[AMDGPU] Split dot-insts feature
Differential Revision: https://reviews.llvm.org/D57971
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@353587 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/AMDGPU/AMDGPU.td b/lib/Target/AMDGPU/AMDGPU.td
index c67d203..e37fd0a 100644
--- a/lib/Target/AMDGPU/AMDGPU.td
+++ b/lib/Target/AMDGPU/AMDGPU.td
@@ -269,10 +269,16 @@
"Has v_fmac_f32 and v_xnor_b32 instructions"
>;
-def FeatureDotInsts : SubtargetFeature<"dot-insts",
- "HasDotInsts",
+def FeatureDot1Insts : SubtargetFeature<"dot1-insts",
+ "HasDot1Insts",
"true",
- "Has v_dot* instructions"
+ "Has v_dot4_i32_i8 and v_dot8_i32_i4 instructions"
+>;
+
+def FeatureDot2Insts : SubtargetFeature<"dot2-insts",
+ "HasDot2Insts",
+ "true",
+ "Has v_dot2_f32_f16, v_dot2_i32_i16, v_dot2_u32_u16, v_dot4_u32_u8, v_dot8_u32_u4 instructions"
>;
def FeatureSRAMECC : SubtargetFeature<"sram-ecc",
@@ -570,7 +576,8 @@
FeatureFmaMixInsts,
FeatureLDSBankCount32,
FeatureDLInsts,
- FeatureDotInsts,
+ FeatureDot1Insts,
+ FeatureDot2Insts,
FeatureSRAMECC,
FeatureCodeObjectV3]>;
@@ -769,8 +776,11 @@
def HasDLInsts : Predicate<"Subtarget->hasDLInsts()">,
AssemblerPredicate<"FeatureDLInsts">;
-def HasDotInsts : Predicate<"Subtarget->hasDotInsts()">,
- AssemblerPredicate<"FeatureDotInsts">;
+def HasDot1Insts : Predicate<"Subtarget->hasDot1Insts()">,
+ AssemblerPredicate<"FeatureDot1Insts">;
+
+def HasDot2Insts : Predicate<"Subtarget->hasDot2Insts()">,
+ AssemblerPredicate<"FeatureDot2Insts">;
def EnableLateCFGStructurize : Predicate<
diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index c08d1aa..cdfbfb5 100644
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -206,7 +206,8 @@
HasDPP(false),
HasR128A16(false),
HasDLInsts(false),
- HasDotInsts(false),
+ HasDot1Insts(false),
+ HasDot2Insts(false),
EnableSRAMECC(false),
FlatAddressSpace(false),
FlatInstOffsets(false),
diff --git a/lib/Target/AMDGPU/AMDGPUSubtarget.h b/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 5bc14f6..789fd2a 100644
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -332,7 +332,8 @@
bool HasDPP;
bool HasR128A16;
bool HasDLInsts;
- bool HasDotInsts;
+ bool HasDot1Insts;
+ bool HasDot2Insts;
bool EnableSRAMECC;
bool FlatAddressSpace;
bool FlatInstOffsets;
@@ -666,8 +667,12 @@
return HasDLInsts;
}
- bool hasDotInsts() const {
- return HasDotInsts;
+ bool hasDot1Insts() const {
+ return HasDot1Insts;
+ }
+
+ bool hasDot2Insts() const {
+ return HasDot2Insts;
}
bool isSRAMECCEnabled() const {
diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index d6abd18..17332ed 100644
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -8708,7 +8708,7 @@
EVT VT = N->getValueType(0);
SDLoc SL(N);
- if (!Subtarget->hasDotInsts() || VT != MVT::f32)
+ if (!Subtarget->hasDot2Insts() || VT != MVT::f32)
return SDValue();
// FMA((F32)S0.x, (F32)S1. x, FMA((F32)S0.y, (F32)S1.y, (F32)z)) ->
diff --git a/lib/Target/AMDGPU/VOP3PInstructions.td b/lib/Target/AMDGPU/VOP3PInstructions.td
index 3c3ac93..be17f13 100644
--- a/lib/Target/AMDGPU/VOP3PInstructions.td
+++ b/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -238,29 +238,39 @@
(AMDGPUmul_u24_oneuse (and i32:$src0, (i32 65535)),
(and i32:$src1, (i32 65535)))
),
- (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))
->;
+ (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
+ let SubtargetPredicate = !cast<VOP_Pseudo>(Inst).SubtargetPredicate;
+}
class SDot2Pat<Instruction Inst> : GCNPat <
(add (add_oneuse (AMDGPUmul_i24_oneuse (sra i32:$src0, (i32 16)),
(sra i32:$src1, (i32 16))), i32:$src2),
(AMDGPUmul_i24_oneuse (sext_inreg i32:$src0, i16),
(sext_inreg i32:$src1, i16))),
- (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))
->;
+ (Inst (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))> {
