1206 lines
48 KiB
C++
1206 lines
48 KiB
C++
#include "rendering/sdl3gpu/sdl3gpu_shader.hpp"
|
|
|
|
#include <SDL3/SDL_log.h>
|
|
|
|
#include <algorithm> // std::min, std::max, std::floor
|
|
#include <cmath> // std::floor, std::ceil
|
|
#include <cstring> // memcpy, strlen
|
|
|
|
#ifndef __APPLE__
|
|
#include "rendering/sdl3gpu/crtpi_frag_spv.h"
|
|
#include "rendering/sdl3gpu/downscale_frag_spv.h"
|
|
#include "rendering/sdl3gpu/postfx_frag_spv.h"
|
|
#include "rendering/sdl3gpu/postfx_vert_spv.h"
|
|
#include "rendering/sdl3gpu/upscale_frag_spv.h"
|
|
#endif
|
|
|
|
#ifdef __APPLE__
|
|
// ============================================================================
|
|
// MSL shaders (Metal Shading Language) — macOS
|
|
// ============================================================================
|
|
|
|
// NOLINTBEGIN(readability-identifier-naming)
|
|
static const char* POSTFX_VERT_MSL = R"(
|
|
#include <metal_stdlib>
|
|
using namespace metal;
|
|
|
|
struct PostVOut {
|
|
float4 pos [[position]];
|
|
float2 uv;
|
|
};
|
|
|
|
vertex PostVOut postfx_vs(uint vid [[vertex_id]]) {
|
|
const float2 positions[3] = { {-1.0, -1.0}, {3.0, -1.0}, {-1.0, 3.0} };
|
|
const float2 uvs[3] = { { 0.0, 1.0}, {2.0, 1.0}, { 0.0,-1.0} };
|
|
PostVOut out;
|
|
out.pos = float4(positions[vid], 0.0, 1.0);
|
|
out.uv = uvs[vid];
|
|
return out;
|
|
}
|
|
)";
|
|
|
|
static const char* POSTFX_FRAG_MSL = R"(
|
|
#include <metal_stdlib>
|
|
using namespace metal;
|
|
|
|
struct PostVOut {
|
|
float4 pos [[position]];
|
|
float2 uv;
|
|
};
|
|
|
|
struct PostFXUniforms {
|
|
float vignette_strength;
|
|
float chroma_strength;
|
|
float scanline_strength;
|
|
float screen_height;
|
|
float mask_strength;
|
|
float gamma_strength;
|
|
float curvature;
|
|
float bleeding;
|
|
float pixel_scale;
|
|
float time;
|
|
float oversample;
|
|
float flicker;
|
|
};
|
|
|
|
static float3 rgb_to_ycc(float3 rgb) {
|
|
return float3(
|
|
0.299f*rgb.r + 0.587f*rgb.g + 0.114f*rgb.b,
|
|
-0.169f*rgb.r - 0.331f*rgb.g + 0.500f*rgb.b + 0.5f,
|
|
0.500f*rgb.r - 0.419f*rgb.g - 0.081f*rgb.b + 0.5f
|
|
);
|
|
}
|
|
static float3 ycc_to_rgb(float3 ycc) {
|
|
float y = ycc.x;
|
|
float cb = ycc.y - 0.5f;
|
|
float cr = ycc.z - 0.5f;
|
|
return clamp(float3(
|
|
y + 1.402f*cr,
|
|
y - 0.344f*cb - 0.714f*cr,
|
|
y + 1.772f*cb
|
|
), 0.0f, 1.0f);
|
|
}
|
|
|
|
fragment float4 postfx_fs(PostVOut in [[stage_in]],
|
|
texture2d<float> scene [[texture(0)]],
|
|
sampler samp [[sampler(0)]],
|
|
constant PostFXUniforms& u [[buffer(0)]]) {
|
|
float2 uv = in.uv;
|
|
|
|
if (u.curvature > 0.0f) {
|
|
float2 c = uv - 0.5f;
|
|
float rsq = dot(c, c);
|
|
float2 dist = float2(0.05f, 0.1f) * u.curvature;
|
|
float2 barrelScale = 1.0f - 0.23f * dist;
|
|
c += c * (dist * rsq);
|
|
c *= barrelScale;
|
|
if (abs(c.x) >= 0.5f || abs(c.y) >= 0.5f) {
|
|
return float4(0.0f, 0.0f, 0.0f, 1.0f);
|
|
}
|
|
uv = c + 0.5f;
|
|
}
|
|
|
|
float3 base = scene.sample(samp, uv).rgb;
|
|
|
|
float3 colour;
|
|
if (u.bleeding > 0.0f) {
|
|
float tw = float(scene.get_width());
|
|
float step = u.oversample / tw;
|
|
float3 ycc = rgb_to_ycc(base);
|
|
float3 ycc_l2 = rgb_to_ycc(scene.sample(samp, uv - float2(2.0f*step, 0.0f)).rgb);
|
|
float3 ycc_l1 = rgb_to_ycc(scene.sample(samp, uv - float2(1.0f*step, 0.0f)).rgb);
|
|
float3 ycc_r1 = rgb_to_ycc(scene.sample(samp, uv + float2(1.0f*step, 0.0f)).rgb);
|
|
float3 ycc_r2 = rgb_to_ycc(scene.sample(samp, uv + float2(2.0f*step, 0.0f)).rgb);
|
|
ycc.yz = (ycc_l2.yz + ycc_l1.yz*2.0f + ycc.yz*2.0f + ycc_r1.yz*2.0f + ycc_r2.yz) / 8.0f;
|
|
colour = mix(base, ycc_to_rgb(ycc), u.bleeding);
|
|
} else {
|
|
colour = base;
|
|
}
|
|
|
|
float ca = u.chroma_strength * 0.005f * (1.0f + 0.15f * sin(u.time * 7.3f));
|
|
colour.r = scene.sample(samp, uv + float2(ca, 0.0f)).r;
|
|
colour.b = scene.sample(samp, uv - float2(ca, 0.0f)).b;
|
|
|
|
if (u.gamma_strength > 0.0f) {
|
|
float3 lin = pow(colour, float3(2.4f));
|
|
colour = mix(colour, lin, u.gamma_strength);
|
|
}
|
|
|
|
const float SCAN_DARK_RATIO = 0.333f;
|
|
const float SCAN_DARK_FLOOR = 0.42f;
|
|
if (u.scanline_strength > 0.0f) {
|
|
float ps = max(1.0f, round(u.pixel_scale));
|
|
float frac_in_row = fract(uv.y * u.screen_height);
|
|
float row_pos = floor(frac_in_row * ps);
|
|
float bright_rows = (ps < 2.0f) ? ps : ((ps < 3.0f) ? 1.0f : floor(ps * (1.0f - SCAN_DARK_RATIO)));
|
|
float is_dark = step(bright_rows, row_pos);
|
|
float scan = mix(1.0f, SCAN_DARK_FLOOR, is_dark);
|
|
colour *= mix(1.0f, scan, u.scanline_strength);
|
|
}
|
|
|
|
if (u.gamma_strength > 0.0f) {
|
|
float3 enc = pow(colour, float3(1.0f/2.2f));
|
|
colour = mix(colour, enc, u.gamma_strength);
|
|
}
|
|
|
|
float2 d = uv - 0.5f;
|
|
float vignette = 1.0f - dot(d, d) * u.vignette_strength;
|
|
colour *= clamp(vignette, 0.0f, 1.0f);
|
|
|
|
if (u.mask_strength > 0.0f) {
|
|
float whichMask = fract(in.pos.x * 0.3333333f);
|
|
float3 mask = float3(0.80f);
|
|
if (whichMask < 0.3333333f) mask.x = 1.0f;
|
|
else if (whichMask < 0.6666667f) mask.y = 1.0f;
|
|
else mask.z = 1.0f;
|
|
colour = mix(colour, colour * mask, u.mask_strength);
|
|
}
|
|
|
|
if (u.flicker > 0.0f) {
|
|
float flicker_wave = sin(u.time * 100.0f) * 0.5f + 0.5f;
|
|
colour *= 1.0f - u.flicker * 0.04f * flicker_wave;
|
|
}
|
|
|
|
return float4(colour, 1.0f);
|
|
}
|
|
)";
|
|
static const char* UPSCALE_FRAG_MSL = R"(
|
|
#include <metal_stdlib>
|
|
using namespace metal;
|
|
struct VertOut { float4 pos [[position]]; float2 uv; };
|
|
fragment float4 upscale_fs(VertOut in [[stage_in]],
|
|
texture2d<float> scene [[texture(0)]],
|
|
sampler smp [[sampler(0)]])
|
|
{
|
|
return scene.sample(smp, in.uv);
|
|
}
|
|
)";
|
