[AArch64] Adopt updated B16B16 target flags #104602
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
llvmbot
added
clang
|
@llvm/pr-subscribers-clang-driver @llvm/pr-subscribers-backend-aarch64 Author: None (SpencerAbson) ChangesThe enablement of SVE/SME non-widening BFloat16 instructions was recently changed in response to an architecture update, in which:
New flags, 'sve-b16b16' and 'sme-b16b16' were introduced to replace the existing 'b16b16'. This was acheived in the below two patches.
Ideally, the interface change introduced here will be valid in LLVM-19. We do not see it necessary to back-port the entire change, but just to add 'sme-b16b16' and 'sve-b16b16' as aliases to the existing (and unchanged) 'b16b16' and 'sme2' flags which together cover all of these features. The predication of Bf16 variants of svmin/svminnm and svmax/svmaxnm is also fixed in this change. Full diff: https://github.com/llvm/llvm-project/pull/104602.diff 4 Files Affected:
diff --git a/clang/include/clang/Basic/arm_sve.td b/clang/include/clang/Basic/arm_sve.td
index 94c093d8911562..fb11d743fd6479 100644
--- a/clang/include/clang/Basic/arm_sve.td
+++ b/clang/include/clang/Basic/arm_sve.td
@@ -2116,7 +2116,7 @@ def SVFCLAMP_BF : SInst<"svclamp[_{d}]", "dddd", "b", MergeNone, "aarch64_sve_
multiclass MinMaxIntr<string i, string zm, string mul, string t> {
def SVS # NAME : SInst<"sv" # i # "[" # zm # "_{d}_" # mul # "]", t, "csil", MergeNone, "aarch64_sve_s" # i # zm # "_" # mul, [IsStreaming], []>;
def SVU # NAME : SInst<"sv" # i # "[" # zm # "_{d}_" # mul # "]", t, "UcUsUiUl", MergeNone, "aarch64_sve_u" # i # zm # "_" # mul, [IsStreaming], []>;
- def SVF # NAME : SInst<"sv" # i # "[" # zm # "_{d}_" # mul # "]", t, "bhfd", MergeNone, "aarch64_sve_f" # i # zm # "_" # mul, [IsStreaming], []>;
+ def SVF # NAME : SInst<"sv" # i # "[" # zm # "_{d}_" # mul # "]", t, "hfd", MergeNone, "aarch64_sve_f" # i # zm # "_" # mul, [IsStreaming], []>;
}
let SVETargetGuard = InvalidMode, SMETargetGuard = "sme2" in {
@@ -2134,11 +2134,11 @@ let SVETargetGuard = InvalidMode, SMETargetGuard = "sme2" in {
}
multiclass SInstMinMaxByVector<string name> {
- def NAME # _SINGLE_X2 : SInst<"sv" # name # "nm[_single_{d}_x2]", "22d", "bhfd", MergeNone, "aarch64_sve_f" # name # "nm_single_x2", [IsStreaming], []>;
- def NAME # _SINGLE_X4 : SInst<"sv" # name # "nm[_single_{d}_x4]", "44d", "bhfd", MergeNone, "aarch64_sve_f" # name # "nm_single_x4", [IsStreaming], []>;
+ def NAME # _SINGLE_X2 : SInst<"sv" # name # "nm[_single_{d}_x2]", "22d", "hfd", MergeNone, "aarch64_sve_f" # name # "nm_single_x2", [IsStreaming], []>;
