[ARM,AArch64] Add intrinsics for dot product instructions

The ACLE spec which describes these intrinsics hasn't been published yet, but this is based on the final draft which will be published soon, and these have already been implemented by GCC. Differential revision: https://reviews.llvm.org/D46109 git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@331039 91177308-0d34-0410-b5e6-96231b3b80d8
anukulrm · Apr 27, 2018 · acc1201 · acc1201
1 parent 6769f8d
commit acc1201
Show file tree

Hide file tree

Showing 6 changed files with 238 additions and 0 deletions.
diff --git a/include/clang/Basic/arm_neon.td b/include/clang/Basic/arm_neon.td
@@ -199,6 +199,13 @@ def OP_SCALAR_HALF_SET_LNQ : Op<(bitcast "float16x8_t",
                                           (bitcast "int16_t", $p0),
                                           (bitcast "int16x8_t", $p1), $p2))>;
 
+def OP_DOT_LN
+    : Op<(call "vdot", $p0, $p1,
+          (bitcast $p1, (splat(bitcast "uint32x2_t", $p2), $p3)))>;
+def OP_DOT_LNQ
+    : Op<(call "vdot", $p0, $p1,
+          (bitcast $p1, (splat(bitcast "uint32x4_t", $p2), $p3)))>;
+
 //===----------------------------------------------------------------------===//
 // Instructions
 //===----------------------------------------------------------------------===//
@@ -1579,3 +1586,13 @@ let ArchGuard = "defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(__aarc
   def SCALAR_VDUP_LANEH  : IInst<"vdup_lane", "sdi", "Sh">;
   def SCALAR_VDUP_LANEQH : IInst<"vdup_laneq", "sji", "Sh">;
 }
+
+// v8.2-A dot product instructions.
+let ArchGuard = "defined(__ARM_FEATURE_DOTPROD)" in {
+  def DOT : SInst<"vdot", "dd88", "iQiUiQUi">;
+  def DOT_LANE : SOpInst<"vdot_lane", "dd87i", "iUiQiQUi", OP_DOT_LN>;
+}
+let ArchGuard = "defined(__ARM_FEATURE_DOTPROD) && defined(__aarch64__)" in {
+  // Variants indexing into a 128-bit vector are A64 only.
+  def UDOT_LANEQ : SOpInst<"vdot_laneq", "dd89i", "iUiQiQUi", OP_DOT_LNQ>;
+}
diff --git a/include/clang/Basic/arm_neon_incl.td b/include/clang/Basic/arm_neon_incl.td
@@ -253,6 +253,9 @@ def OP_UNAVAILABLE : Operation {
 // B,C,D: array of default elts, force 'Q' size modifier.
 // p: pointer type
 // c: const pointer type
+// 7: vector of 8-bit elements, ignore 'Q' size modifier
+// 8: vector of 8-bit elements, same width as default type
+// 9: vector of 8-bit elements, force 'Q' size modifier
 
 // Every intrinsic subclasses Inst.
 class Inst <string n, string p, string t, Operation o> {

diff --git a/lib/CodeGen/CGBuiltin.cpp b/lib/CodeGen/CGBuiltin.cpp
@@ -3867,6 +3867,8 @@ static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = {
   NEONMAP0(vcvtq_u16_v),
   NEONMAP0(vcvtq_u32_v),
   NEONMAP0(vcvtq_u64_v),
+  NEONMAP2(vdot_v, arm_neon_udot, arm_neon_sdot, 0),
+  NEONMAP2(vdotq_v, arm_neon_udot, arm_neon_sdot, 0),
   NEONMAP0(vext_v),
   NEONMAP0(vextq_v),
   NEONMAP0(vfma_v),
@@ -4061,6 +4063,8 @@ static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = {
   NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0),
   NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType),
+  NEONMAP2(vdot_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
+  NEONMAP2(vdotq_v, aarch64_neon_udot, aarch64_neon_sdot, 0),
   NEONMAP0(vext_v),
   NEONMAP0(vextq_v),
   NEONMAP0(vfma_v),
@@ -4974,6 +4978,14 @@ Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(
     }
     return SV;
   }
+  case NEON::BI__builtin_neon_vdot_v:
+  case NEON::BI__builtin_neon_vdotq_v: {
+    llvm::Type *InputTy =
+        llvm::VectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8);
+    llvm::Type *Tys[2] = { Ty, InputTy };
+    Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic;
+    return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vdot");
+  }
   }
 
