renderer_software: Add jit compiler for tev stages

* Profiling shows this has a significant impact

src/video_core/CMakeLists.txt

@@ -95,6 +95,8 @@ add_library(video_core STATIC
     renderer_software/sw_proctex.h
     renderer_software/sw_rasterizer.cpp
     renderer_software/sw_rasterizer.h
+    renderer_software/sw_tev_jit.cpp
+    renderer_software/sw_tev_jit.h
     renderer_software/sw_texturing.cpp
     renderer_software/sw_texturing.h
     renderer_vulkan/pica_to_vk.h

src/video_core/renderer_software/renderer_software.cpp

@@ -8,6 +8,7 @@
 #include "core/hw/hw.h"
 #include "core/hw/lcd.h"
 #include "video_core/renderer_software/renderer_software.h"
+#include "video_core/renderer_software/sw_rasterizer.h"
 
 namespace SwRenderer {
 
@@ -17,6 +18,10 @@ RendererSoftware::RendererSoftware(Core::System& system, Frontend::EmuWindow& wi
 
 RendererSoftware::~RendererSoftware() = default;
 
+VideoCore::RasterizerInterface* RendererSoftware::Rasterizer() const {
+    return rasterizer.get();
+}
+
 void RendererSoftware::SwapBuffers() {
     PrepareRenderTarget();
     EndFrame();

src/video_core/renderer_software/renderer_software.h

@@ -5,7 +5,6 @@
 #pragma once
 
 #include "video_core/renderer_base.h"
-#include "video_core/renderer_software/sw_rasterizer.h"
 
 namespace Core {
 class System;
@@ -19,19 +18,18 @@ struct ScreenInfo {
     std::vector<u8> pixels;
 };
 
+class RasterizerSoftware;
+
 class RendererSoftware : public VideoCore::RendererBase {
 public:
     explicit RendererSoftware(Core::System& system, Frontend::EmuWindow& window);
     ~RendererSoftware() override;
 
-    [[nodiscard]] VideoCore::RasterizerInterface* Rasterizer() const override {
-        return rasterizer.get();
-    }
-
     [[nodiscard]] const ScreenInfo& Screen(VideoCore::ScreenId id) const noexcept {
         return screen_infos[static_cast<u32>(id)];
     }
 
+    VideoCore::RasterizerInterface* Rasterizer() const override;
     void SwapBuffers() override;
     void TryPresent(int timeout_ms, bool is_secondary) override {}
     void Sync() override {}

src/video_core/renderer_software/sw_framebuffer.cpp

@@ -41,10 +41,22 @@ Framebuffer::Framebuffer(Memory::MemorySystem& memory_, const Pica::FramebufferR
 
 Framebuffer::~Framebuffer() = default;
 
-void Framebuffer::DrawPixel(int x, int y, const Common::Vec4<u8>& color) const {
-    const auto& framebuffer = regs.framebuffer;
-    const PAddr addr = framebuffer.GetColorBufferPhysicalAddress();
+void Framebuffer::Bind() {
+    PAddr addr = regs.framebuffer.GetColorBufferPhysicalAddress();
+    if (color_addr != addr) [[unlikely]] {
+        color_addr = addr;
+        color_buffer = memory.GetPhysicalPointer(color_addr);
+    }
 
+    addr = regs.framebuffer.GetDepthBufferPhysicalAddress();
+    if (depth_addr != addr) [[unlikely]] {
+        depth_addr = addr;
+        depth_buffer = memory.GetPhysicalPointer(depth_addr);
+    }
+}
+
+void Framebuffer::DrawPixel(u32 x, u32 y, const Common::Vec4<u8>& color) const {
+    const auto& framebuffer = regs.framebuffer;
     // Similarly to textures, the render framebuffer is laid out from bottom to top, too.
     // NOTE: The framebuffer height register contains the actual FB height minus one.
     y = framebuffer.height - y;
@@ -54,8 +66,7 @@ void Framebuffer::DrawPixel(int x, int y, const Common::Vec4<u8>& color) const {
         GPU::Regs::BytesPerPixel(GPU::Regs::PixelFormat(framebuffer.color_format.Value()));
     const u32 dst_offset = VideoCore::GetMortonOffset(x, y, bytes_per_pixel) +
                            coarse_y * framebuffer.width * bytes_per_pixel;
-    u8* depth_buffer = memory.GetPhysicalPointer(addr);
-    u8* dst_pixel = depth_buffer + dst_offset;
+    u8* dst_pixel = color_buffer + dst_offset;
 
     switch (framebuffer.color_format) {
     case FramebufferRegs::ColorFormat::RGBA8:
@@ -80,10 +91,8 @@ void Framebuffer::DrawPixel(int x, int y, const Common::Vec4<u8>& color) const {
     }
 }
 
