Vulkan: Reimplement FSR constant generation functions to avoid GCC warnings

author: Marshall Mohror 2021-10-23 00:11:12 -0500
committer: Fernando Sahmkow 2021-11-16 22:11:33 +0100
commit: bb0367548591a66cd963b14373ed5b9c61adabb1 (patch)
tree: 988222e55a86b1023678f68b42617f55648c63d7 /src
parent: vk_blit_screen: Fix AA destruction order (diff)
download: yuzu-bb0367548591a66cd963b14373ed5b9c61adabb1.tar.gz
yuzu-bb0367548591a66cd963b14373ed5b9c61adabb1.tar.xz
yuzu-bb0367548591a66cd963b14373ed5b9c61adabb1.zip
2 files changed, 145 insertions, 9 deletions
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index 07b94dcc8..91a30fef7 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -237,7 +237,6 @@ target_include_directories(video_core PRIVATE ${FFmpeg_INCLUDE_DIR})
 target_link_libraries(video_core PRIVATE ${FFmpeg_LIBRARIES})
 target_link_options(video_core PRIVATE ${FFmpeg_LDFLAGS})
-target_link_libraries(video_core PRIVATE ffx-fsr)
 add_dependencies(video_core host_shaders)
 target_include_directories(video_core PRIVATE ${HOST_SHADERS_INCLUDE})
 target_include_directories(video_core PRIVATE sirit ../../externals/Vulkan-Headers/include)
diff --git a/src/video_core/renderer_vulkan/vk_fsr.cpp b/src/video_core/renderer_vulkan/vk_fsr.cpp
index 9288aa7c2..2feaa9d37 100644
--- a/src/video_core/renderer_vulkan/vk_fsr.cpp
+++ b/src/video_core/renderer_vulkan/vk_fsr.cpp
@@ -13,9 +13,146 @@
 #include "video_core/renderer_vulkan/vk_shader_util.h"
 #include "video_core/vulkan_common/vulkan_device.h"
-#define A_CPU
+// Reimplementations of the constant generating functions in ffx_fsr1.h
-#include <ffx_a.h>
+// GCC generated a lot of warnings when using the official header.
-#include <ffx_fsr1.h>
+static u32 AU1_AH1_AF1(f32 f) {
+    static constexpr u32 base[512]{
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
+        0x0000, 0x0000, 0x0000, 0x0000, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040,
+        0x0080, 0x0100, 0x0200, 0x0400, 0x0800, 0x0c00, 0x1000, 0x1400, 0x1800, 0x1c00, 0x2000,
+        0x2400, 0x2800, 0x2c00, 0x3000, 0x3400, 0x3800, 0x3c00, 0x4000, 0x4400, 0x4800, 0x4c00,
+        0x5000, 0x5400, 0x5800, 0x5c00, 0x6000, 0x6400, 0x6800, 0x6c00, 0x7000, 0x7400, 0x7800,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
+        0x7bff, 0x7bff, 0x7bff, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
+        0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8001, 0x8002, 0x8004, 0x8008,
+        0x8010, 0x8020, 0x8040, 0x8080, 0x8100, 0x8200, 0x8400, 0x8800, 0x8c00, 0x9000, 0x9400,
+        0x9800, 0x9c00, 0xa000, 0xa400, 0xa800, 0xac00, 0xb000, 0xb400, 0xb800, 0xbc00, 0xc000,
+        0xc400, 0xc800, 0xcc00, 0xd000, 0xd400, 0xd800, 0xdc00, 0xe000, 0xe400, 0xe800, 0xec00,
+        0xf000, 0xf400, 0xf800, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+        0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
+    };
+    static constexpr s8 shift[512]{
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17, 0x16,
+        0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
+        0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
+        0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17,
+        0x16, 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
+        0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
+        0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
+        0x18, 0x18,
+    };
+    auto u = std::bit_cast<u32>(f);
+    u32 i = u >> 23;
+    return base[i] + ((u & 0x7fffff) >> shift[i]);
+}
+static u32 AU1_AH2_AF2(f32 a[2]) {
+    return AU1_AH1_AF1(a[0]) + (AU1_AH1_AF1(a[1]) << 16);
+}
+static void FsrEasuCon(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4],
+                       f32 inputViewportInPixelsX, f32 inputViewportInPixelsY,
+                       f32 inputSizeInPixelsX, f32 inputSizeInPixelsY, f32 outputSizeInPixelsX,
+                       f32 outputSizeInPixelsY) {
+    con0[0] = std::bit_cast<u32>(inputViewportInPixelsX / outputSizeInPixelsX);
+    con0[1] = std::bit_cast<u32>(inputViewportInPixelsY / outputSizeInPixelsY);
+    con0[2] = std::bit_cast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f);
+    con0[3] = std::bit_cast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f);
+    con1[0] = std::bit_cast<u32>(1.0f / inputSizeInPixelsX);
+    con1[1] = std::bit_cast<u32>(1.0f / inputSizeInPixelsY);
+    con1[2] = std::bit_cast<u32>(1.0f / inputSizeInPixelsX);
+    con1[3] = std::bit_cast<u32>(-1.0f / inputSizeInPixelsY);
+    con2[0] = std::bit_cast<u32>(-1.0f / inputSizeInPixelsX);
+    con2[1] = std::bit_cast<u32>(2.0f / inputSizeInPixelsY);
+    con2[2] = std::bit_cast<u32>(1.0f / inputSizeInPixelsX);
