[AArch64] Implement intrinsics for FP8 FCVT/FCVTN/BFCVT

llvm · Dec 10, 2024 · 65b749f · 65b749f
1 parent 24162bd
commit 65b749f
Show file tree

Hide file tree

Showing 7 changed files with 166 additions and 8 deletions.
diff --git a/clang/include/clang/Basic/arm_sve.td b/clang/include/clang/Basic/arm_sve.td
@@ -2436,6 +2436,12 @@ let SVETargetGuard = InvalidMode, SMETargetGuard = "sme2,fp8" in {
   // Convert from FP8 to deinterleaved half-precision/BFloat16 multi-vector
   def SVF1CVTL_X2 : Inst<"svcvtl1_{d}[_mf8]_x2_fpm",  "2~>", "bh", MergeNone, "aarch64_sve_fp8_cvtl1_x2",  [IsStreaming, SetsFPMR], []>;
   def SVF2CVTL_X2 : Inst<"svcvtl2_{d}[_mf8]_x2_fpm",  "2~>", "bh", MergeNone, "aarch64_sve_fp8_cvtl2_x2",  [IsStreaming, SetsFPMR], []>;
+
+  // Convert from single/half/bfloat multivector to FP8
+  def SVFCVT_X2 : Inst<"svcvt_mf8[_{d}_x2]_fpm", "~2>", "bh", MergeNone, "aarch64_sve_fp8_cvt_x2", [IsStreaming, SetsFPMR], []>;
+  def SVFCVT_X4 : Inst<"svcvt_mf8[_{d}_x4]_fpm", "~4>", "f",  MergeNone, "aarch64_sve_fp8_cvt_x4", [IsOverloadNone, IsStreaming, SetsFPMR], []>;
+  // interleaved
+  def SVFCVTN_X4 : Inst<"svcvtn_mf8[_{d}_x4]_fpm", "~4>", "f", MergeNone, "aarch64_sve_fp8_cvtn_x4", [IsOverloadNone, IsStreaming, SetsFPMR], []>;
 }
 
 let SVETargetGuard = "sve2p1", SMETargetGuard = "sme2" in {

diff --git a/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sme2_fp8_cvt.c b/clang/test/CodeGen/AArch64/fp8-intrinsics/acle_sme2_fp8_cvt.c
@@ -16,6 +16,70 @@
 #define SVE_ACLE_FUNC(A1,A2,A3) A1##A2##A3
 #endif
 
+// CHECK-LABEL: @test_cvt_f16_x2(
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x2.nxv8f16(<vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CPP-CHECK-LABEL: @_Z15test_cvt_f16_x213svfloat16x2_tm(
+// CPP-CHECK-NEXT:  entry:
+// CPP-CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x2.nxv8f16(<vscale x 8 x half> [[ZN_COERCE0:%.*]], <vscale x 8 x half> [[ZN_COERCE1:%.*]])
+// CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_cvt_f16_x2(svfloat16x2_t zn, fpm_t fpmr)  __arm_streaming {
+  return SVE_ACLE_FUNC(svcvt_mf8,_f16_x2,_fpm)(zn, fpmr);
+}
+
+// CHECK-LABEL: @test_cvt_f32_x4(
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x4(<vscale x 4 x float> [[ZN_COERCE0:%.*]], <vscale x 4 x float> [[ZN_COERCE1:%.*]], <vscale x 4 x float> [[ZN_COERCE2:%.*]], <vscale x 4 x float> [[ZN_COERCE3:%.*]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CPP-CHECK-LABEL: @_Z15test_cvt_f32_x413svfloat32x4_tm(
+// CPP-CHECK-NEXT:  entry:
+// CPP-CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x4(<vscale x 4 x float> [[ZN_COERCE0:%.*]], <vscale x 4 x float> [[ZN_COERCE1:%.*]], <vscale x 4 x float> [[ZN_COERCE2:%.*]], <vscale x 4 x float> [[ZN_COERCE3:%.*]])
+// CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_cvt_f32_x4(svfloat32x4_t zn, fpm_t fpmr)  __arm_streaming {
+  return SVE_ACLE_FUNC(svcvt_mf8,_f32_x4,_fpm)(zn, fpmr);
+}
+
+// CHECK-LABEL: @test_cvtn_f32_x4(
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.x4(<vscale x 4 x float> [[ZN_COERCE0:%.*]], <vscale x 4 x float> [[ZN_COERCE1:%.*]], <vscale x 4 x float> [[ZN_COERCE2:%.*]], <vscale x 4 x float> [[ZN_COERCE3:%.*]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CPP-CHECK-LABEL: @_Z16test_cvtn_f32_x413svfloat32x4_tm(
+// CPP-CHECK-NEXT:  entry:
+// CPP-CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.x4(<vscale x 4 x float> [[ZN_COERCE0:%.*]], <vscale x 4 x float> [[ZN_COERCE1:%.*]], <vscale x 4 x float> [[ZN_COERCE2:%.*]], <vscale x 4 x float> [[ZN_COERCE3:%.*]])
+// CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_cvtn_f32_x4(svfloat32x4_t zn, fpm_t fpmr)  __arm_streaming {
+  return SVE_ACLE_FUNC(svcvtn_mf8,_f32_x4,_fpm)(zn, fpmr);
+}
+
+// CHECK-LABEL: @test_cvt_bf16_x2(
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x2.nxv8bf16(<vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]])
+// CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+// CPP-CHECK-LABEL: @_Z16test_cvt_bf16_x214svbfloat16x2_tm(
+// CPP-CHECK-NEXT:  entry:
+// CPP-CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])
+// CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x2.nxv8bf16(<vscale x 8 x bfloat> [[ZN_COERCE0:%.*]], <vscale x 8 x bfloat> [[ZN_COERCE1:%.*]])
+// CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
+//
+svmfloat8_t test_cvt_bf16_x2(svbfloat16x2_t zn, fpm_t fpmr)  __arm_streaming {
+  return SVE_ACLE_FUNC(svcvt_mf8,_bf16_x2,_fpm)(zn, fpmr);
+}
+
 // CHECK-LABEL: @test_cvt1_f16_x2(
 // CHECK-NEXT:  entry:
 // CHECK-NEXT:    tail call void @llvm.aarch64.set.fpmr(i64 [[FPMR:%.*]])

