renderer_software: Multi-thread processing

* Doubles the performance in most cases
This commit is contained in:
GPUCode 2023-07-16 03:02:55 +03:00
parent d702915624
commit 531d280461
2 changed files with 165 additions and 151 deletions

View file

@ -95,8 +95,14 @@ private:
} // Anonymous namespace
// Kirby Blowout Blast relies on the combiner output of a previous draw
// in order to render the sky correctly.
static thread_local Common::Vec4<u8> combiner_output{};
RasterizerSoftware::RasterizerSoftware(Memory::MemorySystem& memory_)
: memory{memory_}, state{Pica::g_state}, regs{state.regs}, fb{memory, regs.framebuffer} {}
: memory{memory_}, state{Pica::g_state}, regs{state.regs},
num_sw_threads{std::max(std::thread::hardware_concurrency(), 2U)},
sw_workers{num_sw_threads, "SwRenderer workers"}, fb{memory, regs.framebuffer} {}
void RasterizerSoftware::AddTriangle(const Pica::Shader::OutputVertex& v0,
const Pica::Shader::OutputVertex& v1,
@ -295,161 +301,171 @@ void RasterizerSoftware::ProcessTriangle(const Vertex& v0, const Vertex& v1, con
// Enter rasterization loop, starting at the center of the topleft bounding box corner.
// TODO: Not sure if looping through x first might be faster
for (u16 y = min_y + 8; y < max_y; y += 0x10) {
for (u16 x = min_x + 8; x < max_x; x += 0x10) {
// Do not process the pixel if it's inside the scissor box and the scissor mode is set
// to Exclude.
if (regs.rasterizer.scissor_test.mode == RasterizerRegs::ScissorMode::Exclude) {
if (x >= scissor_x1 && x < scissor_x2 && y >= scissor_y1 && y < scissor_y2) {
const auto process_scanline = [&, y] {
for (u16 x = min_x + 8; x < max_x; x += 0x10) {
// Do not process the pixel if it's inside the scissor box and the scissor mode is
// set to Exclude.
if (regs.rasterizer.scissor_test.mode == RasterizerRegs::ScissorMode::Exclude) {
if (x >= scissor_x1 && x < scissor_x2 && y >= scissor_y1 && y < scissor_y2) {
continue;
}
}
// Calculate the barycentric coordinates w0, w1 and w2
const s32 w0 = bias0 + SignedArea(vtxpos[1].xy(), vtxpos[2].xy(), {x, y});
const s32 w1 = bias1 + SignedArea(vtxpos[2].xy(), vtxpos[0].xy(), {x, y});
const s32 w2 = bias2 + SignedArea(vtxpos[0].xy(), vtxpos[1].xy(), {x, y});
const s32 wsum = w0 + w1 + w2;
// If current pixel is not covered by the current primitive
if (w0 < 0 || w1 < 0 || w2 < 0) {
continue;
}
}
// Calculate the barycentric coordinates w0, w1 and w2
const s32 w0 = bias0 + SignedArea(vtxpos[1].xy(), vtxpos[2].xy(), {x, y});
const s32 w1 = bias1 + SignedArea(vtxpos[2].xy(), vtxpos[0].xy(), {x, y});
const s32 w2 = bias2 + SignedArea(vtxpos[0].xy(), vtxpos[1].xy(), {x, y});
const s32 wsum = w0 + w1 + w2;
const auto baricentric_coordinates = Common::MakeVec(
f24::FromFloat32(static_cast<f32>(w0)), f24::FromFloat32(static_cast<f32>(w1)),
f24::FromFloat32(static_cast<f32>(w2)));
const f24 interpolated_w_inverse =
f24::One() / Common::Dot(w_inverse, baricentric_coordinates);
// If current pixel is not covered by the current primitive
if (w0 < 0 || w1 < 0 || w2 < 0) {
continue;
}
// interpolated_z = z / w
const float interpolated_z_over_w =
(v0.screenpos[2].ToFloat32() * w0 + v1.screenpos[2].ToFloat32() * w1 +
v2.screenpos[2].ToFloat32() * w2) /
wsum;
const auto baricentric_coordinates = Common::MakeVec(
f24::FromFloat32(static_cast<f32>(w0)), f24::FromFloat32(static_cast<f32>(w1)),
f24::FromFloat32(static_cast<f32>(w2)));
const f24 interpolated_w_inverse =
f24::One() / Common::Dot(w_inverse, baricentric_coordinates);
// Not fully accurate. About 3 bits in precision are missing.