-const Common::Vec4<u8> Framebuffer::GetPixel(int x, int y) const {
+const Common::Vec4<u8> Framebuffer::GetPixel(u32 x, u32 y) const {
     const auto& framebuffer = regs.framebuffer;
-    const PAddr addr = framebuffer.GetColorBufferPhysicalAddress();
-
     y = framebuffer.height - y;
 
     const u32 coarse_y = y & ~7;
@@ -91,7 +100,6 @@ const Common::Vec4<u8> Framebuffer::GetPixel(int x, int y) const {
         GPU::Regs::BytesPerPixel(GPU::Regs::PixelFormat(framebuffer.color_format.Value()));
     const u32 src_offset = VideoCore::GetMortonOffset(x, y, bytes_per_pixel) +
                            coarse_y * framebuffer.width * bytes_per_pixel;
-    const u8* color_buffer = memory.GetPhysicalPointer(addr);
     const u8* src_pixel = color_buffer + src_offset;
 
     switch (framebuffer.color_format) {
@@ -114,10 +122,8 @@ const Common::Vec4<u8> Framebuffer::GetPixel(int x, int y) const {
     return {0, 0, 0, 0};
 }
 
-u32 Framebuffer::GetDepth(int x, int y) const {
+u32 Framebuffer::GetDepth(u32 x, u32 y) const {
     const auto& framebuffer = regs.framebuffer;
-    const PAddr addr = framebuffer.GetDepthBufferPhysicalAddress();
-
     y = framebuffer.height - y;
 
     const u32 coarse_y = y & ~7;
@@ -125,7 +131,6 @@ u32 Framebuffer::GetDepth(int x, int y) const {
     const u32 stride = framebuffer.width * bytes_per_pixel;
 
     const u32 src_offset = VideoCore::GetMortonOffset(x, y, bytes_per_pixel) + coarse_y * stride;
-    const u8* depth_buffer = memory.GetPhysicalPointer(addr);
     const u8* src_pixel = depth_buffer + src_offset;
 
     switch (framebuffer.depth_format) {
@@ -143,10 +148,8 @@ u32 Framebuffer::GetDepth(int x, int y) const {
     }
 }
 
-u8 Framebuffer::GetStencil(int x, int y) const {
+u8 Framebuffer::GetStencil(u32 x, u32 y) const {
     const auto& framebuffer = regs.framebuffer;
-    const PAddr addr = framebuffer.GetDepthBufferPhysicalAddress();
-
     y = framebuffer.height - y;
 
     const u32 coarse_y = y & ~7;
@@ -154,7 +157,6 @@ u8 Framebuffer::GetStencil(int x, int y) const {
     const u32 stride = framebuffer.width * bytes_per_pixel;
 
     const u32 src_offset = VideoCore::GetMortonOffset(x, y, bytes_per_pixel) + coarse_y * stride;
-    const u8* depth_buffer = memory.GetPhysicalPointer(addr);
     const u8* src_pixel = depth_buffer + src_offset;
 
     switch (framebuffer.depth_format) {
@@ -169,10 +171,8 @@ u8 Framebuffer::GetStencil(int x, int y) const {
     }
 }
 
-void Framebuffer::SetDepth(int x, int y, u32 value) const {
+void Framebuffer::SetDepth(u32 x, u32 y, u32 value) const {
     const auto& framebuffer = regs.framebuffer;
-    const PAddr addr = framebuffer.GetDepthBufferPhysicalAddress();
-
     y = framebuffer.height - y;
 
     const u32 coarse_y = y & ~7;
@@ -180,7 +180,6 @@ void Framebuffer::SetDepth(int x, int y, u32 value) const {
     const u32 stride = framebuffer.width * bytes_per_pixel;
 
     const u32 dst_offset = VideoCore::GetMortonOffset(x, y, bytes_per_pixel) + coarse_y * stride;
-    u8* depth_buffer = memory.GetPhysicalPointer(addr);
     u8* dst_pixel = depth_buffer + dst_offset;
 
     switch (framebuffer.depth_format) {
@@ -201,10 +200,8 @@ void Framebuffer::SetDepth(int x, int y, u32 value) const {
     }
 }
 
-void Framebuffer::SetStencil(int x, int y, u8 value) const {
+void Framebuffer::SetStencil(u32 x, u32 y, u8 value) const {
     const auto& framebuffer = regs.framebuffer;
-    const PAddr addr = framebuffer.GetDepthBufferPhysicalAddress();
-
     y = framebuffer.height - y;
 
     const u32 coarse_y = y & ~7;
@@ -212,7 +209,6 @@ void Framebuffer::SetStencil(int x, int y, u8 value) const {
     const u32 stride = framebuffer.width * bytes_per_pixel;
 
     const u32 dst_offset = VideoCore::GetMortonOffset(x, y, bytes_per_pixel) + coarse_y * stride;
-    u8* depth_buffer = memory.GetPhysicalPointer(addr);
     u8* dst_pixel = depth_buffer + dst_offset;
 
     switch (framebuffer.depth_format) {
@@ -231,7 +227,7 @@ void Framebuffer::SetStencil(int x, int y, u8 value) const {
     }
 }
 
