X86: Implement the floating point media max microop.
authorGabe Black <gblack@eecs.umich.edu>
Tue, 18 Aug 2009 01:33:25 +0000 (18:33 -0700)
committerGabe Black <gblack@eecs.umich.edu>
Tue, 18 Aug 2009 01:33:25 +0000 (18:33 -0700)
src/arch/x86/isa/microops/mediaop.isa

index 15bae742a87cb7b99c39195e8837838d013c2c02..5cd7bd4f57f5554562b48605ede2522f194fc04b 100644 (file)
@@ -479,4 +479,54 @@ let {{
             }
             FpDestReg.uqw = result;
         '''
+
+    class Mmaxf(MediaOp):
+        code = '''
+            union floatInt
+            {
+                float f;
+                uint32_t i;
+            };
+            union doubleInt
+            {
+                double d;
+                uint64_t i;
+            };
+
+            assert(srcSize == destSize);
+            int size = srcSize;
+            int sizeBits = size * 8;
+            assert(srcSize == 4 || srcSize == 8);
+            int items = (ext & 0x1) ? 1: (sizeof(FloatRegBits) / size);
+            uint64_t result = FpDestReg.uqw;
+
+            for (int i = 0; i < items; i++) {
+                double arg1, arg2;
+                int hiIndex = (i + 1) * sizeBits - 1;
+                int loIndex = (i + 0) * sizeBits;
+                uint64_t arg1Bits = bits(FpSrcReg1.uqw, hiIndex, loIndex);
+                uint64_t arg2Bits = bits(FpSrcReg2.uqw, hiIndex, loIndex);
+
+                if (size == 4) {
+                    floatInt fi;
+                    fi.i = arg1Bits;
+                    arg1 = fi.f;
+                    fi.i = arg2Bits;
+                    arg2 = fi.f;
+                } else {
+                    doubleInt di;
+                    di.i = arg1Bits;
+                    arg1 = di.d;
+                    di.i = arg2Bits;
+                    arg2 = di.d;
+                }
+
+                if (arg1 > arg2) {
+                    result = insertBits(result, hiIndex, loIndex, arg1Bits);
+                } else {
+                    result = insertBits(result, hiIndex, loIndex, arg2Bits);
+                }
+            }
+            FpDestReg.uqw = result;
+        '''
 }};