diff --git a/clang/test/Sema/aarch64-fp8-intrinsics/acle_sme2_fp8_cvt.c b/clang/test/Sema/aarch64-fp8-intrinsics/acle_sme2_fp8_cvt.c
@@ -5,7 +5,8 @@
 #include <arm_sve.h>
 
 
-void test_features_sme2_fp8(svmfloat8_t zn, fpm_t fpmr) __arm_streaming {
+void test_features_sme2_fp8(svmfloat8_t zn, svfloat16x2_t znf16, svbfloat16x2_t znbf16,
+                            svfloat32x4_t znf32,  fpm_t fpmr) __arm_streaming {
     // expected-error@+1 {{'svcvtl1_f16_mf8_x2_fpm' needs target feature sme,sme2,fp8}}
     svcvtl1_f16_mf8_x2_fpm(zn, fpmr);
     // expected-error@+1 {{'svcvtl2_f16_mf8_x2_fpm' needs target feature sme,sme2,fp8}}
@@ -23,4 +24,13 @@ void test_features_sme2_fp8(svmfloat8_t zn, fpm_t fpmr) __arm_streaming {
     svcvt1_bf16_mf8_x2_fpm(zn, fpmr);
     // expected-error@+1 {{'svcvt2_bf16_mf8_x2_fpm' needs target feature sme,sme2,fp8}}
     svcvt2_bf16_mf8_x2_fpm(zn, fpmr);
+
+    // expected-error@+1 {{'svcvt_mf8_f16_x2_fpm' needs target feature sme,sme2,fp8}}
+    svcvt_mf8_f16_x2_fpm(znf16, fpmr);
+    // expected-error@+1 {{'svcvt_mf8_bf16_x2_fpm' needs target feature sme,sme2,fp8}}
+    svcvt_mf8_bf16_x2_fpm(znbf16, fpmr);
+    // expected-error@+1 {{'svcvt_mf8_f32_x4_fpm' needs target feature sme,sme2,fp8}}
+    svcvt_mf8_f32_x4_fpm(znf32, fpmr);
+    // expected-error@+1 {{'svcvtn_mf8_f32_x4_fpm' needs target feature sme,sme2,fp8}}
+    svcvtn_mf8_f32_x4_fpm(znf32, fpmr);
 }
diff --git a/llvm/include/llvm/IR/IntrinsicsAArch64.td b/llvm/include/llvm/IR/IntrinsicsAArch64.td
@@ -3812,6 +3812,7 @@ let TargetPrefix = "aarch64" in {
                             [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>,
                              LLVMVectorOfBitcastsToInt<0>, LLVMVectorOfBitcastsToInt<0>, LLVMVectorOfBitcastsToInt<0>, LLVMVectorOfBitcastsToInt<0>],
                             [IntrNoMem]>;
+
 }
 
