Mirrors/yuzu
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Presentation: Only use FP16 in scaling shaders on supported devices in Vulkan

Browse source
This commit is contained in:
Marshall Mohror 2021-10-22 23:09:29 -05:00 • committed by Fernando Sahmkow
parent a39e867c73
commit dcc5b4f6b0
15 changed files with 199 additions and 115 deletions
Download patch file Download diff file Expand all files Collapse all files

10
src/video_core/host_shaders/CMakeLists.txt
View file

@ -18,16 +18,20 @@ set(SHADER_FILES
opengl_copy_bc4.comp
opengl_present.frag
opengl_present.vert
opengl_present_scaleforce.frag
pitch_unswizzle.comp
present_scaleforce.frag
present_bicubic.frag
present_gaussian.frag
vulkan_blit_color_float.frag
vulkan_blit_depth_stencil.frag
vulkan_fidelityfx_fsr_easu.comp
vulkan_fidelityfx_fsr_rcas.comp
vulkan_fidelityfx_fsr_easu_fp16.comp
vulkan_fidelityfx_fsr_easu_fp32.comp
vulkan_fidelityfx_fsr_rcas_fp16.comp
vulkan_fidelityfx_fsr_rcas_fp32.comp
vulkan_present.frag
vulkan_present.vert
vulkan_present_scaleforce_fp16.frag
vulkan_present_scaleforce_fp32.frag
vulkan_quad_indexed.comp
vulkan_uint8.comp
)

