diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 02a3de140e..0f16faf1f6 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -288,6 +288,13 @@ jobs:
           args: ./b.sh --loongarch64 PPSSPPHeadless
           id: loongarch64
 
+        - os: ubuntu-26.04
+          extra: riscv64
+          cc: gcc
+          cxx: g++
+          args: ./b.sh --riscv64 PPSSPPHeadless
+          id: riscv64
+
         - os: macos-latest
           extra: test
           cc: clang
@@ -327,7 +334,7 @@ jobs:
         ndk-version: r29
 
     - name: Install Linux dependencies
-      if: runner.os == 'Linux' && matrix.extra != 'android' && matrix.extra != 'loongarch64'
+      if: runner.os == 'Linux' && matrix.extra != 'android' && matrix.extra != 'loongarch64' && matrix.extra != 'riscv64'
       run: |
         sudo apt-get update -y -qq
         sudo apt-get install libsdl3-dev libgl1-mesa-dev libglu1-mesa-dev libsdl3-ttf-dev libfontconfig1-dev libcurl4-openssl-dev
@@ -336,12 +343,12 @@ jobs:
       if: runner.os == 'macOS' && matrix.id == 'macos'
       run: brew install sdl3 sdl3_ttf
 
-    - name: Install loongarch64 cross-compilation dependencies
-      if: matrix.extra == 'loongarch64'
+    - name: Install cross-compilation dependencies
+      if: matrix.extra == 'loongarch64' || matrix.extra == 'riscv64'
       run: |
         sudo apt-get update -y -qq
-        sudo apt-get install -y gcc-14-loongarch64-linux-gnu g++-14-loongarch64-linux-gnu binutils-loongarch64-linux-gnu libgl-dev libglu1-mesa-dev
-        sudo cmake/scripts/setup-loongarch64-cross.sh
+        sudo apt-get install -y gcc-14-${{ matrix.extra }}-linux-gnu g++-14-${{ matrix.extra }}-linux-gnu binutils-${{ matrix.extra }}-linux-gnu libgl-dev libglu1-mesa-dev
+        sudo cmake/scripts/setup-cross.sh ${{ matrix.extra }}
 
     - name: Install iOS dependencies
       if: matrix.id == 'ios'
diff --git a/CMakeLists.txt b/CMakeLists.txt
index c81d415524..65e2b49ca7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -184,6 +184,8 @@ option(USE_VULKAN_DISPLAY_KHR "Enable or disable full screen display of Vulkan"
 # :: Frontends
 option(MOBILE_DEVICE "Set to ON when targeting a mobile device" ${MOBILE_DEVICE})
 option(HEADLESS "Set to OFF to not generate the PPSSPPHeadless target" ${HEADLESS})
+# Set by b.sh for the loongarch64/riscv64 cross builds, whose sysroots have no SDL3.
+option(HEADLESS_CROSS "Set to ON for a headless-only cross build (no SDL frontend)" ${HEADLESS_CROSS})
 option(ATLAS_TOOL "Set to OFF to not generate the ATLAS_TOOL target" ${ATLAS_TOOL})
 option(UNITTEST "Set to ON to generate the unittest target" ${UNITTEST})
 option(SIMULATOR "Set to ON when targeting an x86 simulator of an ARM platform" ${SIMULATOR})
@@ -331,7 +333,7 @@ set(SDL_LIB_TARGET "")
 set(SDL_TTF_LIB_TARGET "")
 set(FONTCONFIG_LIB_TARGET "")
 
-if(NOT LIBRETRO AND NOT IOS AND NOT LOONGARCH64_DEVICE AND NOT ANDROID)
+if(NOT LIBRETRO AND NOT IOS AND NOT HEADLESS_CROSS AND NOT ANDROID)
 	find_package(SDL3 QUIET)
 	if(NOT SDL3_FOUND)
 		if(CMAKE_SYSTEM_NAME STREQUAL "Linux" AND (EXISTS "/usr/bin/apt" OR EXISTS "/usr/bin/apt-get" OR EXISTS "/etc/debian_version"))
@@ -1001,9 +1003,11 @@ elseif(WIN32)
 	# Don't care about SDL.
 	set(TargetBin PPSSPPWindows)
 elseif(LIBRETRO)
-elseif(LOONGARCH64_DEVICE)
-	# Cross-compiling for loongarch64 currently only builds PPSSPPHeadless, and there's
-	# no SDL3 available for that target, so skip the desktop SDL frontend entirely.
+elseif(HEADLESS_CROSS)
+	# The loongarch64/riscv64 cross builds only build PPSSPPHeadless, and there's no
+	# SDL3 available in those sysroots, so skip the desktop SDL frontend entirely.
+	# Note this is deliberately not keyed off the target architecture: a native build
+	# on such a machine can have SDL3 and should still get the normal frontend.
 else()
 	if(GOLD)
 		set(TargetBin PPSSPPGold)
@@ -1362,6 +1366,7 @@ if(UNITTEST)
 		unittest/TestShaderGenerators.cpp
 		unittest/TestArmEmitter.cpp
 		unittest/TestArm64Emitter.cpp
+		unittest/TestCrossSIMD.cpp
 		unittest/TestIRPassSimplify.cpp
 		unittest/TestX64Emitter.cpp
 		unittest/TestVertexJit.cpp
diff --git a/Common/LoongArch64Emitter.h b/Common/LoongArch64Emitter.h
index 134b5de61d..d5abda3809 100644
--- a/Common/LoongArch64Emitter.h
+++ b/Common/LoongArch64Emitter.h
@@ -166,7 +166,8 @@ public:
 		QuickCallFunction((const u8 *)func, scratchreg);
 	}
 	void QuickCallFunctionR(const u8 *func, LoongArch64Reg arg, LoongArch64Reg scratchreg = R_RA) {
-		MOVE(LoongArch64Reg::X4, arg);
+		// The first integer argument goes in a0, which is R4 - X4 is an LASX vector register.
+		MOVE(LoongArch64Reg::R4, arg);
 		QuickJump(scratchreg, R_RA, func);
 	}
 	template <typename T>
diff --git a/Common/Math/CrossSIMD.h b/Common/Math/CrossSIMD.h
index af30c25b06..8c9316b671 100644
--- a/Common/Math/CrossSIMD.h
+++ b/Common/Math/CrossSIMD.h
@@ -6,6 +6,7 @@
 
 #pragma once
 
+#include <cmath>
 #include <cstring>
 #include "Common/Math/SIMDHeaders.h"
 
@@ -1187,8 +1188,11 @@ struct Vec4F32 {
 	}
 	void StoreConvertToU8(uint8_t *dest) {
 		__m128i ivalue32 = __lsx_vftintrz_w_s(v);
-		__m128i ivalue16 = __lsx_vssrlrni_hu_w(ivalue32, ivalue32, 0);
-		__m128i ivalue8 = __lsx_vssrlrni_bu_h(ivalue16, ivalue16, 0);
+		// Narrow signed->signed first, then signed->unsigned, matching the packs/packus pair the
+		// SSE version uses. The logical (unsigned) narrowing shifts turn a negative value into a
+		// huge unsigned one, which saturates to 255 instead of clamping to 0.
+		__m128i ivalue16 = __lsx_vssrani_h_w(ivalue32, ivalue32, 0);
+		__m128i ivalue8 = __lsx_vssrani_bu_h(ivalue16, ivalue16, 0);
 		uint32_t value = __lsx_vpickve2gr_wu(ivalue8, 0);
 		memcpy(dest, &value, sizeof(uint32_t));
 	}
@@ -1198,10 +1202,10 @@ struct Vec4F32 {
 		// index is a compile-time constant parameter to the intrinsic.
 		int ival;
 		switch (index) {
-		case 0: ival = __lsx_vpickve2gr_w((__m128i)v, 0);
-		case 1: ival = __lsx_vpickve2gr_w((__m128i)v, 1);
-		case 2: ival = __lsx_vpickve2gr_w((__m128i)v, 2);
-		default: ival = __lsx_vpickve2gr_w((__m128i)v, 3);
+		case 0: ival = __lsx_vpickve2gr_w((__m128i)v, 0); break;
+		case 1: ival = __lsx_vpickve2gr_w((__m128i)v, 1); break;
+		case 2: ival = __lsx_vpickve2gr_w((__m128i)v, 2); break;
+		default: ival = __lsx_vpickve2gr_w((__m128i)v, 3); break;
 		}
 		float fval;
 		memcpy(&fval, &ival, sizeof(float));
@@ -1671,15 +1675,38 @@ struct Vec4F32 {
 		return temp;
 	}
 