 // SVE2.1 - ZIPQ1, ZIPQ2, UZPQ1, UZPQ2
@@ -3876,6 +3877,11 @@ let TargetPrefix = "aarch64" in {
                             [llvm_nxv16i8_ty],
                             [IntrReadMem, IntrInaccessibleMemOnly]>;
 
+  class SME2_FP8_CVT_Single_X4_Intrinsic
+    : DefaultAttrsIntrinsic<[llvm_nxv16i8_ty],
+                            [llvm_nxv4f32_ty, llvm_nxv4f32_ty, llvm_nxv4f32_ty, llvm_nxv4f32_ty],
+                            [IntrReadMem, IntrInaccessibleMemOnly]>;
+
   class SME_FP8_OuterProduct_Intrinsic
       : DefaultAttrsIntrinsic<[],
           [llvm_i32_ty,
@@ -3894,6 +3900,17 @@ let TargetPrefix = "aarch64" in {
   def int_aarch64_sve_fp8_cvtl1_x2 : SME2_FP8_CVT_X2_Single_Intrinsic;
   def int_aarch64_sve_fp8_cvtl2_x2 : SME2_FP8_CVT_X2_Single_Intrinsic;
 
+  //
+  // CVT to FP8 from half-precision/BFloat16/single-precision multi-vector
+  //
+  def int_aarch64_sve_fp8_cvt_x2
+    : DefaultAttrsIntrinsic<[llvm_nxv16i8_ty],
+                            [llvm_anyvector_ty, LLVMMatchType<0>],
+                            [IntrReadMem, IntrInaccessibleMemOnly]>;
+
+  def int_aarch64_sve_fp8_cvt_x4  : SME2_FP8_CVT_Single_X4_Intrinsic;
+  def int_aarch64_sve_fp8_cvtn_x4 : SME2_FP8_CVT_Single_X4_Intrinsic;
+
   // FP8 outer product
   def int_aarch64_sme_fp8_fmopa_za16 : SME_FP8_OuterProduct_Intrinsic;
   def int_aarch64_sme_fp8_fmopa_za32 : SME_FP8_OuterProduct_Intrinsic;

diff --git a/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td b/llvm/lib/Target/AArch64/AArch64SMEInstrInfo.td
@@ -954,10 +954,10 @@ defm F2CVTL_2ZZ_BtoH   : sme2p1_fp8_cvt_vector_vg2_single<"f2cvtl",  0b10, 0b1>;
 defm BF2CVT_2ZZ_BtoH   : sme2p1_fp8_cvt_vector_vg2_single<"bf2cvt",  0b11, 0b0>;
 defm BF2CVTL_2ZZ_BtoH  : sme2p1_fp8_cvt_vector_vg2_single<"bf2cvtl", 0b11, 0b1>;
 
-defm FCVT_Z2Z_HtoB  : sme2_fp8_cvt_vg2_single<"fcvt",   0b0>;
-defm BFCVT_Z2Z_HtoB : sme2_fp8_cvt_vg2_single<"bfcvt",  0b1>;
-defm FCVT_Z4Z_StoB  : sme2_fp8_cvt_vg4_single<"fcvt",   0b0>;
-defm FCVTN_Z4Z_StoB : sme2_fp8_cvt_vg4_single<"fcvtn",  0b1>;
+defm FCVT_Z2Z_HtoB  : sme2_fp8_cvt_vg2_single<"fcvt",  0b0, nxv8f16,  int_aarch64_sve_fp8_cvt_x2>;
+defm BFCVT_Z2Z_HtoB : sme2_fp8_cvt_vg2_single<"bfcvt", 0b1, nxv8bf16, int_aarch64_sve_fp8_cvt_x2>;
+defm FCVT_Z4Z_StoB  : sme2_fp8_cvt_vg4_single<"fcvt",  0b0, int_aarch64_sve_fp8_cvt_x4>;
+defm FCVTN_Z4Z_StoB : sme2_fp8_cvt_vg4_single<"fcvtn", 0b1, int_aarch64_sve_fp8_cvtn_x4>;
 
 defm FSCALE_2ZZ   : sme2_fp_sve_destructive_vector_vg2_single<"fscale", 0b0011000>;
 defm FSCALE_4ZZ   : sme2_fp_sve_destructive_vector_vg4_single<"fscale", 0b0011000>;

diff --git a/llvm/lib/Target/AArch64/SMEInstrFormats.td b/llvm/lib/Target/AArch64/SMEInstrFormats.td
@@ -2398,10 +2398,14 @@ multiclass sme2_cvt_vg2_single<string mnemonic, bits<5> op, ValueType out_vt,
 }
 