+    con2[3] = std::bit_cast<u32>(2.0f / inputSizeInPixelsY);
+    con3[0] = std::bit_cast<u32>(0.0f / inputSizeInPixelsX);
+    con3[1] = std::bit_cast<u32>(4.0f / inputSizeInPixelsY);
+    con3[2] = con3[3] = 0;
+}
+static void FsrEasuConOffset(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4],
+                             f32 inputViewportInPixelsX, f32 inputViewportInPixelsY,
+                             f32 inputSizeInPixelsX, f32 inputSizeInPixelsY,
+                             f32 outputSizeInPixelsX, f32 outputSizeInPixelsY,
+                             f32 inputOffsetInPixelsX, f32 inputOffsetInPixelsY) {
+    FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY,
+               inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY);
+    con0[2] = std::bit_cast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f +
+                                 inputOffsetInPixelsX);
+    con0[3] = std::bit_cast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f +
+                                 inputOffsetInPixelsY);
+}
+static void FsrRcasCon(u32* con, f32 sharpness) {
+    sharpness = std::exp2f(-sharpness);
+    f32 hSharp[2]{sharpness, sharpness};
+    con[0] = std::bit_cast<u32>(sharpness);
+    con[1] = AU1_AH2_AF2(hSharp);
+    con[2] = 0;
+    con[3] = 0;
+}
 namespace Vulkan {
@@ -62,15 +199,15 @@ VkImageView FSR::Draw(VKScheduler& scheduler, size_t image_index, VkImageView im
        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *easu_pipeline);
-        std::array<AU1, 4 * 4> push_constants;
+        std::array<u32, 4 * 4> push_constants;
        FsrEasuConOffset(
            push_constants.data() + 0, push_constants.data() + 4, push_constants.data() + 8,
            push_constants.data() + 12,
-            static_cast<AF1>(crop_rect.GetWidth()), static_cast<AF1>(crop_rect.GetHeight()),
+            static_cast<f32>(crop_rect.GetWidth()), static_cast<f32>(crop_rect.GetHeight()),
-            static_cast<AF1>(input_image_extent.width), static_cast<AF1>(input_image_extent.height),
+            static_cast<f32>(input_image_extent.width), static_cast<f32>(input_image_extent.height),
-            static_cast<AF1>(output_size.width), static_cast<AF1>(output_size.height),
+            static_cast<f32>(output_size.width), static_cast<f32>(output_size.height),
-            static_cast<AF1>(crop_rect.left), static_cast<AF1>(crop_rect.top));
+            static_cast<f32>(crop_rect.left), static_cast<f32>(crop_rect.top));
        cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, push_constants);
        {
author	Marshall Mohror	2021-10-23 00:11:12 -0500
committer	Fernando Sahmkow	2021-11-16 22:11:33 +0100
commit	bb0367548591a66cd963b14373ed5b9c61adabb1 (patch)
tree	988222e55a86b1023678f68b42617f55648c63d7 /src
parent	vk_blit_screen: Fix AA destruction order (diff)
download	yuzu-bb0367548591a66cd963b14373ed5b9c61adabb1.tar.gz yuzu-bb0367548591a66cd963b14373ed5b9c61adabb1.tar.xz yuzu-bb0367548591a66cd963b14373ed5b9c61adabb1.zip

diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt index 07b94dcc8..91a30fef7 100644 --- a/src/video_core/CMakeLists.txt +++ b/src/video_core/CMakeLists.txt
@@ -237,7 +237,6 @@ target_include_directories(video_core PRIVATE ${FFmpeg_INCLUDE_DIR})
237	target_link_libraries(video_core PRIVATE ${FFmpeg_LIBRARIES})	237	target_link_libraries(video_core PRIVATE ${FFmpeg_LIBRARIES})
238	target_link_options(video_core PRIVATE ${FFmpeg_LDFLAGS})	238	target_link_options(video_core PRIVATE ${FFmpeg_LDFLAGS})
239		239
240	target_link_libraries(video_core PRIVATE ffx-fsr)
241	add_dependencies(video_core host_shaders)	240	add_dependencies(video_core host_shaders)
242	target_include_directories(video_core PRIVATE ${HOST_SHADERS_INCLUDE})	241	target_include_directories(video_core PRIVATE ${HOST_SHADERS_INCLUDE})
243	target_include_directories(video_core PRIVATE sirit ../../externals/Vulkan-Headers/include)	242	target_include_directories(video_core PRIVATE sirit ../../externals/Vulkan-Headers/include)


diff --git a/src/video_core/renderer_vulkan/vk_fsr.cpp b/src/video_core/renderer_vulkan/vk_fsr.cpp index 9288aa7c2..2feaa9d37 100644 --- a/src/video_core/renderer_vulkan/vk_fsr.cpp +++ b/src/video_core/renderer_vulkan/vk_fsr.cpp
@@ -13,9 +13,146 @@
13	#include "video_core/renderer_vulkan/vk_shader_util.h"	13	#include "video_core/renderer_vulkan/vk_shader_util.h"
14	#include "video_core/vulkan_common/vulkan_device.h"	14	#include "video_core/vulkan_common/vulkan_device.h"
15		15
16	#define A_CPU	16	// Reimplementations of the constant generating functions in ffx_fsr1.h
17	#include <ffx_a.h>	17	// GCC generated a lot of warnings when using the official header.