|
|
|
static const char* DOWNSCALE_FRAG_MSL = R"(
|
|
#include <metal_stdlib>
|
|
using namespace metal;
|
|
struct VertOut { float4 pos [[position]]; float2 uv; };
|
|
struct DownscaleUniforms { int algorithm; float pad0; float pad1; float pad2; };
|
|
|
|
static float lanczos_w(float t, float a) {
|
|
t = abs(t);
|
|
if (t < 0.0001f) { return 1.0f; }
|
|
if (t >= a) { return 0.0f; }
|
|
const float PI = 3.14159265358979f;
|
|
float pt = PI * t;
|
|
return (a * sin(pt) * sin(pt / a)) / (pt * pt);
|
|
}
|
|
|
|
fragment float4 downscale_fs(VertOut in [[stage_in]],
|
|
texture2d<float> source [[texture(0)]],
|
|
sampler smp [[sampler(0)]],
|
|
constant DownscaleUniforms& u [[buffer(0)]])
|
|
{
|
|
float2 src_size = float2(source.get_width(), source.get_height());
|
|
float2 p = in.uv * src_size;
|
|
float2 p_floor = floor(p);
|
|
float a = (u.algorithm == 0) ? 2.0f : 3.0f;
|
|
int win = int(a);
|
|
float4 color = float4(0.0f);
|
|
float weight_sum = 0.0f;
|
|
for (int j = -win; j <= win; j++) {
|
|
for (int i = -win; i <= win; i++) {
|
|
float2 tap_center = p_floor + float2(float(i), float(j)) + 0.5f;
|
|
float2 offset = tap_center - p;
|
|
float w = lanczos_w(offset.x, a) * lanczos_w(offset.y, a);
|
|
color += source.sample(smp, tap_center / src_size) * w;
|
|
weight_sum += w;
|
|
}
|
|
}
|
|
return (weight_sum > 0.0f) ? (color / weight_sum) : float4(0.0f, 0.0f, 0.0f, 1.0f);
|
|
}
|
|
)";
|
|
static const char* CRTPI_FRAG_MSL = R"(
|
|
#include <metal_stdlib>
|
|
using namespace metal;
|
|
|
|
struct PostVOut {
|
|
float4 pos [[position]];
|
|
float2 uv;
|
|
};
|
|
|
|
struct CrtPiUniforms {
|
|
float scanline_weight;
|
|
float scanline_gap_brightness;
|
|
float bloom_factor;
|
|
float input_gamma;
|
|
float output_gamma;
|
|
float mask_brightness;
|
|
float curvature_x;
|
|
float curvature_y;
|
|
int mask_type;
|
|
int enable_scanlines;
|
|
int enable_multisample;
|
|
int enable_gamma;
|
|
int enable_curvature;
|
|
int enable_sharper;
|
|
float texture_width;
|
|
float texture_height;
|
|
};
|
|
|
|
static float2 crtpi_distort(float2 coord, float2 screen_scale, float cx, float cy) {
|
|
float2 curvature = float2(cx, cy);
|
|
float2 barrel_scale = 1.0f - (0.23f * curvature);
|
|
coord *= screen_scale;
|
|
coord -= 0.5f;
|
|
float rsq = coord.x * coord.x + coord.y * coord.y;
|
|
coord += coord * (curvature * rsq);
|
|
coord *= barrel_scale;
|
|
if (abs(coord.x) >= 0.5f || abs(coord.y) >= 0.5f) { return float2(-1.0f); }
|
|
coord += 0.5f;
|
|
coord /= screen_scale;
|
|
return coord;
|
|
}
|
|
|
|
static float crtpi_scan_weight(float dist, float sw, float gap) {
|
|
return max(1.0f - dist * dist * sw, gap);
|
|
}
|
|
|
|
static float crtpi_scan_line(float dy, float filter_w, float sw, float gap, bool ms) {
|
|
float w = crtpi_scan_weight(dy, sw, gap);
|
|
if (ms) {
|
|
w += crtpi_scan_weight(dy - filter_w, sw, gap);
|
|
w += crtpi_scan_weight(dy + filter_w, sw, gap);
|
|
w *= 0.3333333f;
|
|
}
|
|
return w;
|
|
}
|
|
|
|
fragment float4 crtpi_fs(PostVOut in [[stage_in]],
|
|
texture2d<float> tex [[texture(0)]],
|
|
sampler samp [[sampler(0)]],
|
|
constant CrtPiUniforms& u [[buffer(0)]]) {
|
|
float2 tex_size = float2(u.texture_width, u.texture_height);
|
|
float filter_width = (768.0f / u.texture_height) / 3.0f;
|
|
float2 texcoord = in.uv;
|
|
|
|
if (u.enable_curvature != 0) {
|
|
texcoord = crtpi_distort(texcoord, float2(1.0f, 1.0f), u.curvature_x, u.curvature_y);
|
|
if (texcoord.x < 0.0f) { return float4(0.0f, 0.0f, 0.0f, 1.0f); }
|
|
}
|
|
|
|
float2 coord_in_pixels = texcoord * tex_size;
|
|
float2 tc;
|
|
float scan_weight;
|
|
|
|
if (u.enable_sharper != 0) {
|
|
float2 temp = floor(coord_in_pixels) + 0.5f;
|
|
tc = temp / tex_size;
|
|
float2 deltas = coord_in_pixels - temp;
|
|
scan_weight = crtpi_scan_line(deltas.y, filter_width, u.scanline_weight, u.scanline_gap_brightness, u.enable_multisample != 0);
|
|
float2 signs = sign(deltas);
|
|
deltas.x *= 2.0f;
|
|
deltas = deltas * deltas;
|
|
deltas.y = deltas.y * deltas.y;
|
|
deltas.x *= 0.5f;
|
|
deltas.y *= 8.0f;
|
|
deltas /= tex_size;
|
|
deltas *= signs;
|
|
tc = tc + deltas;
|
|
} else {
|
|
float temp_y = floor(coord_in_pixels.y) + 0.5f;
|
|
float y_coord = temp_y / tex_size.y;
|
|
float dy = coord_in_pixels.y - temp_y;
|
|
scan_weight = crtpi_scan_line(dy, filter_width, u.scanline_weight, u.scanline_gap_brightness, u.enable_multisample != 0);
|
|
float sign_y = sign(dy);
|
|
dy = dy * dy;
|
|
dy = dy * dy;
|
|
dy *= 8.0f;
|
|
dy /= tex_size.y;
|
|
dy *= sign_y;
|
|
tc = float2(texcoord.x, y_coord + dy);
|
|
}
|
|
|
|
float3 colour = tex.sample(samp, tc).rgb;
|
|
|
|
if (u.enable_scanlines != 0) {
|
|
if (u.enable_gamma != 0) { colour = pow(colour, float3(u.input_gamma)); }
|
|
colour *= scan_weight * u.bloom_factor;
|
|
if (u.enable_gamma != 0) { colour = pow(colour, float3(1.0f / u.output_gamma)); }
|
|
}
|
|
|
|
if (u.mask_type == 1) {
|
|
float wm = fract(in.pos.x * 0.5f);
|
|
float3 mask = (wm < 0.5f) ? float3(u.mask_brightness, 1.0f, u.mask_brightness)
|
|
: float3(1.0f, u.mask_brightness, 1.0f);
|
|
colour *= mask;
|
|
} else if (u.mask_type == 2) {
|
|
float wm = fract(in.pos.x * 0.3333333f);
|
|
float3 mask = float3(u.mask_brightness);
|
|
if (wm < 0.3333333f) mask.x = 1.0f;
|
|
else if (wm < 0.6666666f) mask.y = 1.0f;
|
|
else mask.z = 1.0f;
|
|
colour *= mask;
|
|
}
|
|
|
|
return float4(colour, 1.0f);
|
|
}
|
|
)";
|
|
// NOLINTEND(readability-identifier-naming)
|
|
|
|
#endif // __APPLE__
|
|
|
|
namespace Rendering {
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Destructor
|
|
// ---------------------------------------------------------------------------
|
|
SDL3GPUShader::~SDL3GPUShader() {
