(a));} + AF3 APrxMedRcpF3(AF3 a){AF3 b=AF3_AU3(AU3_(0x7ef19fff)-AU3_AF3(a));return b*(-b*a+AF3_(2.0));} + AF3 APrxLoRsqF3(AF3 a){return AF3_AU3(AU3_(0x5f347d74)-(AU3_AF3(a)>>AU3_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxLoSqrtF4(AF4 a){return AF4_AU4((AU4_AF4(a)>>AU4_(1))+AU4_(0x1fbc4639));} + AF4 APrxLoRcpF4(AF4 a){return AF4_AU4(AU4_(0x7ef07ebb)-AU4_AF4(a));} + AF4 APrxMedRcpF4(AF4 a){AF4 b=AF4_AU4(AU4_(0x7ef19fff)-AU4_AF4(a));return b*(-b*a+AF4_(2.0));} + AF4 APrxLoRsqF4(AF4 a){return AF4_AU4(AU4_(0x5f347d74)-(AU4_AF4(a)>>AU4_(1)));} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PQ APPROXIMATIONS +//------------------------------------------------------------------------------------------------------------------------------ +// PQ is very close to x^(1/8). The functions below Use the fast float approximation method to do +// PQ<~>Gamma2 (4th power and fast 4th root) and PQ<~>Linear (8th power and fast 8th root). Maximum error is ~0.2%. +//============================================================================================================================== +// Helpers + AF1 Quart(AF1 a) { a = a * a; return a * a;} + AF1 Oct(AF1 a) { a = a * a; a = a * a; return a * a; } + AF2 Quart(AF2 a) { a = a * a; return a * a; } + AF2 Oct(AF2 a) { a = a * a; a = a * a; return a * a; } + AF3 Quart(AF3 a) { a = a * a; return a * a; } + AF3 Oct(AF3 a) { a = a * a; a = a * a; return a * a; } + AF4 Quart(AF4 a) { a = a * a; return a * a; } + AF4 Oct(AF4 a) { a = a * a; a = a * a; return a * a; } + //------------------------------------------------------------------------------------------------------------------------------ + AF1 APrxPQToGamma2(AF1 a) { return Quart(a); } + AF1 APrxPQToLinear(AF1 a) { return Oct(a); } + AF1 APrxLoGamma2ToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); } + AF1 APrxMedGamma2ToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(2)) + AU1_(0x2F9A4E46)); AF1 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF1 APrxHighGamma2ToPQ(AF1 a) { return sqrt(sqrt(a)); } + AF1 APrxLoLinearToPQ(AF1 a) { return AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); } + AF1 APrxMedLinearToPQ(AF1 a) { AF1 b = AF1_AU1((AU1_AF1(a) >> AU1_(3)) + AU1_(0x378D8723)); AF1 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF1 APrxHighLinearToPQ(AF1 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF2 APrxPQToGamma2(AF2 a) { return Quart(a); } + AF2 APrxPQToLinear(AF2 a) { return Oct(a); } + AF2 APrxLoGamma2ToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); } + AF2 APrxMedGamma2ToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(2)) + AU2_(0x2F9A4E46)); AF2 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF2 APrxHighGamma2ToPQ(AF2 a) { return sqrt(sqrt(a)); } + AF2 APrxLoLinearToPQ(AF2 a) { return AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); } + AF2 APrxMedLinearToPQ(AF2 a) { AF2 b = AF2_AU2((AU2_AF2(a) >> AU2_(3)) + AU2_(0x378D8723)); AF2 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF2 APrxHighLinearToPQ(AF2 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF3 APrxPQToGamma2(AF3 a) { return Quart(a); } + AF3 APrxPQToLinear(AF3 a) { return Oct(a); } + AF3 APrxLoGamma2ToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); } + AF3 APrxMedGamma2ToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(2)) + AU3_(0x2F9A4E46)); AF3 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF3 APrxHighGamma2ToPQ(AF3 a) { return sqrt(sqrt(a)); } + AF3 APrxLoLinearToPQ(AF3 a) { return AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); } + AF3 APrxMedLinearToPQ(AF3 a) { AF3 b = AF3_AU3((AU3_AF3(a) >> AU3_(3)) + AU3_(0x378D8723)); AF3 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF3 APrxHighLinearToPQ(AF3 a) { return sqrt(sqrt(sqrt(a))); } + //------------------------------------------------------------------------------------------------------------------------------ + AF4 APrxPQToGamma2(AF4 a) { return Quart(a); } + AF4 APrxPQToLinear(AF4 a) { return Oct(a); } + AF4 APrxLoGamma2ToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); } + AF4 APrxMedGamma2ToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(2)) + AU4_(0x2F9A4E46)); AF4 b4 = Quart(b); return b - b * (b4 - a) / (AF1_(4.0) * b4); } + AF4 APrxHighGamma2ToPQ(AF4 a) { return sqrt(sqrt(a)); } + AF4 APrxLoLinearToPQ(AF4 a) { return AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); } + AF4 APrxMedLinearToPQ(AF4 a) { AF4 b = AF4_AU4((AU4_AF4(a) >> AU4_(3)) + AU4_(0x378D8723)); AF4 b8 = Oct(b); return b - b * (b8 - a) / (AF1_(8.0) * b8); } + AF4 APrxHighLinearToPQ(AF4 a) { return sqrt(sqrt(sqrt(a))); } //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //_____________________________________________________________/\_______________________________________________________________ @@ -1464,21 +1920,195 @@ //------------------------------------------------------------------------------------------------------------------------------ // Approximate answers to transcendental questions. //------------------------------------------------------------------------------------------------------------------------------ -// TODO -// ==== -// - Verify packed math ABS is correctly doing an AND. //============================================================================================================================== - // Valid input range is {-1 to 1} representing {0 to 2 pi}. - // Output range is {-1/4 to -1/4} representing {-1 to 1}. - AF1 APSinF1(AF1 x){return x*abs(x)-x;} // MAD. - AF1 APCosF1(AF1 x){x=AFractF1(x*AF1_(0.5)+AF1_(0.75));x=x*AF1_(2.0)-AF1_(1.0);return APSinF1(x);} // 3x MAD, FRACT + #if 1 + // Valid input range is {-1 to 1} representing {0 to 2 pi}. + // Output range is {-1/4 to 1/4} representing {-1 to 1}. + AF1 APSinF1(AF1 x){return x*abs(x)-x;} // MAD. + AF2 APSinF2(AF2 x){return x*abs(x)-x;} + AF1 APCosF1(AF1 x){x=AFractF1(x*AF1_(0.5)+AF1_(0.75));x=x*AF1_(2.0)-AF1_(1.0);return APSinF1(x);} // 3x MAD, FRACT + AF2 APCosF2(AF2 x){x=AFractF2(x*AF2_(0.5)+AF2_(0.75));x=x*AF2_(2.0)-AF2_(1.0);return APSinF2(x);} + AF2 APSinCosF1(AF1 x){AF1 y=AFractF1(x*AF1_(0.5)+AF1_(0.75));y=y*AF1_(2.0)-AF1_(1.0);return APSinF2(AF2(x,y));} + #endif //------------------------------------------------------------------------------------------------------------------------------ #ifdef A_HALF // For a packed {sin,cos} pair, // - Native takes 16 clocks and 4 issue slots (no packed transcendentals). // - Parabolic takes 8 clocks and 8 issue slots (only fract is non-packed). + AH1 APSinH1(AH1 x){return x*abs(x)-x;} AH2 APSinH2(AH2 x){return x*abs(x)-x;} // AND,FMA + AH1 APCosH1(AH1 x){x=AFractH1(x*AH1_(0.5)+AH1_(0.75));x=x*AH1_(2.0)-AH1_(1.0);return APSinH1(x);} AH2 APCosH2(AH2 x){x=AFractH2(x*AH2_(0.5)+AH2_(0.75));x=x*AH2_(2.0)-AH2_(1.0);return APSinH2(x);} // 3x FMA, 2xFRACT, AND + AH2 APSinCosH1(AH1 x){AH1 y=AFractH1(x*AH1_(0.5)+AH1_(0.75));y=y*AH1_(2.0)-AH1_(1.0);return APSinH2(AH2(x,y));} + #endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// [ZOL] ZERO ONE LOGIC +//------------------------------------------------------------------------------------------------------------------------------ +// Conditional free logic designed for easy 16-bit packing, and backwards porting to 32-bit. +//------------------------------------------------------------------------------------------------------------------------------ +// 0 := false +// 1 := true +//------------------------------------------------------------------------------------------------------------------------------ +// AndNot(x,y) -> !(x&y) .... One op. +// AndOr(x,y,z) -> (x&y)|z ... One op. +// GtZero(x) -> x>0.0 ..... One op. +// Sel(x,y,z) -> x?y:z ..... Two ops, has no precision loss. +// Signed(x) -> x<0.0 ..... One op. +// ZeroPass(x,y) -> x?0:y ..... Two ops, 'y' is a pass through safe for aliasing as integer. +//------------------------------------------------------------------------------------------------------------------------------ +// OPTIMIZATION NOTES +// ================== +// - On Vega to use 2 constants in a packed op, pass in as one AW2 or one AH2 'k.xy' and use as 'k.xx' and 'k.yy'. +// For example 'a.xy*k.xx+k.yy'. +//============================================================================================================================== + #if 1 + AU1 AZolAndU1(AU1 x,AU1 y){return min(x,y);} + AU2 AZolAndU2(AU2 x,AU2 y){return min(x,y);} + AU3 AZolAndU3(AU3 x,AU3 y){return min(x,y);} + AU4 AZolAndU4(AU4 x,AU4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolNotU1(AU1 x){return x^AU1_(1);} + AU2 AZolNotU2(AU2 x){return x^AU2_(1);} + AU3 AZolNotU3(AU3 x){return x^AU3_(1);} + AU4 AZolNotU4(AU4 x){return x^AU4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AU1 AZolOrU1(AU1 x,AU1 y){return max(x,y);} + AU2 AZolOrU2(AU2 x,AU2 y){return max(x,y);} + AU3 AZolOrU3(AU3 x,AU3 y){return max(x,y);} + AU4 AZolOrU4(AU4 x,AU4 y){return max(x,y);} +//============================================================================================================================== + AU1 AZolF1ToU1(AF1 x){return AU1(x);} + AU2 AZolF2ToU2(AF2 x){return AU2(x);} + AU3 AZolF3ToU3(AF3 x){return AU3(x);} + AU4 AZolF4ToU4(AF4 x){return AU4(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // 2 ops, denormals don't work in 32-bit on PC (and if they are enabled, OMOD is disabled). + AU1 AZolNotF1ToU1(AF1 x){return AU1(AF1_(1.0)-x);} + AU2 AZolNotF2ToU2(AF2 x){return AU2(AF2_(1.0)-x);} + AU3 AZolNotF3ToU3(AF3 x){return AU3(AF3_(1.0)-x);} + AU4 AZolNotF4ToU4(AF4 x){return AU4(AF4_(1.0)-x);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolU1ToF1(AU1 x){return AF1(x);} + AF2 AZolU2ToF2(AU2 x){return AF2(x);} + AF3 AZolU3ToF3(AU3 x){return AF3(x);} + AF4 AZolU4ToF4(AU4 x){return