diff --git a/GPU/Software/SoftGpu.cpp b/GPU/Software/SoftGpu.cpp
index 3c218d2d4f..0d6c84791f 100644
--- a/GPU/Software/SoftGpu.cpp
+++ b/GPU/Software/SoftGpu.cpp
@@ -36,28 +36,6 @@
 #include "GPU/Software/Rasterizer.h"
 #include "GPU/Common/FramebufferCommon.h"
 
-// This is horrible, sorry.
-class Matrix4x4 {
-public:
-	union {
-		struct {
-			float xx, xy, xz, xw;
-			float yx, yy, yz, yw;
-			float zx, zy, zz, zw;
-			float wx, wy, wz, ww;
-		};
-		float m[16];
-	};
-
-	void setIdentity() {
-		memset(this, 0, 16 * sizeof(float));
-		xx = 1.0f;
-		yy = 1.0f;
-		zz = 1.0f;
-		ww = 1.0f;
-	}
-};
-
 const int FB_WIDTH = 480;
 const int FB_HEIGHT = 272;
 FormatBuffer fb;
@@ -220,9 +198,14 @@ void SoftGPU::CopyToCurrentFboFromDisplayRam(int srcwidth, int srcheight)
 
 	thin3d->SetTexture(0, fbTex);
 	Thin3DShaderSet *texColor = thin3d->GetShaderSetPreset(SS_TEXTURE_COLOR_2D);
-	Matrix4x4 identity;
-	identity.setIdentity();
-	texColor->SetMatrix4x4("WorldViewProj", identity);
+
+	static const float identity4x4[16] = {
+		1.0f, 0.0f, 0.0f, 0.0f,
+		0.0f, 1.0f, 0.0f, 0.0f,
+		0.0f, 0.0f, 1.0f, 0.0f,
+		0.0f, 0.0f, 0.0f, 1.0f,
+	};
+	texColor->SetMatrix4x4("WorldViewProj", identity4x4);
 	thin3d->DrawIndexed(T3DPrimitive::PRIM_TRIANGLES, texColor, vformat, vdata, idata, 6, 0);
 }
 
diff --git a/ext/native/gfx_es2/draw_buffer.cpp b/ext/native/gfx_es2/draw_buffer.cpp
index 527fee59d5..2989977430 100644
--- a/ext/native/gfx_es2/draw_buffer.cpp
+++ b/ext/native/gfx_es2/draw_buffer.cpp
@@ -80,7 +80,7 @@ void DrawBuffer::Flush(bool set_blend_state) {
 	if (count_ == 0)
 		return;
 
-	shaderSet_->SetMatrix4x4("WorldViewProj", drawMatrix_);
+	shaderSet_->SetMatrix4x4("WorldViewProj", drawMatrix_.getReadPtr());
 
 	if (vbuf_) {
 		vbuf_->SubData((const uint8_t *)verts_, 0, sizeof(Vertex) * count_);
diff --git a/ext/native/thin3d/thin3d.h b/ext/native/thin3d/thin3d.h
index bdf63c2120..5a71dfa528 100644
--- a/ext/native/thin3d/thin3d.h
+++ b/ext/native/thin3d/thin3d.h
@@ -282,7 +282,7 @@ class Thin3DShaderSet : public Thin3DObject {
 public:
 	// TODO: Make some faster way of doing these. Support uniform buffers (and fake them on GL 2.0?)
 	virtual void SetVector(const char *name, float *value, int n) = 0;
-	virtual void SetMatrix4x4(const char *name, const Matrix4x4 &value) = 0;
+	virtual void SetMatrix4x4(const char *name, const float value[16]) = 0;
 };
 
 struct T3DBlendStateDesc {
diff --git a/ext/native/thin3d/thin3d_d3d9.cpp b/ext/native/thin3d/thin3d_d3d9.cpp
index cc49f7d9c7..a4ede14352 100644
--- a/ext/native/thin3d/thin3d_d3d9.cpp
+++ b/ext/native/thin3d/thin3d_d3d9.cpp
@@ -242,7 +242,7 @@ public:
 		}
 	}
 	void SetVector(LPDIRECT3DDEVICE9 device, const char *name, float *value, int n);
-	void SetMatrix4x4(LPDIRECT3DDEVICE9 device, const char *name, const Matrix4x4 &value);
+	void SetMatrix4x4(LPDIRECT3DDEVICE9 device, const char *name, const float value[16]);
 
 private:
 	bool isPixelShader_;
@@ -258,7 +258,7 @@ public:
 	Thin3DDX9Shader *pshader;
 	void Apply(LPDIRECT3DDEVICE9 device);
 	void SetVector(const char *name, float *value, int n) { vshader->SetVector(device_, name, value, n); pshader->SetVector(device_, name, value, n); }
-	void SetMatrix4x4(const char *name, const Matrix4x4 &value) { vshader->SetMatrix4x4(device_, name, value); }  // pshaders don't usually have matrices
+	void SetMatrix4x4(const char *name, const float value[16]) { vshader->SetMatrix4x4(device_, name, value); }  // pshaders don't usually have matrices
 private:
 	LPDIRECT3DDEVICE9 device_;
 };
@@ -792,10 +792,10 @@ void Thin3DDX9Shader::SetVector(LPDIRECT3DDEVICE9 device, const char *name, floa
 	}
 }
 
-void Thin3DDX9Shader::SetMatrix4x4(LPDIRECT3DDEVICE9 device, const char *name, const Matrix4x4 &value) {
+void Thin3DDX9Shader::SetMatrix4x4(LPDIRECT3DDEVICE9 device, const char *name, const float value[16]) {
 	D3DXHANDLE handle = constantTable_->GetConstantByName(NULL, name);
 	if (handle) {
-		constantTable_->SetFloatArray(device, handle, value.getReadPtr(), 16);
+		constantTable_->SetFloatArray(device, handle, value, 16);
 	}
 }
 
diff --git a/ext/native/thin3d/thin3d_gl.cpp b/ext/native/thin3d/thin3d_gl.cpp
index 141ed0b49d..4d72ac85ce 100644
--- a/ext/native/thin3d/thin3d_gl.cpp
+++ b/ext/native/thin3d/thin3d_gl.cpp
@@ -327,7 +327,7 @@ public:
 	int GetUniformLoc(const char *name);
 
 	void SetVector(const char *name, float *value, int n) override;
-	void SetMatrix4x4(const char *name, const Matrix4x4 &value) override;
+	void SetMatrix4x4(const char *name, const float value[16]) override;
 
 	void GLLost() override {
 		vshader->Compile(vshader->GetSource().c_str());
@@ -835,11 +835,11 @@ void Thin3DGLShaderSet::SetVector(const char *name, float *value, int n) {
 	}
 }
 
-void Thin3DGLShaderSet::SetMatrix4x4(const char *name, const Matrix4x4 &value) {
+void Thin3DGLShaderSet::SetMatrix4x4(const char *name, const float value[16]) {
 	glUseProgram(program_);
 	int loc = GetUniformLoc(name);
 	if (loc != -1) {
-		glUniformMatrix4fv(loc, 1, false, value.getReadPtr());
+		glUniformMatrix4fv(loc, 1, false, value);
 	}
 }