// Z-Buffer (z / w * scale + offset)
const float depth_scale =
f24::FromRaw(regs.rasterizer.viewport_depth_range).ToFloat32();
const float depth_offset =
f24::FromRaw(regs.rasterizer.viewport_depth_near_plane).ToFloat32();
float depth = interpolated_z_over_w * depth_scale + depth_offset;
// interpolated_z = z / w
const float interpolated_z_over_w =
(v0.screenpos[2].ToFloat32() * w0 + v1.screenpos[2].ToFloat32() * w1 +
v2.screenpos[2].ToFloat32() * w2) /
wsum;
// Potentially switch to W-Buffer
if (regs.rasterizer.depthmap_enable ==
Pica::RasterizerRegs::DepthBuffering::WBuffering) {
// W-Buffer (z * scale + w * offset = (z / w * scale + offset) * w)
depth *= interpolated_w_inverse.ToFloat32() * wsum;
}
// Not fully accurate. About 3 bits in precision are missing.
// Z-Buffer (z / w * scale + offset)
const float depth_scale =
f24::FromRaw(regs.rasterizer.viewport_depth_range).ToFloat32();
const float depth_offset =
f24::FromRaw(regs.rasterizer.viewport_depth_near_plane).ToFloat32();
float depth = interpolated_z_over_w * depth_scale + depth_offset;
// Clamp the result
depth = std::clamp(depth, 0.0f, 1.0f);
// Potentially switch to W-Buffer
if (regs.rasterizer.depthmap_enable ==
Pica::RasterizerRegs::DepthBuffering::WBuffering) {
// W-Buffer (z * scale + w * offset = (z / w * scale + offset) * w)
depth *= interpolated_w_inverse.ToFloat32() * wsum;
}
// Clamp the result
depth = std::clamp(depth, 0.0f, 1.0f);
/**
* Perspective correct attribute interpolation:
* Attribute values cannot be calculated by simple linear interpolation since
* they are not linear in screen space. For example, when interpolating a
* texture coordinate across two vertices, something simple like
* u = (u0*w0 + u1*w1)/(w0+w1)
* will not work. However, the attribute value divided by the
* clipspace w-coordinate (u/w) and and the inverse w-coordinate (1/w) are linear
* in screenspace. Hence, we can linearly interpolate these two independently and
* calculate the interpolated attribute by dividing the results.
* I.e.
* u_over_w = ((u0/v0.pos.w)*w0 + (u1/v1.pos.w)*w1)/(w0+w1)
* one_over_w = (( 1/v0.pos.w)*w0 + ( 1/v1.pos.w)*w1)/(w0+w1)
* u = u_over_w / one_over_w
*
* The generalization to three vertices is straightforward in baricentric coordinates.