+	static Vec4F32 LoadConvertU8(const uint8_t *src) {  // Note: will load 8 bytes, not 4. Only the first 4 bytes will be used.
+		Vec4F32 temp;
+		for (int i = 0; i < 4; i++) {
+			temp.v[i] = (float)src[i];
+		}
+		return temp;
+	}
+
+	// Truncates towards zero and saturates to 0-255, like the packing instructions the other
+	// implementations use.
+	void StoreConvertToU8(uint8_t *dest) {
+		for (int i = 0; i < 4; i++) {
+			int value = (int)v[i];
+			if (value < 0) value = 0;
+			if (value > 255) value = 255;
+			dest[i] = (uint8_t)value;
+		}
+	}
+
 	static Vec4F32 LoadF24x3_One(const uint32_t *src) {
-		uint32_t shifted[4] = { src[0] << 8, src[1] << 8, src[2] << 8, 0 };
+		constexpr uint32_t kOneF32Bits = 0x3F800000;
+		uint32_t shifted[4] = { src[0] << 8, src[1] << 8, src[2] << 8, kOneF32Bits };
 		Vec4F32 temp;
 		memcpy(temp.v, shifted, sizeof(temp.v));
 		return temp;
 	}
 
 	static Vec4F32 LoadF24x4(const uint32_t *src) {
-		return LoadR24x3_One(src);
+		uint32_t shifted[4] = { src[0] << 8, src[1] << 8, src[2] << 8, src[3] << 8 };
+		Vec4F32 temp;
+		memcpy(temp.v, shifted, sizeof(temp.v));
+		return temp;
 	}
 
 	static Vec4F32 FromVec4S32(Vec4S32 src) {
@@ -1695,7 +1722,7 @@ struct Vec4F32 {
 	Vec4F32 ZeroNaNs() const {
 		Vec4F32 temp;
 		for (int i = 0; i < 4; i++) {
-			temp.v[i] = isnan(v[i]) ? 0.0f : v[i];
+			temp.v[i] = std::isnan(v[i]) ? 0.0f : v[i];
 		}
 		return temp;
 	}
@@ -1703,7 +1730,7 @@ struct Vec4F32 {
 	Vec4F32 CleanNaNInfs() {
 		Vec4F32 temp;
 		for (int i = 0; i < 4; i++) {
-			temp.v[i] = (isnan(v[i]) || isinf(v[i])) ? 0.0f : v[i];
+			temp.v[i] = (std::isnan(v[i]) || std::isinf(v[i])) ? 0.0f : v[i];
 		}
 		return temp;
 	}
@@ -1798,6 +1825,10 @@ struct Vec4F32 {
 		return Vec4F32{ { v[0], v[1], v[2], 1.0f } };
 	}
 
+	Vec4F32 WithLane3From(Vec4F32 other) const {
+		return Vec4F32{ { v[0], v[1], v[2], other.v[3] } };
+	}
+
 	Vec4S32 CompareEq(Vec4F32 other) const {
 		Vec4S32 temp;
 		for (int i = 0; i < 4; i++) {
@@ -1857,6 +1888,15 @@ struct Vec4F32 {
 		return Vec4F32{{ v[3], v[3], v[3], v[3] }};
 	}
 
+	// Loads four rows of four floats and transposes them into columns.
+	static void LoadTranspose(const float *src, Vec4F32 &col0, Vec4F32 &col1, Vec4F32 &col2, Vec4F32 &col3) {
+		col0 = Vec4F32::Load(src);
+		col1 = Vec4F32::Load(src + 4);
+		col2 = Vec4F32::Load(src + 8);
+		col3 = Vec4F32::Load(src + 12);
+		Transpose(col0, col1, col2, col3);
+	}
+
 	// In-place transpose.
 	static void Transpose(Vec4F32 &col0, Vec4F32 &col1, Vec4F32 &col2, Vec4F32 &col3) {
 		float m[16];
@@ -1919,6 +1959,10 @@ inline bool AllCompareBitsSet(Vec4S32 value) {
 	return true;
 }
 
+inline bool AnyCompareBitsSet(Vec4S32 value) {
+	return value.v[0] != 0 || value.v[1] != 0 || value.v[2] != 0 || value.v[3] != 0;
+}
+
 struct Vec4U16 {
 	uint16_t v[4];  // 64 bits.
 
diff --git a/Common/RiscVEmitter.cpp b/Common/RiscVEmitter.cpp
index 17dda6ffed..8232eb1d51 100644
--- a/Common/RiscVEmitter.cpp
+++ b/Common/RiscVEmitter.cpp
@@ -842,7 +842,7 @@ static inline u32 EncodeFVF(RiscVReg vd, RiscVReg rs1, RiscVReg vs2, VUseMask vm
 
 static inline u16 EncodeCR(Opcode16 op, RiscVReg rs2, RiscVReg rd, Funct4 funct4) {
 	_assert_msg_(SupportsCompressed(), "Compressed instructions unsupported");
-	return (u16)op | ((u16)rs2 << 2) | ((u16)rd << 7) | ((u16)funct4 << 12);
+	return (u16)op | ((u16)DecodeReg(rs2) << 2) | ((u16)DecodeReg(rd) << 7) | ((u16)funct4 << 12);
 }
 
 static inline u16 EncodeCI(Opcode16 op, u8 uimm6, RiscVReg rd, Funct3 funct3) {
@@ -850,13 +850,13 @@ static inline u16 EncodeCI(Opcode16 op, u8 uimm6, RiscVReg rd, Funct3 funct3) {
 	_assert_msg_(uimm6 <= 0x3F, "CI immediate overflow: %04x", uimm6);
 	u16 imm4_0 = ImmBits16(uimm6, 0, 5);
 	u16 imm5 = ImmBit16(uimm6, 5);
-	return (u16)op | (imm4_0 << 2) | ((u16)rd << 7) | (imm5 << 12) | ((u16)funct3 << 13);
+	return (u16)op | (imm4_0 << 2) | ((u16)DecodeReg(rd) << 7) | (imm5 << 12) | ((u16)funct3 << 13);
 }
 
 static inline u16 EncodeCSS(Opcode16 op, RiscVReg rs2, u8 uimm6, Funct3 funct3) {
 	_assert_msg_(SupportsCompressed(), "Compressed instructions unsupported");
 	_assert_msg_(uimm6 <= 0x3F, "CI immediate overflow: %04x", uimm6);
-	return (u16)op | ((u16)rs2 << 2) | ((u16)uimm6 << 7) | ((u16)funct3 << 13);
+	return (u16)op | ((u16)DecodeReg(rs2) << 2) | ((u16)uimm6 << 7) | ((u16)funct3 << 13);
 }
 