|
|
destroy();
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// init
|
|
// ---------------------------------------------------------------------------
|
|
auto SDL3GPUShader::init(SDL_Window* window,
|
|
SDL_Texture* texture,
|
|
const std::string& /*vertex_source*/,
|
|
const std::string& /*fragment_source*/) -> bool {
|
|
if (is_initialized_) {
|
|
cleanup();
|
|
}
|
|
|
|
window_ = window;
|
|
|
|
float fw = 0.0F;
|
|
float fh = 0.0F;
|
|
SDL_GetTextureSize(texture, &fw, &fh);
|
|
game_width_ = static_cast<int>(fw);
|
|
game_height_ = static_cast<int>(fh);
|
|
uniforms_.screen_height = static_cast<float>(game_height_);
|
|
uniforms_.oversample = static_cast<float>(oversample_);
|
|
|
|
// 1. GPU disabled by config
|
|
if (preferred_driver_ == "none") {
|
|
SDL_Log("SDL3GPUShader: GPU disabled by config, using SDL_Renderer fallback");
|
|
driver_name_ = "";
|
|
return false;
|
|
}
|
|
|
|
// 2. Create GPU device (solo si no existe ya)
|
|
if (device_ == nullptr) {
|
|
#ifdef __APPLE__
|
|
const SDL_GPUShaderFormat PREFERRED = SDL_GPU_SHADERFORMAT_MSL | SDL_GPU_SHADERFORMAT_METALLIB;
|
|
#else
|
|
const SDL_GPUShaderFormat PREFERRED = SDL_GPU_SHADERFORMAT_SPIRV;
|
|
#endif
|
|
const char* preferred = preferred_driver_.empty() ? nullptr : preferred_driver_.c_str();
|
|
device_ = SDL_CreateGPUDevice(PREFERRED, false, preferred);
|
|
if (device_ == nullptr && preferred != nullptr) {
|
|
SDL_Log("SDL3GPUShader: driver '%s' not available, falling back to auto", preferred);
|
|
device_ = SDL_CreateGPUDevice(PREFERRED, false, nullptr);
|
|
}
|
|
if (device_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: SDL_CreateGPUDevice failed: %s", SDL_GetError());
|
|
return false;
|
|
}
|
|
driver_name_ = SDL_GetGPUDeviceDriver(device_);
|
|
SDL_Log("SDL3GPUShader: driver = %s", driver_name_.c_str());
|
|
|
|
if (!SDL_ClaimWindowForGPUDevice(device_, window_)) {
|
|
SDL_Log("SDL3GPUShader: SDL_ClaimWindowForGPUDevice failed: %s", SDL_GetError());
|
|
SDL_DestroyGPUDevice(device_);
|
|
device_ = nullptr;
|
|
return false;
|
|
}
|
|
SDL_SetGPUSwapchainParameters(device_, window_, SDL_GPU_SWAPCHAINCOMPOSITION_SDR, bestPresentMode(vsync_));
|
|
}
|
|
|
|
// 3. Create scene texture (always game resolution)
|
|
SDL_GPUTextureCreateInfo tex_info = {};
|
|
tex_info.type = SDL_GPU_TEXTURETYPE_2D;
|
|
tex_info.format = SDL_GPU_TEXTUREFORMAT_B8G8R8A8_UNORM;
|
|
tex_info.usage = SDL_GPU_TEXTUREUSAGE_SAMPLER;
|
|
tex_info.width = static_cast<Uint32>(game_width_);
|
|
tex_info.height = static_cast<Uint32>(game_height_);
|
|
tex_info.layer_count_or_depth = 1;
|
|
tex_info.num_levels = 1;
|
|
scene_texture_ = SDL_CreateGPUTexture(device_, &tex_info);
|
|
if (scene_texture_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: failed to create scene texture: %s", SDL_GetError());
|
|
cleanup();
|
|
return false;
|
|
}
|
|
|
|
ss_factor_ = 0;
|
|
|
|
// 4. Create upload transfer buffer (always game resolution)
|
|
SDL_GPUTransferBufferCreateInfo tb_info = {};
|
|
tb_info.usage = SDL_GPU_TRANSFERBUFFERUSAGE_UPLOAD;
|
|
tb_info.size = static_cast<Uint32>(game_width_ * game_height_ * 4);
|
|
upload_buffer_ = SDL_CreateGPUTransferBuffer(device_, &tb_info);
|
|
if (upload_buffer_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: failed to create upload buffer: %s", SDL_GetError());
|
|
cleanup();
|
|
return false;
|
|
}
|
|
|
|
// 5. Create samplers
|
|
SDL_GPUSamplerCreateInfo samp_info = {};
|
|
samp_info.min_filter = SDL_GPU_FILTER_NEAREST;
|
|
samp_info.mag_filter = SDL_GPU_FILTER_NEAREST;
|
|
samp_info.mipmap_mode = SDL_GPU_SAMPLERMIPMAPMODE_NEAREST;
|
|
samp_info.address_mode_u = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
|
|
samp_info.address_mode_v = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
|
|
samp_info.address_mode_w = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
|
|
sampler_ = SDL_CreateGPUSampler(device_, &samp_info);
|
|
if (sampler_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: failed to create sampler: %s", SDL_GetError());
|
|
cleanup();
|
|
return false;
|
|
}
|
|
|
|
SDL_GPUSamplerCreateInfo lsamp_info = {};
|
|
lsamp_info.min_filter = SDL_GPU_FILTER_LINEAR;
|
|
lsamp_info.mag_filter = SDL_GPU_FILTER_LINEAR;
|
|
lsamp_info.mipmap_mode = SDL_GPU_SAMPLERMIPMAPMODE_NEAREST;
|
|
lsamp_info.address_mode_u = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
|
|
lsamp_info.address_mode_v = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
|
|
lsamp_info.address_mode_w = SDL_GPU_SAMPLERADDRESSMODE_CLAMP_TO_EDGE;
|
|
linear_sampler_ = SDL_CreateGPUSampler(device_, &lsamp_info);
|
|
if (linear_sampler_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: failed to create linear sampler: %s", SDL_GetError());
|
|
cleanup();
|
|
return false;
|
|
}
|
|
|
|
// 6. Create pipelines
|
|
if (!createPipeline()) {
|
|
cleanup();
|
|
return false;
|
|
}
|
|
if (!createCrtPiPipeline()) {
|
|
cleanup();
|
|
return false;
|
|
}
|
|
|
|
is_initialized_ = true;
|
|
SDL_Log("SDL3GPUShader: initialized OK (%dx%d)", game_width_, game_height_);
|
|
return true;
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// createPipeline — PostFX + Upscale + PostFX offscreen + Downscale pipelines
|
|
// ---------------------------------------------------------------------------
|
|
auto SDL3GPUShader::createPipeline() -> bool { // NOLINT(readability-function-cognitive-complexity)
|
|
const SDL_GPUTextureFormat SWAPCHAIN_FMT = SDL_GetGPUSwapchainTextureFormat(device_, window_);
|
|
|
|
// ---- PostFX pipeline (→ swapchain) ----
|
|
#ifdef __APPLE__
|
|
