summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat (limited to '')
-rw-r--r--src/video_core/CMakeLists.txt2
-rw-r--r--src/video_core/clipper.cpp3
-rw-r--r--src/video_core/pica.h52
-rw-r--r--src/video_core/rasterizer.cpp180
-rw-r--r--src/video_core/rasterizer.h21
-rw-r--r--src/video_core/video_core.vcxproj2
-rw-r--r--src/video_core/video_core.vcxproj.filters2
7 files changed, 260 insertions, 2 deletions
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index 828bf30fc..8e7b93acb 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -1,6 +1,7 @@
set(SRCS clipper.cpp
command_processor.cpp
primitive_assembly.cpp
+ rasterizer.cpp
utils.cpp
vertex_shader.cpp
video_core.cpp
@@ -10,6 +11,7 @@ set(HEADERS clipper.h
command_processor.h
math.h
primitive_assembly.h
+ rasterizer.h
utils.h
video_core.h
renderer_base.h
diff --git a/src/video_core/clipper.cpp b/src/video_core/clipper.cpp
index e9ab6242c..b7180328c 100644
--- a/src/video_core/clipper.cpp
+++ b/src/video_core/clipper.cpp
@@ -6,6 +6,7 @@
#include "clipper.h"
#include "pica.h"
+#include "rasterizer.h"
#include "vertex_shader.h"
namespace Pica {
@@ -168,7 +169,7 @@ void ProcessTriangle(OutputVertex &v0, OutputVertex &v1, OutputVertex &v2) {
vtx1.screenpos.x.ToFloat32(), vtx1.screenpos.y.ToFloat32(), vtx1.screenpos.z.ToFloat32(),
vtx2.screenpos.x.ToFloat32(), vtx2.screenpos.y.ToFloat32(), vtx2.screenpos.z.ToFloat32());
- // TODO: Send triangle to rasterizer
+ Rasterizer::ProcessTriangle(vtx0, vtx1, vtx2);
}
}
diff --git a/src/video_core/pica.h b/src/video_core/pica.h
index 1ced0d323..81af57336 100644
--- a/src/video_core/pica.h
+++ b/src/video_core/pica.h
@@ -94,7 +94,55 @@ struct Regs {
BitField<16, 16, u32> y;
} viewport_corner;
- INSERT_PADDING_WORDS(0x197);
+ INSERT_PADDING_WORDS(0xa7);
+
+ struct {
+ enum ColorFormat : u32 {
+ RGBA8 = 0,
+ RGB8 = 1,
+ RGBA5551 = 2,
+ RGB565 = 3,
+ RGBA4 = 4,
+ };
+
+ INSERT_PADDING_WORDS(0x6);
+
+ u32 depth_format;
+ u32 color_format;
+
+ INSERT_PADDING_WORDS(0x4);
+
+ u32 depth_buffer_address;
+ u32 color_buffer_address;
+
+ union {
+ // Apparently, the framebuffer width is stored as expected,
+ // while the height is stored as the actual height minus one.
+ // Hence, don't access these fields directly but use the accessors
+ // GetWidth() and GetHeight() instead.
+ BitField< 0, 11, u32> width;
+ BitField<12, 10, u32> height;
+ };
+
+ INSERT_PADDING_WORDS(0x1);
+
+ inline u32 GetColorBufferAddress() const {
+ return Memory::PhysicalToVirtualAddress(DecodeAddressRegister(color_buffer_address));
+ }
+ inline u32 GetDepthBufferAddress() const {
+ return Memory::PhysicalToVirtualAddress(DecodeAddressRegister(depth_buffer_address));
+ }
+
+ inline u32 GetWidth() const {
+ return width;
+ }
+
+ inline u32 GetHeight() const {
+ return height + 1;
+ }
+ } framebuffer;
+
+ INSERT_PADDING_WORDS(0xe0);
struct {
enum class Format : u64 {
@@ -355,6 +403,7 @@ struct Regs {
ADD_FIELD(viewport_depth_range);
ADD_FIELD(viewport_depth_far_plane);
ADD_FIELD(viewport_corner);
+ ADD_FIELD(framebuffer);
ADD_FIELD(vertex_attributes);
ADD_FIELD(index_array);
ADD_FIELD(num_vertices);
@@ -411,6 +460,7 @@ ASSERT_REG_POSITION(viewport_depth_far_plane, 0x4e);
ASSERT_REG_POSITION(vs_output_attributes[0], 0x50);
ASSERT_REG_POSITION(vs_output_attributes[1], 0x51);
ASSERT_REG_POSITION(viewport_corner, 0x68);
+ASSERT_REG_POSITION(framebuffer, 0x110);
ASSERT_REG_POSITION(vertex_attributes, 0x200);
ASSERT_REG_POSITION(index_array, 0x227);
ASSERT_REG_POSITION(num_vertices, 0x228);
diff --git a/src/video_core/rasterizer.cpp b/src/video_core/rasterizer.cpp
new file mode 100644
index 000000000..a7c1bab3e
--- /dev/null
+++ b/src/video_core/rasterizer.cpp
@@ -0,0 +1,180 @@
+// Copyright 2014 Citra Emulator Project
+// Licensed under GPLv2
+// Refer to the license.txt file included.
