diff --git a/bin/dev_hdd0/game/TEST12345/USRDIR/mandelbrot_set_no_pads.self b/bin/dev_hdd0/game/TEST12345/USRDIR/mandelbrot_set_no_pads.self
new file mode 100644
index 0000000000..ab6694184e
Binary files /dev/null and b/bin/dev_hdd0/game/TEST12345/USRDIR/mandelbrot_set_no_pads.self differ
diff --git a/bin/dev_hdd0/game/TEST12345/USRDIR/mandelbrot_set_spu.self b/bin/dev_hdd0/game/TEST12345/USRDIR/mandelbrot_set_spu.self
new file mode 100644
index 0000000000..c2899f38b7
Binary files /dev/null and b/bin/dev_hdd0/game/TEST12345/USRDIR/mandelbrot_set_spu.self differ
diff --git a/rpcs3/Emu/Cell/SPUInterpreter.h b/rpcs3/Emu/Cell/SPUInterpreter.h
index f4e88eccdc..f6b7e9266f 100644
--- a/rpcs3/Emu/Cell/SPUInterpreter.h
+++ b/rpcs3/Emu/Cell/SPUInterpreter.h
@@ -191,18 +191,8 @@ private:
 	{
 		const u32 s = i7 & 0x3f;
 
-		for(u32 j = 0; j < 4; ++j)
-		{
-			const u32 t = CPU.GPR[ra]._u32[j];
-			u32 r = 0;
-
-			for(u32 b = 0; b + s < 32; ++b)
-			{
-				r |= t & (1 << (b + s));
-			}
-
-			CPU.GPR[rt]._u32[j] = r;
-		}
+		for (u32 j = 0; j < 4; ++j)
+			CPU.GPR[rt]._u32[j] = CPU.GPR[ra]._u32[j] << s;
 	}
 	void ROTHI(u32 rt, u32 ra, s32 i7)
 	{
@@ -1412,7 +1402,7 @@ private:
 			}
 		}
 	}
-	void MPYA(u32 rc, u32 ra, u32 rb, u32 rt)
+	void MPYA(u32 rt, u32 ra, u32 rb, u32 rc)
 	{
 		for (int w = 0; w < 4; w++)
 			CPU.GPR[rt]._i32[w] = CPU.GPR[ra]._i16[w*2] * CPU.GPR[rb]._i16[w*2] + CPU.GPR[rc]._i32[w];
@@ -1424,14 +1414,14 @@ private:
 		CPU.GPR[rt]._f[2] = CPU.GPR[rc]._f[2] - CPU.GPR[ra]._f[2] * CPU.GPR[rb]._f[2];
 		CPU.GPR[rt]._f[3] = CPU.GPR[rc]._f[3] - CPU.GPR[ra]._f[3] * CPU.GPR[rb]._f[3];
 	}
-	void FMA(u32 rc, u32 ra, u32 rb, u32 rt)
+	void FMA(u32 rt, u32 ra, u32 rb, u32 rc)
 	{
 		CPU.GPR[rt]._f[0] = CPU.GPR[ra]._f[0] * CPU.GPR[rb]._f[0] + CPU.GPR[rc]._f[0];
 		CPU.GPR[rt]._f[1] = CPU.GPR[ra]._f[1] * CPU.GPR[rb]._f[1] + CPU.GPR[rc]._f[1];
 		CPU.GPR[rt]._f[2] = CPU.GPR[ra]._f[2] * CPU.GPR[rb]._f[2] + CPU.GPR[rc]._f[2];
 		CPU.GPR[rt]._f[3] = CPU.GPR[ra]._f[3] * CPU.GPR[rb]._f[3] + CPU.GPR[rc]._f[3];
 	}
-	void FMS(u32 rc, u32 ra, u32 rb, u32 rt)
+	void FMS(u32 rt, u32 ra, u32 rb, u32 rc)
 	{
 		CPU.GPR[rt]._f[0] = CPU.GPR[ra]._f[0] * CPU.GPR[rb]._f[0] - CPU.GPR[rc]._f[0];
 		CPU.GPR[rt]._f[1] = CPU.GPR[ra]._f[1] * CPU.GPR[rb]._f[1] - CPU.GPR[rc]._f[1];
diff --git a/rpcs3/Emu/Cell/SPUThread.h b/rpcs3/Emu/Cell/SPUThread.h
index de0b0f17a8..e434b2b2cc 100644
--- a/rpcs3/Emu/Cell/SPUThread.h
+++ b/rpcs3/Emu/Cell/SPUThread.h
@@ -444,11 +444,11 @@ public:
 		case MFC_PUT_CMD:
 		case MFC_GET_CMD:
 		{
-			ConLog.Warning("DMA %s%s%s: lsa = 0x%x, ea = 0x%llx, tag = 0x%x, size = 0x%x, cmd = 0x%x", 
+			/* ConLog.Warning("DMA %s%s%s: lsa = 0x%x, ea = 0x%llx, tag = 0x%x, size = 0x%x, cmd = 0x%x", 
 				op & MFC_PUT_CMD ? "PUT" : "GET",
 				op & MFC_BARRIER_MASK ? "B" : "", 
 				op & MFC_FENCE_MASK ? "F" : "", 
-				lsa, ea, tag, size, cmd);
+				lsa, ea, tag, size, cmd); */
 			MFCArgs.CMDStatus.SetValue(dmac.Cmd(cmd, tag, lsa, ea, size));
 		}
 		break;