float and int compare

squash! WIP: intel/compiler: Import Gen8 / Gen9 ALU machine description

1 files changed, 125 insertions, 0 deletions

diff --git a/src/intel/compiler/gen8_md.py b/src/intel/compiler/gen8_md.py
index 5605ef1498ae..bbd9aa7a61d4 100644
--- a/src/intel/compiler/gen8_md.py
+++ b/src/intel/compiler/gen8_md.py
@@ -167,6 +167,131 @@ gen8_md = [
                                       Instruction('ADD', r, r, b).predicate()))
     ),
 
+    # Comparisons
+    (('feq32', 'a@64', 'b@64'), InstructionList([(t0, DF),],
+                                                (Instruction('CMP', t0, a, b).cmod('Z'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('feq32', 'a@32', 'b@32'), Instruction('CMP', retype(r, F), a, b).cmod('Z')),
+    (('feq32', 'a@16', 'b@16'), InstructionList([(t0, HF),],
+                                                (Instruction('CMP', t0, a, b).cmod('Z'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+
+    (('fge32', 'a@64', 'b@64'), InstructionList([(t0, DF),],
+                                                (Instruction('CMP', t0, a, b).cmod('GE'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('fge32', 'a@32', 'b@32'), Instruction('CMP', retype(r, F), a, b).cmod('GE')),
+    (('fge32', 'a@16', 'b@16'), InstructionList([(t0, HF),],
+                                                (Instruction('CMP', t0, a, b).cmod('GE'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+
+    (('flt32', 'a@64', 'b@64'), InstructionList([(t0, DF),],
+                                                (Instruction('CMP', t0, a, b).cmod('L'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('flt32', 'a@32', 'b@32'), Instruction('CMP', retype(r, F), a, b).cmod('L')),
+    (('flt32', 'a@16', 'b@16'), InstructionList([(t0, HF),],
+                                                (Instruction('CMP', t0, a, b).cmod('L'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+
+    (('fne32', 'a@64', 'b@64'), InstructionList([(t0, DF),],
+                                                (Instruction('CMP', t0, a, b).cmod('NZ'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('fne32', 'a@32', 'b@32'), Instruction('CMP', retype(r, F), a, b).cmod('NZ')),
+    (('fne32', 'a@16', 'b@16'), InstructionList([(t0, HF),],
+                                                (Instruction('CMP', t0, a, b).cmod('NZ'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+
+    (('ieq32', 'a@64', 'b@64'), InstructionList([(t0, Q),],
+                                                (Instruction('CMP', t0, retype(a, Q), retype(b, Q)).cmod('Z'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('ieq32', 'a@32', 'b@32'), Instruction('CMP', retype(r, D), retype(a, D), retype(b, D)).cmod('Z')),
+    (('ieq32', 'a@16', 'b@16'), InstructionList([(t0, W),],
+                                                (Instruction('CMP', t0, retype(a, W), retype(b, W)).cmod('Z'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+    (('ieq32', 'a@8',  'b@8' ), InstructionList([(t0, B),],
+                                                (Instruction('CMP', t0, retype(a, B), retype(b, B)).cmod('Z'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, B))))
+    ),
+
+    (('ige32', 'a@64', 'b@64'), InstructionList([(t0, Q),],
+                                                (Instruction('CMP', t0, retype(a, Q), retype(b, Q)).cmod('GE'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('ige32', 'a@32', 'b@32'), Instruction('CMP', retype(r, D), retype(a, D), retype(b, D)).cmod('GE')),
+    (('ige32', 'a@16', 'b@16'), InstructionList([(t0, W),],
+                                                (Instruction('CMP', t0, retype(a, W), retype(b, W)).cmod('GE'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+    (('ige32', 'a@8',  'b@8' ), InstructionList([(t0, B),],
+                                                (Instruction('CMP', t0, retype(a, B), retype(b, B)).cmod('GE'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, B))))
+    ),
+
+    (('ilt32', 'a@64', 'b@64'), InstructionList([(t0, Q),],
+                                                (Instruction('CMP', t0, retype(a, Q), retype(b, Q)).cmod('L'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('ilt32', 'a@32', 'b@32'), Instruction('CMP', retype(r, D), retype(a, D), retype(b, D)).cmod('L')),
+    (('ilt32', 'a@16', 'b@16'), InstructionList([(t0, W),],
+                                                (Instruction('CMP', t0, retype(a, W), retype(b, W)).cmod('L'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+    (('ilt32', 'a@8',  'b@8' ), InstructionList([(t0, B),],
+                                                (Instruction('CMP', t0, retype(a, B), retype(b, B)).cmod('L'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, B))))
+    ),
+
+    (('uge32', 'a@64', 'b@64'), InstructionList([(t0, UQ),],
+                                                (Instruction('CMP', t0, retype(a, UQ), retype(b, UQ)).cmod('GE'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('uge32', 'a@32', 'b@32'), Instruction('CMP', retype(r, UD), retype(a, UD), retype(b, UD)).cmod('GE')),
+    (('uge32', 'a@16', 'b@16'), InstructionList([(t0, UW),],
+                                                (Instruction('CMP', t0, retype(a, UW), retype(b, UW)).cmod('GE'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+    (('uge32', 'a@8',  'b@8' ), InstructionList([(t0, UB),],
+                                                (Instruction('CMP', t0, retype(a, UB), retype(b, UB)).cmod('GE'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, B))))
+    ),
+
+    (('ult32', 'a@64', 'b@64'), InstructionList([(t0, UQ),],
+                                                (Instruction('CMP', t0, retype(a, UQ), retype(b, UQ)).cmod('L'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('ult32', 'a@32', 'b@32'), Instruction('CMP', retype(r, UD), retype(a, UD), retype(b, UD)).cmod('L')),
+    (('ult32', 'a@16', 'b@16'), InstructionList([(t0, UW),],
+                                                (Instruction('CMP', t0, retype(a, UW), retype(b, UW)).cmod('L'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+    (('ult32', 'a@8',  'b@8' ), InstructionList([(t0, UB),],
+                                                (Instruction('CMP', t0, retype(a, UB), retype(b, UB)).cmod('L'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, B))))
+    ),
+
+    (('ine32', 'a@64', 'b@64'), InstructionList([(t0, Q),],
+                                                (Instruction('CMP', t0, retype(a, Q), retype(b, Q)).cmod('NZ'),
+                                                 Instruction('MOV', r, subscript(t0, UD, 0))))
+    ),
+    (('ine32', 'a@32', 'b@32'), Instruction('CMP', retype(r, D), retype(a, D), retype(b, D)).cmod('NZ')),
+    (('ine32', 'a@16', 'b@16'), InstructionList([(t0, W),],
+                                                (Instruction('CMP', t0, retype(a, W), retype(b, W)).cmod('NZ'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, W))))
+    ),
+    (('ine32', 'a@8',  'b@8' ), InstructionList([(t0, B),],
+                                                (Instruction('CMP', t0, retype(a, B), retype(b, B)).cmod('NZ'),
+                                                 Instruction('MOV', retype(r, D), retype(t0, B))))
+    ),
+
     # 3-source arithmetic
     (('ffma', a, b, c), Instruction('MAD', r, c, b, a)),
     (('flrp', a, b, c), Instruction('LRP', r, c, b, a)),