-void Framebuffer::DrawShadowMapPixel(int x, int y, u32 depth, u8 stencil) const {
+void Framebuffer::DrawShadowMapPixel(u32 x, u32 y, u32 depth, u8 stencil) const {
     const auto& framebuffer = regs.framebuffer;
     const auto& shadow = regs.shadow;
     const PAddr addr = framebuffer.GetColorBufferPhysicalAddress();

src/video_core/renderer_software/sw_framebuffer.h

@@ -23,30 +23,37 @@ public:
     explicit Framebuffer(Memory::MemorySystem& memory, const Pica::FramebufferRegs& framebuffer);
     ~Framebuffer();
 
+    /// Updates the framebuffer addresses from the PICA registers.
+    void Bind();
+
     /// Draws a pixel at the specified coordinates.
-    void DrawPixel(int x, int y, const Common::Vec4<u8>& color) const;
+    void DrawPixel(u32 x, u32 y, const Common::Vec4<u8>& color) const;
 
     /// Returns the current color at the specified coordinates.
-    [[nodiscard]] const Common::Vec4<u8> GetPixel(int x, int y) const;
+    [[nodiscard]] const Common::Vec4<u8> GetPixel(u32 x, u32 y) const;
 
     /// Returns the depth value at the specified coordinates.
-    [[nodiscard]] u32 GetDepth(int x, int y) const;
+    [[nodiscard]] u32 GetDepth(u32 x, u32 y) const;
 
     /// Returns the stencil value at the specified coordinates.
-    [[nodiscard]] u8 GetStencil(int x, int y) const;
+    [[nodiscard]] u8 GetStencil(u32 x, u32 y) const;
 
     /// Stores the provided depth value at the specified coordinates.
-    void SetDepth(int x, int y, u32 value) const;
+    void SetDepth(u32 x, u32 y, u32 value) const;
 
     /// Stores the provided stencil value at the specified coordinates.
-    void SetStencil(int x, int y, u8 value) const;
+    void SetStencil(u32 x, u32 y, u8 value) const;
 
     /// Draws a pixel to the shadow buffer.
-    void DrawShadowMapPixel(int x, int y, u32 depth, u8 stencil) const;
+    void DrawShadowMapPixel(u32 x, u32 y, u32 depth, u8 stencil) const;
 
 private:
     Memory::MemorySystem& memory;
     const Pica::FramebufferRegs& regs;
+    PAddr color_addr;
+    u8* color_buffer{};
+    PAddr depth_addr;
+    u8* depth_buffer{};
 };
 
 u8 PerformStencilAction(Pica::FramebufferRegs::StencilAction action, u8 old_stencil, u8 ref);

src/video_core/renderer_software/sw_rasterizer.cpp

@@ -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,
@@ -289,167 +295,194 @@ void RasterizerSoftware::ProcessTriangle(const Vertex& v0, const Vertex& v1, con
 
     const auto w_inverse = Common::MakeVec(v0.pos.w, v1.pos.w, v2.pos.w);
 
-    auto textures = regs.texturing.GetTextures();
+    const auto textures = regs.texturing.GetTextures();
     const auto tev_stages = regs.texturing.GetTevStages();
+    for (u32 i = 0; i < texture_data.size(); i++) {
+        const PAddr addr = textures[i].config.GetPhysicalAddress();
+        if (addr) {
+            texture_data[i] = memory.GetPhysicalPointer(addr);
+        }
+    }
+
+    fb.Bind();
+
+    if (use_jit) {
+        const TevConfigKey key{regs.texturing};
+        auto [it, new_fun] = tev_cache.try_emplace(key.Hash());
+        if (new_fun) {
+            it->second = std::make_unique<TevConfig>(regs, key);
+        }
+        tev_config = it->second.get();
+    }
 
     // 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);
+                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(
@@ -538,11 +571,10 @@ std::array<Common::Vec4<u8>, 4> RasterizerSoftware::TextureColor(
             t = texture.config.height - 1 -
                 GetWrappedTexCoord(texture.config.wrap_t, t, texture.config.height);
 
-            const u8* texture_data = memory.GetPhysicalPointer(texture_address);
             const auto info = TextureInfo::FromPicaRegister(texture.config, texture.format);
 
             // TODO: Apply the min and mag filters to the texture
-            texture_color[i] = LookupTexture(texture_data, s, t, info);
+            texture_color[i] = LookupTexture(texture_data[i], s, t, info);
         }
 
         if (i == 0 && (texture.config.type == TexturingRegs::TextureConfig::Shadow2D ||
@@ -572,8 +604,7 @@ 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;
 