18	#include <ffx_fsr1.h>	18	static u32 AU1_AH1_AF1(f32 f) {
		19	static constexpr u32 base[512]{
		20	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		21	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		22	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		23	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		24	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		25	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		26	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		27	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		28	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
		29	0x0000, 0x0000, 0x0000, 0x0000, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040,
		30	0x0080, 0x0100, 0x0200, 0x0400, 0x0800, 0x0c00, 0x1000, 0x1400, 0x1800, 0x1c00, 0x2000,
		31	0x2400, 0x2800, 0x2c00, 0x3000, 0x3400, 0x3800, 0x3c00, 0x4000, 0x4400, 0x4800, 0x4c00,
		32	0x5000, 0x5400, 0x5800, 0x5c00, 0x6000, 0x6400, 0x6800, 0x6c00, 0x7000, 0x7400, 0x7800,
		33	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		34	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		35	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		36	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		37	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		38	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		39	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		40	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		41	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		42	0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff, 0x7bff,
		43	0x7bff, 0x7bff, 0x7bff, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		44	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		45	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		46	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		47	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		48	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		49	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		50	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		51	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000,
		52	0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8001, 0x8002, 0x8004, 0x8008,
		53	0x8010, 0x8020, 0x8040, 0x8080, 0x8100, 0x8200, 0x8400, 0x8800, 0x8c00, 0x9000, 0x9400,
		54	0x9800, 0x9c00, 0xa000, 0xa400, 0xa800, 0xac00, 0xb000, 0xb400, 0xb800, 0xbc00, 0xc000,
		55	0xc400, 0xc800, 0xcc00, 0xd000, 0xd400, 0xd800, 0xdc00, 0xe000, 0xe400, 0xe800, 0xec00,
		56	0xf000, 0xf400, 0xf800, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		57	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		58	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		59	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		60	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		61	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		62	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		63	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		64	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		65	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		66	0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff, 0xfbff,
		67	};
		68	static constexpr s8 shift[512]{
		69	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		70	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		71	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		72	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		73	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		74	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		75	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17, 0x16,
		76	0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
		77	0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
		78	0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		79	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		80	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		81	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		82	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		83	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		84	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		85	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		86	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		87	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		88	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		89	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		90	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		91	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		92	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x17,