SDL_GPUShader* vert = createShaderMSL(device_, POSTFX_VERT_MSL, "postfx_vs", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* frag = createShaderMSL(device_, POSTFX_FRAG_MSL, "postfx_fs", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#else
|
|
SDL_GPUShader* vert = createShaderSPIRV(device_, kpostfx_vert_spv, kpostfx_vert_spv_size, "main", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* frag = createShaderSPIRV(device_, kpostfx_frag_spv, kpostfx_frag_spv_size, "main", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#endif
|
|
|
|
if ((vert == nullptr) || (frag == nullptr)) {
|
|
SDL_Log("SDL3GPUShader: failed to compile PostFX shaders");
|
|
if (vert != nullptr) { SDL_ReleaseGPUShader(device_, vert); }
|
|
if (frag != nullptr) { SDL_ReleaseGPUShader(device_, frag); }
|
|
return false;
|
|
}
|
|
|
|
SDL_GPUColorTargetBlendState no_blend = {};
|
|
no_blend.enable_blend = false;
|
|
no_blend.enable_color_write_mask = false;
|
|
|
|
SDL_GPUColorTargetDescription color_target = {};
|
|
color_target.format = SWAPCHAIN_FMT;
|
|
color_target.blend_state = no_blend;
|
|
|
|
SDL_GPUVertexInputState no_input = {};
|
|
|
|
SDL_GPUGraphicsPipelineCreateInfo pipe_info = {};
|
|
pipe_info.vertex_shader = vert;
|
|
pipe_info.fragment_shader = frag;
|
|
pipe_info.vertex_input_state = no_input;
|
|
pipe_info.primitive_type = SDL_GPU_PRIMITIVETYPE_TRIANGLELIST;
|
|
pipe_info.target_info.num_color_targets = 1;
|
|
pipe_info.target_info.color_target_descriptions = &color_target;
|
|
|
|
pipeline_ = SDL_CreateGPUGraphicsPipeline(device_, &pipe_info);
|
|
SDL_ReleaseGPUShader(device_, vert);
|
|
SDL_ReleaseGPUShader(device_, frag);
|
|
|
|
if (pipeline_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: PostFX pipeline creation failed: %s", SDL_GetError());
|
|
return false;
|
|
}
|
|
|
|
// ---- Upscale pipeline (scene → scaled_texture_) ----
|
|
#ifdef __APPLE__
|
|
SDL_GPUShader* uvert = createShaderMSL(device_, POSTFX_VERT_MSL, "postfx_vs", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* ufrag = createShaderMSL(device_, UPSCALE_FRAG_MSL, "upscale_fs", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 0);
|
|
#else
|
|
SDL_GPUShader* uvert = createShaderSPIRV(device_, kpostfx_vert_spv, kpostfx_vert_spv_size, "main", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* ufrag = createShaderSPIRV(device_, kupscale_frag_spv, kupscale_frag_spv_size, "main", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 0);
|
|
#endif
|
|
|
|
if ((uvert == nullptr) || (ufrag == nullptr)) {
|
|
if (uvert != nullptr) { SDL_ReleaseGPUShader(device_, uvert); }
|
|
if (ufrag != nullptr) { SDL_ReleaseGPUShader(device_, ufrag); }
|
|
return false;
|
|
}
|
|
|
|
SDL_GPUColorTargetDescription upscale_ct = {};
|
|
upscale_ct.format = SDL_GPU_TEXTUREFORMAT_B8G8R8A8_UNORM;
|
|
upscale_ct.blend_state = no_blend;
|
|
|
|
SDL_GPUGraphicsPipelineCreateInfo upscale_pi = {};
|
|
upscale_pi.vertex_shader = uvert;
|
|
upscale_pi.fragment_shader = ufrag;
|
|
upscale_pi.vertex_input_state = no_input;
|
|
upscale_pi.primitive_type = SDL_GPU_PRIMITIVETYPE_TRIANGLELIST;
|
|
upscale_pi.target_info.num_color_targets = 1;
|
|
upscale_pi.target_info.color_target_descriptions = &upscale_ct;
|
|
|
|
upscale_pipeline_ = SDL_CreateGPUGraphicsPipeline(device_, &upscale_pi);
|
|
SDL_ReleaseGPUShader(device_, uvert);
|
|
SDL_ReleaseGPUShader(device_, ufrag);
|
|
|
|
if (upscale_pipeline_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: upscale pipeline creation failed: %s", SDL_GetError());
|
|
return false;
|
|
}
|
|
|
|
// ---- PostFX offscreen pipeline (→ B8G8R8A8 texture, for Lanczos path) ----
|
|
#ifdef __APPLE__
|
|
SDL_GPUShader* ofvert = createShaderMSL(device_, POSTFX_VERT_MSL, "postfx_vs", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* offrag = createShaderMSL(device_, POSTFX_FRAG_MSL, "postfx_fs", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#else
|
|
SDL_GPUShader* ofvert = createShaderSPIRV(device_, kpostfx_vert_spv, kpostfx_vert_spv_size, "main", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* offrag = createShaderSPIRV(device_, kpostfx_frag_spv, kpostfx_frag_spv_size, "main", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#endif
|
|
|
|
if ((ofvert == nullptr) || (offrag == nullptr)) {
|
|
if (ofvert != nullptr) { SDL_ReleaseGPUShader(device_, ofvert); }
|
|
if (offrag != nullptr) { SDL_ReleaseGPUShader(device_, offrag); }
|
|
return false;
|
|
}
|
|
|
|
SDL_GPUColorTargetDescription offscreen_ct = {};
|
|
offscreen_ct.format = SDL_GPU_TEXTUREFORMAT_B8G8R8A8_UNORM;
|
|
offscreen_ct.blend_state = no_blend;
|
|
|
|
SDL_GPUGraphicsPipelineCreateInfo offscreen_pi = {};
|
|
offscreen_pi.vertex_shader = ofvert;
|
|
offscreen_pi.fragment_shader = offrag;
|
|
offscreen_pi.vertex_input_state = no_input;
|
|
offscreen_pi.primitive_type = SDL_GPU_PRIMITIVETYPE_TRIANGLELIST;
|
|
offscreen_pi.target_info.num_color_targets = 1;
|
|
offscreen_pi.target_info.color_target_descriptions = &offscreen_ct;
|
|
|
|
postfx_offscreen_pipeline_ = SDL_CreateGPUGraphicsPipeline(device_, &offscreen_pi);
|
|
SDL_ReleaseGPUShader(device_, ofvert);
|
|
SDL_ReleaseGPUShader(device_, offrag);
|
|
|
|
if (postfx_offscreen_pipeline_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: PostFX offscreen pipeline creation failed: %s", SDL_GetError());
|
|
return false;
|
|
}
|
|
|
|
// ---- Downscale pipeline (Lanczos → swapchain) ----
|
|
#ifdef __APPLE__
|
|
SDL_GPUShader* dvert = createShaderMSL(device_, POSTFX_VERT_MSL, "postfx_vs", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* dfrag = createShaderMSL(device_, DOWNSCALE_FRAG_MSL, "downscale_fs", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#else
|
|