 static inline u16 EncodeCIW(Opcode16 op, RiscCReg rd, u8 uimm8, Funct3 funct3) {
diff --git a/android/jni/Android.mk b/android/jni/Android.mk
index 2e0a7ab5a9..b308bc058d 100644
--- a/android/jni/Android.mk
+++ b/android/jni/Android.mk
@@ -1051,6 +1051,7 @@ ifeq ($(UNITTEST),1)
 	$(SRC)/unittest/TestX64Emitter.cpp \
     $(SRC)/unittest/TestRiscVEmitter.cpp \
     $(SRC)/unittest/TestLoongArch64Emitter.cpp \
+    $(SRC)/unittest/TestCrossSIMD.cpp \
     $(SRC)/unittest/TestIRPassSimplify.cpp \
     $(SRC)/unittest/TestShaderGenerators.cpp \
     $(SRC)/unittest/TestSoftwareGPUJit.cpp \
diff --git a/b.sh b/b.sh
index a7683d0fb0..ff80e9bdd1 100755
--- a/b.sh
+++ b/b.sh
@@ -31,9 +31,14 @@ do
 			CMAKE_ARGS="-DCMAKE_TOOLCHAIN_FILE=cmake/Toolchains/raspberry.armv8.cmake ${CMAKE_ARGS}"
 			;;
 		--loongarch64)
-			CMAKE_ARGS="-DCMAKE_TOOLCHAIN_FILE=cmake/Toolchains/loongarch64-linux-gnu.cmake -DHEADLESS=ON -DUSE_SYSTEM_LIBPNG=OFF -DUSE_SYSTEM_LIBSDL2=OFF ${CMAKE_ARGS}"
+			CMAKE_ARGS="-DCMAKE_TOOLCHAIN_FILE=cmake/Toolchains/loongarch64-linux-gnu.cmake -DHEADLESS=ON -DHEADLESS_CROSS=ON -DUSE_SYSTEM_LIBPNG=OFF -DUSE_SYSTEM_LIBSDL2=OFF ${CMAKE_ARGS}"
 			TARGET_OS=loongarch64
-			LOONGARCH64_BUILD=1
+			CROSS_STUB_CC=loongarch64-linux-gnu-gcc-14
+			;;
+		--riscv64)
+			CMAKE_ARGS="-DCMAKE_TOOLCHAIN_FILE=cmake/Toolchains/riscv64-linux-gnu.cmake -DHEADLESS=ON -DHEADLESS_CROSS=ON -DUSE_SYSTEM_LIBPNG=OFF -DUSE_SYSTEM_LIBSDL2=OFF ${CMAKE_ARGS}"
+			TARGET_OS=riscv64
+			CROSS_STUB_CC=riscv64-linux-gnu-gcc-14
 			;;
 		--android) CMAKE_ARGS="-DCMAKE_TOOLCHAIN_FILE=android/android.toolchain.cmake ${CMAKE_ARGS}"
 			TARGET_OS=Android
@@ -122,28 +127,29 @@ echo Building with $CORES_COUNT threads
 
 mkdir -p ${BUILD_DIR}
 
-# For loongarch64 cross-compilation, build a comprehensive GL/GLX stub into
+# For the headless cross targets, build a comprehensive GL/GLX stub into
 # <build>/stublibs/libGL.so so GLEW's static archive can resolve its symbols
-# via PLT entries (a direct B26 branch to address 0 overflows on LoongArch).
+# via PLT entries (a direct branch to address 0 overflows the relocation).
 # This stub is always (re)generated to pick up any new needed symbols.
-if [ ! -z "$LOONGARCH64_BUILD" ]; then
+if [ ! -z "$CROSS_STUB_CC" ]; then
 	STUB_DIR=${BUILD_DIR}/stublibs
 	STUB_GL=${STUB_DIR}/libGL.so
 	mkdir -p "${STUB_DIR}"
 	STUB_C=$(mktemp /tmp/gl_stub_XXXXXX.c)
-	echo "/* Loongarch64 GL/GLX stub - cross-compilation only */" > "$STUB_C"
+	echo "/* ${TARGET_OS} GL/GLX stub - cross-compilation only */" > "$STUB_C"
 	# Collect all T (exported) symbols from GL/GLX libs, deduplicate, emit stubs
+	HOST_MULTIARCH=$(gcc -print-multiarch 2>/dev/null || dpkg-architecture -qDEB_HOST_MULTIARCH 2>/dev/null)
 	{
-		for lib in /usr/lib/x86_64-linux-gnu/libGL.so.1 \
-		           /usr/lib/x86_64-linux-gnu/libGLX.so.0 \
-		           /usr/lib/x86_64-linux-gnu/libGLdispatch.so.0; do
+		for lib in /usr/lib/$HOST_MULTIARCH/libGL.so.1 \
+		           /usr/lib/$HOST_MULTIARCH/libGLX.so.0 \
+		           /usr/lib/$HOST_MULTIARCH/libGLdispatch.so.0; do
 			[ -f "$lib" ] && nm -D "$lib" 2>/dev/null | awk '/^[0-9a-f]+ T /{ print $3 }'
 		done
 		# Always include the minimal GLX symbols GLEW directly references
 		printf '%s\n' glXGetProcAddressARB glXGetClientString glXQueryVersion \
 		              glBindTexture glGetString glGetIntegerv
 	} | sort -u | awk '{ print "void "$1"(void){}" }' >> "$STUB_C"
-	loongarch64-linux-gnu-gcc-14 -shared -fPIC -Wno-implicit-function-declaration \
+	$CROSS_STUB_CC -shared -fPIC -Wno-implicit-function-declaration \
 		-o "${STUB_GL}" "$STUB_C"
 	rm "$STUB_C"
 fi
diff --git a/cmake/Toolchains/loongarch64-linux-gnu.cmake b/cmake/Toolchains/loongarch64-linux-gnu.cmake
index b39ed83905..014630439c 100644
--- a/cmake/Toolchains/loongarch64-linux-gnu.cmake
+++ b/cmake/Toolchains/loongarch64-linux-gnu.cmake
@@ -17,8 +17,8 @@ set(USING_X11_VULKAN OFF CACHE BOOL "" FORCE)
 # Use legacy GL preference so find_package(OpenGL) looks for a single libGL.so
 # rather than the GLVND split (libOpenGL.so + libGLX.so), which we don't have.
 # A stub libGL.so + GL headers must be present in the sysroot; they are placed
-# there by cmake/scripts/setup-loongarch64-cross.sh (run once with sudo, or
-# automatically by the CI install step).
+# there by cmake/scripts/setup-cross.sh (run once with sudo, or automatically
+# by the CI install step).
 set(OpenGL_GL_PREFERENCE LEGACY CACHE STRING "" FORCE)
 
 # b.sh also creates a stub in <build-dir>/stublibs/ as a local fallback so the
@@ -27,14 +27,18 @@ set(CMAKE_EXE_LINKER_FLAGS
     "${CMAKE_EXE_LINKER_FLAGS} -L${CMAKE_BINARY_DIR}/stublibs"
     CACHE STRING "" FORCE)
 