@@ -664,10 +695,20 @@ Common::Vec4<u8> RasterizerSoftware::PixelColor(u16 x, u16 y,
 }
 
 void RasterizerSoftware::WriteTevConfig(
-    std::span<const Common::Vec4<u8>, 4> texture_color,
+    std::span<Common::Vec4<u8>, 4> texture_color,
     std::span<const Pica::TexturingRegs::TevStageConfig, 6> tev_stages,
     Common::Vec4<u8> primary_color, Common::Vec4<u8> primary_fragment_color,
     Common::Vec4<u8> secondary_fragment_color) {
+
+#if CITRA_ARCH(x86_64)
+    if (use_jit) {
+        const u32 tev_combiner_buffer_color = regs.texturing.tev_combiner_buffer_color.raw;
+        combiner_output = tev_config->Run(texture_color, primary_color, primary_fragment_color,
+                                          secondary_fragment_color, tev_combiner_buffer_color);
+        return;
+    }
+#endif
+
     /**
      * Texture environment - consists of 6 stages of color and alpha combining.
      * Color combiners take three input color values from some source (e.g. interpolated
@@ -731,6 +772,7 @@ void RasterizerSoftware::WriteTevConfig(
             GetColorModifier(tev_stage.color_modifier2, get_source(tev_stage.color_source2)),
             GetColorModifier(tev_stage.color_modifier3, get_source(tev_stage.color_source3)),
         };
+
         const Common::Vec3<u8> color_output = ColorCombine(tev_stage.color_op, color_result);
 
         u8 alpha_output;
@@ -768,7 +810,7 @@ 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.

src/video_core/renderer_software/sw_rasterizer.h

@@ -4,13 +4,20 @@
 
 #pragma once
 
+#include <memory>
 #include <span>
+#include <unordered_map>
 
+#include "common/arch.h"
+#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"
 #include "video_core/renderer_software/sw_framebuffer.h"
 
+#if CITRA_ARCH(x86_64)
+#include "video_core/renderer_software/sw_tev_jit.h"
+#endif
+
 namespace Pica::Shader {
 struct OutputVertex;
 }
@@ -52,16 +59,16 @@ 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,
+    void WriteTevConfig(std::span<Common::Vec4<u8>, 4> texture_color,
                         std::span<const Pica::TexturingRegs::TevStageConfig, 6> tev_stages,
                         Common::Vec4<u8> primary_color, Common::Vec4<u8> primary_fragment_color,
                         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 +80,13 @@ private:
     Memory::MemorySystem& memory;
     Pica::State& state;
     const Pica::Regs& regs;
+    bool use_jit{true};
+    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{};
+    TevCache tev_cache;
+    TevConfig* tev_config{};
+    std::array<const u8*, 3> texture_data{};
 };
 