SDL_GPUShader* dvert = createShaderSPIRV(device_, kpostfx_vert_spv, kpostfx_vert_spv_size, "main", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* dfrag = createShaderSPIRV(device_, kdownscale_frag_spv, kdownscale_frag_spv_size, "main", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#endif
|
|
|
|
if ((dvert == nullptr) || (dfrag == nullptr)) {
|
|
if (dvert != nullptr) { SDL_ReleaseGPUShader(device_, dvert); }
|
|
if (dfrag != nullptr) { SDL_ReleaseGPUShader(device_, dfrag); }
|
|
return false;
|
|
}
|
|
|
|
SDL_GPUColorTargetDescription downscale_ct = {};
|
|
downscale_ct.format = SWAPCHAIN_FMT;
|
|
downscale_ct.blend_state = no_blend;
|
|
|
|
SDL_GPUGraphicsPipelineCreateInfo downscale_pi = {};
|
|
downscale_pi.vertex_shader = dvert;
|
|
downscale_pi.fragment_shader = dfrag;
|
|
downscale_pi.vertex_input_state = no_input;
|
|
downscale_pi.primitive_type = SDL_GPU_PRIMITIVETYPE_TRIANGLELIST;
|
|
downscale_pi.target_info.num_color_targets = 1;
|
|
downscale_pi.target_info.color_target_descriptions = &downscale_ct;
|
|
|
|
downscale_pipeline_ = SDL_CreateGPUGraphicsPipeline(device_, &downscale_pi);
|
|
SDL_ReleaseGPUShader(device_, dvert);
|
|
SDL_ReleaseGPUShader(device_, dfrag);
|
|
|
|
if (downscale_pipeline_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: downscale pipeline creation failed: %s", SDL_GetError());
|
|
return false;
|
|
}
|
|
|
|
return true;
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// createCrtPiPipeline
|
|
// ---------------------------------------------------------------------------
|
|
auto SDL3GPUShader::createCrtPiPipeline() -> bool {
|
|
const SDL_GPUTextureFormat SWAPCHAIN_FMT = SDL_GetGPUSwapchainTextureFormat(device_, window_);
|
|
|
|
#ifdef __APPLE__
|
|
SDL_GPUShader* vert = createShaderMSL(device_, POSTFX_VERT_MSL, "postfx_vs", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* frag = createShaderMSL(device_, CRTPI_FRAG_MSL, "crtpi_fs", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#else
|
|
SDL_GPUShader* vert = createShaderSPIRV(device_, kpostfx_vert_spv, kpostfx_vert_spv_size, "main", SDL_GPU_SHADERSTAGE_VERTEX, 0, 0);
|
|
SDL_GPUShader* frag = createShaderSPIRV(device_, kcrtpi_frag_spv, kcrtpi_frag_spv_size, "main", SDL_GPU_SHADERSTAGE_FRAGMENT, 1, 1);
|
|
#endif
|
|
|
|
if ((vert == nullptr) || (frag == nullptr)) {
|
|
if (vert != nullptr) { SDL_ReleaseGPUShader(device_, vert); }
|
|
if (frag != nullptr) { SDL_ReleaseGPUShader(device_, frag); }
|
|
return false;
|
|
}
|
|
|
|
SDL_GPUColorTargetBlendState no_blend = {};
|
|
SDL_GPUColorTargetDescription ct = {};
|
|
ct.format = SWAPCHAIN_FMT;
|
|
ct.blend_state = no_blend;
|
|
|
|
SDL_GPUVertexInputState no_input = {};
|
|
SDL_GPUGraphicsPipelineCreateInfo pi = {};
|
|
pi.vertex_shader = vert;
|
|
pi.fragment_shader = frag;
|
|
pi.vertex_input_state = no_input;
|
|
pi.primitive_type = SDL_GPU_PRIMITIVETYPE_TRIANGLELIST;
|
|
pi.target_info.num_color_targets = 1;
|
|
pi.target_info.color_target_descriptions = &ct;
|
|
|
|
crtpi_pipeline_ = SDL_CreateGPUGraphicsPipeline(device_, &pi);
|
|
SDL_ReleaseGPUShader(device_, vert);
|
|
SDL_ReleaseGPUShader(device_, frag);
|
|
|
|
if (crtpi_pipeline_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: CrtPi pipeline creation failed: %s", SDL_GetError());
|
|
return false;
|
|
}
|
|
return true;
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// uploadPixels — direct memcpy, GPU handles upscale
|
|
// ---------------------------------------------------------------------------
|
|
void SDL3GPUShader::uploadPixels(const Uint32* pixels, int width, int height) {
|
|
if (!is_initialized_ || (upload_buffer_ == nullptr)) { return; }
|
|
|
|
void* mapped = SDL_MapGPUTransferBuffer(device_, upload_buffer_, false);
|
|
if (mapped == nullptr) {
|
|
SDL_Log("SDL3GPUShader: SDL_MapGPUTransferBuffer failed: %s", SDL_GetError());
|
|
return;
|
|
}
|
|
|
|
std::memcpy(mapped, pixels, static_cast<size_t>(width * height * 4));
|
|
SDL_UnmapGPUTransferBuffer(device_, upload_buffer_);
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// render — 3-path: CrtPi direct, PostFX+Lanczos, PostFX direct
|
|
// ---------------------------------------------------------------------------
|
|
void SDL3GPUShader::render() { // NOLINT(readability-function-cognitive-complexity)
|
|
if (!is_initialized_) { return; }
|
|
|
|
// SS factor calculation
|
|
if (oversample_ > 1 && game_height_ > 0) {
|
|
int win_w = 0;
|
|
int win_h = 0;
|
|
SDL_GetWindowSizeInPixels(window_, &win_w, &win_h);
|
|
const float ZOOM = static_cast<float>(win_h) / static_cast<float>(game_height_);
|
|
const int NEED_FACTOR = calcSsFactor(ZOOM);
|
|
if (NEED_FACTOR != ss_factor_) {
|
|
SDL_WaitForGPUIdle(device_);
|
|
recreateScaledTexture(NEED_FACTOR);
|
|
}
|
|
}
|
|
|
|
SDL_GPUCommandBuffer* cmd = SDL_AcquireGPUCommandBuffer(device_);
|
|
if (cmd == nullptr) {
|
|
SDL_Log("SDL3GPUShader: SDL_AcquireGPUCommandBuffer failed: %s", SDL_GetError());
|
|
return;
|
|
}
|
|
|
|
// ---- Copy pass: transfer buffer → scene texture ----
|
|
SDL_GPUCopyPass* copy = SDL_BeginGPUCopyPass(cmd);
|
|
if (copy != nullptr) {
|
|
SDL_GPUTextureTransferInfo src = {};
|
|
src.transfer_buffer = upload_buffer_;
|
|
src.offset = 0;
|
|
src.pixels_per_row = static_cast<Uint32>(game_width_);
|
|
src.rows_per_layer = static_cast<Uint32>(game_height_);
|
|
|
|
SDL_GPUTextureRegion dst = {};
|
|
dst.texture = scene_texture_;
|
|
dst.w = static_cast<Uint32>(game_width_);
|
|
dst.h = static_cast<Uint32>(game_height_);
|
|
dst.d = 1;
|
|
|
|
SDL_UploadToGPUTexture(copy, &src, &dst, false);
|
|
SDL_EndGPUCopyPass(copy);
|
|
}
|
|
|
|
// ---- Upscale pass: scene → scaled_texture_ ----
|
|
if (oversample_ > 1 && scaled_texture_ != nullptr && upscale_pipeline_ != nullptr) {
|
|