+
+#include <algorithm>
+
+#include "common/common_types.h"
+
+#include "math.h"
+#include "pica.h"
+#include "rasterizer.h"
+#include "vertex_shader.h"
+
+namespace Pica {
+
+namespace Rasterizer {
+
+static void DrawPixel(int x, int y, const Math::Vec4<u8>& color) {
+ u32* color_buffer = (u32*)Memory::GetPointer(registers.framebuffer.GetColorBufferAddress());
+ u32 value = (color.a() << 24) | (color.r() << 16) | (color.g() << 8) | color.b();
+
+ // Assuming RGBA8 format until actual framebuffer format handling is implemented
+ *(color_buffer + x + y * registers.framebuffer.GetWidth() / 2) = value;
+}
+
+static u32 GetDepth(int x, int y) {
+ u16* depth_buffer = (u16*)Memory::GetPointer(registers.framebuffer.GetDepthBufferAddress());
+
+ // Assuming 16-bit depth buffer format until actual format handling is implemented
+ return *(depth_buffer + x + y * registers.framebuffer.GetWidth() / 2);
+}
+
+static void SetDepth(int x, int y, u16 value) {
+ u16* depth_buffer = (u16*)Memory::GetPointer(registers.framebuffer.GetDepthBufferAddress());
+
+ // Assuming 16-bit depth buffer format until actual format handling is implemented
+ *(depth_buffer + x + y * registers.framebuffer.GetWidth() / 2) = value;
+}
+
+void ProcessTriangle(const VertexShader::OutputVertex& v0,
+ const VertexShader::OutputVertex& v1,
+ const VertexShader::OutputVertex& v2)
+{
+ // NOTE: Assuming that rasterizer coordinates are 12.4 fixed-point values
+ struct Fix12P4 {
+ Fix12P4() {}
+ Fix12P4(u16 val) : val(val) {}
+
+ static u16 FracMask() { return 0xF; }
+ static u16 IntMask() { return (u16)~0xF; }
+
+ operator u16() const {
+ return val;
+ }
+
+ bool operator < (const Fix12P4& oth) const {
+ return (u16)*this < (u16)oth;
+ }
+
+ private:
+ u16 val;
+ };
+
+ // vertex positions in rasterizer coordinates
+ auto FloatToFix = [](float24 flt) {
+ return Fix12P4(flt.ToFloat32() * 16.0f);
+ };
+ auto ScreenToRasterizerCoordinates = [FloatToFix](const Math::Vec3<float24> vec) {
+ return Math::Vec3<Fix12P4>{FloatToFix(vec.x), FloatToFix(vec.y), FloatToFix(vec.z)};
+ };
+ Math::Vec3<Fix12P4> vtxpos[3]{ ScreenToRasterizerCoordinates(v0.screenpos),
+ ScreenToRasterizerCoordinates(v1.screenpos),
+ ScreenToRasterizerCoordinates(v2.screenpos) };
+
+ // TODO: Proper scissor rect test!