-# Allow CMake's compile-check programs to run via QEMU.
-# The /lib64 symlink is created by the setup script; without it, pass -L
-# explicitly so QEMU can find the loongarch64 dynamic linker.
-if(EXISTS "/lib64/ld-linux-loongarch-lp64d.so.1")
-    set(CMAKE_CROSSCOMPILING_EMULATOR "qemu-loongarch64-static"
-        CACHE STRING "" FORCE)
-else()
-    set(CMAKE_CROSSCOMPILING_EMULATOR
-        "qemu-loongarch64-static;-L;/usr/loongarch64-linux-gnu"
-        CACHE STRING "" FORCE)
+# Allow CMake's compile-check programs to run via QEMU, if it's installed.
+# Debian/Ubuntu ship the static binaries in qemu-user-static and the dynamic
+# ones in qemu-user; either will do, so take whichever is present.
+find_program(QEMU_LOONGARCH64 NAMES qemu-loongarch64-static qemu-loongarch64)
+if(QEMU_LOONGARCH64)
+    # The /lib64 symlink is created by the setup script; without it, pass -L
+    # explicitly so QEMU can find the loongarch64 dynamic linker.
+    if(EXISTS "/lib64/ld-linux-loongarch-lp64d.so.1")
+        set(CMAKE_CROSSCOMPILING_EMULATOR "${QEMU_LOONGARCH64}" CACHE STRING "" FORCE)
+    else()
+        set(CMAKE_CROSSCOMPILING_EMULATOR
+            "${QEMU_LOONGARCH64};-L;/usr/loongarch64-linux-gnu"
+            CACHE STRING "" FORCE)
+    endif()
 endif()
diff --git a/cmake/Toolchains/riscv64-linux-gnu.cmake b/cmake/Toolchains/riscv64-linux-gnu.cmake
new file mode 100644
index 0000000000..a80150c6e8
--- /dev/null
+++ b/cmake/Toolchains/riscv64-linux-gnu.cmake
@@ -0,0 +1,44 @@
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR riscv64)
+
+set(CMAKE_C_COMPILER riscv64-linux-gnu-gcc-14)
+set(CMAKE_CXX_COMPILER riscv64-linux-gnu-g++-14)
+set(CMAKE_ASM_COMPILER riscv64-linux-gnu-gcc-14)
+
+set(CMAKE_FIND_ROOT_PATH /usr/riscv64-linux-gnu)
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
+
+# No X11 in the sysroot; disable the X11 Vulkan path to avoid include-dir errors.
+set(USING_X11_VULKAN OFF CACHE BOOL "" FORCE)
+
+# Use legacy GL preference so find_package(OpenGL) looks for a single libGL.so
+# rather than the GLVND split (libOpenGL.so + libGLX.so), which we don't have.
+# A stub libGL.so + GL headers must be present in the sysroot; they are placed
+# there by cmake/scripts/setup-cross.sh (run once with sudo, or automatically
+# by the CI install step).
+set(OpenGL_GL_PREFERENCE LEGACY CACHE STRING "" FORCE)
+
+# b.sh also creates a stub in <build-dir>/stublibs/ as a local fallback so the
+# linker can resolve GL/GLX calls from GLEW's static archive via PLT entries.
+set(CMAKE_EXE_LINKER_FLAGS
+    "${CMAKE_EXE_LINKER_FLAGS} -L${CMAKE_BINARY_DIR}/stublibs"
+    CACHE STRING "" FORCE)
+
+# Allow CMake's compile-check programs to run via QEMU, if it's installed.
+# Debian/Ubuntu ship the static binaries in qemu-user-static and the dynamic
+# ones in qemu-user; either will do, so take whichever is present.
+find_program(QEMU_RISCV64 NAMES qemu-riscv64-static qemu-riscv64)
+if(QEMU_RISCV64)
+    # The /lib64 symlink is created by the setup script; without it, pass -L
+    # explicitly so QEMU can find the riscv64 dynamic linker.
+    if(EXISTS "/lib64/ld-linux-riscv64-lp64d.so.1")
+        set(CMAKE_CROSSCOMPILING_EMULATOR "${QEMU_RISCV64}" CACHE STRING "" FORCE)
+    else()
+        set(CMAKE_CROSSCOMPILING_EMULATOR
+            "${QEMU_RISCV64};-L;/usr/riscv64-linux-gnu"
+            CACHE STRING "" FORCE)
+    endif()
+endif()
diff --git a/cmake/scripts/setup-loongarch64-cross.sh b/cmake/scripts/setup-cross.sh
similarity index 58%
rename from cmake/scripts/setup-loongarch64-cross.sh
rename to cmake/scripts/setup-cross.sh
index 978dbcebc1..ea9b8df86b 100755
--- a/cmake/scripts/setup-loongarch64-cross.sh
+++ b/cmake/scripts/setup-cross.sh
@@ -1,15 +1,34 @@
 #!/bin/bash
-# Sets up the loongarch64 cross-compilation environment.
-# Run once with sudo.  After this, ./b.sh --loongarch64 works on a fresh checkout.
+# Sets up a cross-compilation environment for one of the headless-only targets.
+# Run once with sudo.  After this, ./b.sh --<target> works on a fresh checkout.
+#
+# Usage: sudo cmake/scripts/setup-cross.sh <loongarch64|riscv64>
 
 set -e
 
-SYSROOT=/usr/loongarch64-linux-gnu
-CROSS_GCC=loongarch64-linux-gnu-gcc-14
+TARGET="$1"
+
+case "$TARGET" in
+    loongarch64)
+        TRIPLE=loongarch64-linux-gnu
+        LD_SO=ld-linux-loongarch-lp64d.so.1
+        ;;
+    riscv64)
+        TRIPLE=riscv64-linux-gnu
+        LD_SO=ld-linux-riscv64-lp64d.so.1
+        ;;
+    *)
+        echo "Usage: sudo $0 <loongarch64|riscv64>"
+        exit 1
+        ;;
+esac
+
+SYSROOT=/usr/$TRIPLE
+CROSS_GCC=$TRIPLE-gcc-14
 
 if ! command -v $CROSS_GCC &>/dev/null; then
     echo "Error: $CROSS_GCC not found. Install it first:"
-    echo "  sudo apt install gcc-14-loongarch64-linux-gnu g++-14-loongarch64-linux-gnu binutils-loongarch64-linux-gnu"
+    echo "  sudo apt install gcc-14-$TRIPLE g++-14-$TRIPLE binutils-$TRIPLE"
     exit 1
 fi
 
@@ -18,6 +37,8 @@ if [ "$(id -u)" -ne 0 ]; then
     exit 1
 fi
 
+HOST_MULTIARCH=$(gcc -print-multiarch 2>/dev/null || dpkg-architecture -qDEB_HOST_MULTIARCH 2>/dev/null)
+
 # ── OpenGL headers ───────────────────────────────────────────────────────────
 # GL headers are platform-independent C headers; copy from the host.
 echo "Installing GL headers into sysroot..."
@@ -27,22 +48,22 @@ for h in gl.h glext.h glcorearb.h glu.h; do
 done
 
 # ── GL stub library ──────────────────────────────────────────────────────────
-# Generate a loongarch64 libGL.so that exports all symbols from the host
+# Generate a libGL.so for the target that exports all symbols from the host
 # libGL so that GLEW's static archive can be linked via PLT entries (a direct
-# B26 branch to address 0 would overflow on LoongArch).
+# branch to address 0 would overflow the relocation on these targets).
 echo "Generating libGL.so stub..."
 
 HOST_GL=""
 for candidate in \
-        /usr/lib/x86_64-linux-gnu/libGL.so.1 \
-        /usr/lib/x86_64-linux-gnu/libGL.so \
+        /usr/lib/$HOST_MULTIARCH/libGL.so.1 \
+        /usr/lib/$HOST_MULTIARCH/libGL.so \
         /usr/lib/libGL.so.1 ; do
     [ -f "$candidate" ] && HOST_GL="$candidate" && break
 done
 
 STUB_C=$(mktemp /tmp/gl_stub_XXXXXX.c)
-echo "/* Loongarch64 GL/GLX stub - for cross-compilation only */" > "$STUB_C"
-echo "void __loongarch64_gl_placeholder(void) {}" >> "$STUB_C"
+echo "/* $TARGET GL/GLX stub - for cross-compilation only */" > "$STUB_C"
+echo "void __cross_gl_placeholder(void) {}" >> "$STUB_C"
 
 if [ -n "$HOST_GL" ]; then
     nm -D "$HOST_GL" 2>/dev/null | awk '/^[0-9a-f]+ T /{ print "void "$3"(void){}" }' >> "$STUB_C"
@@ -60,16 +81,16 @@ rm "$STUB_C"
 echo "  Installed to $SYSROOT/lib/libGL.so"
 
 # ── Dynamic linker symlink ───────────────────────────────────────────────────
-# Without this, binfmt_misc can't find the loongarch64 dynamic linker.
-LD_LINUX=$SYSROOT/lib/ld-linux-loongarch-lp64d.so.1
+# Without this, binfmt_misc can't find the target's dynamic linker.
+LD_LINUX=$SYSROOT/lib/$LD_SO
 if [ -f "$LD_LINUX" ]; then
     mkdir -p /lib64
-    ln -sf "$LD_LINUX" /lib64/ld-linux-loongarch-lp64d.so.1
-    echo "Created /lib64/ld-linux-loongarch-lp64d.so.1 -> $LD_LINUX"
+    ln -sf "$LD_LINUX" /lib64/$LD_SO
+    echo "Created /lib64/$LD_SO -> $LD_LINUX"
 fi
 