 } // namespace SwRenderer

src/video_core/renderer_software/sw_tev_jit.cpp

@@ -0,0 +1,473 @@
+// Copyright 2023 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <bit>
+#include <emmintrin.h>
+#include "common/x64/xbyak_abi.h"
+#include "video_core/regs.h"
+#include "video_core/renderer_software/sw_tev_jit.h"
+
+namespace SwRenderer {
+
+namespace {
+
+using namespace Common::X64;
+using namespace Xbyak::util;
+using Pica::TexturingRegs;
+using Xbyak::Reg32;
+using Xbyak::Reg64;
+using Xbyak::Xmm;
+using TevStageConfig = Pica::TexturingRegs::TevStageConfig;
+
+constexpr Reg32 A0 = r11d;
+constexpr Reg32 A1 = r12d;
+constexpr Reg32 A2 = r13d;
+constexpr Reg32 ALPHA_OUTPUT = r14d;
+constexpr Xmm COMBINER_OUTPUT = xmm0;
+constexpr Xmm COMBINER_BUFFER = xmm1;
+constexpr Xmm NEXT_COMBINER_BUFFER = xmm2;
+constexpr Xmm VEC0 = xmm3;
+constexpr Xmm VEC1 = xmm4;
+constexpr Xmm VEC2 = xmm5;
+constexpr Xmm COLOR_OUTPUT = xmm6;
+constexpr Xmm ZERO = xmm13;
+constexpr Xmm MID_COLOR = xmm14;
+constexpr Xmm MAX_COLOR = xmm15;
+
+bool IsPassThroughTevStage(const TevStageConfig& stage) {
+    return (stage.color_op == TevStageConfig::Operation::Replace &&
+            stage.alpha_op == TevStageConfig::Operation::Replace &&
+            stage.color_source1 == TevStageConfig::Source::Previous &&
+            stage.alpha_source1 == TevStageConfig::Source::Previous &&
+            stage.color_modifier1 == TevStageConfig::ColorModifier::SourceColor &&
+            stage.alpha_modifier1 == TevStageConfig::AlphaModifier::SourceAlpha &&
+            stage.GetColorMultiplier() == 1 && stage.GetAlphaMultiplier() == 1);
+}
+
+} // Anonymous namespace
+
+TevConfigKey::TevConfigKey(const Pica::TexturingRegs& regs) {
+    const auto& tev_stages = regs.GetTevStages();
+    for (size_t i = 0; i < tev_stages.size(); i++) {
+        const auto& tev_stage = tev_stages[i];
+        stages[i].sources_raw = tev_stage.sources_raw;
+        stages[i].modifiers_raw = tev_stage.modifiers_raw;
+        stages[i].ops_raw = tev_stage.ops_raw;
+        stages[i].const_color = tev_stage.const_color;
+        stages[i].scales_raw = tev_stage.scales_raw;
+    }
+}
+
+TevConfig::TevConfig(const Pica::Regs& regs_, const TevConfigKey& key) : regs{regs_} {
+    WriteTevConfig(key);
+}
+
+TevConfig::~TevConfig() = default;
+
+Common::Vec4<u8> TevConfig::Run(std::span<Common::Vec4<u8>, 4> texture_color_,
+                                Common::Vec4<u8> primary_color_,
+                                Common::Vec4<u8> primary_fragment_color_,
+                                Common::Vec4<u8> secondary_fragment_color_,
+                                u64 tev_combiner_buffer_color) {
+    u32* texture_color = reinterpret_cast<u32*>(texture_color_.data());
+    const u32 primary_color = std::bit_cast<u32>(primary_color_);
+    const u32 primary_fragment_color = std::bit_cast<u32>(primary_fragment_color_);
+    const u32 secondary_fragment_color = std::bit_cast<u32>(secondary_fragment_color_);
+    const u64 secondary_fragment_color_and_tev_combiner_buffer_color =
+        secondary_fragment_color | (tev_combiner_buffer_color << 32);
+    const u32 result = program(texture_color, primary_color, primary_fragment_color,
+                               secondary_fragment_color_and_tev_combiner_buffer_color);
+    return std::bit_cast<Common::Vec4<u8>>(result);
+}
+
+void TevConfig::WriteTevConfig(const TevConfigKey& key) {
+    program = (CompiledTevFun*)getCurr();
+
+    constexpr Xbyak::Reg TEXTURE_COLOR = ABI_PARAM1;
+    constexpr Xbyak::Reg PRIMARY_COLOR = ABI_PARAM2;
+    constexpr Xbyak::Reg PRIMARY_FRAGMENT_COLOR = ABI_PARAM3;
+    constexpr Xbyak::Reg SECONDARY_FRAGMENT_COLOR = ABI_PARAM4;
+
+    // Save calle state
+    ABI_PushRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8, 16);
+
+    // Clear the combiner registers and zero constant
+    pxor(COMBINER_OUTPUT, COMBINER_OUTPUT);
+    pxor(COMBINER_BUFFER, COMBINER_BUFFER);
+    pxor(ZERO, ZERO);
+
+    // Used to set an xmm register to the max color
+    static const __m128i max = _mm_set1_epi32(255);
+    mov(rax, reinterpret_cast<size_t>(&max));
+    movdqu(MAX_COLOR, xword[rax]);
+
+    // Used to set an xmm register to the mid color
+    static const __m128i mid = _mm_set1_epi32(128);
+    mov(rax, reinterpret_cast<size_t>(&mid));
+    movdqu(MID_COLOR, xword[rax]);
+
+    // Load next_combiner_buffer
+    mov(rax, ABI_PARAM4);
+    shr(rax, 32);
+    vmovd(NEXT_COMBINER_BUFFER, eax);
+    pmovzxbd(NEXT_COMBINER_BUFFER, NEXT_COMBINER_BUFFER);
+
+    for (u32 tev_stage_index = 0; tev_stage_index < key.stages.size(); ++tev_stage_index) {
+        const auto& tev_stage = key.stages[tev_stage_index];
+        if (!IsPassThroughTevStage(tev_stage)) {
+            using Source = TexturingRegs::TevStageConfig::Source;
+
+            const auto get_source = [&](const Xbyak::Xmm& dest, Source source) {
+                switch (source) {
+                case Source::PrimaryColor:
+                    vmovd(dest, PRIMARY_COLOR.cvt32());
+                    pmovzxbd(dest, dest);
+                    break;
+                case Source::PrimaryFragmentColor:
+                    vmovd(dest, PRIMARY_FRAGMENT_COLOR.cvt32());
+                    pmovzxbd(dest, dest);
+                    break;
+                case Source::SecondaryFragmentColor:
+                    vmovd(dest, SECONDARY_FRAGMENT_COLOR.cvt32());
+                    pmovzxbd(dest, dest);
+                    break;
+                case Source::Texture0:
+                case Source::Texture1:
+                case Source::Texture2:
+                case Source::Texture3: {
+                    const u32 index = static_cast<u32>(source) - static_cast<u32>(Source::Texture0);
+                    vmovd(dest, dword[TEXTURE_COLOR + index * sizeof(u32)]);
+                    pmovzxbd(dest, dest);
+                    break;
+                }
+                case Source::PreviousBuffer:
+                    vmovdqa(dest, COMBINER_BUFFER);
+                    break;
+                case Source::Constant:
+                    mov(eax, tev_stage.const_color);
+                    vmovd(dest, eax);
+                    pmovzxbd(dest, dest);
+                    break;
+                case Source::Previous:
+                    vmovdqa(dest, COMBINER_OUTPUT);
+                    break;
+                default:
+                    LOG_ERROR(HW_GPU, "Unknown color combiner source {}", source);
+                    UNIMPLEMENTED();
+                    vmovdqa(dest, ZERO);
+                }
+                return dest;
+            };
+
+            // Load the color modifiers to VEC0/1/2.
+            GetColorModifier(get_source(VEC0, tev_stage.color_source1), tev_stage.color_modifier1);
+            GetColorModifier(get_source(VEC1, tev_stage.color_source2), tev_stage.color_modifier2);
+            GetColorModifier(get_source(VEC2, tev_stage.color_source3), tev_stage.color_modifier3);
+
+            // Combine the texture colors to COLOR_OUTPUT.
+            ColorCombine(COLOR_OUTPUT, tev_stage.color_op);
+
+            if (tev_stage.color_op == TexturingRegs::TevStageConfig::Operation::Dot3_RGBA) {
+                // Result of Dot3_RGBA operation is also placed to the alpha component
+                vmovd(ALPHA_OUTPUT.cvt32(), COLOR_OUTPUT);
+            } else {
+                // Load the alpha modifers to VEC0/1/2.
+                GetAlphaModifier(get_source(VEC0, tev_stage.alpha_source1), A0,
+                                 tev_stage.alpha_modifier1);
+                GetAlphaModifier(get_source(VEC1, tev_stage.alpha_source2), A1,
+                                 tev_stage.alpha_modifier2);
+                GetAlphaModifier(get_source(VEC2, tev_stage.alpha_source3), A2,
+                                 tev_stage.alpha_modifier3);
+
+                // Combine the alpha values to ALPHA_OUTPUT.
+                AlphaCombine(ALPHA_OUTPUT, tev_stage.alpha_op);
+            }
+
+            // Load the color multipler to an SSE vector.
+            mov(eax, tev_stage.GetColorMultiplier());
+            movd(VEC0, eax);
+            pshufd(VEC0, VEC0, 0);
+
+            // Multiply color output with the multiplier and take the minimum.
+            pmulld(COLOR_OUTPUT, VEC0);
+            pminsd(COLOR_OUTPUT, MAX_COLOR);
+
+            // Load the alpha multiplier, multiply it with the alpha output.
+            mov(eax, tev_stage.GetAlphaMultiplier());
+            imul(ALPHA_OUTPUT, eax);
+
+            // Load result to a vector and take the minimum
+            movd(VEC0, ALPHA_OUTPUT);
+            pshufd(VEC0, VEC0, 0);
+            pminsd(VEC0, MAX_COLOR);
+
+            // Blend vectors to get the combiner output
+            vpblendd(COMBINER_OUTPUT, COLOR_OUTPUT, VEC0, 0b1000);
+        }
+
+        // Set combiner buffer to the next buffer
+        movq(COMBINER_BUFFER, NEXT_COMBINER_BUFFER);
+
+        if (regs.texturing.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferColor(
+                tev_stage_index)) {
+            vpblendd(NEXT_COMBINER_BUFFER, COMBINER_OUTPUT, NEXT_COMBINER_BUFFER, 0b1000);
+        }
+
+        if (regs.texturing.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferAlpha(
+                tev_stage_index)) {
+            vpblendd(NEXT_COMBINER_BUFFER, COMBINER_OUTPUT, NEXT_COMBINER_BUFFER, 0b0111);
+        }
+    }
+
+    // Pack combiner output to a u32 to be returned.
+    vpextrd(edx, COMBINER_OUTPUT, 3);
+    vpextrd(eax, COMBINER_OUTPUT, 2);
+    sal(edx, 8);
+    or_(eax, edx);
+    vpextrd(edx, COMBINER_OUTPUT, 1);
+    sal(eax, 8);
+    or_(edx, eax);
+    vmovd(eax, COMBINER_OUTPUT);
+    sal(edx, 8);
+    or_(eax, edx);
+
+    ABI_PopRegistersAndAdjustStack(*this, ABI_ALL_CALLEE_SAVED, 8, 16);
+    ret();
+    ready();
+}
+
+void TevConfig::GetColorModifier(const Xbyak::Xmm& dest, TevStageConfig::ColorModifier factor) {
+    using ColorModifier = TevStageConfig::ColorModifier;
+
+    const auto broadcast = [&](u32 comp) {
+        const u8 mask = comp | (comp << 2) | (comp << 4);
+        vpshufd(dest, dest, mask);
+    };
+
+    switch (factor) {
+    case ColorModifier::SourceColor:
+        vpblendd(dest, dest, ZERO, 0b1000);
+        break;
+    case ColorModifier::OneMinusSourceColor:
+        vpsubd(dest, MAX_COLOR, dest);
+        break;
+    case ColorModifier::SourceAlpha:
+        broadcast(3);
+        break;
+    case ColorModifier::OneMinusSourceAlpha:
+        broadcast(3);
+        vpsubd(dest, MAX_COLOR, dest);
+        break;
+    case ColorModifier::SourceRed:
+        broadcast(0);
+        break;
+    case ColorModifier::OneMinusSourceRed:
+        broadcast(0);
+        vpsubd(dest, MAX_COLOR, dest);
+        break;
+    case ColorModifier::SourceGreen:
+        broadcast(1);
+        break;
+    case ColorModifier::OneMinusSourceGreen:
+        broadcast(1);
+        vpsubd(dest, MAX_COLOR, dest);
+        break;
+    case ColorModifier::SourceBlue:
+        broadcast(2);