106
src/video_core/host_shaders/fidelityfx_fsr.comp
View file

@ -28,80 +28,82 @@
// THE SOFTWARE.
layout( push_constant ) uniform constants {
u32vec2 input_size;
uvec4 Const0;
uvec4 Const1;
uvec4 Const2;
uvec4 Const3;
};
uvec4 Const0;
uvec4 Const1;
uvec4 Const2;
uvec4 Const3;
layout(set=0,binding=0) uniform sampler2D InputTexture;
layout(set=0,binding=1,rgba16f) uniform image2D OutputTexture;
#define A_GPU 1
#define A_GLSL 1
#define A_HALF
#include "ffx_a.h"
#ifndef YUZU_USE_FP16
#include "ffx_a.h"
f16vec4 LinearToSRGB(f16vec4 linear) {
bvec4 selector = greaterThan(linear, f16vec4(0.00313066844250063));
f16vec4 low = linear * float16_t(12.92);
f16vec4 high = float16_t(1.055) * pow(linear, f16vec4(1 / 2.4)) - float16_t(0.055);
return mix(low, high, selector);
}
#if USE_EASU
#define FSR_EASU_F 1
AF4 FsrEasuRF(AF2 p) { AF4 res = textureGather(InputTexture, p, 0); return res; }
AF4 FsrEasuGF(AF2 p) { AF4 res = textureGather(InputTexture, p, 1); return res; }
AF4 FsrEasuBF(AF2 p) { AF4 res = textureGather(InputTexture, p, 2); return res; }
#endif
#if USE_RCAS
#define FSR_RCAS_F 1
AF4 FsrRcasLoadF(ASU2 p) { return texelFetch(InputTexture, ASU2(p), 0); }
void FsrRcasInputF(inout AF1 r, inout AF1 g, inout AF1 b) {}
#endif
#else
#define A_HALF
#include "ffx_a.h"
f16vec4 SRGBToLinear(f16vec4 srgb) {
bvec4 selector = greaterThan(srgb, f16vec4(0.0404482362771082));
f16vec4 low = srgb * float16_t(1.0 / 12.92);
f16vec4 high = pow((srgb + float16_t(0.055)) * float16_t(1.0 / 1.055), f16vec4(2.4));
return mix(low, high, selector);
}
#if USE_EASU
#define FSR_EASU_H 1
f16vec4 FsrEasuRH(vec2 p) { f16vec4 res = f16vec4(textureGather(InputTexture, p, 0)); return res; }
f16vec4 FsrEasuGH(vec2 p) { f16vec4 res = f16vec4(textureGather(InputTexture, p, 1)); return res; }
f16vec4 FsrEasuBH(vec2 p) { f16vec4 res = f16vec4(textureGather(InputTexture, p, 2)); return res; }
#endif
#if USE_RCAS
#define FSR_RCAS_H 1
f16vec4 FsrRcasLoadH(ASW2 p) { return f16vec4(texelFetch(InputTexture, ASU2(p), 0)); }
void FsrRcasInputH(inout float16_t r, inout float16_t g, inout float16_t b) {}
#if USE_EASU
#define FSR_EASU_H 1
AH4 FsrEasuRH(AF2 p) { AH4 res = AH4(textureGather(InputTexture, p, 0)); return res; }
AH4 FsrEasuGH(AF2 p) { AH4 res = AH4(textureGather(InputTexture, p, 1)); return res; }
AH4 FsrEasuBH(AF2 p) { AH4 res = AH4(textureGather(InputTexture, p, 2)); return res; }
#endif
#if USE_RCAS
#define FSR_RCAS_H 1
AH4 FsrRcasLoadH(ASW2 p) { return AH4(texelFetch(InputTexture, ASU2(p), 0)); }
void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b){}
#endif
#endif
#include "ffx_fsr1.h"
void CurrFilter(u32vec2 pos) {
// For debugging
void CurrFilter(AU2 pos) {
#if USE_BILINEAR
vec2 pp = (vec2(pos) * vec2_AU2(Const0.xy) + vec2_AU2(Const0.zw)) * vec2_AU2(Const1.xy) + vec2(0.5, -0.5) * vec2_AU2(Const1.zw);
imageStore(OutputTexture, ivec2(pos), textureLod(InputTexture, pp, 0.0));
AF2 pp = (AF2(pos) * AF2_AU2(Const0.xy) + AF2_AU2(Const0.zw)) * AF2_AU2(Const1.xy) + AF2(0.5, -0.5) * AF2_AU2(Const1.zw);
imageStore(OutputTexture, ASU2(pos), textureLod(InputTexture, pp, 0.0));
#endif
#if USE_EASU
f16vec3 c;
FsrEasuH(c, pos, Const0, Const1, Const2, Const3);
imageStore(OutputTexture, ivec2(pos), f16vec4(c, 1));
#ifndef YUZU_USE_FP16
AF3 c;
FsrEasuF(c, pos, Const0, Const1, Const2, Const3);
imageStore(OutputTexture, ASU2(pos), AF4(c, 1));
#else
AH3 c;
FsrEasuH(c, pos, Const0, Const1, Const2, Const3);
imageStore(OutputTexture, ASU2(pos), AH4(c, 1));
#endif
#endif
#if USE_RCAS
f16vec3 c;
FsrRcasH(c.r, c.g, c.b, pos, Const0);
imageStore(OutputTexture, ivec2(pos), f16vec4(c, 1));
#ifndef YUZU_USE_FP16
AF3 c;
FsrRcasF(c.r, c.g, c.b, pos, Const0);
imageStore(OutputTexture, ASU2(pos), AF4(c, 1));
#else
AH3 c;
FsrRcasH(c.r, c.g, c.b, pos, Const0);
imageStore(OutputTexture, ASU2(pos), AH4(c, 1));
#endif
#endif
}
layout(local_size_x=64) in;
void main() {
#if USE_EASU || USE_BILINEAR
vec2 ires = vec2(input_size);
vec2 tres = textureSize(InputTexture, 0);
vec2 ores = imageSize(OutputTexture);
FsrEasuCon(Const0, Const1, Const2, Const3, ires.x, ires.y, tres.x, tres.y, ores.x, ores.y);
#endif
#if USE_RCAS
FsrRcasCon(Const0, 0.25f);
#endif
// Do remapping of local xy in workgroup for a more PS-like swizzle pattern.
AU2 gxy = ARmp8x8(gl_LocalInvocationID.x) + AU2(gl_WorkGroupID.x << 4u, gl_WorkGroupID.y << 4u);
CurrFilter(gxy);

84
src/video_core/host_shaders/present_scaleforce.frag → src/video_core/host_shaders/opengl_present_scaleforce.frag
View file