SDL_GPUColorTargetInfo upscale_target = {};
|
|
upscale_target.texture = scaled_texture_;
|
|
upscale_target.load_op = SDL_GPU_LOADOP_DONT_CARE;
|
|
upscale_target.store_op = SDL_GPU_STOREOP_STORE;
|
|
|
|
SDL_GPURenderPass* upass = SDL_BeginGPURenderPass(cmd, &upscale_target, 1, nullptr);
|
|
if (upass != nullptr) {
|
|
SDL_BindGPUGraphicsPipeline(upass, upscale_pipeline_);
|
|
SDL_GPUTextureSamplerBinding ubinding = {};
|
|
ubinding.texture = scene_texture_;
|
|
ubinding.sampler = (linear_upscale_ && linear_sampler_ != nullptr) ? linear_sampler_ : sampler_;
|
|
SDL_BindGPUFragmentSamplers(upass, 0, &ubinding, 1);
|
|
SDL_DrawGPUPrimitives(upass, 3, 1, 0, 0);
|
|
SDL_EndGPURenderPass(upass);
|
|
}
|
|
}
|
|
|
|
// ---- Acquire swapchain texture ----
|
|
SDL_GPUTexture* swapchain = nullptr;
|
|
Uint32 sw = 0;
|
|
Uint32 sh = 0;
|
|
if (!SDL_AcquireGPUSwapchainTexture(cmd, window_, &swapchain, &sw, &sh)) {
|
|
SDL_Log("SDL3GPUShader: SDL_AcquireGPUSwapchainTexture failed: %s", SDL_GetError());
|
|
SDL_SubmitGPUCommandBuffer(cmd);
|
|
return;
|
|
}
|
|
if (swapchain == nullptr) {
|
|
SDL_SubmitGPUCommandBuffer(cmd);
|
|
return;
|
|
}
|
|
|
|
// ---- Viewport calculation ----
|
|
float vx = 0.0F;
|
|
float vy = 0.0F;
|
|
float vw = 0.0F;
|
|
float vh = 0.0F;
|
|
if (integer_scale_) {
|
|
const int SCALE = std::max(1, std::min(static_cast<int>(sw) / game_width_, static_cast<int>(sh) / game_height_));
|
|
vw = static_cast<float>(game_width_ * SCALE);
|
|
vh = static_cast<float>(game_height_ * SCALE);
|
|
} else {
|
|
const float SCALE = std::min(
|
|
static_cast<float>(sw) / static_cast<float>(game_width_),
|
|
static_cast<float>(sh) / static_cast<float>(game_height_));
|
|
vw = static_cast<float>(game_width_) * SCALE;
|
|
vh = static_cast<float>(game_height_) * SCALE;
|
|
}
|
|
vx = std::floor((static_cast<float>(sw) - vw) * 0.5F);
|
|
vy = std::floor((static_cast<float>(sh) - vh) * 0.5F);
|
|
|
|
if (oversample_ > 1 && ss_factor_ > 0) {
|
|
uniforms_.pixel_scale = static_cast<float>(ss_factor_);
|
|
} else {
|
|
uniforms_.pixel_scale = (game_height_ > 0) ? (vh / static_cast<float>(game_height_)) : 1.0F;
|
|
}
|
|
uniforms_.time = static_cast<float>(SDL_GetTicks()) / 1000.0F;
|
|
uniforms_.oversample = (oversample_ > 1 && ss_factor_ > 0)
|
|
? static_cast<float>(ss_factor_)
|
|
: 1.0F;
|
|
|
|
// ---- Path A: CrtPi direct ----
|
|
if (active_shader_ == ShaderType::CRTPI && crtpi_pipeline_ != nullptr) {
|
|
SDL_GPUColorTargetInfo ct = {};
|
|
ct.texture = swapchain;
|
|
ct.load_op = SDL_GPU_LOADOP_CLEAR;
|
|
ct.store_op = SDL_GPU_STOREOP_STORE;
|
|
ct.clear_color = {.r = 0.0F, .g = 0.0F, .b = 0.0F, .a = 1.0F};
|
|
|
|
SDL_GPURenderPass* pass = SDL_BeginGPURenderPass(cmd, &ct, 1, nullptr);
|
|
if (pass != nullptr) {
|
|
SDL_BindGPUGraphicsPipeline(pass, crtpi_pipeline_);
|
|
SDL_GPUViewport vp = {.x = vx, .y = vy, .w = vw, .h = vh, .min_depth = 0.0F, .max_depth = 1.0F};
|
|
SDL_SetGPUViewport(pass, &vp);
|
|
|
|
SDL_GPUTextureSamplerBinding binding = {};
|
|
binding.texture = scene_texture_;
|
|
binding.sampler = sampler_;
|
|
SDL_BindGPUFragmentSamplers(pass, 0, &binding, 1);
|
|
|
|
crtpi_uniforms_.texture_width = static_cast<float>(game_width_);
|
|
crtpi_uniforms_.texture_height = static_cast<float>(game_height_);
|
|
SDL_PushGPUFragmentUniformData(cmd, 0, &crtpi_uniforms_, sizeof(CrtPiUniforms));
|
|
|
|
SDL_DrawGPUPrimitives(pass, 3, 1, 0, 0);
|
|
SDL_EndGPURenderPass(pass);
|
|
}
|
|
|
|
SDL_SubmitGPUCommandBuffer(cmd);
|
|
return;
|
|
}
|
|
|
|
// ---- Path B: PostFX + Lanczos (SS active + algo selected) ----
|
|
const bool USE_LANCZOS = (oversample_ > 1 && downscale_algo_ > 0 && scaled_texture_ != nullptr && postfx_texture_ != nullptr && postfx_offscreen_pipeline_ != nullptr && downscale_pipeline_ != nullptr);
|
|
|
|
if (USE_LANCZOS) {
|
|
// Pass B1: PostFX → postfx_texture_
|
|
SDL_GPUColorTargetInfo postfx_ct = {};
|
|
postfx_ct.texture = postfx_texture_;
|
|
postfx_ct.load_op = SDL_GPU_LOADOP_CLEAR;
|
|
postfx_ct.store_op = SDL_GPU_STOREOP_STORE;
|
|
postfx_ct.clear_color = {.r = 0.0F, .g = 0.0F, .b = 0.0F, .a = 1.0F};
|
|
|
|
SDL_GPURenderPass* ppass = SDL_BeginGPURenderPass(cmd, &postfx_ct, 1, nullptr);
|
|
if (ppass != nullptr) {
|
|
SDL_BindGPUGraphicsPipeline(ppass, postfx_offscreen_pipeline_);
|
|
SDL_GPUTextureSamplerBinding pbinding = {};
|
|
pbinding.texture = scaled_texture_;
|
|
pbinding.sampler = sampler_;
|
|
SDL_BindGPUFragmentSamplers(ppass, 0, &pbinding, 1);
|
|
SDL_PushGPUFragmentUniformData(cmd, 0, &uniforms_, sizeof(PostFXUniforms));
|
|
SDL_DrawGPUPrimitives(ppass, 3, 1, 0, 0);
|
|
SDL_EndGPURenderPass(ppass);
|
|
}
|
|
|
|
// Pass B2: Lanczos downscale → swapchain
|
|
SDL_GPUColorTargetInfo ds_ct = {};
|
|
ds_ct.texture = swapchain;
|
|
ds_ct.load_op = SDL_GPU_LOADOP_CLEAR;
|
|
ds_ct.store_op = SDL_GPU_STOREOP_STORE;
|
|
ds_ct.clear_color = {.r = 0.0F, .g = 0.0F, .b = 0.0F, .a = 1.0F};
|
|
|
|
SDL_GPURenderPass* dpass = SDL_BeginGPURenderPass(cmd, &ds_ct, 1, nullptr);
|
|
if (dpass != nullptr) {
|
|
SDL_BindGPUGraphicsPipeline(dpass, downscale_pipeline_);
|
|
SDL_GPUViewport vp = {.x = vx, .y = vy, .w = vw, .h = vh, .min_depth = 0.0F, .max_depth = 1.0F};
|
|
SDL_SetGPUViewport(dpass, &vp);
|
|
SDL_GPUTextureSamplerBinding dbinding = {};
|
|
dbinding.texture = postfx_texture_;
|
|
dbinding.sampler = sampler_;
|
|
SDL_BindGPUFragmentSamplers(dpass, 0, &dbinding, 1);
|
|
DownscaleUniforms du = {.algorithm = downscale_algo_ - 1, .pad0 = 0.0F, .pad1 = 0.0F, .pad2 = 0.0F};
|
|
SDL_PushGPUFragmentUniformData(cmd, 0, &du, sizeof(DownscaleUniforms));
|
|
SDL_DrawGPUPrimitives(dpass, 3, 1, 0, 0);
|
|
SDL_EndGPURenderPass(dpass);
|
|
}
|
|
} else {
|
|