**/
const auto get_interpolated_attribute = [&](f24 attr0, f24 attr1, f24 attr2) {
auto attr_over_w = Common::MakeVec(attr0, attr1, attr2);
f24 interpolated_attr_over_w = Common::Dot(attr_over_w, baricentric_coordinates);
return interpolated_attr_over_w * interpolated_w_inverse;
};
const Common::Vec4<u8> primary_color{
static_cast<u8>(
round(get_interpolated_attribute(v0.color.r(), v1.color.r(), v2.color.r())
.ToFloat32() *
255)),
static_cast<u8>(
round(get_interpolated_attribute(v0.color.g(), v1.color.g(), v2.color.g())
.ToFloat32() *
255)),
static_cast<u8>(
round(get_interpolated_attribute(v0.color.b(), v1.color.b(), v2.color.b())
.ToFloat32() *
255)),
static_cast<u8>(
round(get_interpolated_attribute(v0.color.a(), v1.color.a(), v2.color.a())
.ToFloat32() *
255)),
};
std::array<Common::Vec2<f24>, 3> uv;
uv[0].u() = get_interpolated_attribute(v0.tc0.u(), v1.tc0.u(), v2.tc0.u());
uv[0].v() = get_interpolated_attribute(v0.tc0.v(), v1.tc0.v(), v2.tc0.v());
uv[1].u() = get_interpolated_attribute(v0.tc1.u(), v1.tc1.u(), v2.tc1.u());
uv[1].v() = get_interpolated_attribute(v0.tc1.v(), v1.tc1.v(), v2.tc1.v());
uv[2].u() = get_interpolated_attribute(v0.tc2.u(), v1.tc2.u(), v2.tc2.u());
uv[2].v() = get_interpolated_attribute(v0.tc2.v(), v1.tc2.v(), v2.tc2.v());
// Sample bound texture units.
const f24 tc0_w = get_interpolated_attribute(v0.tc0_w, v1.tc0_w, v2.tc0_w);
const auto texture_color = TextureColor(uv, textures, tc0_w);
Common::Vec4<u8> primary_fragment_color = {0, 0, 0, 0};
Common::Vec4<u8> secondary_fragment_color = {0, 0, 0, 0};
if (!regs.lighting.disable) {
const auto normquat =
Common::Quaternion<f32>{
{get_interpolated_attribute(v0.quat.x, v1.quat.x, v2.quat.x).ToFloat32(),
get_interpolated_attribute(v0.quat.y, v1.quat.y, v2.quat.y).ToFloat32(),
get_interpolated_attribute(v0.quat.z, v1.quat.z, v2.quat.z).ToFloat32()},
get_interpolated_attribute(v0.quat.w, v1.quat.w, v2.quat.w).ToFloat32(),
}
.Normalized();
const Common::Vec3f view{
get_interpolated_attribute(v0.view.x, v1.view.x, v2.view.x).ToFloat32(),
get_interpolated_attribute(v0.view.y, v1.view.y, v2.view.y).ToFloat32(),
get_interpolated_attribute(v0.view.z, v1.view.z, v2.view.z).ToFloat32(),
/**
* Perspective correct attribute interpolation:
* Attribute values cannot be calculated by simple linear interpolation since
* they are not linear in screen space. For example, when interpolating a
* texture coordinate across two vertices, something simple like
* u = (u0*w0 + u1*w1)/(w0+w1)
* will not work. However, the attribute value divided by the
* clipspace w-coordinate (u/w) and and the inverse w-coordinate (1/w) are linear
* in screenspace. Hence, we can linearly interpolate these two independently and
* calculate the interpolated attribute by dividing the results.
* I.e.
* u_over_w = ((u0/v0.pos.w)*w0 + (u1/v1.pos.w)*w1)/(w0+w1)
* one_over_w = (( 1/v0.pos.w)*w0 + ( 1/v1.pos.w)*w1)/(w0+w1)
* u = u_over_w / one_over_w
*
* The generalization to three vertices is straightforward in baricentric
*coordinates.
**/
const auto get_interpolated_attribute = [&](f24 attr0, f24 attr1, f24 attr2) {
auto attr_over_w = Common::MakeVec(attr0, attr1, attr2);
f24 interpolated_attr_over_w =
Common::Dot(attr_over_w, baricentric_coordinates);
return interpolated_attr_over_w * interpolated_w_inverse;
};
std::tie(primary_fragment_color, secondary_fragment_color) = ComputeFragmentsColors(
regs.lighting, state.lighting, normquat, view, texture_color);
}
// Write the TEV stages.