+ def NAME # _SINGLE_X4 : SInst<"sv" # name # "nm[_single_{d}_x4]", "44d", "hfd", MergeNone, "aarch64_sve_f" # name # "nm_single_x4", [IsStreaming], []>;
- def NAME # _X2 : SInst<"sv" # name # "nm[_{d}_x2]", "222", "bhfd", MergeNone, "aarch64_sve_f" # name # "nm_x2", [IsStreaming], []>;
- def NAME # _X4 : SInst<"sv" # name # "nm[_{d}_x4]", "444", "bhfd", MergeNone, "aarch64_sve_f" # name # "nm_x4", [IsStreaming], []>;
+ def NAME # _X2 : SInst<"sv" # name # "nm[_{d}_x2]", "222", "hfd", MergeNone, "aarch64_sve_f" # name # "nm_x2", [IsStreaming], []>;
+ def NAME # _X4 : SInst<"sv" # name # "nm[_{d}_x4]", "444", "hfd", MergeNone, "aarch64_sve_f" # name # "nm_x4", [IsStreaming], []>;
}
let SVETargetGuard = InvalidMode, SMETargetGuard = "sme2" in {
@@ -2172,9 +2172,25 @@ let SVETargetGuard = InvalidMode, SMETargetGuard = "sme2" in {
def SVFCLAMP_X4 : SInst<"svclamp[_single_{d}_x4]", "44dd", "hfd", MergeNone, "aarch64_sve_fclamp_single_x4", [IsStreaming], []>;
}
+multiclass BfSingleMultiVector<string name> {
+ def NAME # _SINGLE_X2 : SInst<"sv" # name # "[_single_{d}_x2]", "22d", "b", MergeNone, "aarch64_sve_f" # name # "_single_x2", [IsStreaming], []>;
+ def NAME # _SINGLE_X4 : SInst<"sv" # name # "[_single_{d}_x4]", "44d", "b", MergeNone, "aarch64_sve_f" # name # "_single_x4", [IsStreaming], []>;
+
+ def NAME # _X2 : SInst<"sv" # name # "[_{d}_x2]", "222", "b", MergeNone, "aarch64_sve_f" # name # "_x2", [IsStreaming], []>;
+ def NAME # _X4 : SInst<"sv" # name # "[_{d}_x4]", "444", "b", MergeNone, "aarch64_sve_f" # name # "_x4", [IsStreaming], []>;
+}
+
let SVETargetGuard = InvalidMode, SMETargetGuard = "sme2,b16b16"in {
def SVBFCLAMP_X2 : SInst<"svclamp[_single_{d}_x2]", "22dd", "b", MergeNone, "aarch64_sve_bfclamp_single_x2", [IsStreaming], []>;
def SVBFCLAMP_X4 : SInst<"svclamp[_single_{d}_x4]", "44dd", "b", MergeNone, "aarch64_sve_bfclamp_single_x4", [IsStreaming], []>;
+
+ // bfmin, bfmax (single, multi)
+ defm SVBFMIN : BfSingleMultiVector<"min">;
+ defm SVBFMAX : BfSingleMultiVector<"max">;
+
+ // bfminnm, bfmaxnm (single, multi)
+ defm SVBFMINNM : BfSingleMultiVector<"minnm">;
+ defm SVBFMAXNM : BfSingleMultiVector<"maxnm">;
}
let SVETargetGuard = InvalidMode, SMETargetGuard = "sme2" in {
diff --git a/clang/test/Driver/print-supported-extensions-aarch64.c b/clang/test/Driver/print-supported-extensions-aarch64.c
index 6b969d50610f8b..cbd76a5062a594 100644
--- a/clang/test/Driver/print-supported-extensions-aarch64.c
+++ b/clang/test/Driver/print-supported-extensions-aarch64.c
@@ -55,6 +55,7 @@
// CHECK-NEXT: sha3 FEAT_SHA3, FEAT_SHA512 Enable SHA512 and SHA3 support
// CHECK-NEXT: sm4 FEAT_SM4, FEAT_SM3 Enable SM3 and SM4 support
// CHECK-NEXT: sme FEAT_SME Enable Scalable Matrix Extension (SME)