AF4(x);} +//============================================================================================================================== + AF1 AZolAndF1(AF1 x,AF1 y){return min(x,y);} + AF2 AZolAndF2(AF2 x,AF2 y){return min(x,y);} + AF3 AZolAndF3(AF3 x,AF3 y){return min(x,y);} + AF4 AZolAndF4(AF4 x,AF4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 ASolAndNotF1(AF1 x,AF1 y){return (-x)*y+AF1_(1.0);} + AF2 ASolAndNotF2(AF2 x,AF2 y){return (-x)*y+AF2_(1.0);} + AF3 ASolAndNotF3(AF3 x,AF3 y){return (-x)*y+AF3_(1.0);} + AF4 ASolAndNotF4(AF4 x,AF4 y){return (-x)*y+AF4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolAndOrF1(AF1 x,AF1 y,AF1 z){return ASatF1(x*y+z);} + AF2 AZolAndOrF2(AF2 x,AF2 y,AF2 z){return ASatF2(x*y+z);} + AF3 AZolAndOrF3(AF3 x,AF3 y,AF3 z){return ASatF3(x*y+z);} + AF4 AZolAndOrF4(AF4 x,AF4 y,AF4 z){return ASatF4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolGtZeroF1(AF1 x){return ASatF1(x*AF1_(A_INFP_F));} + AF2 AZolGtZeroF2(AF2 x){return ASatF2(x*AF2_(A_INFP_F));} + AF3 AZolGtZeroF3(AF3 x){return ASatF3(x*AF3_(A_INFP_F));} + AF4 AZolGtZeroF4(AF4 x){return ASatF4(x*AF4_(A_INFP_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolNotF1(AF1 x){return AF1_(1.0)-x;} + AF2 AZolNotF2(AF2 x){return AF2_(1.0)-x;} + AF3 AZolNotF3(AF3 x){return AF3_(1.0)-x;} + AF4 AZolNotF4(AF4 x){return AF4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolOrF1(AF1 x,AF1 y){return max(x,y);} + AF2 AZolOrF2(AF2 x,AF2 y){return max(x,y);} + AF3 AZolOrF3(AF3 x,AF3 y){return max(x,y);} + AF4 AZolOrF4(AF4 x,AF4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSelF1(AF1 x,AF1 y,AF1 z){AF1 r=(-x)*z+z;return x*y+r;} + AF2 AZolSelF2(AF2 x,AF2 y,AF2 z){AF2 r=(-x)*z+z;return x*y+r;} + AF3 AZolSelF3(AF3 x,AF3 y,AF3 z){AF3 r=(-x)*z+z;return x*y+r;} + AF4 AZolSelF4(AF4 x,AF4 y,AF4 z){AF4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolSignedF1(AF1 x){return ASatF1(x*AF1_(A_INFN_F));} + AF2 AZolSignedF2(AF2 x){return ASatF2(x*AF2_(A_INFN_F));} + AF3 AZolSignedF3(AF3 x){return ASatF3(x*AF3_(A_INFN_F));} + AF4 AZolSignedF4(AF4 x){return ASatF4(x*AF4_(A_INFN_F));} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AZolZeroPassF1(AF1 x,AF1 y){return AF1_AU1((AU1_AF1(x)!=AU1_(0))?AU1_(0):AU1_AF1(y));} + AF2 AZolZeroPassF2(AF2 x,AF2 y){return AF2_AU2((AU2_AF2(x)!=AU2_(0))?AU2_(0):AU2_AF2(y));} + AF3 AZolZeroPassF3(AF3 x,AF3 y){return AF3_AU3((AU3_AF3(x)!=AU3_(0))?AU3_(0):AU3_AF3(y));} + AF4 AZolZeroPassF4(AF4 x,AF4 y){return AF4_AU4((AU4_AF4(x)!=AU4_(0))?AU4_(0):AU4_AF4(y));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AW1 AZolAndW1(AW1 x,AW1 y){return min(x,y);} + AW2 AZolAndW2(AW2 x,AW2 y){return min(x,y);} + AW3 AZolAndW3(AW3 x,AW3 y){return min(x,y);} + AW4 AZolAndW4(AW4 x,AW4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolNotW1(AW1 x){return x^AW1_(1);} + AW2 AZolNotW2(AW2 x){return x^AW2_(1);} + AW3 AZolNotW3(AW3 x){return x^AW3_(1);} + AW4 AZolNotW4(AW4 x){return x^AW4_(1);} +//------------------------------------------------------------------------------------------------------------------------------ + AW1 AZolOrW1(AW1 x,AW1 y){return max(x,y);} + AW2 AZolOrW2(AW2 x,AW2 y){return max(x,y);} + AW3 AZolOrW3(AW3 x,AW3 y){return max(x,y);} + AW4 AZolOrW4(AW4 x,AW4 y){return max(x,y);} +//============================================================================================================================== + // Uses denormal trick. + AW1 AZolH1ToW1(AH1 x){return AW1_AH1(x*AH1_AW1(AW1_(1)));} + AW2 AZolH2ToW2(AH2 x){return AW2_AH2(x*AH2_AW2(AW2_(1)));} + AW3 AZolH3ToW3(AH3 x){return AW3_AH3(x*AH3_AW3(AW3_(1)));} + AW4 AZolH4ToW4(AH4 x){return AW4_AH4(x*AH4_AW4(AW4_(1)));} +//------------------------------------------------------------------------------------------------------------------------------ + // AMD arch lacks a packed conversion opcode. + AH1 AZolW1ToH1(AW1 x){return AH1_AW1(x*AW1_AH1(AH1_(1.0)));} + AH2 AZolW2ToH2(AW2 x){return AH2_AW2(x*AW2_AH2(AH2_(1.0)));} + AH3 AZolW1ToH3(AW3 x){return AH3_AW3(x*AW3_AH3(AH3_(1.0)));} + AH4 AZolW2ToH4(AW4 x){return AH4_AW4(x*AW4_AH4(AH4_(1.0)));} +//============================================================================================================================== + AH1 AZolAndH1(AH1 x,AH1 y){return min(x,y);} + AH2 AZolAndH2(AH2 x,AH2 y){return min(x,y);} + AH3 AZolAndH3(AH3 x,AH3 y){return min(x,y);} + AH4 AZolAndH4(AH4 x,AH4 y){return min(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 ASolAndNotH1(AH1 x,AH1 y){return (-x)*y+AH1_(1.0);} + AH2 ASolAndNotH2(AH2 x,AH2 y){return (-x)*y+AH2_(1.0);} + AH3 ASolAndNotH3(AH3 x,AH3 y){return (-x)*y+AH3_(1.0);} + AH4 ASolAndNotH4(AH4 x,AH4 y){return (-x)*y+AH4_(1.0);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolAndOrH1(AH1 x,AH1 y,AH1 z){return ASatH1(x*y+z);} + AH2 AZolAndOrH2(AH2 x,AH2 y,AH2 z){return ASatH2(x*y+z);} + AH3 AZolAndOrH3(AH3 x,AH3 y,AH3 z){return ASatH3(x*y+z);} + AH4 AZolAndOrH4(AH4 x,AH4 y,AH4 z){return ASatH4(x*y+z);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolGtZeroH1(AH1 x){return ASatH1(x*AH1_(A_INFP_H));} + AH2 AZolGtZeroH2(AH2 x){return ASatH2(x*AH2_(A_INFP_H));} + AH3 AZolGtZeroH3(AH3 x){return ASatH3(x*AH3_(A_INFP_H));} + AH4 AZolGtZeroH4(AH4 x){return ASatH4(x*AH4_(A_INFP_H));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolNotH1(AH1 x){return AH1_(1.0)-x;} + AH2 AZolNotH2(AH2 x){return AH2_(1.0)-x;} + AH3 AZolNotH3(AH3 x){return AH3_(1.0)-x;} + AH4 AZolNotH4(AH4 x){return AH4_(1.0)-x;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolOrH1(AH1 x,AH1 y){return max(x,y);} + AH2 AZolOrH2(AH2 x,AH2 y){return max(x,y);} + AH3 AZolOrH3(AH3 x,AH3 y){return max(x,y);} + AH4 AZolOrH4(AH4 x,AH4 y){return max(x,y);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSelH1(AH1 x,AH1 y,AH1 z){AH1 r=(-x)*z+z;return x*y+r;} + AH2 AZolSelH2(AH2 x,AH2 y,AH2 z){AH2 r=(-x)*z+z;return x*y+r;} + AH3 AZolSelH3(AH3 x,AH3 y,AH3 z){AH3 r=(-x)*z+z;return x*y+r;} + AH4 AZolSelH4(AH4 x,AH4 y,AH4 z){AH4 r=(-x)*z+z;return x*y+r;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AZolSignedH1(AH1 x){return ASatH1(x*AH1_(A_INFN_H));} + AH2 AZolSignedH2(AH2 x){return ASatH2(x*AH2_(A_INFN_H));} + AH3 AZolSignedH3(AH3 x){return ASatH3(x*AH3_(A_INFN_H));} + AH4 AZolSignedH4(AH4 x){return ASatH4(x*AH4_(A_INFN_H));} #endif //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -1498,6 +2128,21 @@ // Pq ...... PQ native for HDR10 // Srgb .... The sRGB output, typical of PC displays, useful for 10-bit output, or storing to 8-bit UNORM without SRGB type // Two ..... Gamma 2.0, fastest conversion (useful for intermediate pass approximations) +// Three ... Gamma 3.0, less fast, but good for HDR. +//------------------------------------------------------------------------------------------------------------------------------ +// KEEPING TO SPEC +// =============== +// Both Rec.709 and sRGB have a linear segment which as spec'ed would intersect the curved segment 2 times. +// (a.) For 8-bit sRGB, steps {0 to 10.3} are in the linear region (4% of the encoding range). +// (b.) For 8-bit 709, steps {0 to 20.7} are in the linear region (8% of the encoding range). +// Also there is a slight step in the transition regions. +// Precision of the coefficients in the spec being the likely cause. +// Main usage case of the sRGB code is to do the linear->sRGB converstion in a compute shader before store. +// This is to work around lack of hardware (typically only ROP does the conversion for free). +// To "correct" the linear segment, would be to introduce error, because hardware decode of sRGB->linear is fixed (and free). +// So this header keeps with the spec. +// For linear->sRGB transforms, the linear segment in some respects reduces error, because rounding in that region is linear. +// Rounding in the curved region in hardware (and fast software code) introduces error due to rounding in non-linear. //------------------------------------------------------------------------------------------------------------------------------ // FOR PQ // ====== @@ -1517,51 +2162,133 @@ // ===== // Could be faster for PQ conversions to be in ALU or a texture lookup depending on usage case. //============================================================================================================================== - AF1 ATo709F1(AF1 c){return max(min(c*AF1_(4.5),AF1_(0.018)),AF1_(1.099)*pow(c,AF1_(0.45))-AF1_(0.099));} + #if 1 + AF1 ATo709F1(AF1 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 ATo709F2(AF2 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 ATo709F3(AF3 c){AF3 j=AF3(0.018*4.5,4.5,0.45);AF2 k=AF2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} //------------------------------------------------------------------------------------------------------------------------------ - // Note 'rcpX' is '1/x', where the 'x' is what would be used in AFromGamma(). - AF1 AToGammaF1(AF1 c,AF1 rcpX){return pow(c,rcpX);} + // Note 'rcpX' is '1/x', where the 'x' is what would be used in AFromGamma(). + AF1 AToGammaF1(AF1 c,AF1 rcpX){return pow(c,AF1_(rcpX));} + AF2 AToGammaF2(AF2 c,AF1 rcpX){return pow(c,AF2_(rcpX));} + AF3 AToGammaF3(AF3 c,AF1 rcpX){return pow(c,AF3_(rcpX));} //------------------------------------------------------------------------------------------------------------------------------ - AF1 AToPqF1(AF1 x){AF1 p=pow(x,AF1_(0.159302)); - return pow((AF1_(0.835938)+AF1_(18.8516)*p)/(AF1_(1.0)+AF1_(18.6875)*p),AF1_(78.8438));} + AF1 AToPqF1(AF1 x){AF1 p=pow(x,AF1_(0.159302)); + return pow((AF1_(0.835938)+AF1_(18.8516)*p)/(AF1_(1.0)+AF1_(18.6875)*p),AF1_(78.8438));} + AF2 AToPqF1(AF2 x){AF2 p=pow(x,AF2_(0.159302)); + return pow((AF2_(0.835938)+AF2_(18.8516)*p)/(AF2_(1.0)+AF2_(18.6875)*p),AF2_(78.8438));} + AF3 AToPqF1(AF3 