+        break;
+    case ColorModifier::OneMinusSourceBlue:
+        broadcast(2);
+        vpsubd(dest, MAX_COLOR, dest);
+        break;
+    default:
+        UNREACHABLE();
+    }
+    pand(dest, MAX_COLOR);
+};
+
+void TevConfig::ColorCombine(const Xbyak::Xmm& dest, TevStageConfig::Operation op) {
+    using Operation = TevStageConfig::Operation;
+
+    switch (op) {
+    case Operation::Replace:
+        vmovdqa(dest, VEC0);
+        break;
+    case Operation::Modulate:
+        pmulld(VEC0, VEC1);
+        vpsrlq(dest, VEC0, 8); // TODO: This is a very crude approximation of division by 255
+        break;
+    case Operation::Add:
+        vpaddd(VEC0, VEC0, VEC1);
+        vpminsd(dest, MAX_COLOR, VEC0);
+        break;
+    case Operation::AddSigned:
+        vpaddd(VEC0, VEC0, VEC1);
+        vpsubd(VEC0, VEC0, MID_COLOR);
+        vpminsd(VEC0, VEC0, MAX_COLOR);
+        vpmaxsd(dest, VEC0, ZERO);
+        break;
+    case Operation::Lerp:
+        pmulld(VEC0, VEC2);
+        psubd(VEC2, MAX_COLOR);
+        pmulld(VEC1, VEC2);
+        vpaddd(dest, VEC0, VEC1);
+        vpsrlq(dest, VEC0, 8); // TODO: This is a very crude approximation of division by 255
+        break;
+    case Operation::Subtract:
+        psubd(VEC0, VEC1);
+        vpmaxsd(dest, VEC0, ZERO);
+        break;
+    case Operation::MultiplyThenAdd:
+        pmulld(VEC0, VEC1);
+        pmulld(VEC2, MAX_COLOR);
+        paddd(VEC0, VEC2);
+        pminsd(VEC0, MAX_COLOR);
+        vpsrlq(dest, VEC0, 8); // TODO: This is a very crude approximation of division by 255
+        break;
+    case Operation::AddThenMultiply:
+        paddd(VEC0, VEC1);
+        pminsd(VEC0, MAX_COLOR);
+        pmulld(VEC0, VEC2);
+        vpsrlq(dest, VEC0, 8); // TODO: This is a very crude approximation of division by 255
+        break;
+    case Operation::Dot3_RGB:
+    case Operation::Dot3_RGBA:
+        pslld(VEC0, 1);
+        psubd(VEC0, MAX_COLOR);
+        pslld(VEC1, 1);
+        psubd(VEC1, MAX_COLOR);
+        pmulld(VEC0, VEC1);
+        paddd(VEC0, MID_COLOR);
+        psrld(VEC0, 8);
+        vpblendd(VEC0, VEC0, ZERO, 0b1000);
+        phaddd(VEC0, VEC0);
+        phaddd(VEC0, VEC0);
+        pminsd(VEC0, MAX_COLOR);
+        pmaxsd(VEC0, ZERO);
+        pshufd(dest, VEC0, 0);
+        break;
+    default:
+        LOG_ERROR(HW_GPU, "Unknown color combiner operation {}", (int)op);
+        UNIMPLEMENTED();
+    }
+    pand(dest, MAX_COLOR);
+};
+
+void TevConfig::GetAlphaModifier(const Xbyak::Xmm& src, const Xbyak::Reg32& dest,
+                                 TevStageConfig::AlphaModifier factor) {
+    using AlphaModifier = TevStageConfig::AlphaModifier;
+
+    const auto get_comp = [&](u32 comp, bool minus = false) {
+        const auto& reg = minus ? eax : dest;
+        vpextrd(reg, src, comp);
+        if (minus) {
+            mov(dest, 255);
+            sub(dest, reg);
+        }
+    };
+
+    switch (factor) {
+    case AlphaModifier::SourceAlpha:
+        get_comp(3);
+        break;
+    case AlphaModifier::OneMinusSourceAlpha:
+        get_comp(3, true);
+        break;
+    case AlphaModifier::SourceRed:
+        get_comp(0);
+        break;
+    case AlphaModifier::OneMinusSourceRed:
+        get_comp(0, true);
+        break;
+    case AlphaModifier::SourceGreen:
+        get_comp(1);
+        break;
+    case AlphaModifier::OneMinusSourceGreen:
+        get_comp(1, true);
+        break;
+    case AlphaModifier::SourceBlue:
+        get_comp(2);
+        break;
+    case AlphaModifier::OneMinusSourceBlue:
+        get_comp(2, true);
+        break;
+    default:
+        UNREACHABLE();
+    }
+};
+
+void TevConfig::AlphaCombine(const Xbyak::Reg32& dest, TevStageConfig::Operation op) {
+    using Operation = TevStageConfig::Operation;
+
+    const auto div_255 = [&](const Reg32& dst, const Reg32& src) {
+        mov(dst, 0x80808081);
+        imul(dst.cvt64(), src.cvt64());
+        shr(dst.cvt64(), 39);
+    };
+
+    switch (op) {
+    case Operation::Replace:
+        mov(dest, A0);
+        break;
+    case Operation::Modulate:
+        imul(A0, A1);
+        div_255(dest, A0);
+        break;
+    case Operation::Add:
+        add(A0, A1);
+        cmp(A0, 255);
+        mov(eax, 255);
+        cmovb(A0, eax);
+        break;
+    case Operation::AddSigned:
+        xor_(eax, eax);
+        add(A0, A1);
+        sub(A0, 128);
+        test(A0, A0);
+        cmovg(eax, A0);
+        cmp(eax, 255);
+        mov(A0, 255);
+        cmovb(A0, eax);
+        break;
+    case Operation::Lerp:
+        imul(A0, A2);
+        mov(eax, 255);
+        sub(eax, A2);
+        imul(A1, eax);
+        add(A0, A1);
+        div_255(dest, A0);
+        break;
+    case Operation::Subtract:
+        sub(A0, A1);
+        xor_(eax, eax);
+        test(A0, A0);
+        cmovl(A0, eax);
+        mov(dest, A0);
+        break;
+    case Operation::MultiplyThenAdd:
+        imul(A0, A1);
+        mov(dest, A2);
+        shl(dest, 8);
+        sub(dest, A2);
+        add(dest, A0);
+        div_255(eax, dest);
+        cmp(eax, 255);
+        mov(dest, 255);
+        cmovb(dest, eax);
+        break;
+    case Operation::AddThenMultiply:
+        add(A0, A1);
+        cmp(A0, 255);
+        mov(eax, 255);
+        cmovg(A0, eax);
+        imul(A0, A2);
+        div_255(dest, A0);
+        break;
+    default:
+        LOG_ERROR(HW_GPU, "Unknown alpha combiner operation {}", (int)op);
+        UNIMPLEMENTED();
+    }
+};
+
+} // namespace SwRenderer