 echo ""
 echo "Setup complete."
-echo "Build:  ./b.sh --loongarch64"
-echo "Run:    qemu-loongarch64-static -L $SYSROOT <binary>"
-echo "  (after setup the /lib64 symlink lets you run loongarch64 binaries directly)"
+echo "Build:  ./b.sh --$TARGET"
+echo "Run:    qemu-$TARGET -L $SYSROOT <binary>"
+echo "  (after setup the /lib64 symlink lets you run $TARGET binaries directly)"
diff --git a/headless/Headless.cpp b/headless/Headless.cpp
index dfc5b25268..0bdd823435 100644
--- a/headless/Headless.cpp
+++ b/headless/Headless.cpp
@@ -289,10 +289,10 @@ static GraphicsContext *CreateGraphicsContext(GPUCore gpuCore, std::string **dev
 	default:
 		return nullptr;
 	}
-#elif PPSSPP_ARCH(LOONGARCH64)
-	// The loongarch64 cross-compilation toolchain has no SDL3 packages available (see the
-	// LOONGARCH64_DEVICE branch in CMakeLists.txt), so this build is compile-tested only and
-	// never actually needs to create a graphics context at runtime.
+#elif PPSSPP_ARCH(LOONGARCH64) || PPSSPP_ARCH(RISCV64)
+	// The loongarch64 and riscv64 cross-compilation sysroots have no SDL3 (see the HEADLESS_CROSS
+	// branch in CMakeLists.txt). These builds still run fine under qemu with --graphics=software,
+	// which needs no graphics context.
 	*deviceSetting = nullptr;
 	return nullptr;
 #elif PPSSPP_PLATFORM(ANDROID)
@@ -904,7 +904,7 @@ int main(int argc, const char* argv[]) {
 		// We don't bother with a window.
 		graphicsContext = new NullGraphicsContext();
 	} else {
-#if PPSSPP_PLATFORM(ANDROID) || PPSSPP_ARCH(LOONGARCH64)
+#if PPSSPP_PLATFORM(ANDROID) || PPSSPP_ARCH(LOONGARCH64) || PPSSPP_ARCH(RISCV64)
 		fprintf(stderr, "Headless graphics context creation is not supported on this platform.\n");
 		return 1;
 #else
@@ -1134,7 +1134,7 @@ int main(int argc, const char* argv[]) {
 		ShutdownWebServer();
 	}
 