+// CHECK-NEXT: sme-b16b16 FEAT_SME_B16B16 Enable SME2.1 ZA-targeting non-widening BFloat16 instructions
// CHECK-NEXT: sme-f16f16 FEAT_SME_F16F16 Enable SME non-widening Float16 instructions
// CHECK-NEXT: sme-f64f64 FEAT_SME_F64F64 Enable Scalable Matrix Extension (SME) F64F64 instructions
// CHECK-NEXT: sme-f8f16 FEAT_SME_F8F16 Enable Scalable Matrix Extension (SME) F8F16 instructions
@@ -71,6 +72,7 @@
// CHECK-NEXT: ssve-fp8dot4 FEAT_SSVE_FP8DOT4 Enable SVE2 FP8 4-way dot product instructions
// CHECK-NEXT: ssve-fp8fma FEAT_SSVE_FP8FMA Enable SVE2 FP8 multiply-add instructions
// CHECK-NEXT: sve FEAT_SVE Enable Scalable Vector Extension (SVE) instructions
+// CHECK-NEXT: sve-b16b16 FEAT_SVE_B16B16 Enable SVE2 non-widening and SME2 Z-targeting non-widening BFloat16 instructions
// CHECK-NEXT: sve2 FEAT_SVE2 Enable Scalable Vector Extension 2 (SVE2) instructions
// CHECK-NEXT: sve2-aes FEAT_SVE_AES, FEAT_SVE_PMULL128 Enable AES SVE2 instructions
// CHECK-NEXT: sve2-bitperm FEAT_SVE_BitPerm Enable bit permutation SVE2 instructions
diff --git a/llvm/lib/Target/AArch64/AArch64Features.td b/llvm/lib/Target/AArch64/AArch64Features.td
index a1ae0873fc1902..e8ce88a6bd64f0 100644
--- a/llvm/lib/Target/AArch64/AArch64Features.td
+++ b/llvm/lib/Target/AArch64/AArch64Features.td
@@ -438,6 +438,15 @@ def FeatureSVE2p1: ExtensionWithMArch<"sve2p1", "SVE2p1", "FEAT_SVE2p1",
def FeatureB16B16 : ExtensionWithMArch<"b16b16", "B16B16", "FEAT_SVE_B16B16",
"Enable SVE2.1 or SME2.1 non-widening BFloat16 to BFloat16 instructions", [FeatureBF16]>;
+// FeatureSVEB16B16 and FeatureSMEB16B16 act as aliases for {FeatureB16B16}, and
+// {FeatureB16B16, FeatureSME2} respectively. This allows LLVM-20 interfacing programs
+// that use '+sve-b16b16' and '+sme-b16b16' to compile in LLVM-19.
+def FeatureSVEB16B16 : ExtensionWithMArch<"sve-b16b16", "SVEB16B16", "FEAT_SVE_B16B16",
+ "Enable SVE2 non-widening and SME2 Z-targeting non-widening BFloat16 instructions", [FeatureB16B16]>;
+
+def FeatureSMEB16B16 : ExtensionWithMArch<"sme-b16b16", "SMEB16B16", "FEAT_SME_B16B16",
+ "Enable SME2.1 ZA-targeting non-widening BFloat16 instructions", [FeatureSME2, FeatureB16B16]>;
+
def FeatureSMEF16F16 : ExtensionWithMArch<"sme-f16f16", "SMEF16F16", "FEAT_SME_F16F16",
"Enable SME non-widening Float16 instructions", [FeatureSME2]>;
diff --git a/llvm/unittests/TargetParser/TargetParserTest.cpp b/llvm/unittests/TargetParser/TargetParserTest.cpp
index 3d55b0309d26fd..9d08dd83684c90 100644
--- a/llvm/unittests/TargetParser/TargetParserTest.cpp
+++ b/llvm/unittests/TargetParser/TargetParserTest.cpp
@@ -2005,6 +2005,7 @@ TEST(TargetParserTest, AArch64ExtensionFeatures) {
AArch64::AEK_CPA, AArch64::AEK_PAUTHLR,