x){AF3 p=pow(x,AF3_(0.159302)); + return pow((AF3_(0.835938)+AF3_(18.8516)*p)/(AF3_(1.0)+AF3_(18.6875)*p),AF3_(78.8438));} //------------------------------------------------------------------------------------------------------------------------------ - AF1 AToSrgbF1(AF1 c){return max(min(c*AF1_(12.92),AF1_(0.0031308)),AF1_(1.055)*pow(c,AF1_(0.41666))-AF1_(0.055));} + AF1 AToSrgbF1(AF1 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AF2 AToSrgbF2(AF2 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AF3 AToSrgbF3(AF3 c){AF3 j=AF3(0.0031308*12.92,12.92,1.0/2.4);AF2 k=AF2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} //------------------------------------------------------------------------------------------------------------------------------ - AF1 AToTwoF1(AF1 c){return sqrt(c);} + AF1 AToTwoF1(AF1 c){return sqrt(c);} + AF2 AToTwoF2(AF2 c){return sqrt(c);} + AF3 AToTwoF3(AF3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AF1 AToThreeF1(AF1 c){return pow(c,AF1_(1.0/3.0));} + AF2 AToThreeF2(AF2 c){return pow(c,AF2_(1.0/3.0));} + AF3 AToThreeF3(AF3 c){return pow(c,AF3_(1.0/3.0));} + #endif //============================================================================================================================== - AF1 AFrom709F1(AF1 c){return max(min(c*AF1_(1.0/4.5),AF1_(0.081)), - pow((c+AF1_(0.099))*(AF1_(1.0)/(AF1_(1.099))),AF1_(1.0/0.45)));} + #if 1 + // Unfortunately median won't work here. + AF1 AFrom709F1(AF1 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFrom709F2(AF2 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFrom709F3(AF3 c){AF3 j=AF3(0.081/4.5,1.0/4.5,1.0/0.45);AF2 k=AF2(1.0/1.099,0.099/1.099); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} //------------------------------------------------------------------------------------------------------------------------------ - AF1 AFromGammaF1(AF1 c,AF1 x){return pow(c,x);} + AF1 AFromGammaF1(AF1 c,AF1 x){return pow(c,AF1_(x));} + AF2 AFromGammaF2(AF2 c,AF1 x){return pow(c,AF2_(x));} + AF3 AFromGammaF3(AF3 c,AF1 x){return pow(c,AF3_(x));} //------------------------------------------------------------------------------------------------------------------------------ - AF1 AFromPqF1(AF1 x){AF1 p=pow(x,AF1_(0.0126833)); - return pow(ASatF1(p-AF1_(0.835938))/(AF1_(18.8516)-AF1_(18.6875)*p),AF1_(6.27739));} + AF1 AFromPqF1(AF1 x){AF1 p=pow(x,AF1_(0.0126833)); + return pow(ASatF1(p-AF1_(0.835938))/(AF1_(18.8516)-AF1_(18.6875)*p),AF1_(6.27739));} + AF2 AFromPqF1(AF2 x){AF2 p=pow(x,AF2_(0.0126833)); + return pow(ASatF2(p-AF2_(0.835938))/(AF2_(18.8516)-AF2_(18.6875)*p),AF2_(6.27739));} + AF3 AFromPqF1(AF3 x){AF3 p=pow(x,AF3_(0.0126833)); + return pow(ASatF3(p-AF3_(0.835938))/(AF3_(18.8516)-AF3_(18.6875)*p),AF3_(6.27739));} //------------------------------------------------------------------------------------------------------------------------------ - AF1 AFromSrgbF1(AF1 c){return max(min(c*AF1_(1.0/12.92),AF1_(0.04045)), - pow((c+AF1_(0.055))*(AF1_(1.0)/AF1_(1.055)),AF1_(2.4)));} + // Unfortunately median won't work here. + AF1 AFromSrgbF1(AF1 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF1(AZolSignedF1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AF2 AFromSrgbF2(AF2 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF2(AZolSignedF2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AF3 AFromSrgbF3(AF3 c){AF3 j=AF3(0.04045/12.92,1.0/12.92,2.4);AF2 k=AF2(1.0/1.055,0.055/1.055); + return AZolSelF3(AZolSignedF3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} //------------------------------------------------------------------------------------------------------------------------------ - AF1 AFromTwoF1(AF1 c){return c*c;} -//============================================================================================================================== - #ifdef A_HALF - AH2 ATo709H2(AH2 c){return max(min(c*AH2_(4.5),AH2_(0.018)),AH2_(1.099)*pow(c,AH2_(0.45))-AH2_(0.099));} + AF1 AFromTwoF1(AF1 c){return c*c;} + AF2 AFromTwoF2(AF2 c){return c*c;} + AF3 AFromTwoF3(AF3 c){return c*c;} //------------------------------------------------------------------------------------------------------------------------------ - AH2 AToGammaH2(AH2 c,AH1 rcpX){return pow(c,AH2_(rcpX));} -//------------------------------------------------------------------------------------------------------------------------------ - AH2 AToSrgbH2(AH2 c){return max(min(c*AH2_(12.92),AH2_(0.0031308)),AH2_(1.055)*pow(c,AH2_(0.41666))-AH2_(0.055));} -//------------------------------------------------------------------------------------------------------------------------------ - AH2 AToTwoH2(AH2 c){return sqrt(c);} + AF1 AFromThreeF1(AF1 c){return c*c*c;} + AF2 AFromThreeF2(AF2 c){return c*c*c;} + AF3 AFromThreeF3(AF3 c){return c*c*c;} #endif //============================================================================================================================== #ifdef A_HALF - AH2 AFrom709H2(AH2 c){return max(min(c*AH2_(1.0/4.5),AH2_(0.081)), - pow((c+AH2_(0.099))*(AH2_(1.0)/(AH2_(1.099))),AH2_(1.0/0.45)));} + AH1 ATo709H1(AH1 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 ATo709H2(AH2 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 ATo709H3(AH3 c){AH3 j=AH3(0.018*4.5,4.5,0.45);AH2 k=AH2(1.099,-0.099); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} //------------------------------------------------------------------------------------------------------------------------------ + AH1 AToGammaH1(AH1 c,AH1 rcpX){return pow(c,AH1_(rcpX));} + AH2 AToGammaH2(AH2 c,AH1 rcpX){return pow(c,AH2_(rcpX));} + AH3 AToGammaH3(AH3 c,AH1 rcpX){return pow(c,AH3_(rcpX));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToSrgbH1(AH1 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.x ,c*j.y ,pow(c,j.z )*k.x +k.y );} + AH2 AToSrgbH2(AH2 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xx ,c*j.yy ,pow(c,j.zz )*k.xx +k.yy );} + AH3 AToSrgbH3(AH3 c){AH3 j=AH3(0.0031308*12.92,12.92,1.0/2.4);AH2 k=AH2(1.055,-0.055); + return clamp(j.xxx,c*j.yyy,pow(c,j.zzz)*k.xxx+k.yyy);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToTwoH1(AH1 c){return sqrt(c);} + AH2 AToTwoH2(AH2 c){return sqrt(c);} + AH3 AToTwoH3(AH3 c){return sqrt(c);} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AToThreeF1(AH1 c){return pow(c,AH1_(1.0/3.0));} + AH2 AToThreeF2(AH2 c){return pow(c,AH2_(1.0/3.0));} + AH3 AToThreeF3(AH3 c){return pow(c,AH3_(1.0/3.0));} + #endif +//============================================================================================================================== + #ifdef A_HALF + AH1 AFrom709H1(AH1 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AFrom709H2(AH2 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AFrom709H3(AH3 c){AH3 j=AH3(0.081/4.5,1.0/4.5,1.0/0.45);AH2 k=AH2(1.0/1.099,0.099/1.099); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromGammaH1(AH1 c,AH1 x){return pow(c,AH1_(x));} AH2 AFromGammaH2(AH2 c,AH1 x){return pow(c,AH2_(x));} + AH3 AFromGammaH3(AH3 c,AH1 x){return pow(c,AH3_(x));} //------------------------------------------------------------------------------------------------------------------------------ - AH2 AFromSrgbH2(AH2 c){return max(min(c*AH2_(1.0/12.92),AH2_(0.04045)), - pow((c+AH2_(0.055))*(AH2_(1.0)/AH2_(1.055)),AH2_(2.4)));} + AH1 AHromSrgbF1(AH1 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH1(AZolSignedH1(c-j.x ),c*j.y ,pow(c*k.x +k.y ,j.z ));} + AH2 AHromSrgbF2(AH2 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH2(AZolSignedH2(c-j.xx ),c*j.yy ,pow(c*k.xx +k.yy ,j.zz ));} + AH3 AHromSrgbF3(AH3 c){AH3 j=AH3(0.04045/12.92,1.0/12.92,2.4);AH2 k=AH2(1.0/1.055,0.055/1.055); + return AZolSelH3(AZolSignedH3(c-j.xxx),c*j.yyy,pow(c*k.xxx+k.yyy,j.zzz));} //------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromTwoH1(AH1 c){return c*c;} AH2 AFromTwoH2(AH2 c){return c*c;} + AH3 AFromTwoH3(AH3 c){return c*c;} +//------------------------------------------------------------------------------------------------------------------------------ + AH1 AFromThreeH1(AH1 c){return c*c*c;} + AH2 AFromThreeH2(AH2 c){return c*c*c;} + AH3 AFromThreeH3(AH3 c){return c*c*c;} #endif //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -1584,15 +2311,20 @@ // Details, // LANE TO 8x8 MAPPING // =================== - // 00 01 08 09 10 11 18 19 + // 00 01 08 09 10 11 18 19 // 02 03 0a 0b 12 13 1a 1b // 04 05 0c 0d 14 15 1c 1d - // 06 07 0e 0f 16 17 1e 1f - // 20 21 28 29 30 31 38 39 + // 06 07 0e 0f 16 17 1e 1f + // 20 21 28 29 30 31 38 39 // 22 23 2a 2b 32 33 3a 3b // 24 25 2c 2d 34 35 3c 3d - // 26 27 2e 2f 36 37 3e 3f + // 26 27 2e 2f 36 37 3e 3f AU2 ARmpRed8x8(AU1 a){return AU2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} +//============================================================================================================================== + #ifdef A_HALF + AW2 ARmp8x8H(AU1 a){return AW2(ABfe(a,1u,3u),ABfiM(ABfe(a,3u,3u),a,1u));} + AW2 ARmpRed8x8H(AU1 a){return AW2(ABfiM(ABfe(a,2u,3u),a,1u),ABfiM(ABfe(a,3u,3u),ABfe(a,1u,2u),2u));} + #endif #endif //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -1643,13 +2375,22 @@ // ... // 1023 = 2^(-14)*(1-2^(-10)) = 2^(-14)*(1-1/1024) ... last denormal value // 1024 = 2^(-14) = 1/16384 .......................... first normal value that still maps to integers -// 2047 .............................................. last normal value that still maps to integers +// 2047 .............................................. last normal value that still maps to integers // Scaling limits, // 2^15 = 32768 ...................................... largest power of 2 scaling // Largest pow2 conversion mapping is at *32768, -// 1 : 2^(-9) = 1/128 -// 1024 : 8 -// 2047 : a little less than 16 +// 1 : 2^(-9) = 