+ let SubtargetPredicate = !cast<VOP_Pseudo>(Inst).SubtargetPredicate;
+}
-let SubtargetPredicate = HasDotInsts in {
+let SubtargetPredicate = HasDot2Insts in {
def V_DOT2_F32_F16 : VOP3PInst<"v_dot2_f32_f16", VOP3_Profile<VOP_F32_V2F16_V2F16_F32>>;
def V_DOT2_I32_I16 : VOP3PInst<"v_dot2_i32_i16", VOP3_Profile<VOP_I32_V2I16_V2I16_I32>>;
def V_DOT2_U32_U16 : VOP3PInst<"v_dot2_u32_u16", VOP3_Profile<VOP_I32_V2I16_V2I16_I32>>;
-def V_DOT4_I32_I8 : VOP3PInst<"v_dot4_i32_i8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
def V_DOT4_U32_U8 : VOP3PInst<"v_dot4_u32_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
-def V_DOT8_I32_I4 : VOP3PInst<"v_dot8_i32_i4", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
def V_DOT8_U32_U4 : VOP3PInst<"v_dot8_u32_u4", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
+} // End SubtargetPredicate = HasDot2Insts
+
+let SubtargetPredicate = HasDot1Insts in {
+
+def V_DOT4_I32_I8 : VOP3PInst<"v_dot4_i32_i8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
+def V_DOT8_I32_I4 : VOP3PInst<"v_dot8_i32_i4", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_PACKED>>;
+
+} // End SubtargetPredicate = HasDot1Insts
+
multiclass DotPats<SDPatternOperator dot_op,
VOP3PInst dot_inst> {
+ let SubtargetPredicate = dot_inst.SubtargetPredicate in
def : GCNPat <
(dot_op (dot_inst.Pfl.Src0VT (VOP3PMods0 dot_inst.Pfl.Src0VT:$src0, i32:$src0_modifiers)),
(dot_inst.Pfl.Src1VT (VOP3PMods dot_inst.Pfl.Src1VT:$src1, i32:$src1_modifiers)),
@@ -280,12 +290,14 @@
def : SDot2Pat<V_DOT2_I32_I16>;
foreach Type = ["U", "I"] in
+ let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT4_"#Type#"32_"#Type#8).SubtargetPredicate in
def : GCNPat <
!cast<dag>(!foldl((i32 i32:$src2), [0, 1, 2, 3], lhs, y,
(add_oneuse lhs, (!cast<PatFrag>("Mul"#Type#"_Elt"#y) i32:$src0, i32:$src1)))),
(!cast<VOP3PInst>("V_DOT4_"#Type#"32_"#Type#8) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
foreach Type = ["U", "I"] in
+ let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).SubtargetPredicate in
def : GCNPat <
!cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
[1, 2, 3, 4, 5, 6, 7], lhs, y,
@@ -295,14 +307,13 @@
// Different variants of dot8 code-gen dag patterns are not generated through table-gen due to a huge increase
// in the compile time. Directly handle the pattern generated by the FE here.
foreach Type = ["U", "I"] in
+ let SubtargetPredicate = !cast<VOP_Pseudo>("V_DOT8_"#Type#"32_"#Type#4).SubtargetPredicate in
def : GCNPat <
!cast<dag>(!foldl((add_oneuse i32:$src2, (!cast<PatFrag>("Mul"#Type#"0_4bit") i32:$src0, i32:$src1)),
[7, 1, 2, 3, 4, 5, 6], lhs, y,
(NonACAdd_oneuse lhs, (!cast<PatFrag>("Mul"#Type#y#"_4bit") i32:$src0, i32:$src1)))),
(!cast<VOP3PInst>("V_DOT8_"#Type#"32_"#Type#4) (i32 8), $src0, (i32 8), $src1, (i32 8), $src2, (i1 0))>;
-} // End SubtargetPredicate = HasDotInsts
-
multiclass VOP3P_Real_vi<bits<10> op> {
def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.VI>,
VOP3Pe <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
@@ -351,14 +362,19 @@
}
-let SubtargetPredicate = HasDotInsts in {
+let SubtargetPredicate = HasDot2Insts in {
defm V_DOT2_F32_F16 : VOP3P_Real_vi <0x3a3>;
defm V_DOT2_I32_I16 : VOP3P_Real_vi <0x3a6>;
defm V_DOT2_U32_U16 : VOP3P_Real_vi <0x3a7>;
-defm V_DOT4_I32_I8 : VOP3P_Real_vi <0x3a8>;
defm V_DOT4_U32_U8 : VOP3P_Real_vi <0x3a9>;
-defm V_DOT8_I32_I4 : VOP3P_Real_vi <0x3aa>;
defm V_DOT8_U32_U4 : VOP3P_Real_vi <0x3ab>;
-} // End SubtargetPredicate = HasDotInsts
+} // End SubtargetPredicate = HasDot2Insts
+
+let SubtargetPredicate = HasDot1Insts in {
+
+defm V_DOT4_I32_I8 : VOP3P_Real_vi <0x3a8>;
+defm V_DOT8_I32_I4 : VOP3P_Real_vi <0x3aa>;
+
+} // End SubtargetPredicate = HasDot1Insts