-#if PPSSPP_PLATFORM(ANDROID) || PPSSPP_ARCH(LOONGARCH64)
+#if PPSSPP_PLATFORM(ANDROID) || PPSSPP_ARCH(LOONGARCH64) || PPSSPP_ARCH(RISCV64)
 	// ... see above
 #else
 	if (window) {
diff --git a/unittest/TestCrossSIMD.cpp b/unittest/TestCrossSIMD.cpp
new file mode 100644
index 0000000000..9ad066a315
--- /dev/null
+++ b/unittest/TestCrossSIMD.cpp
@@ -0,0 +1,551 @@
+// Copyright (c) 2012- PPSSPP Project.
+
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, version 2.0 or later versions.
+
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License 2.0 for more details.
+
+// A copy of the GPL 2.0 should have been included with the program.
+// If not, see http://www.gnu.org/licenses/
+
+// Official git repository and contact information can be found at
+// https://github.com/hrydgard/ppsspp and http://www.ppsspp.org/.
+
+// Tests for Common/Math/CrossSIMD.h.
+//
+// CrossSIMD has four independent implementations - SSE2, NEON, LoongArch LSX and a plain scalar
+// fallback - and only one of them is compiled on any given machine. So the point of these tests is
+// to check whatever got compiled against results worked out by hand, rather than against each other;
+// that way every architecture we build for verifies its own copy. Running the unit tests under
+// qemu covers the ones we have no hardware for.
+//
+// Only functions that all four implementations provide are tested here, otherwise this wouldn't
+// build everywhere.
+
+#include "ppsspp_config.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+
+#include "Common/Common.h"
+#include "Common/CommonTypes.h"
+#include "Common/Math/CrossSIMD.h"
+#include "unittest/UnitTest.h"
+
+static bool CompareFloats(const float *values, const float *known_good, int count, int line) {
+	int wrongCount = 0;
+	for (int i = 0; i < count; i++) {
+		if (values[i] != known_good[i]) {
+			wrongCount++;
+		}
+	}
+	if (wrongCount > 0) {
+		for (int i = 0; i < count; i++) {
+			bool wrong = values[i] != known_good[i];
+			printf("%d: %0.3f vs %0.3f %s\n", i + 1, values[i], known_good[i], wrong ? "!! MISMATCH" : "");
+		}
+		printf("At TestCrossSIMD.cpp:%d: %d / %d were wrong\n", line, wrongCount, count);
+		return false;
+	}
+	return true;
+}
+
+// For the approximate operations (reciprocals), which are allowed to differ between architectures.
+static bool CompareFloatsApprox(const float *values, const float *known_good, int count, float tolerance, int line) {
+	for (int i = 0; i < count; i++) {
+		const float diff = fabsf(values[i] - known_good[i]);
+		const float scale = fabsf(known_good[i]) > 1.0f ? fabsf(known_good[i]) : 1.0f;
+		if (diff / scale > tolerance) {
+			printf("At TestCrossSIMD.cpp:%d: lane %d: %0.6f vs %0.6f\n", line, i, values[i], known_good[i]);
+			return false;
+		}
+	}
+	return true;
+}
+
+static bool TestVec4S32() {
+	const int a_values[4] = { 3, -7, 0x4000, -1 };
+	const int b_values[4] = { 5, 11, -2, 0x7FFF };
+
+	Vec4S32 a = Vec4S32::Load(a_values);
+	Vec4S32 b = Vec4S32::Load(b_values);
+
+	int result[4];
+
+	(a + b).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], a_values[i] + b_values[i]);
+	}
+
+	(a - b).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], a_values[i] - b_values[i]);
+	}
+
+	a.Mul(b).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], a_values[i] * b_values[i]);
+	}
+
+	// Mul16 only promises correct results when both sides fit in 16 bits.
+	const int small_a[4] = { 3, -7, 300, -1 };
+	const int small_b[4] = { 5, 11, -2, 1000 };
+	Vec4S32 sa = Vec4S32::Load(small_a);
+	Vec4S32 sb = Vec4S32::Load(small_b);
+	sa.Mul16(sb).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], small_a[i] * small_b[i]);
+	}
+
+	Vec4S32::Splat(-5).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], -5);
+	}
+
+	Vec4S32::Zero().Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], 0);
+	}
+
+	a.Shl<2>().Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], a_values[i] << 2);
+	}
+
+	// operator[] should agree with Store.
+	a.Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(a[i], result[i]);
+	}
+
+	// Min16/Max16 likewise operate on 16-bit lanes.
+	sa.Min16(sb).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], small_a[i] < small_b[i] ? small_a[i] : small_b[i]);
+	}
+	sa.Max16(sb).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_INT(result[i], small_a[i] > small_b[i] ? small_a[i] : small_b[i]);
+	}
+
+	const int sext_values[4] = { 0x0000FFFF, 0x00007FFF, 0x00008000, 0x00000001 };
+	Vec4S32::Load(sext_values).SignExtend16().Store(result);
+	EXPECT_EQ_INT(result[0], -1);
+	EXPECT_EQ_INT(result[1], 32767);
+	EXPECT_EQ_INT(result[2], -32768);
+	EXPECT_EQ_INT(result[3], 1);
+
+	return true;
+}
+
+static bool TestVec4S32Compares() {
+	const int a_values[4] = { 1, 2, 3, 4 };
+	const int b_values[4] = { 1, 5, 0, 4 };
+	Vec4S32 a = Vec4S32::Load(a_values);
+	Vec4S32 b = Vec4S32::Load(b_values);
+
+	int result[4];
+
+	// Compares produce all-ones or all-zeros per lane.
+	a.CompareEq(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[0], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[1], 0u);
+	EXPECT_EQ_HEX((u32)result[2], 0u);
+	EXPECT_EQ_HEX((u32)result[3], 0xFFFFFFFFu);
+
+	a.CompareLt(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[0], 0u);
+	EXPECT_EQ_HEX((u32)result[1], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[2], 0u);
+	EXPECT_EQ_HEX((u32)result[3], 0u);
+
+	a.CompareGt(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[0], 0u);
+	EXPECT_EQ_HEX((u32)result[1], 0u);
+	EXPECT_EQ_HEX((u32)result[2], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[3], 0u);
+
+	// AllCompareBitsSet / AnyCompareBitsSet over those masks.
+	EXPECT_FALSE(AllCompareBitsSet(a.CompareEq(b)));
+	EXPECT_TRUE(AnyCompareBitsSet(a.CompareEq(b)));
+	EXPECT_TRUE(AllCompareBitsSet(a.CompareEq(a)));
+	EXPECT_FALSE(AnyCompareBitsSet(a.CompareLt(a)));
+
+	return true;
+}
+
+static bool TestVec4F32Arith() {
+	const float a_values[4] = { 1.0f, -2.0f, 3.5f, 0.25f };
+	const float b_values[4] = { 4.0f, 8.0f, -0.5f, 2.0f };
+
+	Vec4F32 a = Vec4F32::Load(a_values);
+	Vec4F32 b = Vec4F32::Load(b_values);
+
+	float result[4];
+	float expected[4];
+
+	(a + b).Store(result);
+	for (int i = 0; i < 4; i++) expected[i] = a_values[i] + b_values[i];
+	if (!CompareFloats(result, expected, 4, __LINE__)) return false;
+
+	(a - b).Store(result);
+	for (int i = 0; i < 4; i++) expected[i] = a_values[i] - b_values[i];
+	if (!CompareFloats(result, expected, 4, __LINE__)) return false;
+
+	(a * b).Store(result);
+	for (int i = 0; i < 4; i++) expected[i] = a_values[i] * b_values[i];
+	if (!CompareFloats(result, expected, 4, __LINE__)) return false;
+
+	a.Mul(2.0f).Store(result);
+	for (int i = 0; i < 4; i++) expected[i] = a_values[i] * 2.0f;
+	if (!CompareFloats(result, expected, 4, __LINE__)) return false;
+
+	a.Min(b).Store(result);
+	for (int i = 0; i < 4; i++) expected[i] = a_values[i] < b_values[i] ? a_values[i] : b_values[i];
+	if (!CompareFloats(result, expected, 4, __LINE__)) return false;
+
+	a.Max(b).Store(result);
+	for (int i = 0; i < 4; i++) expected[i] = a_values[i] > b_values[i] ? a_values[i] : b_values[i];
+	if (!CompareFloats(result, expected, 4, __LINE__)) return false;
+
+	a.Clamp(0.0f, 3.0f).Store(result);
+	static const float known_clamp[4] = { 1.0f, 0.0f, 3.0f, 0.25f };
+	if (!CompareFloats(result, known_clamp, 4, __LINE__)) return false;
+
+	// Recip is allowed to be approximate on some architectures, RecipApprox definitely is.
+	b.Recip().Store(result);
+	for (int i = 0; i < 4; i++) expected[i] = 1.0f / b_values[i];
+	if (!CompareFloatsApprox(result, expected, 4, 0.001f, __LINE__)) return false;
+
+	b.RecipApprox().Store(result);
+	if (!CompareFloatsApprox(result, expected, 4, 0.01f, __LINE__)) return false;
+
+	Vec4F32::Splat(1.5f).Store(result);
+	static const float known_splat[4] = { 1.5f, 1.5f, 1.5f, 1.5f };
+	if (!CompareFloats(result, known_splat, 4, __LINE__)) return false;
+
+	Vec4F32::Zero().Store(result);
+	static const float known_zero[4] = { 0.0f, 0.0f, 0.0f, 0.0f };
+	if (!CompareFloats(result, known_zero, 4, __LINE__)) return false;
+
+	// Dot products. Dot3 ignores lane 3.
+	EXPECT_EQ_FLOAT(a.Dot3(b), a_values[0] * b_values[0] + a_values[1] * b_values[1] + a_values[2] * b_values[2]);
+	EXPECT_EQ_FLOAT(a.Dot4(b), a_values[0] * b_values[0] + a_values[1] * b_values[1] + a_values[2] * b_values[2] + a_values[3] * b_values[3]);
+
+	// operator[] and GetLane<> should agree with Store.
+	a.Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(a[i], result[i]);
+	}
+	EXPECT_EQ_FLOAT(a.GetLane<0>(), a_values[0]);
+	EXPECT_EQ_FLOAT(a.GetLane<3>(), a_values[3]);
+
+	return true;
+}
+
+static bool TestVec4F32Compares() {
+	const float a_values[4] = { 1.0f, 2.0f, 3.0f, 4.0f };
+	const float b_values[4] = { 1.0f, 5.0f, 0.0f, 4.0f };
+	Vec4F32 a = Vec4F32::Load(a_values);
+	Vec4F32 b = Vec4F32::Load(b_values);
+
+	int result[4];
+
+	a.CompareEq(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[0], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[1], 0u);
+
+	a.CompareLt(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[1], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[2], 0u);
+
+	a.CompareGt(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[2], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[1], 0u);
+
+	a.CompareLe(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[0], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[2], 0u);
+
+	a.CompareGe(b).Store(result);
+	EXPECT_EQ_HEX((u32)result[0], 0xFFFFFFFFu);
+	EXPECT_EQ_HEX((u32)result[1], 0u);
+
+	return true;
+}
+
+static bool TestVec4F32Lanes() {
+	const float values[4] = { 1.0f, 2.0f, 3.0f, 4.0f };
+	const float other_values[4] = { 9.0f, 9.0f, 9.0f, 42.0f };
+	Vec4F32 v = Vec4F32::Load(values);
+	Vec4F32 other = Vec4F32::Load(other_values);
+
+	float result[4];
+
+	v.WithLane3Zero().Store(result);
+	static const float known_zero3[4] = { 1.0f, 2.0f, 3.0f, 0.0f };
+	if (!CompareFloats(result, known_zero3, 4, __LINE__)) return false;
+
+	v.WithLane3One().Store(result);
+	static const float known_one3[4] = { 1.0f, 2.0f, 3.0f, 1.0f };
+	if (!CompareFloats(result, known_one3, 4, __LINE__)) return false;
+
+	v.WithLane3From(other).Store(result);
+	static const float known_from3[4] = { 1.0f, 2.0f, 3.0f, 42.0f };
+	if (!CompareFloats(result, known_from3, 4, __LINE__)) return false;
+
+	v.ShuffleXXXX().Store(result);
+	static const float known_xxxx[4] = { 1.0f, 1.0f, 1.0f, 1.0f };
+	if (!CompareFloats(result, known_xxxx, 4, __LINE__)) return false;
+
+	v.ShuffleYYYY().Store(result);
+	static const float known_yyyy[4] = { 2.0f, 2.0f, 2.0f, 2.0f };
+	if (!CompareFloats(result, known_yyyy, 4, __LINE__)) return false;
+
+	v.ShuffleZZZZ().Store(result);
+	static const float known_zzzz[4] = { 3.0f, 3.0f, 3.0f, 3.0f };
+	if (!CompareFloats(result, known_zzzz, 4, __LINE__)) return false;
+
+	v.ShuffleWWWW().Store(result);
+	static const float known_wwww[4] = { 4.0f, 4.0f, 4.0f, 4.0f };
+	if (!CompareFloats(result, known_wwww, 4, __LINE__)) return false;
+
+	v.ShuffleXXYY().Store(result);
+	static const float known_xxyy[4] = { 1.0f, 1.0f, 2.0f, 2.0f };
+	if (!CompareFloats(result, known_xxyy, 4, __LINE__)) return false;
+
+	v.ShuffleZZWW().Store(result);
+	static const float known_zzww[4] = { 3.0f, 3.0f, 4.0f, 4.0f };
+	if (!CompareFloats(result, known_zzww, 4, __LINE__)) return false;
+
+	// Store2/Store3 must leave the rest of the destination alone.
+	float partial[4] = { -1.0f, -1.0f, -1.0f, -1.0f };
+	v.Store2(partial);
+	static const float known_store2[4] = { 1.0f, 2.0f, -1.0f, -1.0f };
+	if (!CompareFloats(partial, known_store2, 4, __LINE__)) return false;
+
+	float partial3[4] = { -1.0f, -1.0f, -1.0f, -1.0f };
+	v.Store3(partial3);
+	static const float known_store3[4] = { 1.0f, 2.0f, 3.0f, -1.0f };
+	if (!CompareFloats(partial3, known_store3, 4, __LINE__)) return false;
+
+	// Transpose of four columns.
+	static const float col_values[4][4] = {
+		{ 0.0f, 1.0f, 2.0f, 3.0f },
+		{ 4.0f, 5.0f, 6.0f, 7.0f },
+		{ 8.0f, 9.0f, 10.0f, 11.0f },
+		{ 12.0f, 13.0f, 14.0f, 15.0f },
+	};
+	Vec4F32 c0 = Vec4F32::Load(col_values[0]);
+	Vec4F32 c1 = Vec4F32::Load(col_values[1]);
+	Vec4F32 c2 = Vec4F32::Load(col_values[2]);
+	Vec4F32 c3 = Vec4F32::Load(col_values[3]);
+	Vec4F32::Transpose(c0, c1, c2, c3);
+	float transposed[16];
+	c0.Store(transposed);
+	c1.Store(transposed + 4);
+	c2.Store(transposed + 8);
+	c3.Store(transposed + 12);
+	for (int row = 0; row < 4; row++) {
+		for (int col = 0; col < 4; col++) {
+			EXPECT_EQ_FLOAT(transposed[row * 4 + col], col_values[col][row]);
+		}
+	}
+
+	// LoadTranspose should match a plain Load of each row followed by Transpose.
+	static const float flat[16] = {
+		0.0f, 1.0f, 2.0f, 3.0f,
+		4.0f, 5.0f, 6.0f, 7.0f,
+		8.0f, 9.0f, 10.0f, 11.0f,
+		12.0f, 13.0f, 14.0f, 15.0f,
+	};
+	Vec4F32 t0, t1, t2, t3;
+	Vec4F32::LoadTranspose(flat, t0, t1, t2, t3);
+	float loadTransposed[16];
+	t0.Store(loadTransposed);
+	t1.Store(loadTransposed + 4);
+	t2.Store(loadTransposed + 8);
+	t3.Store(loadTransposed + 12);
+	if (!CompareFloats(loadTransposed, transposed, 16, __LINE__)) return false;
+
+	return true;
+}
+
+static bool TestVec4F32NaNInf() {
+	const float inf = INFINITY;
+	const float nan = NAN;
+	const float values[4] = { 1.0f, nan, inf, -inf };
+	Vec4F32 v = Vec4F32::Load(values);
+
+	float result[4];
+
+	v.ZeroNaNs().Store(result);
+	EXPECT_EQ_FLOAT(result[0], 1.0f);
+	EXPECT_EQ_FLOAT(result[1], 0.0f);
+	// ZeroNaNs leaves infinities alone.
+	EXPECT_TRUE(std::isinf(result[2]));
+	EXPECT_TRUE(std::isinf(result[3]));
+
+	// The contract is just "a value that yields zero when multiplied by zero" - whatever is
+	// cheapest to get there. So the implementations legitimately differ on what a bad lane
+	// becomes: SSE2 clamps to +-FLT_MAX, NEON, LSX and the scalar fallback zero it. Check the
+	// property rather than the value, and that good lanes are left alone.
+	v.CleanNaNInfs().Store(result);
+	EXPECT_EQ_FLOAT(result[0], 1.0f);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(result[i] * 0.0f, 0.0f);
+	}
+
+	return true;
+}
+
+static bool TestVec4F32Loads() {
+	float result[4];
+
+	// LoadF24x3_One: three 24-bit values shifted up into floats, lane 3 forced to 1.0f.
+	const uint32_t f24_values[4] = { 0x3F8000 >> 0, 0x400000, 0x3F0000, 0x123456 };
+	Vec4F32::LoadF24x3_One(f24_values).Store(result);
+	for (int i = 0; i < 3; i++) {
+		float expected;
+		uint32_t bits = f24_values[i] << 8;
+		memcpy(&expected, &bits, 4);
+		EXPECT_EQ_FLOAT(result[i], expected);
+	}
+	EXPECT_EQ_FLOAT(result[3], 1.0f);
+
+	// LoadF24x4: same, but all four lanes come from the source.
+	Vec4F32::LoadF24x4(f24_values).Store(result);
+	for (int i = 0; i < 4; i++) {
+		float expected;
+		uint32_t bits = f24_values[i] << 8;
+		memcpy(&expected, &bits, 4);
+		EXPECT_EQ_FLOAT(result[i], expected);
+	}
+
+	// The normalizing loads. Some of these read 8 bytes, so give them room.
+	const int8_t s8_values[8] = { -1, -128, 127, 45, 0, 0, 0, 0 };
+	Vec4F32::LoadS8Norm(s8_values).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(result[i], (float)s8_values[i] / 128.0f);
+	}
+
+	const int16_t s16_values[8] = { -1, -32768, 32767, 1234, 0, 0, 0, 0 };
+	Vec4F32::LoadS16Norm(s16_values).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(result[i], (float)s16_values[i] / 32768.0f);
+	}
+
+	const uint8_t u8_values[8] = { 0, 255, 128, 7, 0, 0, 0, 0 };
+	Vec4F32::LoadU8Norm(u8_values).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_APPROX_EQ_FLOAT(result[i], (float)u8_values[i] / 255.0f);
+	}
+
+	// The converting (non-normalizing) loads.
+	Vec4F32::LoadConvertS16(s16_values).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(result[i], (float)s16_values[i]);
+	}
+
+	Vec4F32::LoadConvertS8(s8_values).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(result[i], (float)s8_values[i]);
+	}
+
+	Vec4F32::LoadConvertU8(u8_values).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(result[i], (float)u8_values[i]);
+	}
+
+	// StoreConvertToU8 truncates towards zero and saturates to 0-255.
+	const float to_u8_values[4] = { 0.0f, 255.0f, 12.75f, 200.0f };
+	uint8_t u8_out[4];
+	Vec4F32::Load(to_u8_values).StoreConvertToU8(u8_out);
+	EXPECT_EQ_INT(u8_out[0], 0);
+	EXPECT_EQ_INT(u8_out[1], 255);
+	EXPECT_EQ_INT(u8_out[2], 12);
+	EXPECT_EQ_INT(u8_out[3], 200);
+
+	const float saturate_values[4] = { -1.0f, 300.0f, -0.5f, 1000.0f };
+	Vec4F32::Load(saturate_values).StoreConvertToU8(u8_out);
+	EXPECT_EQ_INT(u8_out[0], 0);
+	EXPECT_EQ_INT(u8_out[1], 255);
+	EXPECT_EQ_INT(u8_out[2], 0);
+	EXPECT_EQ_INT(u8_out[3], 255);
+
+	// Int to float.
+	const int int_values[4] = { 0, -7, 1000, -32768 };
+	Vec4F32::FromVec4S32(Vec4S32::Load(int_values)).Store(result);
+	for (int i = 0; i < 4; i++) {
+		EXPECT_EQ_FLOAT(result[i], (float)int_values[i]);
+	}
+
+	// Load2 only fills the first two lanes.
+	const float two_values[2] = { 3.0f, 4.0f };
+	Vec4F32::Load2(two_values).Store(result);
+	EXPECT_EQ_FLOAT(result[0], 3.0f);
+	EXPECT_EQ_FLOAT(result[1], 4.0f);
+
+	return true;
+}
+
+static bool TestMatrices() {
+	static const float a_values[16] = { 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 7.0f, 8.0f, 9.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f };
+	static const float b_values[16] = { -12.0f, 3.0f, -2.5f, 5.0f, 31.0f, 0.5f, 4.0f, 6.0f, 7.0f, 13.0f, 12.0f, 51.0f, 81.0f, 32.0f };
+	static const float known_result[16] = { 395.0f, 171.0f, 41.5f, 170.0f, 942.0f, 410.5f, 111.5f, 475.0f, 1358.0f, 607.5f, 163.0f, 728.0f, 297.0f, 49.5f, 25.0f, 160.0f, };
+	float result[16];
+	Mat4F32 a(a_values);
+	Mat4F32 b(b_values);
+
+	Mul4x4By4x4(a, b).Store(result);
+	if (!CompareFloats(result, known_result, 16, __LINE__)) {
+		return false;
+	}
+
+	Mat4x3F32 d = Mat4x3F32(b_values + 2);
+	Mul4x3By4x4(d, a).Store(result);
+
+	static const float known_4x3_result[16] = { 332.5f, 371.0f, 404.5f, 438.0f, 80.5f, 95.0f, 105.5f, 116.0f, 192.0f, 237.0f, 269.0f, 301.0f, 790.0f, 1036.0f, 1185.0f, 1349.0f, };
+	if (!CompareFloats(result, known_4x3_result, 16, __LINE__)) {
+		return false;
+	}
+
+	static const float vec_values[4] = { 3.0f, 5.0f, 7.0f, 10000000.0f };
+	Vec4F32 v = Vec4F32::Load(vec_values);
+
+	v.AsVec3ByMatrix44(b).Store3(result);
+
+	static const float known_vec_result[3] = { 249.0f, 134.5f, 96.5f, };
+	if (!CompareFloats(result, known_vec_result, ARRAY_SIZE(known_vec_result), __LINE__)) {
+		return false;
+	}
+	Vec4F32 scale = Vec4F32::Load(a_values);
+	Vec4F32 translate = Vec4F32::Load(b_values);
+
+	TranslateAndScaleInplace(a, scale, translate);
+	a.Store(result);
+
+	static const float known_scale_result[16] = { -47.0f, 16.0f, -1.0f, 36.0f, -103.0f, 41.0f, 1.5f, 81.0f, -146.0f, 61.0f, 3.5f, 117.0f, 14.0f, 30.0f, 0.0f, 0.0f,};
+	if (!CompareFloats(result, known_scale_result, ARRAY_SIZE(known_scale_result), __LINE__)) {
+		return false;
+	}
+
+	return true;
+}
+
+bool TestCrossSIMD() {
+	if (!TestVec4S32()) return false;
+	if (!TestVec4S32Compares()) return false;
+	if (!TestVec4F32Arith()) return false;
+	if (!TestVec4F32Compares()) return false;
+	if (!TestVec4F32Lanes()) return false;
+	if (!TestVec4F32NaNInf()) return false;
+	if (!TestVec4F32Loads()) return false;
+	if (!TestMatrices()) return false;
+	return true;
+}
diff --git a/unittest/UnitTest.cpp b/unittest/UnitTest.cpp
index ebb9dc7cd7..383fe445b0 100644
--- a/unittest/UnitTest.cpp
+++ b/unittest/UnitTest.cpp
@@ -2764,88 +2764,6 @@ bool TestSIMD() {
 	return true;
 }
 