1/512 +// 2 : 1/256 +// 4 : 1/128 +// 8 : 1/64 +// 16 : 1/32 +// 32 : 1/16 +// 64 : 1/8 +// 128 : 1/4 +// 256 : 1/2 +// 512 : 1 +// 1024 : 2 +// 2047 : a little less than 4 //============================================================================================================================== //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// @@ -1783,10 +2524,10 @@ #define ALog2D1(a) log2(AD1(a)) #define ALog2F1(a) log2(AF1(a)) //------------------------------------------------------------------------------------------------------------------------------ - #define AMaxD1(a,b) min(a,b) - #define AMaxF1(a,b) min(a,b) - #define AMaxL1(a,b) min(a,b) - #define AMaxU1(a,b) min(a,b) + #define AMaxD1(a,b) max(a,b) + #define AMaxF1(a,b) max(a,b) + #define AMaxL1(a,b) max(a,b) + #define AMaxU1(a,b) max(a,b) //------------------------------------------------------------------------------------------------------------------------------ #define AMinD1(a,b) min(a,b) #define AMinF1(a,b) min(a,b) @@ -1823,6 +2564,10 @@ AD2 opAAddD2(outAD2 d,inAD2 a,inAD2 b){d=a+b;return d;} AD3 opAAddD3(outAD3 d,inAD3 a,inAD3 b){d=a+b;return d;} AD4 opAAddD4(outAD4 d,inAD4 a,inAD4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AD2 opAAddOneD2(outAD2 d,inAD2 a,AD1 b){d=a+AD2_(b);return d;} + AD3 opAAddOneD3(outAD3 d,inAD3 a,AD1 b){d=a+AD3_(b);return d;} + AD4 opAAddOneD4(outAD4 d,inAD4 a,AD1 b){d=a+AD4_(b);return d;} //------------------------------------------------------------------------------------------------------------------------------ AD2 opACpyD2(outAD2 d,inAD2 a){d=a;return d;} AD3 opACpyD3(outAD3 d,inAD3 a){d=a;return d;} @@ -1868,6 +2613,10 @@ AF2 opAAddF2(outAF2 d,inAF2 a,inAF2 b){d=a+b;return d;} AF3 opAAddF3(outAF3 d,inAF3 a,inAF3 b){d=a+b;return d;} AF4 opAAddF4(outAF4 d,inAF4 a,inAF4 b){d=a+b;return d;} +//------------------------------------------------------------------------------------------------------------------------------ + AF2 opAAddOneF2(outAF2 d,inAF2 a,AF1 b){d=a+AF2_(b);return d;} + AF3 opAAddOneF3(outAF3 d,inAF3 a,AF1 b){d=a+AF3_(b);return d;} + AF4 opAAddOneF4(outAF4 d,inAF4 a,AF1 b){d=a+AF4_(b);return d;} //------------------------------------------------------------------------------------------------------------------------------ AF2 opACpyF2(outAF2 d,inAF2 a){d=a;return d;} AF3 opACpyF3(outAF3 d,inAF3 a){d=a;return d;} diff --git a/client/renderers/EGL/shader/ffx_cas.frag b/client/renderers/EGL/shader/ffx_cas.frag index 5130151d..5e03ce99 100644 --- a/client/renderers/EGL/shader/ffx_cas.frag +++ b/client/renderers/EGL/shader/ffx_cas.frag @@ -1,7 +1,8 @@ #version 300 es - precision mediump float; +#include "compat.h" + in vec2 iFragCoord; out vec4 fragColor; @@ -10,22 +11,6 @@ uniform uvec2 uInRes[8]; uniform uvec2 uOutRes; uniform float uSharpness; -// the following are not available until verion 400 or later -// so we implement our own versions of them -uint bitfieldExtract(uint val, int off, int size) -{ - uint mask = uint((1 << size) - 1); - return uint(val >> off) & mask; -} - -uint bitfieldInsert(uint a, uint b, int c, int d) -{ - uint mask = ~(0xffffffffu << d) << c; - mask = ~mask; - a &= mask; - return a | (b << c); -} - #define A_GPU 1 #define A_GLSL 1 diff --git a/client/renderers/EGL/shader/ffx_fsr1.h b/client/renderers/EGL/shader/ffx_fsr1.h new file mode 100644 index 00000000..873307fb --- /dev/null +++ b/client/renderers/EGL/shader/ffx_fsr1.h @@ -0,0 +1,1199 @@ +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// +// AMD FidelityFX SUPER RESOLUTION [FSR 1] ::: SPATIAL SCALING & EXTRAS - v1.20210629 +// +// +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FidelityFX Super Resolution Sample +// +// Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files(the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions : +// The above copyright notice and this permission notice shall be included in +// all copies or substantial portions of the Software. +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN +// THE SOFTWARE. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// ABOUT +// ===== +// FSR is a collection of algorithms relating to generating a higher resolution image. +// This specific header focuses on single-image non-temporal image scaling, and related tools. +// +// The core functions are EASU and RCAS: +// [EASU] Edge Adaptive Spatial Upsampling ....... 1x to 4x area range spatial scaling, clamped adaptive elliptical filter. +// [RCAS] Robust Contrast Adaptive Sharpening .... A non-scaling variation on CAS. +// RCAS needs to be applied after EASU as a separate pass. +// +// Optional utility functions are: +// [LFGA] Linear Film Grain Applicator ........... Tool to apply film grain after scaling. +// [SRTM] Simple Reversible Tone-Mapper .......... Linear HDR {0 to FP16_MAX} to {0 to 1} and back. +// [TEPD] Temporal Energy Preserving Dither ...... Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// See each individual sub-section for inline documentation. +//------------------------------------------------------------------------------------------------------------------------------ +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//------------------------------------------------------------------------------------------------------------------------------ +// FUNCTION PERMUTATIONS +// ===================== +// *F() ..... Single item computation with 32-bit. +// *H() ..... Single item computation with 16-bit, with packing (aka two 16-bit ops in parallel) when possible. +// *Hx2() ... Processing two items in parallel with 16-bit, easier packing. +// Not all interfaces in this file have a *Hx2() form. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [EASU] EDGE ADAPTIVE SPATIAL UPSAMPLING +// +//------------------------------------------------------------------------------------------------------------------------------ +// EASU provides a high quality spatial-only scaling at relatively low cost. +// Meaning EASU is appropiate for laptops and other low-end GPUs. +// Quality from 1x to 4x area scaling is good. +//------------------------------------------------------------------------------------------------------------------------------ +// The scalar uses a modified fast approximation to the standard lanczos(size=2) kernel. +// EASU runs in a single pass, so it applies a directionally and anisotropically adaptive radial lanczos. +// This is also kept as simple as possible to have minimum runtime. +//------------------------------------------------------------------------------------------------------------------------------ +// The lanzcos filter has negative lobes, so by itself it will introduce ringing. +// To remove all ringing, the algorithm uses the nearest 2x2 input texels as a neighborhood, +// and limits output to the minimum and maximum of that neighborhood. +//------------------------------------------------------------------------------------------------------------------------------ +// Input image requirements: +// +// Color needs to be encoded as 3 channel[red, green, blue](e.g.XYZ not supported) +// Each channel needs to be in the range[0, 1] +// Any color primaries are supported +// Display / tonemapping curve needs to be as if presenting to sRGB display or similar(e.g.Gamma 2.0) +// There should be no banding in the input +// There should be no high amplitude noise in the input +// There should be no noise in the input that is not at input pixel granularity +// For performance purposes, use 32bpp formats +//------------------------------------------------------------------------------------------------------------------------------ +// Best to apply EASU at the end of the frame after tonemapping +// but before film grain or composite of the UI. +//------------------------------------------------------------------------------------------------------------------------------ +// Example of including this header for D3D HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan GLSL : +// +// #define A_GPU 1 +// #define A_GLSL 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of including this header for Vulkan HLSL : +// +// #define A_GPU 1 +// #define A_HLSL 1 +// #define A_HLSL_6_2 1 +// #define A_NO_16_BIT_CAST 1 +// #define A_HALF 1 +// #include "ffx_a.h" +// #define FSR_EASU_H 1 +// #define FSR_RCAS_H 1 +// //declare input callbacks +// #include "ffx_fsr1.h" +// +// Example of declaring the required input callbacks for GLSL : +// The callbacks need to gather4 for each color channel using the specified texture coordinate 'p'. +// EASU uses gather4 to reduce position computation logic and for free Arrays of Structures to Structures of Arrays conversion. +// +// AH4 FsrEasuRH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,0));} +// AH4 FsrEasuGH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,1));} +// AH4 FsrEasuBH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,2));} +// ... +// The FsrEasuCon function needs to be called from the CPU or GPU to set up constants. +// The difference in viewport and input image size is there to support Dynamic Resolution Scaling. +// To use FsrEasuCon() on the CPU, define A_CPU before including ffx_a and ffx_fsr1. +// Including a GPU example here, the 'con0' through 'con3' values would be stored out to a constant buffer. +// AU4 con0,con1,con2,con3; +// FsrEasuCon(con0,con1,con2,con3, +// 1920.0,1080.0, // Viewport size (top left aligned) in the input image which is to be scaled. +// 3840.0,2160.0, // The size of the input image. +// 2560.0,1440.0); // The output resolution. +//============================================================================================================================== +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrEasuCon( +outAU4 