@ -22,11 +22,29 @@
// Adapted from https://github.com/BreadFish64/ScaleFish/tree/master/scaleforce
#version 460
//! #version 460
#extension GL_ARB_separate_shader_objects : enable
#ifdef YUZU_USE_FP16
#extension GL_AMD_gpu_shader_half_float : enable
#extension GL_NV_gpu_shader5 : enable
#define lfloat float16_t
#define lvec2 f16vec2
#define lvec3 f16vec3
#define lvec4 f16vec4
#else
#define lfloat float
#define lvec2 vec2
#define lvec3 vec3
#define lvec4 vec4
#endif
#ifdef VULKAN
#define BINDING_COLOR_TEXTURE 1
@ -45,25 +63,25 @@ layout (binding = BINDING_COLOR_TEXTURE) uniform sampler2D input_texture;
const bool ignore_alpha = true;
float16_t ColorDist1(f16vec4 a, f16vec4 b) {
lfloat ColorDist1(lvec4 a, lvec4 b) {
// https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.2020_conversion
const f16vec3 K = f16vec3(0.2627, 0.6780, 0.0593);
const float16_t scaleB = float16_t(0.5) / (float16_t(1.0) - K.b);
const float16_t scaleR = float16_t(0.5) / (float16_t(1.0) - K.r);
f16vec4 diff = a - b;
float16_t Y = dot(diff.rgb, K);
float16_t Cb = scaleB * (diff.b - Y);
float16_t Cr = scaleR * (diff.r - Y);
f16vec3 YCbCr = f16vec3(Y, Cb, Cr);
float16_t d = length(YCbCr);
const lvec3 K = lvec3(0.2627, 0.6780, 0.0593);
const lfloat scaleB = lfloat(0.5) / (lfloat(1.0) - K.b);
const lfloat scaleR = lfloat(0.5) / (lfloat(1.0) - K.r);
lvec4 diff = a - b;
lfloat Y = dot(diff.rgb, K);
lfloat Cb = scaleB * (diff.b - Y);
lfloat Cr = scaleR * (diff.r - Y);
lvec3 YCbCr = lvec3(Y, Cb, Cr);
lfloat d = length(YCbCr);
if (ignore_alpha) {
return d;
}
return sqrt(a.a * b.a * d * d + diff.a * diff.a);
}
f16vec4 ColorDist(f16vec4 ref, f16vec4 A, f16vec4 B, f16vec4 C, f16vec4 D) {
return f16vec4(
lvec4 ColorDist(lvec4 ref, lvec4 A, lvec4 B, lvec4 C, lvec4 D) {
return lvec4(
ColorDist1(ref, A),
ColorDist1(ref, B),
ColorDist1(ref, C),
@ -72,36 +90,36 @@ f16vec4 ColorDist(f16vec4 ref, f16vec4 A, f16vec4 B, f16vec4 C, f16vec4 D) {
}
vec4 Scaleforce(sampler2D tex, vec2 tex_coord) {
f16vec4 bl = f16vec4(textureOffset(tex, tex_coord, ivec2(-1, -1)));
f16vec4 bc = f16vec4(textureOffset(tex, tex_coord, ivec2(0, -1)));
f16vec4 br = f16vec4(textureOffset(tex, tex_coord, ivec2(1, -1)));
f16vec4 cl = f16vec4(textureOffset(tex, tex_coord, ivec2(-1, 0)));
f16vec4 cc = f16vec4(texture(tex, tex_coord));
f16vec4 cr = f16vec4(textureOffset(tex, tex_coord, ivec2(1, 0)));
f16vec4 tl = f16vec4(textureOffset(tex, tex_coord, ivec2(-1, 1)));
f16vec4 tc = f16vec4(textureOffset(tex, tex_coord, ivec2(0, 1)));
f16vec4 tr = f16vec4(textureOffset(tex, tex_coord, ivec2(1, 1)));
lvec4 bl = lvec4(textureOffset(tex, tex_coord, ivec2(-1, -1)));
lvec4 bc = lvec4(textureOffset(tex, tex_coord, ivec2(0, -1)));
lvec4 br = lvec4(textureOffset(tex, tex_coord, ivec2(1, -1)));
lvec4 cl = lvec4(textureOffset(tex, tex_coord, ivec2(-1, 0)));
lvec4 cc = lvec4(texture(tex, tex_coord));
lvec4 cr = lvec4(textureOffset(tex, tex_coord, ivec2(1, 0)));
lvec4 tl = lvec4(textureOffset(tex, tex_coord, ivec2(-1, 1)));
lvec4 tc = lvec4(textureOffset(tex, tex_coord, ivec2(0, 1)));
lvec4 tr = lvec4(textureOffset(tex, tex_coord, ivec2(1, 1)));
f16vec4 offset_tl = ColorDist(cc, tl, tc, tr, cr);
f16vec4 offset_br = ColorDist(cc, br, bc, bl, cl);
lvec4 offset_tl = ColorDist(cc, tl, tc, tr, cr);
lvec4 offset_br = ColorDist(cc, br, bc, bl, cl);
// Calculate how different cc is from the texels around it
const float16_t plus_weight = float16_t(1.5);
const float16_t cross_weight = float16_t(1.5);
float16_t total_dist = dot(offset_tl + offset_br, f16vec4(cross_weight, plus_weight, cross_weight, plus_weight));
const lfloat plus_weight = lfloat(1.5);
const lfloat cross_weight = lfloat(1.5);
lfloat total_dist = dot(offset_tl + offset_br, lvec4(cross_weight, plus_weight, cross_weight, plus_weight));
if (total_dist == float16_t(0.0)) {
if (total_dist == lfloat(0.0)) {
return cc;
} else {
// Add together all the distances with direction taken into account
f16vec4 tmp = offset_tl - offset_br;
f16vec2 total_offset = tmp.wy * plus_weight + (tmp.zz + f16vec2(-tmp.x, tmp.x)) * cross_weight;
lvec4 tmp = offset_tl - offset_br;
lvec2 total_offset = tmp.wy * plus_weight + (tmp.zz + lvec2(-tmp.x, tmp.x)) * cross_weight;
// When the image has thin points, they tend to split apart.
// This is because the texels all around are different and total_offset reaches into clear areas.
// This works pretty well to keep the offset in bounds for these cases.
float16_t clamp_val = length(total_offset) / total_dist;
f16vec2 final_offset = clamp(total_offset, -clamp_val, clamp_val) / f16vec2(textureSize(tex, 0));
lfloat clamp_val = length(total_offset) / total_dist;
vec2 final_offset = vec2(clamp(total_offset, -clamp_val, clamp_val)) / textureSize(tex, 0);
return texture(tex, tex_coord - final_offset);
}
@ -109,4 +127,4 @@ vec4 Scaleforce(sampler2D tex, vec2 tex_coord) {
void main() {
frag_color = Scaleforce(input_texture, tex_coord);
}
}