AArch64::AEK_TLBIW, AArch64::AEK_JSCVT,
AArch64::AEK_FCMA, AArch64::AEK_FP8,
+ AArch64::AEK_SMEB16B16, AArch64::AEK_SVEB16B16,
};
@@ -2043,6 +2044,7 @@ TEST(TargetParserTest, AArch64ExtensionFeatures) {
EXPECT_TRUE(llvm::is_contained(Features, "+sve2-sha3"));
EXPECT_TRUE(llvm::is_contained(Features, "+sve2-bitperm"));
EXPECT_TRUE(llvm::is_contained(Features, "+sve2p1"));
+ EXPECT_TRUE(llvm::is_contained(Features, "+sve-b16b16"));
EXPECT_TRUE(llvm::is_contained(Features, "+b16b16"));
EXPECT_TRUE(llvm::is_contained(Features, "+rcpc"));
EXPECT_TRUE(llvm::is_contained(Features, "+rand"));
@@ -2063,6 +2065,7 @@ TEST(TargetParserTest, AArch64ExtensionFeatures) {
EXPECT_TRUE(llvm::is_contained(Features, "+sme-f64f64"));
EXPECT_TRUE(llvm::is_contained(Features, "+sme-i16i64"));
EXPECT_TRUE(llvm::is_contained(Features, "+sme-f16f16"));
+ EXPECT_TRUE(llvm::is_contained(Features, "+sme-b16b16"));
EXPECT_TRUE(llvm::is_contained(Features, "+sme2"));
EXPECT_TRUE(llvm::is_contained(Features, "+sme2p1"));
EXPECT_TRUE(llvm::is_contained(Features, "+hbc"));
@@ -2188,6 +2191,7 @@ TEST(TargetParserTest, AArch64ArchExtFeature) {
{"lse", "nolse", "+lse", "-lse"},
{"rdm", "nordm", "+rdm", "-rdm"},
{"sve", "nosve", "+sve", "-sve"},
+ {"sve-b16b16", "nosve-b16b16", "+sve-b16b16", "-sve-b16b16"},
{"sve2", "nosve2", "+sve2", "-sve2"},
{"sve2-aes", "nosve2-aes", "+sve2-aes", "-sve2-aes"},
{"sve2-sm4", "nosve2-sm4", "+sve2-sm4", "-sve2-sm4"},
@@ -2212,6 +2216,7 @@ TEST(TargetParserTest, AArch64ArchExtFeature) {
{"sme-f64f64", "nosme-f64f64", "+sme-f64f64", "-sme-f64f64"},
{"sme-i16i64", "nosme-i16i64", "+sme-i16i64", "-sme-i16i64"},
{"sme-f16f16", "nosme-f16f16", "+sme-f16f16", "-sme-f16f16"},
+ {"sme-b16b16", "nosme-b16b16", "+sme-b16b16", "-sme-b16b16"},
{"sme2", "nosme2", "+sme2", "-sme2"},
{"sme2p1", "nosme2p1", "+sme2p1", "-sme2p1"},
{"hbc", "nohbc", "+hbc", "-hbc"},
@@ -2452,6 +2457,12 @@ AArch64ExtensionDependenciesBaseArchTestParams
{AArch64::ARMV8A, {"nobf16", "b16b16"}, {"bf16", "b16b16"}, {}},
{AArch64::ARMV8A, {"b16b16", "nobf16"}, {}, {"bf16", "b16b16"}},
+ // b16b16 -> {sve-b16b16, sme-b16b16}
+ {AArch64::ARMV8A, {"nob16b16", "sve-b16b16"}, {"b16b16", "sve-b16b16"}, {}},
+ {AArch64::ARMV8A, {"sve-b16b16", "nob16b16"}, {}, {"sve-b16b16", "b16b16"}},
+ {AArch64::ARMV8A, {"nob16b16", "sme-b16b16"}, {"b16b16", "sme-b16b16"}, {}},
+ {AArch64::ARMV8A, {"sme-b16b16", "nob16b16"}, {}, {"b16b16", "sme-b16b16"}},
+
// sve -> {sve2, f32mm, f64mm}
{AArch64::ARMV8A, {"nosve", "sve2"}, {"sve", "sve2"}, {}},
{AArch64::ARMV8A, {"sve2", "nosve"}, {}, {"sve", "sve2"}},
@@ -2491,7 +2502,7 @@ AArch64ExtensionDependenciesBaseArchTestParams