con0, +outAU4 con1, +outAU4 con2, +outAU4 con3, +// This the rendered image resolution being upscaled +AF1 inputViewportInPixelsX, +AF1 inputViewportInPixelsY, +// This is the resolution of the resource containing the input image (useful for dynamic resolution) +AF1 inputSizeInPixelsX, +AF1 inputSizeInPixelsY, +// This is the display resolution which the input image gets upscaled to +AF1 outputSizeInPixelsX, +AF1 outputSizeInPixelsY){ + // Output integer position to a pixel position in viewport. + con0[0]=AU1_AF1(inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)); + con0[1]=AU1_AF1(inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)); + con0[2]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)-AF1_(0.5)); + con0[3]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)-AF1_(0.5)); + // Viewport pixel position to normalized image space. + // This is used to get upper-left of 'F' tap. + con1[0]=AU1_AF1(ARcpF1(inputSizeInPixelsX)); + con1[1]=AU1_AF1(ARcpF1(inputSizeInPixelsY)); + // Centers of gather4, first offset from upper-left of 'F'. + // +---+---+ + // | | | + // +--(0)--+ + // | b | c | + // +---F---+---+---+ + // | e | f | g | h | + // +--(1)--+--(2)--+ + // | i | j | k | l | + // +---+---+---+---+ + // | n | o | + // +--(3)--+ + // | | | + // +---+---+ + con1[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con1[3]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsY)); + // These are from (0) instead of 'F'. + con2[0]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsX)); + con2[1]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con2[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); + con2[3]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); + con3[0]=AU1_AF1(AF1_( 0.0)*ARcpF1(inputSizeInPixelsX)); + con3[1]=AU1_AF1(AF1_( 4.0)*ARcpF1(inputSizeInPixelsY)); + con3[2]=con3[3]=0u;} + +//If the an offset into the input image resource +A_STATIC void FsrEasuConOffset( + outAU4 con0, + outAU4 con1, + outAU4 con2, + outAU4 con3, + // This the rendered image resolution being upscaled + AF1 inputViewportInPixelsX, + AF1 inputViewportInPixelsY, + // This is the resolution of the resource containing the input image (useful for dynamic resolution) + AF1 inputSizeInPixelsX, + AF1 inputSizeInPixelsY, + // This is the display resolution which the input image gets upscaled to + AF1 outputSizeInPixelsX, + AF1 outputSizeInPixelsY, + // This is the input image offset into the resource containing it (useful for dynamic resolution) + AF1 inputOffsetInPixelsX, + AF1 inputOffsetInPixelsY) { + FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY, inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY); + con0[2] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsX * ARcpF1(outputSizeInPixelsX) - AF1_(0.5) + inputOffsetInPixelsX); + con0[3] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsY * ARcpF1(outputSizeInPixelsY) - AF1_(0.5) + inputOffsetInPixelsY); +} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_EASU_F) + // Input callback prototypes, need to be implemented by calling shader + AF4 FsrEasuRF(AF2 p); + AF4 FsrEasuGF(AF2 p); + AF4 FsrEasuBF(AF2 p); +//------------------------------------------------------------------------------------------------------------------------------ + // Filtering for a given tap for the scalar. + void FsrEasuTapF( + inout AF3 aC, // Accumulated color, with negative lobe. + inout AF1 aW, // Accumulated weight. + AF2 off, // Pixel offset from resolve position to tap. + AF2 dir, // Gradient direction. + AF2 len, // Length. + AF1 lob, // Negative lobe strength. + AF1 clp, // Clipping point. + AF3 c){ // Tap color. + // Rotate offset by direction. + AF2 v; + v.x=(off.x*( dir.x))+(off.y*dir.y); + v.y=(off.x*(-dir.y))+(off.y*dir.x); + // Anisotropy. + v*=len; + // Compute distance^2. + AF1 d2=v.x*v.x+v.y*v.y; + // Limit to the window as at corner, 2 taps can easily be outside. + d2=min(d2,clp); + // Approximation of lancos2 without sin() or rcp(), or sqrt() to get x. + // (25/16 * (2/5 * x^2 - 1)^2 - (25/16 - 1)) * (1/4 * x^2 - 1)^2 + // |_______________________________________| |_______________| + // base window + // The general form of the 'base' is, + // (a*(b*x^2-1)^2-(a-1)) + // Where 'a=1/(2*b-b^2)' and 'b' moves around the negative lobe. + AF1 wB=AF1_(2.0/5.0)*d2+AF1_(-1.0); + AF1 wA=lob*d2+AF1_(-1.0); + wB*=wB; + wA*=wA; + wB=AF1_(25.0/16.0)*wB+AF1_(-(25.0/16.0-1.0)); + AF1 w=wB*wA; + // Do weighted average. + aC+=c*w;aW+=w;} +//------------------------------------------------------------------------------------------------------------------------------ + // Accumulate direction and length. + void FsrEasuSetF( + inout AF2 dir, + inout AF1 len, + AF2 pp, + AP1 biS,AP1 biT,AP1 biU,AP1 biV, + AF1 lA,AF1 lB,AF1 lC,AF1 lD,AF1 lE){ + // Compute bilinear weight, branches factor out as predicates are compiler time immediates. + // s t + // u v + AF1 w = AF1_(0.0); + if(biS)w=(AF1_(1.0)-pp.x)*(AF1_(1.0)-pp.y); + if(biT)w= pp.x *(AF1_(1.0)-pp.y); + if(biU)w=(AF1_(1.0)-pp.x)* pp.y ; + if(biV)w= pp.x * pp.y ; + // Direction is the '+' diff. + // a + // b c d + // e + // Then takes magnitude from abs average of both sides of 'c'. + // Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms. + AF1 dc=lD-lC; + AF1 cb=lC-lB; + AF1 lenX=max(abs(dc),abs(cb)); + lenX=APrxLoRcpF1(lenX); + AF1 dirX=lD-lB; + dir.x+=dirX*w; + lenX=ASatF1(abs(dirX)*lenX); + lenX*=lenX; + len+=lenX*w; + // Repeat for the y axis. + AF1 ec=lE-lC; + AF1 ca=lC-lA; + AF1 lenY=max(abs(ec),abs(ca)); + lenY=APrxLoRcpF1(lenY); + AF1 dirY=lE-lA; + dir.y+=dirY*w; + lenY=ASatF1(abs(dirY)*lenY); + lenY*=lenY; + len+=lenY*w;} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrEasuF( + out AF3 pix, + AU2 ip, // Integer pixel position in output. + AU4 con0, // Constants generated by FsrEasuCon(). + AU4 con1, + AU4 con2, + AU4 con3){ +//------------------------------------------------------------------------------------------------------------------------------ + // Get position of 'f'. + AF2 pp=AF2(ip)*AF2_AU2(con0.xy)+AF2_AU2(con0.zw); + AF2 fp=floor(pp); + pp-=fp; +//------------------------------------------------------------------------------------------------------------------------------ + // 12-tap kernel. + // b c + // e f g h + // i j k l + // n o + // Gather 4 ordering. + // a b + // r g + // For packed FP16, need either {rg} or {ab} so using the following setup for gather in all versions, + // a b <- unused (z) + // r g + // a b a b + // r g r g + // a b + // r g <- unused (z) + // Allowing dead-code removal to remove the 'z's. + AF2 p0=fp*AF2_AU2(con1.xy)+AF2_AU2(con1.zw); + // These are from p0 to avoid pulling two constants on pre-Navi hardware. + AF2 p1=p0+AF2_AU2(con2.xy); + AF2 p2=p0+AF2_AU2(con2.zw); + AF2 p3=p0+AF2_AU2(con3.xy); + AF4 bczzR=FsrEasuRF(p0); + AF4 bczzG=FsrEasuGF(p0); + AF4 bczzB=FsrEasuBF(p0); + AF4 ijfeR=FsrEasuRF(p1); + AF4 ijfeG=FsrEasuGF(p1); + AF4 ijfeB=FsrEasuBF(p1); + AF4 klhgR=FsrEasuRF(p2); + AF4 klhgG=FsrEasuGF(p2); + AF4 klhgB=FsrEasuBF(p2); + AF4 zzonR=FsrEasuRF(p3); + AF4 zzonG=FsrEasuGF(p3); + AF4 zzonB=FsrEasuBF(p3); +//------------------------------------------------------------------------------------------------------------------------------ + // Simplest multi-channel approximate luma possible (luma times 2, in 2 FMA/MAD). + AF4 bczzL=bczzB*AF4_(0.5)+(bczzR*AF4_(0.5)+bczzG); + AF4 ijfeL=ijfeB*AF4_(0.5)+(ijfeR*AF4_(0.5)+ijfeG); + AF4 klhgL=klhgB*AF4_(0.5)+(klhgR*AF4_(0.5)+klhgG); + AF4 zzonL=zzonB*AF4_(0.5)+(zzonR*AF4_(0.5)+zzonG); + // Rename. + AF1 bL=bczzL.x; + AF1 cL=bczzL.y; + AF1 iL=ijfeL.x; + AF1 jL=ijfeL.y; + AF1 fL=ijfeL.z; + AF1 eL=ijfeL.w; + AF1 kL=klhgL.x; + AF1 lL=klhgL.y; + AF1 hL=klhgL.z; + AF1 gL=klhgL.w; + AF1 oL=zzonL.z; + AF1 nL=zzonL.w; + // Accumulate for bilinear interpolation. + AF2 dir=AF2_(0.0); + AF1 len=AF1_(0.0); + FsrEasuSetF(dir,len,pp,true, false,false,false,bL,eL,fL,gL,jL); + FsrEasuSetF(dir,len,pp,false,true ,false,false,cL,fL,gL,hL,kL); + FsrEasuSetF(dir,len,pp,false,false,true ,false,fL,iL,jL,kL,nL); + FsrEasuSetF(dir,len,pp,false,false,false,true ,gL,jL,kL,lL,oL); +//------------------------------------------------------------------------------------------------------------------------------ + // Normalize with approximation, and cleanup close to zero. + AF2 dir2=dir*dir; + AF1 dirR=dir2.x+dir2.y; + AP1 zro=dirR w = -m/(n+e+w+s) +// 1 == (w*(n+e+w+s)+m)/(4*w+1) -> w = (1-m)/(n+e+w+s-4*1) +// Then chooses the 'w' which results in no clipping, limits 'w', and multiplies by the 'sharp' amount. +// This solution above has issues with MSAA input as the steps along the gradient cause edge detection issues. +// So RCAS uses 4x the maximum and 4x the minimum (depending on equation)in place of the individual taps. +// As well as switching from 'm' to either the minimum or maximum (depending on side), to help in energy conservation. +// This stabilizes RCAS. +// RCAS does a simple highpass which is normalized against the local contrast then shaped, +// 0.25 +// 0.25 -1 0.25 +// 0.25 +// This is used as a noise detection filter, to reduce the effect of RCAS on grain, and focus on real edges. +// +// GLSL example for the required callbacks : +// +// AH4 FsrRcasLoadH(ASW2 p){return AH4(imageLoad(imgSrc,ASU2(p)));} +// void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b) +// { +// //do any simple input color conversions here or leave empty if none needed +// } +// +// FsrRcasCon need to be called from the CPU or GPU to set up constants. +// Including a GPU example here, the 'con' value would be stored out to a constant buffer. +// +// AU4 con; +// FsrRcasCon(con, +// 0.0); // The scale is {0.0 := maximum