// ---- Path C: PostFX direct → swapchain ----
|
|
SDL_GPUColorTargetInfo ct = {};
|
|
ct.texture = swapchain;
|
|
ct.load_op = SDL_GPU_LOADOP_CLEAR;
|
|
ct.store_op = SDL_GPU_STOREOP_STORE;
|
|
ct.clear_color = {.r = 0.0F, .g = 0.0F, .b = 0.0F, .a = 1.0F};
|
|
|
|
SDL_GPURenderPass* pass = SDL_BeginGPURenderPass(cmd, &ct, 1, nullptr);
|
|
if (pass != nullptr) {
|
|
SDL_BindGPUGraphicsPipeline(pass, pipeline_);
|
|
SDL_GPUViewport vp = {.x = vx, .y = vy, .w = vw, .h = vh, .min_depth = 0.0F, .max_depth = 1.0F};
|
|
SDL_SetGPUViewport(pass, &vp);
|
|
|
|
SDL_GPUTexture* input_texture = (oversample_ > 1 && scaled_texture_ != nullptr) ? scaled_texture_ : scene_texture_;
|
|
SDL_GPUSampler* active_sampler = (oversample_ > 1 && linear_sampler_ != nullptr) ? linear_sampler_ : sampler_;
|
|
|
|
SDL_GPUTextureSamplerBinding binding = {};
|
|
binding.texture = input_texture;
|
|
binding.sampler = active_sampler;
|
|
SDL_BindGPUFragmentSamplers(pass, 0, &binding, 1);
|
|
|
|
SDL_PushGPUFragmentUniformData(cmd, 0, &uniforms_, sizeof(PostFXUniforms));
|
|
SDL_DrawGPUPrimitives(pass, 3, 1, 0, 0);
|
|
SDL_EndGPURenderPass(pass);
|
|
}
|
|
}
|
|
|
|
SDL_SubmitGPUCommandBuffer(cmd);
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// cleanup
|
|
// ---------------------------------------------------------------------------
|
|
void SDL3GPUShader::cleanup() {
|
|
is_initialized_ = false;
|
|
|
|
if (device_ != nullptr) {
|
|
SDL_WaitForGPUIdle(device_);
|
|
|
|
auto release_pipeline = [this](SDL_GPUGraphicsPipeline*& p) {
|
|
if (p != nullptr) {
|
|
SDL_ReleaseGPUGraphicsPipeline(device_, p);
|
|
p = nullptr;
|
|
}
|
|
};
|
|
auto release_texture = [this](SDL_GPUTexture*& t) {
|
|
if (t != nullptr) {
|
|
SDL_ReleaseGPUTexture(device_, t);
|
|
t = nullptr;
|
|
}
|
|
};
|
|
|
|
release_pipeline(pipeline_);
|
|
release_pipeline(crtpi_pipeline_);
|
|
release_pipeline(postfx_offscreen_pipeline_);
|
|
release_pipeline(upscale_pipeline_);
|
|
release_pipeline(downscale_pipeline_);
|
|
release_texture(scene_texture_);
|
|
release_texture(scaled_texture_);
|
|
release_texture(postfx_texture_);
|
|
ss_factor_ = 0;
|
|
if (upload_buffer_ != nullptr) {
|
|
SDL_ReleaseGPUTransferBuffer(device_, upload_buffer_);
|
|
upload_buffer_ = nullptr;
|
|
}
|
|
if (sampler_ != nullptr) {
|
|
SDL_ReleaseGPUSampler(device_, sampler_);
|
|
sampler_ = nullptr;
|
|
}
|
|
if (linear_sampler_ != nullptr) {
|
|
SDL_ReleaseGPUSampler(device_, linear_sampler_);
|
|
linear_sampler_ = nullptr;
|
|
}
|
|
}
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// destroy
|
|
// ---------------------------------------------------------------------------
|
|
void SDL3GPUShader::destroy() {
|
|
cleanup();
|
|
|
|
if (device_ != nullptr) {
|
|
if (window_ != nullptr) {
|
|
SDL_ReleaseWindowFromGPUDevice(device_, window_);
|
|
}
|
|
SDL_DestroyGPUDevice(device_);
|
|
device_ = nullptr;
|
|
}
|
|
window_ = nullptr;
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// Shader creation helpers
|
|
// ---------------------------------------------------------------------------
|
|
auto SDL3GPUShader::createShaderMSL(SDL_GPUDevice* device,
|
|
const char* msl_source,
|
|
const char* entrypoint,
|
|
SDL_GPUShaderStage stage,
|
|
Uint32 num_samplers,
|
|
Uint32 num_uniform_buffers) -> SDL_GPUShader* {
|
|
SDL_GPUShaderCreateInfo info = {};
|
|
info.code = reinterpret_cast<const Uint8*>(msl_source);
|
|
info.code_size = std::strlen(msl_source) + 1;
|
|
info.entrypoint = entrypoint;
|
|
info.format = SDL_GPU_SHADERFORMAT_MSL;
|
|
info.stage = stage;
|
|
info.num_samplers = num_samplers;
|
|
info.num_uniform_buffers = num_uniform_buffers;
|
|
SDL_GPUShader* shader = SDL_CreateGPUShader(device, &info);
|
|
if (shader == nullptr) {
|
|
SDL_Log("SDL3GPUShader: MSL shader '%s' failed: %s", entrypoint, SDL_GetError());
|
|
}
|
|
return shader;
|
|
}
|
|
|
|
auto SDL3GPUShader::createShaderSPIRV(SDL_GPUDevice* device, // NOLINT(readability-convert-member-functions-to-static)
|
|
const uint8_t* spv_code,
|
|
size_t spv_size,
|
|
const char* entrypoint,
|
|
SDL_GPUShaderStage stage,
|
|
Uint32 num_samplers,
|
|
Uint32 num_uniform_buffers) -> SDL_GPUShader* {
|
|
SDL_GPUShaderCreateInfo info = {};
|
|
info.code = spv_code;
|
|
info.code_size = spv_size;
|
|
info.entrypoint = entrypoint;
|
|
info.format = SDL_GPU_SHADERFORMAT_SPIRV;
|
|
info.stage = stage;
|
|
info.num_samplers = num_samplers;
|
|
info.num_uniform_buffers = num_uniform_buffers;
|
|
SDL_GPUShader* shader = SDL_CreateGPUShader(device, &info);
|
|
if (shader == nullptr) {
|
|
SDL_Log("SDL3GPUShader: SPIRV shader '%s' failed: %s", entrypoint, SDL_GetError());
|
|
}
|
|
return shader;
|
|
}
|
|
|
|
void SDL3GPUShader::setPostFXParams(const PostFXParams& p) {
|
|
uniforms_.vignette_strength = p.vignette;
|
|
uniforms_.chroma_strength = p.chroma;
|
|
uniforms_.scanline_strength = p.scanlines;
|
|
uniforms_.mask_strength = p.mask;
|
|
uniforms_.gamma_strength = p.gamma;
|
|
uniforms_.curvature = p.curvature;
|
|
uniforms_.bleeding = p.bleeding;
|
|
uniforms_.flicker = p.flicker;
|
|
}
|
|
|
|
void SDL3GPUShader::setCrtPiParams(const CrtPiParams& p) {
|
|
crtpi_uniforms_.scanline_weight = p.scanline_weight;
|
|
crtpi_uniforms_.scanline_gap_brightness = p.scanline_gap_brightness;
|
|
crtpi_uniforms_.bloom_factor = p.bloom_factor;
|
|
crtpi_uniforms_.input_gamma = p.input_gamma;
|
|
crtpi_uniforms_.output_gamma = p.output_gamma;
|
|
crtpi_uniforms_.mask_brightness = p.mask_brightness;
|
|
crtpi_uniforms_.curvature_x = p.curvature_x;
|
|
crtpi_uniforms_.curvature_y = p.curvature_y;
|
|
crtpi_uniforms_.mask_type = p.mask_type;
|
|
crtpi_uniforms_.enable_scanlines = p.enable_scanlines ? 1 : 0;
|
|