 // SME2 multi-vec FP8 down convert two registers
-multiclass sme2_fp8_cvt_vg2_single<string mnemonic, bit op> {
+multiclass sme2_fp8_cvt_vg2_single<string mnemonic, bit op, ValueType in_vt, SDPatternOperator intrinsic> {
   def NAME :  sme2_cvt_vg2_single<mnemonic, {op, 0b1000}, ZPR8, ZZ_h_mul_r>{
+    let mayLoad = 1;
+    let mayStore = 0;
     let Uses = [FPMR, FPCR];
   }
+  def : Pat<(nxv16i8 (intrinsic in_vt:$Zn1, in_vt:$Zn2)),
+            (!cast<Instruction>(NAME) (REG_SEQUENCE ZPR2Mul2, in_vt:$Zn1, zsub0, in_vt:$Zn2, zsub1))>;
 }
 
 class sme2_cvt_unpk_vector_vg2<bits<2>sz, bits<3> op, bit u, RegisterOperand first_ty,
@@ -2467,8 +2471,13 @@ multiclass sme2_int_cvt_vg4_single<string mnemonic, bits<3> op, SDPatternOperato
 }
 
 //SME2 multi-vec FP8 down convert four registers
-multiclass sme2_fp8_cvt_vg4_single<string mnemonic, bit N> {
- def _NAME : sme2_cvt_vg4_single<0b0, {0b00, N}, 0b0100, ZPR8, ZZZZ_s_mul_r, mnemonic>;
+multiclass sme2_fp8_cvt_vg4_single<string mnemonic, bit N, SDPatternOperator intrinsic> {
+ def NAME : sme2_cvt_vg4_single<0b0, {0b00, N}, 0b0100, ZPR8, ZZZZ_s_mul_r, mnemonic> {
+    let mayLoad = 1;
+    let mayStore = 0;
+    let Uses = [FPMR, FPCR];
+ }
+ def : SME2_Cvt_VG4_Pat<NAME, intrinsic, nxv16i8, nxv4f32>;
 }
 
 class sme2_unpk_vector_vg4<bits<2>sz, bit u, RegisterOperand first_ty,

diff --git a/llvm/test/CodeGen/AArch64/sme2-fp8-intrinsics-cvt.ll b/llvm/test/CodeGen/AArch64/sme2-fp8-intrinsics-cvt.ll
@@ -1,6 +1,58 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
 ; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme2,+fp8 -verify-machineinstrs -force-streaming < %s | FileCheck %s
 
+; FCVT / FCVTN / BFCVT
+
+define <vscale x 16 x i8> @fcvt_x2(<vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1) {
+; CHECK-LABEL: fcvt_x2:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    fcvt z0.b, { z0.h, z1.h }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x2.nxv8f16(<vscale x 8 x half> %zn0, <vscale x 8 x half> %zn1)
+  ret <vscale x 16 x i8> %res
+}
+
+define <vscale x 16 x i8> @fcvt_x4(<vscale x 4 x float> %zn0, <vscale x 4 x float> %zn1, <vscale x 4 x float> %zn2, <vscale x 4 x float> %zn3) {
+; CHECK-LABEL: fcvt_x4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    fcvt z0.b, { z0.s - z3.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x4(<vscale x 4 x float> %zn0, <vscale x 4 x float> %zn1,
+                                                              <vscale x 4 x float> %zn2, <vscale x 4 x float> %zn3)
+  ret <vscale x 16 x i8> %res
+}
+
+define <vscale x 16 x i8> @fcvtn(<vscale x 4 x float> %zn0, <vscale x 4 x float> %zn1, <vscale x 4 x float> %zn2, <vscale x 4 x float> %zn3) {
+; CHECK-LABEL: fcvtn:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z3
+; CHECK-NEXT:    fcvtn z0.b, { z0.s - z3.s }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvtn.x4(<vscale x 4 x float> %zn0, <vscale x 4 x float> %zn1,
+                                                               <vscale x 4 x float> %zn2, <vscale x 4 x float> %zn3)
+  ret <vscale x 16 x i8> %res
+}
+
+define <vscale x 16 x i8> @bfcvt(<vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1) {
+; CHECK-LABEL: bfcvt:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1
+; CHECK-NEXT:    bfcvt z0.b, { z0.h, z1.h }
+; CHECK-NEXT:    ret
+  %res = call <vscale x 16 x i8> @llvm.aarch64.sve.fp8.cvt.x2.nxv8f16(<vscale x 8 x bfloat> %zn0, <vscale x 8 x bfloat> %zn1)
+  ret <vscale x 16 x i8> %res
+}
+
 ; F1CVT / F2CVT
 
 define { <vscale x 8 x half>, <vscale x 8 x half> } @f1cvt(<vscale x 16 x i8> %zm) {