sharpness, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +// --------------- +// RCAS sharpening supports a CAS-like pass-through alpha via, +// #define FSR_RCAS_PASSTHROUGH_ALPHA 1 +// RCAS also supports a define to enable a more expensive path to avoid some sharpening of noise. +// Would suggest it is better to apply film grain after RCAS sharpening (and after scaling) instead of using this define, +// #define FSR_RCAS_DENOISE 1 +//============================================================================================================================== +// This is set at the limit of providing unnatural results for sharpening. +#define FSR_RCAS_LIMIT (0.25-(1.0/16.0)) +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// CONSTANT SETUP +//============================================================================================================================== +// Call to setup required constant values (works on CPU or GPU). +A_STATIC void FsrRcasCon( +outAU4 con, +// The scale is {0.0 := maximum, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. +AF1 sharpness){ + // Transform from stops to linear value. + sharpness=AExp2F1(-sharpness); + varAF2(hSharp)=initAF2(sharpness,sharpness); + con[0]=AU1_AF1(sharpness); + con[1]=AU1_AH2_AF2(hSharp); + con[2]=0u; + con[3]=0u;} +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 32-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(FSR_RCAS_F) + // Input callback prototypes that need to be implemented by calling shader + AF4 FsrRcasLoadF(ASU2 p); + void FsrRcasInputF(inout AF1 r,inout AF1 g,inout AF1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasF( + out AF1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AF1 pixG, + out AF1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AF1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASU2 sp=ASU2(ip); + AF3 b=FsrRcasLoadF(sp+ASU2( 0,-1)).rgb; + AF3 d=FsrRcasLoadF(sp+ASU2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AF4 ee=FsrRcasLoadF(sp); + AF3 e=ee.rgb;pixA=ee.a; + #else + AF3 e=FsrRcasLoadF(sp).rgb; + #endif + AF3 f=FsrRcasLoadF(sp+ASU2( 1, 0)).rgb; + AF3 h=FsrRcasLoadF(sp+ASU2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AF1 bR=b.r; + AF1 bG=b.g; + AF1 bB=b.b; + AF1 dR=d.r; + AF1 dG=d.g; + AF1 dB=d.b; + AF1 eR=e.r; + AF1 eG=e.g; + AF1 eB=e.b; + AF1 fR=f.r; + AF1 fG=f.g; + AF1 fB=f.b; + AF1 hR=h.r; + AF1 hG=h.g; + AF1 hB=h.b; + // Run optional input transform. + FsrRcasInputF(bR,bG,bB); + FsrRcasInputF(dR,dG,dB); + FsrRcasInputF(eR,eG,eB); + FsrRcasInputF(fR,fG,fB); + FsrRcasInputF(hR,hG,hB); + // Luma times 2. + AF1 bL=bB*AF1_(0.5)+(bR*AF1_(0.5)+bG); + AF1 dL=dB*AF1_(0.5)+(dR*AF1_(0.5)+dG); + AF1 eL=eB*AF1_(0.5)+(eR*AF1_(0.5)+eG); + AF1 fL=fB*AF1_(0.5)+(fR*AF1_(0.5)+fG); + AF1 hL=hB*AF1_(0.5)+(hR*AF1_(0.5)+hG); + // Noise detection. + AF1 nz=AF1_(0.25)*bL+AF1_(0.25)*dL+AF1_(0.25)*fL+AF1_(0.25)*hL-eL; + nz=ASatF1(abs(nz)*APrxMedRcpF1(AMax3F1(AMax3F1(bL,dL,eL),fL,hL)-AMin3F1(AMin3F1(bL,dL,eL),fL,hL))); + nz=AF1_(-0.5)*nz+AF1_(1.0); + // Min and max of ring. + AF1 mn4R=min(AMin3F1(bR,dR,fR),hR); + AF1 mn4G=min(AMin3F1(bG,dG,fG),hG); + AF1 mn4B=min(AMin3F1(bB,dB,fB),hB); + AF1 mx4R=max(AMax3F1(bR,dR,fR),hR); + AF1 mx4G=max(AMax3F1(bG,dG,fG),hG); + AF1 mx4B=max(AMax3F1(bB,dB,fB),hB); + // Immediate constants for peak range. + AF2 peakC=AF2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AF1 hitMinR=mn4R*ARcpF1(AF1_(4.0)*mx4R); + AF1 hitMinG=mn4G*ARcpF1(AF1_(4.0)*mx4G); + AF1 hitMinB=mn4B*ARcpF1(AF1_(4.0)*mx4B); + AF1 hitMaxR=(peakC.x-mx4R)*ARcpF1(AF1_(4.0)*mn4R+peakC.y); + AF1 hitMaxG=(peakC.x-mx4G)*ARcpF1(AF1_(4.0)*mn4G+peakC.y); + AF1 hitMaxB=(peakC.x-mx4B)*ARcpF1(AF1_(4.0)*mn4B+peakC.y); + AF1 lobeR=max(-hitMinR,hitMaxR); + AF1 lobeG=max(-hitMinG,hitMaxG); + AF1 lobeB=max(-hitMinB,hitMaxB); + AF1 lobe=max(AF1_(-FSR_RCAS_LIMIT),min(AMax3F1(lobeR,lobeG,lobeB),AF1_(0.0)))*AF1_AU1(con.x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AF1 rcpL=APrxMedRcpF1(AF1_(4.0)*lobe+AF1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL; + return;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// NON-PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_H) + // Input callback prototypes that need to be implemented by calling shader + AH4 FsrRcasLoadH(ASW2 p); + void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b); +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasH( + out AH1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. + out AH1 pixG, + out AH1 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH1 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // Sharpening algorithm uses minimal 3x3 pixel neighborhood. + // b + // d e f + // h + ASW2 sp=ASW2(ip); + AH3 b=FsrRcasLoadH(sp+ASW2( 0,-1)).rgb; + AH3 d=FsrRcasLoadH(sp+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee=FsrRcasLoadH(sp); + AH3 e=ee.rgb;pixA=ee.a; + #else + AH3 e=FsrRcasLoadH(sp).rgb; + #endif + AH3 f=FsrRcasLoadH(sp+ASW2( 1, 0)).rgb; + AH3 h=FsrRcasLoadH(sp+ASW2( 0, 1)).rgb; + // Rename (32-bit) or regroup (16-bit). + AH1 bR=b.r; + AH1 bG=b.g; + AH1 bB=b.b; + AH1 dR=d.r; + AH1 dG=d.g; + AH1 dB=d.b; + AH1 eR=e.r; + AH1 eG=e.g; + AH1 eB=e.b; + AH1 fR=f.r; + AH1 fG=f.g; + AH1 fB=f.b; + AH1 hR=h.r; + AH1 hG=h.g; + AH1 hB=h.b; + // Run optional input transform. + FsrRcasInputH(bR,bG,bB); + FsrRcasInputH(dR,dG,dB); + FsrRcasInputH(eR,eG,eB); + FsrRcasInputH(fR,fG,fB); + FsrRcasInputH(hR,hG,hB); + // Luma times 2. + AH1 bL=bB*AH1_(0.5)+(bR*AH1_(0.5)+bG); + AH1 dL=dB*AH1_(0.5)+(dR*AH1_(0.5)+dG); + AH1 eL=eB*AH1_(0.5)+(eR*AH1_(0.5)+eG); + AH1 fL=fB*AH1_(0.5)+(fR*AH1_(0.5)+fG); + AH1 hL=hB*AH1_(0.5)+(hR*AH1_(0.5)+hG); + // Noise detection. + AH1 nz=AH1_(0.25)*bL+AH1_(0.25)*dL+AH1_(0.25)*fL+AH1_(0.25)*hL-eL; + nz=ASatH1(abs(nz)*APrxMedRcpH1(AMax3H1(AMax3H1(bL,dL,eL),fL,hL)-AMin3H1(AMin3H1(bL,dL,eL),fL,hL))); + nz=AH1_(-0.5)*nz+AH1_(1.0); + // Min and max of ring. + AH1 mn4R=min(AMin3H1(bR,dR,fR),hR); + AH1 mn4G=min(AMin3H1(bG,dG,fG),hG); + AH1 mn4B=min(AMin3H1(bB,dB,fB),hB); + AH1 mx4R=max(AMax3H1(bR,dR,fR),hR); + AH1 mx4G=max(AMax3H1(bG,dG,fG),hG); + AH1 mx4B=max(AMax3H1(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH1 hitMinR=mn4R*ARcpH1(AH1_(4.0)*mx4R); + AH1 hitMinG=mn4G*ARcpH1(AH1_(4.0)*mx4G); + AH1 hitMinB=mn4B*ARcpH1(AH1_(4.0)*mx4B); + AH1 hitMaxR=(peakC.x-mx4R)*ARcpH1(AH1_(4.0)*mn4R+peakC.y); + AH1 hitMaxG=(peakC.x-mx4G)*ARcpH1(AH1_(4.0)*mn4G+peakC.y); + AH1 hitMaxB=(peakC.x-mx4B)*ARcpH1(AH1_(4.0)*mn4B+peakC.y); + AH1 lobeR=max(-hitMinR,hitMaxR); + AH1 lobeG=max(-hitMinG,hitMaxG); + AH1 lobeB=max(-hitMinB,hitMaxB); + AH1 lobe=max(AH1_(-FSR_RCAS_LIMIT),min(AMax3H1(lobeR,lobeG,lobeB),AH1_(0.0)))*AH2_AU1(con.y).x; + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH1 rcpL=APrxMedRcpH1(AH1_(4.0)*lobe+AH1_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// PACKED 16-BIT VERSION +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_HX2) + // Input callback prototypes that need to be implemented by the calling shader + AH4 FsrRcasLoadHx2(ASW2 p); + void FsrRcasInputHx2(inout AH2 r,inout AH2 g,inout AH2 b); +//------------------------------------------------------------------------------------------------------------------------------ + // Can be used to convert from packed Structures of Arrays to Arrays of Structures for store. + void FsrRcasDepackHx2(out AH4 pix0,out AH4 pix1,AH2 pixR,AH2 pixG,AH2 pixB){ + #ifdef A_HLSL + // Invoke a slower path for DX only, since it won't allow uninitialized values. + pix0.a=pix1.a=0.0; + #endif + pix0.rgb=AH3(pixR.x,pixG.x,pixB.x); + pix1.rgb=AH3(pixR.y,pixG.y,pixB.y);} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrRcasHx2( + // Output values are for 2 8x8 tiles in a 16x8 region. + // pix.x = left 8x8 tile + // pix.y = right 8x8 tile + // This enables later processing to easily be packed as well. + out AH2 pixR, + out AH2 pixG, + out AH2 pixB, + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + out AH2 pixA, + #endif + AU2 ip, // Integer pixel position in output. + AU4 con){ // Constant generated by RcasSetup(). + // No scaling algorithm uses minimal 3x3 pixel neighborhood. + ASW2 sp0=ASW2(ip); + AH3 b0=FsrRcasLoadHx2(sp0+ASW2( 0,-1)).rgb; + AH3 d0=FsrRcasLoadHx2(sp0+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee0=FsrRcasLoadHx2(sp0); + AH3 e0=ee0.rgb;pixA.r=ee0.a; + #else + AH3 e0=FsrRcasLoadHx2(sp0).rgb; + #endif + AH3 f0=FsrRcasLoadHx2(sp0+ASW2( 1, 0)).rgb; + AH3 h0=FsrRcasLoadHx2(sp0+ASW2( 0, 1)).rgb; + ASW2 sp1=sp0+ASW2(8,0); + AH3 b1=FsrRcasLoadHx2(sp1+ASW2( 0,-1)).rgb; + AH3 d1=FsrRcasLoadHx2(sp1+ASW2(-1, 0)).rgb; + #ifdef FSR_RCAS_PASSTHROUGH_ALPHA + AH4 ee1=FsrRcasLoadHx2(sp1); + AH3 e1=ee1.rgb;pixA.g=ee1.a; + #else + AH3 e1=FsrRcasLoadHx2(sp1).rgb; + #endif + AH3 f1=FsrRcasLoadHx2(sp1+ASW2( 1, 0)).rgb; + AH3 h1=FsrRcasLoadHx2(sp1+ASW2( 0, 1)).rgb; + // Arrays of Structures to Structures of Arrays conversion. + AH2 bR=AH2(b0.r,b1.r); + AH2 bG=AH2(b0.g,b1.g); + AH2 bB=AH2(b0.b,b1.b); + AH2 dR=AH2(d0.r,d1.r); + AH2 dG=AH2(d0.g,d1.g); + AH2 dB=AH2(d0.b,d1.b); + AH2 eR=AH2(e0.r,e1.r); + AH2 eG=AH2(e0.g,e1.g); + AH2 eB=AH2(e0.b,e1.b); + AH2 fR=AH2(f0.r,f1.r); + AH2 fG=AH2(f0.g,f1.g); + AH2 fB=AH2(f0.b,f1.b); + AH2 hR=AH2(h0.r,h1.r); + AH2 hG=AH2(h0.g,h1.g); + AH2 hB=AH2(h0.b,h1.b); + // Run optional input transform. + FsrRcasInputHx2(bR,bG,bB); + FsrRcasInputHx2(dR,dG,dB); + FsrRcasInputHx2(eR,eG,eB); + FsrRcasInputHx2(fR,fG,fB); + FsrRcasInputHx2(hR,hG,hB); + // Luma times 2. + AH2 bL=bB*AH2_(0.5)+(bR*AH2_(0.5)+bG); + AH2 dL=dB*AH2_(0.5)+(dR*AH2_(0.5)+dG); + AH2 eL=eB*AH2_(0.5)+(eR*AH2_(0.5)+eG); + AH2 fL=fB*AH2_(0.5)+(fR*AH2_(0.5)+fG); + AH2 hL=hB*AH2_(0.5)+(hR*AH2_(0.5)+hG); + // Noise detection. + AH2 nz=AH2_(0.25)*bL+AH2_(0.25)*dL+AH2_(0.25)*fL+AH2_(0.25)*hL-eL; + nz=ASatH2(abs(nz)*APrxMedRcpH2(AMax3H2(AMax3H2(bL,dL,eL),fL,hL)-AMin3H2(AMin3H2(bL,dL,eL),fL,hL))); + nz=AH2_(-0.5)*nz+AH2_(1.0); + // Min