WriteTevConfig(texture_color, tev_stages, primary_color, primary_fragment_color,
secondary_fragment_color);
const Common::Vec4<u8> primary_color{
static_cast<u8>(
round(get_interpolated_attribute(v0.color.r(), v1.color.r(), v2.color.r())
.ToFloat32() *
255)),
static_cast<u8>(
round(get_interpolated_attribute(v0.color.g(), v1.color.g(), v2.color.g())
.ToFloat32() *
255)),
static_cast<u8>(
round(get_interpolated_attribute(v0.color.b(), v1.color.b(), v2.color.b())
.ToFloat32() *
255)),
static_cast<u8>(
round(get_interpolated_attribute(v0.color.a(), v1.color.a(), v2.color.a())
.ToFloat32() *
255)),
};
const auto& output_merger = regs.framebuffer.output_merger;
if (output_merger.fragment_operation_mode ==
FramebufferRegs::FragmentOperationMode::Shadow) {
u32 depth_int = static_cast<u32>(depth * 0xFFFFFF);
// Use green color as the shadow intensity
u8 stencil = combiner_output.y;
fb.DrawShadowMapPixel(x >> 4, y >> 4, depth_int, stencil);
// Skip the normal output merger pipeline if it is in shadow mode
continue;
}
std::array<Common::Vec2<f24>, 3> uv;
uv[0].u() = get_interpolated_attribute(v0.tc0.u(), v1.tc0.u(), v2.tc0.u());
uv[0].v() = get_interpolated_attribute(v0.tc0.v(), v1.tc0.v(), v2.tc0.v());
uv[1].u() = get_interpolated_attribute(v0.tc1.u(), v1.tc1.u(), v2.tc1.u());
uv[1].v() = get_interpolated_attribute(v0.tc1.v(), v1.tc1.v(), v2.tc1.v());
uv[2].u() = get_interpolated_attribute(v0.tc2.u(), v1.tc2.u(), v2.tc2.u());
uv[2].v() = get_interpolated_attribute(v0.tc2.v(), v1.tc2.v(), v2.tc2.v());
// Does alpha testing happen before or after stencil?
if (!DoAlphaTest(combiner_output.a())) {
continue;
// Sample bound texture units.
const f24 tc0_w = get_interpolated_attribute(v0.tc0_w, v1.tc0_w, v2.tc0_w);
const auto texture_color = TextureColor(uv, textures, tc0_w);
Common::Vec4<u8> primary_fragment_color = {0, 0, 0, 0};
Common::Vec4<u8> secondary_fragment_color = {0, 0, 0, 0};
if (!regs.lighting.disable) {
const auto normquat =
Common::Quaternion<f32>{
{get_interpolated_attribute(v0.quat.x, v1.quat.x, v2.quat.x)
.ToFloat32(),
get_interpolated_attribute(v0.quat.y, v1.quat.y, v2.quat.y)
.ToFloat32(),
get_interpolated_attribute(v0.quat.z, v1.quat.z, v2.quat.z)
.ToFloat32()},
get_interpolated_attribute(v0.quat.w, v1.quat.w, v2.quat.w).ToFloat32(),
}
.Normalized();
const Common::Vec3f view{
get_interpolated_attribute(v0.view.x, v1.view.x, v2.view.x).ToFloat32(),
get_interpolated_attribute(v0.view.y, v1.view.y, v2.view.y).ToFloat32(),
get_interpolated_attribute(v0.view.z, v1.view.z, v2.view.z).ToFloat32(),
};
std::tie(primary_fragment_color, secondary_fragment_color) =
ComputeFragmentsColors(regs.lighting, state.lighting, normquat, view,
texture_color);
}
// Write the TEV stages.