+ u16 min_x = std::min({vtxpos[0].x, vtxpos[1].x, vtxpos[2].x});
+ u16 min_y = std::min({vtxpos[0].y, vtxpos[1].y, vtxpos[2].y});
+ u16 max_x = std::max({vtxpos[0].x, vtxpos[1].x, vtxpos[2].x});
+ u16 max_y = std::max({vtxpos[0].y, vtxpos[1].y, vtxpos[2].y});
+
+ min_x = min_x & Fix12P4::IntMask();
+ min_y = min_y & Fix12P4::IntMask();
+ max_x = (max_x + Fix12P4::FracMask()) & Fix12P4::IntMask();
+ max_y = (max_y + Fix12P4::FracMask()) & Fix12P4::IntMask();
+
+ // Triangle filling rules: Pixels on the right-sided edge or on flat bottom edges are not
+ // drawn. Pixels on any other triangle border are drawn. This is implemented with three bias
+ // values which are added to the barycentric coordinates w0, w1 and w2, respectively.
+ // NOTE: These are the PSP filling rules. Not sure if the 3DS uses the same ones...
+ auto IsRightSideOrFlatBottomEdge = [](const Math::Vec2<Fix12P4>& vtx,
+ const Math::Vec2<Fix12P4>& line1,
+ const Math::Vec2<Fix12P4>& line2)
+ {
+ if (line1.y == line2.y) {
+ // just check if vertex is above us => bottom line parallel to x-axis
+ return vtx.y < line1.y;
+ } else {
+ // check if vertex is on our left => right side
+ // TODO: Not sure how likely this is to overflow
+ return (int)vtx.x < (int)line1.x + ((int)line2.x - (int)line1.x) * ((int)vtx.y - (int)line1.y) / ((int)line2.y - (int)line1.y);
+ }
+ };
+ int bias0 = IsRightSideOrFlatBottomEdge(vtxpos[0].xy(), vtxpos[1].xy(), vtxpos[2].xy()) ? -1 : 0;
+ int bias1 = IsRightSideOrFlatBottomEdge(vtxpos[1].xy(), vtxpos[2].xy(), vtxpos[0].xy()) ? -1 : 0;
+ int bias2 = IsRightSideOrFlatBottomEdge(vtxpos[2].xy(), vtxpos[0].xy(), vtxpos[1].xy()) ? -1 : 0;
+
+ // TODO: Not sure if looping through x first might be faster
+ for (u16 y = min_y; y < max_y; y += 0x10) {
+ for (u16 x = min_x; x < max_x; x += 0x10) {
+
+ // Calculate the barycentric coordinates w0, w1 and w2
+ auto orient2d = [](const Math::Vec2<Fix12P4>& vtx1,
+ const Math::Vec2<Fix12P4>& vtx2,
+ const Math::Vec2<Fix12P4>& vtx3) {
+ const auto vec1 = (vtx2.Cast<int>() - vtx1.Cast<int>()).Append(0);
+ const auto vec2 = (vtx3.Cast<int>() - vtx1.Cast<int>()).Append(0);
+ // TODO: There is a very small chance this will overflow for sizeof(int) == 4
+ return Cross(vec1, vec2).z;
+ };
+
+ int w0 = bias0 + orient2d(vtxpos[1].xy(), vtxpos[2].xy(), {x, y});
+ int w1 = bias1 + orient2d(vtxpos[2].xy(), vtxpos[0].xy(), {x, y});
+ int w2 = bias2 + orient2d(vtxpos[0].xy(), vtxpos[1].xy(), {x, y});
+ int wsum = w0 + w1 + w2;
+
+ // If current pixel is not covered by the current primitive
+ if (w0 < 0 || w1 < 0 || w2 < 0)
+ continue;
+
+ // 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.