4
src/video_core/host_shaders/vulkan_fidelityfx_fsr_easu.comp → src/video_core/host_shaders/vulkan_fidelityfx_fsr_easu_fp16.comp
View file

@ -5,9 +5,7 @@
#version 460 core
#extension GL_GOOGLE_include_directive : enable
layout(set=0,binding=0) uniform sampler2D InputTexture;
layout(set=0,binding=1,rgba16f) uniform image2D OutputTexture;
#define YUZU_USE_FP16
#define USE_EASU 1
#include "fidelityfx_fsr.comp"

10
src/video_core/host_shaders/vulkan_fidelityfx_fsr_easu_fp32.comp Normal file
View file

@ -0,0 +1,10 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#version 460 core
#extension GL_GOOGLE_include_directive : enable
#define USE_EASU 1
#include "fidelityfx_fsr.comp"

4
src/video_core/host_shaders/vulkan_fidelityfx_fsr_rcas.comp → src/video_core/host_shaders/vulkan_fidelityfx_fsr_rcas_fp16.comp
View file

@ -5,9 +5,7 @@
#version 460 core
#extension GL_GOOGLE_include_directive : enable
layout(set=0,binding=0) uniform sampler2D InputTexture;
layout(set=0,binding=1,rgba16f) uniform image2D OutputTexture;
#define YUZU_USE_FP16
#define USE_RCAS 1
#include "fidelityfx_fsr.comp"

10
src/video_core/host_shaders/vulkan_fidelityfx_fsr_rcas_fp32.comp Normal file
View file

@ -0,0 +1,10 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#version 460 core
#extension GL_GOOGLE_include_directive : enable
#define USE_RCAS 1
#include "fidelityfx_fsr.comp"

7
src/video_core/host_shaders/vulkan_present_scaleforce_fp16.frag Normal file
View file

@ -0,0 +1,7 @@
#version 460
#extension GL_GOOGLE_include_directive : enable
#define YUZU_USE_FP16
#include "opengl_present_scaleforce.frag"

5
src/video_core/host_shaders/vulkan_present_scaleforce_fp32.frag Normal file
View file

@ -0,0 +1,5 @@
#version 460
#extension GL_GOOGLE_include_directive : enable
#include "opengl_present_scaleforce.frag"