crtpi_uniforms_.enable_multisample = p.enable_multisample ? 1 : 0;
|
|
crtpi_uniforms_.enable_gamma = p.enable_gamma ? 1 : 0;
|
|
crtpi_uniforms_.enable_curvature = p.enable_curvature ? 1 : 0;
|
|
crtpi_uniforms_.enable_sharper = p.enable_sharper ? 1 : 0;
|
|
}
|
|
|
|
void SDL3GPUShader::setActiveShader(ShaderType type) {
|
|
active_shader_ = type;
|
|
}
|
|
|
|
auto SDL3GPUShader::bestPresentMode(bool vsync) const -> SDL_GPUPresentMode {
|
|
if (vsync) { return SDL_GPU_PRESENTMODE_VSYNC; }
|
|
if (SDL_WindowSupportsGPUPresentMode(device_, window_, SDL_GPU_PRESENTMODE_IMMEDIATE)) {
|
|
return SDL_GPU_PRESENTMODE_IMMEDIATE;
|
|
}
|
|
if (SDL_WindowSupportsGPUPresentMode(device_, window_, SDL_GPU_PRESENTMODE_MAILBOX)) {
|
|
return SDL_GPU_PRESENTMODE_MAILBOX;
|
|
}
|
|
return SDL_GPU_PRESENTMODE_VSYNC;
|
|
}
|
|
|
|
void SDL3GPUShader::setVSync(bool vsync) {
|
|
vsync_ = vsync;
|
|
if (device_ != nullptr && window_ != nullptr) {
|
|
SDL_SetGPUSwapchainParameters(device_, window_, SDL_GPU_SWAPCHAINCOMPOSITION_SDR, bestPresentMode(vsync_));
|
|
}
|
|
}
|
|
|
|
void SDL3GPUShader::setScaleMode(bool integer_scale) {
|
|
integer_scale_ = integer_scale;
|
|
}
|
|
|
|
void SDL3GPUShader::setOversample(int factor) {
|
|
const int NEW_FACTOR = std::max(1, factor);
|
|
if (NEW_FACTOR == oversample_) { return; }
|
|
oversample_ = NEW_FACTOR;
|
|
if (is_initialized_) {
|
|
reinitTexturesAndBuffer();
|
|
}
|
|
}
|
|
|
|
void SDL3GPUShader::setLinearUpscale(bool linear) {
|
|
linear_upscale_ = linear;
|
|
}
|
|
|
|
void SDL3GPUShader::setDownscaleAlgo(int algo) {
|
|
downscale_algo_ = std::max(0, std::min(algo, 2));
|
|
}
|
|
|
|
auto SDL3GPUShader::getSsTextureSize() const -> std::pair<int, int> {
|
|
if (ss_factor_ <= 1) { return {0, 0}; }
|
|
return {game_width_ * ss_factor_, game_height_ * ss_factor_};
|
|
}
|
|
|
|
// ---------------------------------------------------------------------------
|
|
// reinitTexturesAndBuffer
|
|
// ---------------------------------------------------------------------------
|
|
auto SDL3GPUShader::reinitTexturesAndBuffer() -> bool {
|
|
if (device_ == nullptr) { return false; }
|
|
SDL_WaitForGPUIdle(device_);
|
|
|
|
if (scene_texture_ != nullptr) {
|
|
SDL_ReleaseGPUTexture(device_, scene_texture_);
|
|
scene_texture_ = nullptr;
|
|
}
|
|
if (scaled_texture_ != nullptr) {
|
|
SDL_ReleaseGPUTexture(device_, scaled_texture_);
|
|
scaled_texture_ = nullptr;
|
|
}
|
|
ss_factor_ = 0;
|
|
if (upload_buffer_ != nullptr) {
|
|
SDL_ReleaseGPUTransferBuffer(device_, upload_buffer_);
|
|
upload_buffer_ = nullptr;
|
|
}
|
|
|
|
uniforms_.screen_height = static_cast<float>(game_height_);
|
|
uniforms_.oversample = static_cast<float>(oversample_);
|
|
|
|
SDL_GPUTextureCreateInfo tex_info = {};
|
|
tex_info.type = SDL_GPU_TEXTURETYPE_2D;
|
|
tex_info.format = SDL_GPU_TEXTUREFORMAT_B8G8R8A8_UNORM;
|
|
tex_info.usage = SDL_GPU_TEXTUREUSAGE_SAMPLER;
|
|
tex_info.width = static_cast<Uint32>(game_width_);
|
|
tex_info.height = static_cast<Uint32>(game_height_);
|
|
tex_info.layer_count_or_depth = 1;
|
|
tex_info.num_levels = 1;
|
|
scene_texture_ = SDL_CreateGPUTexture(device_, &tex_info);
|
|
if (scene_texture_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: reinit — failed to create scene texture: %s", SDL_GetError());
|
|
return false;
|
|
}
|
|
|
|
SDL_GPUTransferBufferCreateInfo tb_info = {};
|
|
tb_info.usage = SDL_GPU_TRANSFERBUFFERUSAGE_UPLOAD;
|
|
tb_info.size = static_cast<Uint32>(game_width_ * game_height_ * 4);
|
|
upload_buffer_ = SDL_CreateGPUTransferBuffer(device_, &tb_info);
|
|
if (upload_buffer_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: reinit — failed to create upload buffer: %s", SDL_GetError());
|
|
SDL_ReleaseGPUTexture(device_, scene_texture_);
|
|
scene_texture_ = nullptr;
|
|
return false;
|
|
}
|
|
|
|
SDL_Log("SDL3GPUShader: reinit — scene %dx%d, SS %s", game_width_, game_height_, oversample_ > 1 ? "on" : "off");
|
|
return true;
|
|
}
|
|
|
|
auto SDL3GPUShader::calcSsFactor(float zoom) -> int {
|
|
const int MULTIPLE = 3;
|
|
const int N = static_cast<int>(std::ceil(zoom / static_cast<float>(MULTIPLE)));
|
|
return std::max(1, N) * MULTIPLE;
|
|
}
|
|
|
|
auto SDL3GPUShader::recreateScaledTexture(int factor) -> bool {
|
|
if (scaled_texture_ != nullptr) {
|
|
SDL_ReleaseGPUTexture(device_, scaled_texture_);
|
|
scaled_texture_ = nullptr;
|
|
}
|
|
if (postfx_texture_ != nullptr) {
|
|
SDL_ReleaseGPUTexture(device_, postfx_texture_);
|
|
postfx_texture_ = nullptr;
|
|
}
|
|
ss_factor_ = 0;
|
|
|
|
const int W = game_width_ * factor;
|
|
const int H = game_height_ * factor;
|
|
|
|
SDL_GPUTextureCreateInfo info = {};
|
|
info.type = SDL_GPU_TEXTURETYPE_2D;
|
|
info.format = SDL_GPU_TEXTUREFORMAT_B8G8R8A8_UNORM;
|
|
info.usage = SDL_GPU_TEXTUREUSAGE_SAMPLER | SDL_GPU_TEXTUREUSAGE_COLOR_TARGET;
|
|
info.width = static_cast<Uint32>(W);
|
|
info.height = static_cast<Uint32>(H);
|
|
info.layer_count_or_depth = 1;
|
|
info.num_levels = 1;
|
|
|
|
scaled_texture_ = SDL_CreateGPUTexture(device_, &info);
|
|
if (scaled_texture_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: failed to create scaled texture %dx%d: %s", W, H, SDL_GetError());
|
|
return false;
|
|
}
|
|
|
|
postfx_texture_ = SDL_CreateGPUTexture(device_, &info);
|
|
if (postfx_texture_ == nullptr) {
|
|
SDL_Log("SDL3GPUShader: failed to create postfx texture %dx%d: %s", W, H, SDL_GetError());
|
|
SDL_ReleaseGPUTexture(device_, scaled_texture_);
|
|
scaled_texture_ = nullptr;
|
|
return false;
|
|
}
|
|
|
|
ss_factor_ = factor;
|
|
SDL_Log("SDL3GPUShader: scaled+postfx textures %dx%d (factor %dx)", W, H, factor);
|
|
return true;
|
|
}
|
|
|
|
} // namespace Rendering
|