-static void PrintFloats(const float *f, int count) {
-	for (int i = 0; i < count; i++) {
-		printf("%.1ff, ", f[i]);
-	}
-	printf("\n");
-}
-
-static bool CompareFloats(const float *values, const float *known_good, int count, int line) {
-	int wrongCount = 0;
-
-	for (int i = 0; i < count; i++) {
-		if (values[i] != known_good[i]) {
-			wrongCount++;
-		}
-	}
-
-	if (wrongCount > 0) {
-		for (int i = 0; i < count; i++) {
-			bool wrong = values[i] != known_good[i];
-			printf("%d: %0.3f vs %0.3f %s\n", i + 1, values[i], known_good[i], wrong ? "!! MISMATCH" : "");
-		}
-		printf("At UnitTest.cpp:%d: %d / %d were wrong\n", line, wrongCount, count);
-		return false;
-	} else {
-		return true;
-	}
-}
-
-bool TestCrossSIMD() {
-	static const float a_values[16] = { 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 7.0f, 8.0f, 9.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f };
-	static const float b_values[16] = { -12.0f, 3.0f, -2.5f, 5.0f, 31.0f, 0.5f, 4.0f, 6.0f, 7.0f, 13.0f, 12.0f, 51.0f, 81.0f, 32.0f };
-	static const float known_result[16] = { 395.0f, 171.0f, 41.5f, 170.0f, 942.0f, 410.5f, 111.5f, 475.0f, 1358.0f, 607.5f, 163.0f, 728.0f, 297.0f, 49.5f, 25.0f, 160.0f, };
-	float result[16];
-	Mat4F32 a(a_values);
-	Mat4F32 b(b_values);
-
-	Mul4x4By4x4(a, b).Store(result);
-	if (!CompareFloats(result, known_result, 16, __LINE__)) {
-		return false;
-	}
-
-	Mat4x3F32 d = Mat4x3F32(b_values + 2);
-	Mul4x3By4x4(d, a).Store(result);
-
-	static const float known_4x3_result[16] = { 332.5f, 371.0f, 404.5f, 438.0f, 80.5f, 95.0f, 105.5f, 116.0f, 192.0f, 237.0f, 269.0f, 301.0f, 790.0f, 1036.0f, 1185.0f, 1349.0f, };
-	if (!CompareFloats(result, known_4x3_result, 16, __LINE__)) {
-		return false;
-	}
-
-	static const float vec_values[4] = { 3.0f, 5.0f, 7.0f, 10000000.0f };
-	Vec4F32 v = Vec4F32::Load(vec_values);
-
-	v.AsVec3ByMatrix44(b).Store3(result);
-
-	static const float known_vec_result[3] = { 249.0f, 134.5f, 96.5f, };
-	if (!CompareFloats(result, known_vec_result, ARRAY_SIZE(known_vec_result), __LINE__)) {
-		return false;
-	}
-	Vec4F32 scale = Vec4F32::Load(a_values);
-	Vec4F32 translate = Vec4F32::Load(b_values);
-
-	TranslateAndScaleInplace(a, scale, translate);
-	a.Store(result);
-
-	static const float known_scale_result[16] = { -47.0f, 16.0f, -1.0f, 36.0f, -103.0f, 41.0f, 1.5f, 81.0f, -146.0f, 61.0f, 3.5f, 117.0f, 14.0f, 30.0f, 0.0f, 0.0f,};
-	if (!CompareFloats(result, known_scale_result, ARRAY_SIZE(known_scale_result), __LINE__)) {
-		return false;
-	}
-
-	s8 values[4] = {-1, -128, 127, 45};
-	float fvalues[4];
-	Vec4F32::LoadS8Norm(values).Store(fvalues);
-	static const float known_s8norm_result[4] = {(float)values[0]/128.0f, (float)values[1]/128.0f, (float)values[2]/128.0f, (float)values[3]/128.0f,};
-	if (!CompareFloats(fvalues, known_s8norm_result, ARRAY_SIZE(known_s8norm_result), __LINE__)) {
-		return false;
-	}
-
-	// PrintFloats(result, 16);
-
-	return true;
-}
-
 bool TestVolumeFunc() {
 	for (int i = 0; i <= 20; i++) {
 		float mul = Volume10ToMultiplier(i);
@@ -3044,6 +2962,7 @@ struct TestItem {
 