and max of ring. + AH2 mn4R=min(AMin3H2(bR,dR,fR),hR); + AH2 mn4G=min(AMin3H2(bG,dG,fG),hG); + AH2 mn4B=min(AMin3H2(bB,dB,fB),hB); + AH2 mx4R=max(AMax3H2(bR,dR,fR),hR); + AH2 mx4G=max(AMax3H2(bG,dG,fG),hG); + AH2 mx4B=max(AMax3H2(bB,dB,fB),hB); + // Immediate constants for peak range. + AH2 peakC=AH2(1.0,-1.0*4.0); + // Limiters, these need to be high precision RCPs. + AH2 hitMinR=mn4R*ARcpH2(AH2_(4.0)*mx4R); + AH2 hitMinG=mn4G*ARcpH2(AH2_(4.0)*mx4G); + AH2 hitMinB=mn4B*ARcpH2(AH2_(4.0)*mx4B); + AH2 hitMaxR=(peakC.x-mx4R)*ARcpH2(AH2_(4.0)*mn4R+peakC.y); + AH2 hitMaxG=(peakC.x-mx4G)*ARcpH2(AH2_(4.0)*mn4G+peakC.y); + AH2 hitMaxB=(peakC.x-mx4B)*ARcpH2(AH2_(4.0)*mn4B+peakC.y); + AH2 lobeR=max(-hitMinR,hitMaxR); + AH2 lobeG=max(-hitMinG,hitMaxG); + AH2 lobeB=max(-hitMinB,hitMaxB); + AH2 lobe=max(AH2_(-FSR_RCAS_LIMIT),min(AMax3H2(lobeR,lobeG,lobeB),AH2_(0.0)))*AH2_(AH2_AU1(con.y).x); + // Apply noise removal. + #ifdef FSR_RCAS_DENOISE + lobe*=nz; + #endif + // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. + AH2 rcpL=APrxMedRcpH2(AH2_(4.0)*lobe+AH2_(1.0)); + pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; + pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; + pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [LFGA] LINEAR FILM GRAIN APPLICATOR +// +//------------------------------------------------------------------------------------------------------------------------------ +// Adding output-resolution film grain after scaling is a good way to mask both rendering and scaling artifacts. +// Suggest using tiled blue noise as film grain input, with peak noise frequency set for a specific look and feel. +// The 'Lfga*()' functions provide a convenient way to introduce grain. +// These functions limit grain based on distance to signal limits. +// This is done so that the grain is temporally energy preserving, and thus won't modify image tonality. +// Grain application should be done in a linear colorspace. +// The grain should be temporally changing, but have a temporal sum per pixel that adds to zero (non-biased). +//------------------------------------------------------------------------------------------------------------------------------ +// Usage, +// FsrLfga*( +// color, // In/out linear colorspace color {0 to 1} ranged. +// grain, // Per pixel grain texture value {-0.5 to 0.5} ranged, input is 3-channel to support colored grain. +// amount); // Amount of grain (0 to 1} ranged. +//------------------------------------------------------------------------------------------------------------------------------ +// Example if grain texture is monochrome: 'FsrLfgaF(color,AF3_(grain),amount)' +//============================================================================================================================== +#if defined(A_GPU) + // Maximum grain is the minimum distance to the signal limit. + void FsrLfgaF(inout AF3 c,AF3 t,AF1 a){c+=(t*AF3_(a))*min(AF3_(1.0)-c,c);} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + // Half precision version (slower). + void FsrLfgaH(inout AH3 c,AH3 t,AH1 a){c+=(t*AH3_(a))*min(AH3_(1.0)-c,c);} +//------------------------------------------------------------------------------------------------------------------------------ + // Packed half precision version (faster). + void FsrLfgaHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 tR,AH2 tG,AH2 tB,AH1 a){ + cR+=(tR*AH2_(a))*min(AH2_(1.0)-cR,cR);cG+=(tG*AH2_(a))*min(AH2_(1.0)-cG,cG);cB+=(tB*AH2_(a))*min(AH2_(1.0)-cB,cB);} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [SRTM] SIMPLE REVERSIBLE TONE-MAPPER +// +//------------------------------------------------------------------------------------------------------------------------------ +// This provides a way to take linear HDR color {0 to FP16_MAX} and convert it into a temporary {0 to 1} ranged post-tonemapped linear. +// The tonemapper preserves RGB ratio, which helps maintain HDR color bleed during filtering. +//------------------------------------------------------------------------------------------------------------------------------ +// Reversible tonemapper usage, +// FsrSrtm*(color); // {0 to FP16_MAX} converted to {0 to 1}. +// FsrSrtmInv*(color); // {0 to 1} converted into {0 to 32768, output peak safe for FP16}. +//============================================================================================================================== +#if defined(A_GPU) + void FsrSrtmF(inout AF3 c){c*=AF3_(ARcpF1(AMax3F1(c.r,c.g,c.b)+AF1_(1.0)));} + // The extra max solves the c=1.0 case (which is a /0). + void FsrSrtmInvF(inout AF3 c){c*=AF3_(ARcpF1(max(AF1_(1.0/32768.0),AF1_(1.0)-AMax3F1(c.r,c.g,c.b))));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + void FsrSrtmH(inout AH3 c){c*=AH3_(ARcpH1(AMax3H1(c.r,c.g,c.b)+AH1_(1.0)));} + void FsrSrtmInvH(inout AH3 c){c*=AH3_(ARcpH1(max(AH1_(1.0/32768.0),AH1_(1.0)-AMax3H1(c.r,c.g,c.b))));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrSrtmHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(AMax3H2(cR,cG,cB)+AH2_(1.0));cR*=rcp;cG*=rcp;cB*=rcp;} + void FsrSrtmInvHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ + AH2 rcp=ARcpH2(max(AH2_(1.0/32768.0),AH2_(1.0)-AMax3H2(cR,cG,cB)));cR*=rcp;cG*=rcp;cB*=rcp;} +#endif +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// +//_____________________________________________________________/\_______________________________________________________________ +//============================================================================================================================== +// +// FSR - [TEPD] TEMPORAL ENERGY PRESERVING DITHER +// +//------------------------------------------------------------------------------------------------------------------------------ +// Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. +// Gamma 2.0 is used so that the conversion back to linear is just to square the color. +// The conversion comes in 8-bit and 10-bit modes, designed for output to 8-bit UNORM or 10:10:10:2 respectively. +// Given good non-biased temporal blue noise as dither input, +// the output dither will temporally conserve energy. +// This is done by choosing the linear nearest step point instead of perceptual nearest. +// See code below for details. +//------------------------------------------------------------------------------------------------------------------------------ +// DX SPEC RULES FOR FLOAT->UNORM 8-BIT CONVERSION +// =============================================== +// - Output is 'uint(floor(saturate(n)*255.0+0.5))'. +// - Thus rounding is to nearest. +// - NaN gets converted to zero. +// - INF is clamped to {0.0 to 1.0}. +//============================================================================================================================== +#if defined(A_GPU) + // Hand tuned integer position to dither value, with more values than simple checkerboard. + // Only 32-bit has enough precision for this compddation. + // Output is {0 to <1}. + AF1 FsrTepdDitF(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + // The 1.61803 golden ratio. + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + // Number designed to provide a good visual pattern. + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AFractF1(x);} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 8-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC8F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(255.0))*AF3_(1.0/255.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/255.0);b=b*b; + // Ratio of 'a' to 'b' required to produce 'c'. + // APrxLoRcpF1() won't work here (at least for very high dynamic ranges). + // APrxMedRcpF1() is an IADD,FMA,MUL. + AF3 r=(c-b)*APrxMedRcpF3(a-b); + // Use the ratio as a cutoff to choose 'a' or 'b'. + // AGtZeroF1() is a MUL. + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + // This version is 10-bit gamma 2.0. + // The 'c' input is {0 to 1}. + // Output is {0 to 1} ready for image store. + void FsrTepdC10F(inout AF3 c,AF1 dit){ + AF3 n=sqrt(c); + n=floor(n*AF3_(1023.0))*AF3_(1.0/1023.0); + AF3 a=n*n; + AF3 b=n+AF3_(1.0/1023.0);b=b*b; + AF3 r=(c-b)*APrxMedRcpF3(a-b); + c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/1023.0));} +#endif +//============================================================================================================================== +#if defined(A_GPU)&&defined(A_HALF) + AH1 FsrTepdDitH(AU2 p,AU1 f){ + AF1 x=AF1_(p.x+f); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*a+(y*b); + return AH1(AFractF1(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(255.0))*AH3_(1.0/255.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/255.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10H(inout AH3 c,AH1 dit){ + AH3 n=sqrt(c); + n=floor(n*AH3_(1023.0))*AH3_(1.0/1023.0); + AH3 a=n*n; + AH3 b=n+AH3_(1.0/1023.0);b=b*b; + AH3 r=(c-b)*APrxMedRcpH3(a-b); + c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/1023.0));} +//============================================================================================================================== + // This computes dither for positions 'p' and 'p+{8,0}'. + AH2 FsrTepdDitHx2(AU2 p,AU1 f){ + AF2 x; + x.x=AF1_(p.x+f); + x.y=x.x+AF1_(8.0); + AF1 y=AF1_(p.y); + AF1 a=AF1_((1.0+sqrt(5.0))/2.0); + AF1 b=AF1_(1.0/3.69); + x=x*AF2_(a)+AF2_(y*b); + return AH2(AFractF2(x));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC8Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(255.0))*AH2_(1.0/255.0); + nG=floor(nG*AH2_(255.0))*AH2_(1.0/255.0); + nB=floor(nB*AH2_(255.0))*AH2_(1.0/255.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/255.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/255.