WriteTevConfig(texture_color, tev_stages, primary_color, primary_fragment_color,
secondary_fragment_color);
const auto& output_merger = regs.framebuffer.output_merger;
if (output_merger.fragment_operation_mode ==
FramebufferRegs::FragmentOperationMode::Shadow) {
u32 depth_int = static_cast<u32>(depth * 0xFFFFFF);
// Use green color as the shadow intensity
u8 stencil = combiner_output.y;
fb.DrawShadowMapPixel(x >> 4, y >> 4, depth_int, stencil);
// Skip the normal output merger pipeline if it is in shadow mode
continue;
}
// Does alpha testing happen before or after stencil?
if (!DoAlphaTest(combiner_output.a())) {
continue;
}
WriteFog(depth);
if (!DoDepthStencilTest(x, y, depth)) {
continue;
}
const auto result = PixelColor(x, y);
if (regs.framebuffer.framebuffer.allow_color_write != 0) {
fb.DrawPixel(x >> 4, y >> 4, result);
}
}
WriteFog(combiner_output, depth);
if (!DoDepthStencilTest(x, y, depth)) {
continue;
}
const auto result = PixelColor(x, y, combiner_output);
if (regs.framebuffer.framebuffer.allow_color_write != 0) {
fb.DrawPixel(x >> 4, y >> 4, result);
}
}
};
sw_workers.QueueWork(std::move(process_scanline));
}
sw_workers.WaitForRequests();
}
std::array<Common::Vec4<u8>, 4> RasterizerSoftware::TextureColor(
@ -572,8 +588,7 @@ std::array<Common::Vec4<u8>, 4> RasterizerSoftware::TextureColor(
return texture_color;
}
Common::Vec4<u8> RasterizerSoftware::PixelColor(u16 x, u16 y,
Common::Vec4<u8>& combiner_output) const {
Common::Vec4<u8> RasterizerSoftware::PixelColor(u16 x, u16 y) const {
const auto dest = fb.GetPixel(x >> 4, y >> 4);
Common::Vec4<u8> blend_output = combiner_output;
@ -768,7 +783,7 @@ void RasterizerSoftware::WriteTevConfig(
}
}
void RasterizerSoftware::WriteFog(Common::Vec4<u8>& combiner_output, float depth) const {
void RasterizerSoftware::WriteFog(float depth) const {
/**
* Apply fog combiner. Not fully accurate. We'd have to know what data type is used to
* store the depth etc. Using float for now until we know more about Pica datatypes.

View file

@ -5,7 +5,7 @@
#pragma once
#include <span>
#include "common/thread_worker.h"
#include "video_core/rasterizer_interface.h"
#include "video_core/regs_texturing.h"
#include "video_core/renderer_software/sw_clipper.h"
@ -52,7 +52,7 @@ private:
std::span<const Pica::TexturingRegs::FullTextureConfig, 3> textures, f24 tc0_w) const;
/// Returns the final pixel color with blending or logic ops applied.
Common::Vec4<u8> PixelColor(u16 x, u16 y, Common::Vec4<u8>& combiner_output) const;
Common::Vec4<u8> PixelColor(u16 x, u16 y) const;
/// Emulates the TEV configuration and returns the combiner output.
void WriteTevConfig(std::span<const Common::Vec4<u8>, 4> texture_color,
@ -61,7 +61,7 @@ private:
Common::Vec4<u8> secondary_fragment_color);
/// Blends fog to the combiner output if enabled.
void WriteFog(Common::Vec4<u8>& combiner_output, float depth) const;
void WriteFog(float depth) const;
/// Performs the alpha test. Returns false if the test failed.
bool DoAlphaTest(u8 alpha) const;
@ -73,10 +73,9 @@ private:
Memory::MemorySystem& memory;
Pica::State& state;
const Pica::Regs& regs;
size_t num_sw_threads;
Common::ThreadWorker sw_workers;
Framebuffer fb;
// Kirby Blowout Blast relies on the combiner output of a previous draw
// in order to render the sky correctly.
Common::Vec4<u8> combiner_output{};
};
} // namespace SwRenderer