{AArch64::ARMV8A, {"sme-fa64", "nosme"}, {}, {"sme", "sme-fa64"}},
// sme2 -> {sme2p1, ssve-fp8fma, ssve-fp8dot2, ssve-fp8dot4, sme-f8f16,
- // sme-f8f32}
+ // sme-f8f32, sme-b16b16}
{AArch64::ARMV8A, {"nosme2", "sme2p1"}, {"sme2", "sme2p1"}, {}},
{AArch64::ARMV8A, {"sme2p1", "nosme2"}, {}, {"sme2", "sme2p1"}},
{AArch64::ARMV8A,
@@ -2522,6 +2533,8 @@ AArch64ExtensionDependenciesBaseArchTestParams
{AArch64::ARMV8A, {"sme-f8f16", "nosme2"}, {}, {"sme2", "sme-f8f16"}},
{AArch64::ARMV8A, {"nosme2", "sme-f8f32"}, {"sme2", "sme-f8f32"}, {}},
{AArch64::ARMV8A, {"sme-f8f32", "nosme2"}, {}, {"sme2", "sme-f8f32"}},
+ {AArch64::ARMV8A, {"nosme2", "sme-b16b16"}, {"sme2", "sme-b16b16"}, {}},
+ {AArch64::ARMV8A, {"sme-b16b16", "nosme2"}, {}, {"sme2", "sme-b16b16"}},
// fp8 -> {sme-f8f16, sme-f8f32}
{AArch64::ARMV8A, {"nofp8", "sme-f8f16"}, {"fp8", "sme-f8f16"}, {}},
|
sdesmalen-arm
approved these changes
Aug 16, 2024
The enablement of SVE/SME non-widening BFloat16 instructions was recently changed in response to an architecture update, in which: - FEAT_SVE_B16B16 was weakened - FEAT_SME_B16B16 was introduced New flags, 'sve-b16b16' and 'sme-b16b16' were introduced to replace the existing 'b16b16'. This was acheived in the below two patches. - llvm#101480 - llvm#102501 Ideally, the interface change introduced here will be valid in LLVM-19. We do not see it necessary to back-port the entire change, but just to add 'sme-b16b16' and 'sve-b16b16' as aliases to the existing (and unchanged) 'b16b16' and 'sme2' flags which together cover all of these features. The predication of Bf16 variants of svmin/svminnm and svmax/svmaxnm is also fixed in this change.
tru
force-pushed
the
adopt-b16b16-flags
branch
from
9c1a30b to
02cafa8
Compare
|
@SpencerAbson (or anyone else). If you would like to add a note about this fix in the release notes (completely optional). Please reply to this comment with a one or two sentence description of the fix. When you are done, please add the release:note label to this PR. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The enablement of SVE/SME non-widening BFloat16 instructions was recently changed in response to an architecture update, in which:
New flags, 'sve-b16b16' and 'sme-b16b16' were introduced to replace the existing 'b16b16'. This was acheived in the below two patches.
Ideally, the interface change introduced here will be valid in LLVM-19. We do not see it necessary to back-port the entire change, but just to add 'sme-b16b16' and 'sve-b16b16' as aliases to the existing (and unchanged) 'b16b16' and 'sme2' flags which together cover all of these features.
The predication of Bf16 variants of svmin/svminnm and svmax/svmaxnm is also fixed in this change.