 bool TestArmEmitter();
 bool TestArm64Emitter();
+bool TestCrossSIMD();
 bool TestX64Emitter();
 bool TestRiscVEmitter();
 bool TestLoongArch64Emitter();
diff --git a/unittest/UnitTests.vcxproj b/unittest/UnitTests.vcxproj
index 5793f24e1e..e77fb4ba3f 100644
--- a/unittest/UnitTests.vcxproj
+++ b/unittest/UnitTests.vcxproj
@@ -287,6 +287,7 @@
     <ClCompile Include="..\Windows\CaptureDevice.cpp" />
     <ClCompile Include="JitHarness.cpp" />
     <ClCompile Include="TestArm64Emitter.cpp" />
+    <ClCompile Include="TestCrossSIMD.cpp" />
     <ClCompile Include="TestIRPassSimplify.cpp" />
     <ClCompile Include="TestLoongArch64Emitter.cpp" />
     <ClCompile Include="TestRiscVEmitter.cpp" />
diff --git a/unittest/UnitTests.vcxproj.filters b/unittest/UnitTests.vcxproj.filters
index 37af8831d9..9d249d43dd 100644
--- a/unittest/UnitTests.vcxproj.filters
+++ b/unittest/UnitTests.vcxproj.filters
@@ -14,6 +14,7 @@
     <ClCompile Include="TestShaderGenerators.cpp" />
     <ClCompile Include="TestThreadManager.cpp" />
     <ClCompile Include="TestSoftwareGPUJit.cpp" />
+    <ClCompile Include="TestCrossSIMD.cpp" />
     <ClCompile Include="TestIRPassSimplify.cpp" />
     <ClCompile Include="TestDemangle.cpp" />
     <ClCompile Include="TestLzrc.cpp" />