src/video_core/renderer_software/sw_tev_jit.h

@@ -0,0 +1,64 @@
+// Copyright 2023 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <span>
+#include <xbyak/xbyak.h>
+
+#include "common/hash.h"
+#include "common/vector_math.h"
+#include "video_core/regs_texturing.h"
+
+namespace Pica {
+struct State;
+struct Regs;
+} // namespace Pica
+
+namespace SwRenderer {
+
+struct TevConfigKey {
+    explicit TevConfigKey(const Pica::TexturingRegs& regs);
+
+    u64 Hash() const noexcept {
+        return Common::ComputeHash64(this, sizeof(TevConfigKey));
+    }
+
+    std::array<Pica::TexturingRegs::TevStageConfig, 6> stages;
+};
+
+class TevConfig : public Xbyak::CodeGenerator {
+public:
+    explicit TevConfig(const Pica::Regs& regs, const TevConfigKey& key);
+    ~TevConfig();
+
+    Common::Vec4<u8> Run(std::span<Common::Vec4<u8>, 4> texture_color_,
+                         Common::Vec4<u8> primary_color_, Common::Vec4<u8> primary_fragment_color_,
+                         Common::Vec4<u8> secondary_fragment_color_, u64 tev_combiner_buffer_color);
+
+private:
+    void WriteTevConfig(const TevConfigKey& key);
+
+    void GetColorModifier(const Xbyak::Xmm& dest,
+                          Pica::TexturingRegs::TevStageConfig::ColorModifier factor);
+
+    void GetAlphaModifier(const Xbyak::Xmm& src, const Xbyak::Reg32& dest,
+                          Pica::TexturingRegs::TevStageConfig::AlphaModifier factor);
+
+    void ColorCombine(const Xbyak::Xmm& dest, Pica::TexturingRegs::TevStageConfig::Operation op);
+
+    void AlphaCombine(const Xbyak::Reg32& dest, Pica::TexturingRegs::TevStageConfig::Operation op);
+
+private:
+    const Pica::Regs& regs;
+
+    using CompiledTevFun = u32(u32* texture_color, u32 primary_color, u32 primary_fragment_color,
+                               u64 secondary_fragment_color_and_tev_combiner_buffer_color);
+
+    CompiledTevFun* program = nullptr;
+};
+
+using TevCache = std::unordered_map<u64, std::unique_ptr<TevConfig>, Common::IdentityHash<u64>>;
+
+} // namespace SwRenderer