		93	0x16, 0x15, 0x14, 0x13, 0x12, 0x11, 0x10, 0x0f, 0x0e, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
		94	0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d,
		95	0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x0d, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		96	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		97	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		98	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		99	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		100	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		101	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		102	0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18,
		103	0x18, 0x18,
		104	};
		105	auto u = std::bit_cast<u32>(f);
		106	u32 i = u >> 23;
		107	return base[i] + ((u & 0x7fffff) >> shift[i]);
		108	}
		109
		110	static u32 AU1_AH2_AF2(f32 a[2]) {
		111	return AU1_AH1_AF1(a[0]) + (AU1_AH1_AF1(a[1]) << 16);
		112	}
		113
		114	static void FsrEasuCon(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4],
		115	f32 inputViewportInPixelsX, f32 inputViewportInPixelsY,
		116	f32 inputSizeInPixelsX, f32 inputSizeInPixelsY, f32 outputSizeInPixelsX,
		117	f32 outputSizeInPixelsY) {
		118	con0[0] = std::bit_cast<u32>(inputViewportInPixelsX / outputSizeInPixelsX);
		119	con0[1] = std::bit_cast<u32>(inputViewportInPixelsY / outputSizeInPixelsY);
		120	con0[2] = std::bit_cast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f);
		121	con0[3] = std::bit_cast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f);
		122	con1[0] = std::bit_cast<u32>(1.0f / inputSizeInPixelsX);
		123	con1[1] = std::bit_cast<u32>(1.0f / inputSizeInPixelsY);
		124	con1[2] = std::bit_cast<u32>(1.0f / inputSizeInPixelsX);
		125	con1[3] = std::bit_cast<u32>(-1.0f / inputSizeInPixelsY);
		126	con2[0] = std::bit_cast<u32>(-1.0f / inputSizeInPixelsX);
		127	con2[1] = std::bit_cast<u32>(2.0f / inputSizeInPixelsY);
		128	con2[2] = std::bit_cast<u32>(1.0f / inputSizeInPixelsX);
		129	con2[3] = std::bit_cast<u32>(2.0f / inputSizeInPixelsY);
		130	con3[0] = std::bit_cast<u32>(0.0f / inputSizeInPixelsX);
		131	con3[1] = std::bit_cast<u32>(4.0f / inputSizeInPixelsY);
		132	con3[2] = con3[3] = 0;
		133	}
		134
		135	static void FsrEasuConOffset(u32 con0[4], u32 con1[4], u32 con2[4], u32 con3[4],
		136	f32 inputViewportInPixelsX, f32 inputViewportInPixelsY,
		137	f32 inputSizeInPixelsX, f32 inputSizeInPixelsY,
		138	f32 outputSizeInPixelsX, f32 outputSizeInPixelsY,
		139	f32 inputOffsetInPixelsX, f32 inputOffsetInPixelsY) {
		140	FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY,
		141	inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY);
		142	con0[2] = std::bit_cast<u32>(0.5f * inputViewportInPixelsX / outputSizeInPixelsX - 0.5f +
		143	inputOffsetInPixelsX);
		144	con0[3] = std::bit_cast<u32>(0.5f * inputViewportInPixelsY / outputSizeInPixelsY - 0.5f +
		145	inputOffsetInPixelsY);
		146	}
		147
		148	static void FsrRcasCon(u32* con, f32 sharpness) {
		149	sharpness = std::exp2f(-sharpness);
		150	f32 hSharp[2]{sharpness, sharpness};
		151	con[0] = std::bit_cast<u32>(sharpness);
		152	con[1] = AU1_AH2_AF2(hSharp);
		153	con[2] = 0;
		154	con[3] = 0;
		155	}
19		156
20	namespace Vulkan {	157	namespace Vulkan {
21		158
@@ -62,15 +199,15 @@ VkImageView FSR::Draw(VKScheduler& scheduler, size_t image_index, VkImageView im
62		199
63	cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *easu_pipeline);	200	cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *easu_pipeline);
64		201
65	std::array<AU1, 4 * 4> push_constants;	202	std::array<u32, 4 * 4> push_constants;
66	FsrEasuConOffset(	203	FsrEasuConOffset(
67	push_constants.data() + 0, push_constants.data() + 4, push_constants.data() + 8,	204	push_constants.data() + 0, push_constants.data() + 4, push_constants.data() + 8,
68	push_constants.data() + 12,	205	push_constants.data() + 12,
69		206
70	static_cast<AF1>(crop_rect.GetWidth()), static_cast<AF1>(crop_rect.GetHeight()),	207	static_cast<f32>(crop_rect.GetWidth()), static_cast<f32>(crop_rect.GetHeight()),
71	static_cast<AF1>(input_image_extent.width), static_cast<AF1>(input_image_extent.height),	208	static_cast<f32>(input_image_extent.width), static_cast<f32>(input_image_extent.height),
72	static_cast<AF1>(output_size.width), static_cast<AF1>(output_size.height),	209	static_cast<f32>(output_size.width), static_cast<f32>(output_size.height),
73	static_cast<AF1>(crop_rect.left), static_cast<AF1>(crop_rect.top));	210	static_cast<f32>(crop_rect.left), static_cast<f32>(crop_rect.top));
74	cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, push_constants);	211	cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, push_constants);
75		212
76	{	213	{