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/255.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/255.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/255.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/255.0));} +//------------------------------------------------------------------------------------------------------------------------------ + void FsrTepdC10Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ + AH2 nR=sqrt(cR); + AH2 nG=sqrt(cG); + AH2 nB=sqrt(cB); + nR=floor(nR*AH2_(1023.0))*AH2_(1.0/1023.0); + nG=floor(nG*AH2_(1023.0))*AH2_(1.0/1023.0); + nB=floor(nB*AH2_(1023.0))*AH2_(1.0/1023.0); + AH2 aR=nR*nR; + AH2 aG=nG*nG; + AH2 aB=nB*nB; + AH2 bR=nR+AH2_(1.0/1023.0);bR=bR*bR; + AH2 bG=nG+AH2_(1.0/1023.0);bG=bG*bG; + AH2 bB=nB+AH2_(1.0/1023.0);bB=bB*bB; + AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); + AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); + AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); + cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/1023.0)); + cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/1023.0)); + cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/1023.0));} +#endif diff --git a/client/renderers/EGL/shader/ffx_fsr1_easu.frag b/client/renderers/EGL/shader/ffx_fsr1_easu.frag new file mode 100644 index 00000000..4ca34308 --- /dev/null +++ b/client/renderers/EGL/shader/ffx_fsr1_easu.frag @@ -0,0 +1,48 @@ +#version 300 es +precision mediump float; + +#include "compat.h" + +in vec2 iFragCoord; +out vec4 fragColor; + +uniform sampler2D iChannel0; +uniform uvec2 uInRes[8]; +uniform uvec2 uOutRes; + +#define A_GPU 1 +#define A_GLSL 1 +#define A_FULL 1 + +#include "ffx_a.h" + +#define FSR_EASU_F 1 +#define FSR_RCAS_F 1 + +AF4 FsrEasuRF(AF2 p){return AF4(textureGather(iChannel0, p, 0));} +AF4 FsrEasuGF(AF2 p){return AF4(textureGather(iChannel0, p, 1));} +AF4 FsrEasuBF(AF2 p){return AF4(textureGather(iChannel0, p, 2));} + +#include "ffx_fsr1.h" + +void main() +{ + AU4 con0, con1, con2, con3; + vec2 inRes = vec2(uInRes[0]); + vec2 outRes = vec2(uOutRes); + + FsrEasuCon( + con0, + con1, + con2, + con3, + inRes.x , inRes.y, + inRes.x , inRes.y, + outRes.x, outRes.y + ); + + vec3 color; + uvec2 point = uvec2(iFragCoord * outRes); + FsrEasuF(color, point, con0, con1, con2, con3); + fragColor = vec4(color.xyz, 1); +} diff --git a/client/renderers/EGL/shader/ffx_fsr1_rcas.frag b/client/renderers/EGL/shader/ffx_fsr1_rcas.frag new file mode 100644 index 00000000..1efed7a8 --- /dev/null +++ b/client/renderers/EGL/shader/ffx_fsr1_rcas.frag @@ -0,0 +1,36 @@ +#version 300 es +precision mediump float; + +#include "compat.h" + +in vec2 iFragCoord; +out vec4 fragColor; + +uniform sampler2D iChannel0; +uniform uvec2 uInRes[8]; +uniform float uSharpness; + +#define A_GPU 1 +#define A_GLSL 1 +#define A_FULL 1 + +#include "ffx_a.h" + +AF4 FsrRcasLoadF(ASU2 p) { return texelFetch(iChannel0, ASU2(p), 0); } +void FsrRcasInputF(inout AF1 r, inout AF1 g, inout AF1 b) {} + +#define FSR_RCAS_F 1 +#define FSR_RCAS_DENOISE 1 +#include "ffx_fsr1.h" + +void main() +{ + vec2 inRes = vec2(uInRes[0]); + uvec2 point = uvec2(iFragCoord * inRes); + + uvec4 const0; + FsrRcasCon(const0, 1.0f - uSharpness); + + FsrRcasF(fragColor.r, fragColor.g, fragColor.b, point, const0); + fragColor.a = 1.0f; +} diff --git a/client/renderers/EGL/texture.c b/client/renderers/EGL/texture.c index c6a45cdb..d949d69f 100644 --- a/client/renderers/EGL/texture.c +++ b/client/renderers/EGL/texture.c @@ -42,10 +42,10 @@ typedef struct RenderStep EGL_Texture *owner; bool enabled; + bool ready; GLuint fb; GLuint tex; EGL_Shader * shader; - float scale; unsigned int width, height; @@ -92,8 +92,6 @@ bool egl_textureInit(EGL * egl, EGL_Texture ** texture_, glSamplerParameteri(this->sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR); glSamplerParameteri(this->sampler, GL_TEXTURE_WRAP_S , GL_CLAMP_TO_EDGE); glSamplerParameteri(this->sampler, GL_TEXTURE_WRAP_T , GL_CLAMP_TO_EDGE); - - this->scale = 1.0f; return true; } @@ -126,15 +124,15 @@ void egl_textureFree(EGL_Texture ** tex) bool setupRenderStep(EGL_Texture * this, RenderStep * step) { - step->width = this->format.width * step->scale; - step->height = this->format.height * step->scale; + if (step->ready && (step->width > 0 || step->height > 0)) + return true; glBindTexture(GL_TEXTURE_2D, step->tex); glTexImage2D(GL_TEXTURE_2D, 0, this->format.intFormat, - step->width, - step->height, + step->width > 0 ? step->width : this->format.width, + step->height > 0 ? step->height : this->format.height, 0, this->format.format, this->format.dataType, @@ -143,6 +141,7 @@ bool setupRenderStep(EGL_Texture * this, RenderStep * step) glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); glBindTexture(GL_TEXTURE_2D, 0); + step->ready = true; return true; } @@ -168,8 +167,7 @@ bool egl_textureSetup(EGL_Texture * this, enum EGL_PixelFormat pixFmt, { RenderStep * step; for(ll_reset(this->render); ll_walk(this->render, (void **)&step); ) - if (!setupRenderStep(this, step)) - return false; + step->ready = false; } return this->ops.setup(this, &setup); @@ -309,9 +307,25 @@ enum EGL_TexStatus egl_textureBind(EGL_Texture * this) { ringbuffer_reset(this->textures); + /* configure all the filters */ + for(ll_reset(this->render); ll_walk(this->render, (void **)&step); ) + { + if (!step->enabled) + continue; + + if (!step->ready) + setupRenderStep(this, step); + } + if ((status = this->ops.get(this, &tex)) != EGL_TEX_STATUS_OK) return status; + struct Rect finalSz = + { + .x = this->format.width, + .y = this->format.height + }; + ringbuffer_push(this->textures, &(BindInfo) { .tex = tex, .width = this->format.width, @@ -346,7 +360,13 @@ enum EGL_TexStatus egl_textureBind(EGL_Texture * this) else glBindFramebuffer(GL_FRAMEBUFFER, step->fb); - glViewport(0, 0, step->width, step->height); + const struct Rect sz = + { + .x = step->width > 0 ? step->width : this->format.width, + .y = step->height > 0 ? step->height : this->format.height + }; + + glViewport(0, 0, sz.x, sz.y); /* use the shader (also configures it's set uniforms) */ egl_shaderUse(step->shader); @@ -354,16 +374,18 @@ enum EGL_TexStatus egl_textureBind(EGL_Texture * this) /* set the size uniforms */ glUniform2uiv(step->uInRes, ringbuffer_getCount(this->textures), bd->dimensions); - glUniform2ui(step->uOutRes, step->width, step->height); + glUniform2ui(step->uOutRes, sz.x, sz.y); /* render the scene */ egl_modelRender(this->model); + finalSz.x = sz.x; + finalSz.y = sz.y; /* push the details into the ringbuffer for the next pass */ ringbuffer_push(this->textures, &(BindInfo) { .tex = step->tex, - .width = step->width, - .height = step->height + .width = sz.x, + .height = sz.y }); /* bind the textures for the next pass */ @@ -378,6 +400,8 @@ enum EGL_TexStatus egl_textureBind(EGL_Texture * this) egl_resetViewport(this->egl); } + this->finalWidth = finalSz.x; + this->finalHeight = finalSz.y; this->postProcessed = true; } else @@ -393,7 +417,7 @@ enum EGL_TexStatus egl_textureBind(EGL_Texture * this) } PostProcessHandle egl_textureAddFilter(EGL_Texture * this, EGL_Shader * shader, - float outputScale, bool enabled) + bool enabled) { if (!this->render) { @@ -407,21 +431,10 @@ PostProcessHandle egl_textureAddFilter(EGL_Texture * this, EGL_Shader * shader, glGenTextures(1, &step->tex); step->owner = this; step->shader = shader; - step->scale = outputScale; step->uInRes = egl_shaderGetUniform(shader, "uInRes" ); step->uOutRes = egl_shaderGetUniform(shader, "uOutRes"); step->enabled = enabled; - this->scale = outputScale; - - if (this->formatValid) - if (!setupRenderStep(this, step)) - { - glDeleteTextures(1, &step->tex); - free(step); - return NULL; - } - ll_push(this->render, step); return (PostProcessHandle)step; } @@ -436,12 +449,33 @@ void egl_textureEnableFilter(PostProcessHandle * handle, bool enable) egl_textureInvalidate(step->owner); } +void egl_textureSetFilterRes(PostProcessHandle * handle, + unsigned int x, unsigned int y) +{ + RenderStep * step = (RenderStep *)handle; + if (step->width == x && step->height == y) + return; + + step->width = x; + step->height = y; + step->ready = false; + egl_textureInvalidate(step->owner); +} + void egl_textureInvalidate(EGL_Texture * texture) { texture->postProcessed = false; } -float egl_textureGetScale(EGL_Texture * this) +void egl_textureGetFinalSize(EGL_Texture * this, struct Rect * rect) { - return this->scale; + if (!this->render) + { + rect->x = this->format.width; + rect->y = this->format.height; + return; + } + + rect->x = this->finalWidth; + rect->y = this->finalHeight; } diff --git a/client/renderers/EGL/texture.h b/client/renderers/EGL/texture.h index 04b03a1f..62071677 100644 --- a/client/renderers/EGL/texture.h +++ b/client/renderers/EGL/texture.h @@ -26,6 +26,7 @@ #include "model.h" #include "common/framebuffer.h" #include "common/ringbuffer.h" +#include "common/types.h" #include "util.h" #include "ll.h" @@ -144,7 +145,7 @@ struct EGL_Texture _Atomic(bool) updated; bool postProcessed; EGL_Model * model; - float scale; + unsigned int finalWidth, finalHeight; void * bindData; int bindDataSize; @@ -172,10 +173,13 @@ enum EGL_TexStatus egl_textureBind(EGL_Texture * texture); typedef void * PostProcessHandle; PostProcessHandle egl_textureAddFilter(EGL_Texture * texture, - EGL_Shader * shader, float outputScale, bool enabled); + EGL_Shader * shader, bool enabled); void egl_textureEnableFilter(PostProcessHandle * handle, bool enable); +void egl_textureSetFilterRes(PostProcessHandle * handle, + unsigned int x, unsigned int y); + void egl_textureInvalidate(EGL_Texture * texture); -float egl_textureGetScale(EGL_Texture * texture); +void egl_textureGetFinalSize(EGL_Texture * texture, struct Rect * rect);