+ auto GetInterpolatedAttribute = [&](float24 attr0, float24 attr1, float24 attr2) {
+ auto attr_over_w = Math::MakeVec3(attr0 / v0.pos.w,
+ attr1 / v1.pos.w,
+ attr2 / v2.pos.w);
+ auto w_inverse = Math::MakeVec3(float24::FromFloat32(1.f) / v0.pos.w,
+ float24::FromFloat32(1.f) / v1.pos.w,
+ float24::FromFloat32(1.f) / v2.pos.w);
+ auto baricentric_coordinates = Math::MakeVec3(float24::FromFloat32(w0),
+ float24::FromFloat32(w1),
+ float24::FromFloat32(w2));
+
+ float24 interpolated_attr_over_w = Math::Dot(attr_over_w, baricentric_coordinates);
+ float24 interpolated_w_inverse = Math::Dot(w_inverse, baricentric_coordinates);
+ return interpolated_attr_over_w / interpolated_w_inverse;
+ };
+
+ Math::Vec4<u8> primary_color{
+ (u8)(GetInterpolatedAttribute(v0.color.r(), v1.color.r(), v2.color.r()).ToFloat32() * 255),
+ (u8)(GetInterpolatedAttribute(v0.color.g(), v1.color.g(), v2.color.g()).ToFloat32() * 255),
+ (u8)(GetInterpolatedAttribute(v0.color.b(), v1.color.b(), v2.color.b()).ToFloat32() * 255),
+ (u8)(GetInterpolatedAttribute(v0.color.a(), v1.color.a(), v2.color.a()).ToFloat32() * 255)
+ };
+
+ u16 z = (u16)(((float)v0.screenpos[2].ToFloat32() * w0 +
+ (float)v1.screenpos[2].ToFloat32() * w1 +
+ (float)v2.screenpos[2].ToFloat32() * w2) * 65535.f / wsum); // TODO: Shouldn't need to multiply by 65536?
+ SetDepth(x >> 4, y >> 4, z);
+
+ DrawPixel(x >> 4, y >> 4, primary_color);
+ }
+ }
+}
+
+} // namespace Rasterizer
+
+} // namespace Pica
diff --git a/src/video_core/rasterizer.h b/src/video_core/rasterizer.h
new file mode 100644
index 000000000..500be9462
--- /dev/null
+++ b/src/video_core/rasterizer.h
@@ -0,0 +1,21 @@
+// Copyright 2014 Citra Emulator Project
+// Licensed under GPLv2
+// Refer to the license.txt file included.
+
+#pragma once
+
+namespace Pica {
+
+namespace VertexShader {
+ struct OutputVertex;
+}
+
+namespace Rasterizer {
+
+void ProcessTriangle(const VertexShader::OutputVertex& v0,
+ const VertexShader::OutputVertex& v1,
+ const VertexShader::OutputVertex& v2);
+
+} // namespace Rasterizer
+
+} // namespace Pica
diff --git a/src/video_core/video_core.vcxproj b/src/video_core/video_core.vcxproj
index 99ab63dce..48d77cdc4 100644
--- a/src/video_core/video_core.vcxproj
+++ b/src/video_core/video_core.vcxproj
@@ -23,6 +23,7 @@
<ClCompile Include="clipper.cpp" />
<ClCompile Include="command_processor.cpp" />
<ClCompile Include="primitive_assembly.cpp" />
+ <ClCompile Include="rasterizer.cpp" />
<ClCompile Include="utils.cpp" />
<ClCompile Include="vertex_shader.cpp" />
<ClCompile Include="video_core.cpp" />
@@ -34,6 +35,7 @@
<ClInclude Include="math.h" />
<ClInclude Include="pica.h" />
<ClInclude Include="primitive_assembly.h" />
+ <ClInclude Include="rasterizer.h" />
<ClInclude Include="renderer_base.h" />
<ClInclude Include="utils.h" />
<ClInclude Include="vertex_shader.h" />
diff --git a/src/video_core/video_core.vcxproj.filters b/src/video_core/video_core.vcxproj.filters
index 5222f2fa0..31af4f1df 100644
--- a/src/video_core/video_core.vcxproj.filters
+++ b/src/video_core/video_core.vcxproj.filters
@@ -12,6 +12,7 @@
<ClCompile Include="clipper.cpp" />
<ClCompile Include="command_processor.cpp" />
<ClCompile Include="primitive_assembly.cpp" />
+ <ClCompile Include="rasterizer.cpp" />
<ClCompile Include="utils.cpp" />
<ClCompile Include="vertex_shader.cpp" />
<ClCompile Include="video_core.cpp" />
@@ -26,6 +27,7 @@
<ClInclude Include="math.h" />
<ClInclude Include="pica.h" />
<ClInclude Include="primitive_assembly.h" />
+ <ClInclude Include="rasterizer.h" />
<ClInclude Include="renderer_base.h" />
<ClInclude Include="utils.h" />
<ClInclude Include="vertex_shader.h" />
generated by cgit v1.2.3 (git 2.25.1) at 2024-11-02 09:03:39 +0100