   assert(Int && "Expected valid intrinsic number");

diff --git a/test/CodeGen/aarch64-neon-dot-product.c b/test/CodeGen/aarch64-neon-dot-product.c
@@ -0,0 +1,117 @@
+// RUN: %clang_cc1 -triple arm64-none-linux-gnu -target-feature +neon -target-feature +dotprod \
+// RUN: -disable-O0-optnone  -emit-llvm -o - %s | opt -S -instcombine | FileCheck %s
+
+// REQUIRES: aarch64-registered-target
+
+// Test AArch64 Armv8.2-A dot product intrinsics
+
+#include <arm_neon.h>
+
+uint32x2_t test_vdot_u32(uint32x2_t a, uint8x8_t b, uint8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_u32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_u32(a, b, c);
+}
+
+uint32x4_t test_vdotq_u32(uint32x4_t a, uint8x16_t b, uint8x16_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_u32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_u32(a, b, c);
+}
+
+int32x2_t test_vdot_s32(int32x2_t a, int8x8_t b, int8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_s32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_s32(a, b, c);
+}
+
+int32x4_t test_vdotq_s32(int32x4_t a, int8x16_t b, int8x16_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_s32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_s32(a, b, c);
+}
+
+uint32x2_t test_vdot_lane_u32(uint32x2_t a, uint8x8_t b, uint8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_lane_u32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <2 x i32> <i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> [[CAST2]])
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_lane_u32(a, b, c, 1);
+}
+
+uint32x4_t test_vdotq_lane_u32(uint32x4_t a, uint8x16_t b, uint8x8_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_lane_u32(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> [[CAST2]])
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_lane_u32(a, b, c, 1);
+}
+
+uint32x2_t test_vdot_laneq_u32(uint32x2_t a, uint8x8_t b, uint8x16_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_laneq_u32(<2 x i32> %a, <8 x i8> %b, <16 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <16 x i8> %c to <4 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <4 x i32> [[CAST1]], <4 x i32> undef, <2 x i32> <i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.aarch64.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> [[CAST2]])
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_laneq_u32(a, b, c, 1);
+}
+
+uint32x4_t test_vdotq_laneq_u32(uint32x4_t a, uint8x16_t b, uint8x16_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_laneq_u32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <16 x i8> %c to <4 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <4 x i32> [[CAST1]], <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.aarch64.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> [[CAST2]])
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_laneq_u32(a, b, c, 1);
+}
+
+int32x2_t test_vdot_lane_s32(int32x2_t a, int8x8_t b, int8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_lane_s32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <2 x i32> <i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> [[CAST2]])
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_lane_s32(a, b, c, 1);
+}
+
+int32x4_t test_vdotq_lane_s32(int32x4_t a, int8x16_t b, int8x8_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_lane_s32(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> [[CAST2]])
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_lane_s32(a, b, c, 1);
+}
+
+int32x2_t test_vdot_laneq_s32(int32x2_t a, int8x8_t b, int8x16_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_laneq_s32(<2 x i32> %a, <8 x i8> %b, <16 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <16 x i8> %c to <4 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <4 x i32> [[CAST1]], <4 x i32> undef, <2 x i32> <i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.aarch64.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> [[CAST2]])
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_laneq_s32(a, b, c, 1);
+}
+
+int32x4_t test_vdotq_laneq_s32(int32x4_t a, int8x16_t b, int8x16_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_laneq_s32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <16 x i8> %c to <4 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <4 x i32> [[CAST1]], <4 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.aarch64.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> [[CAST2]])
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_laneq_s32(a, b, c, 1);
+}
+
diff --git a/test/CodeGen/arm-neon-dot-product.c b/test/CodeGen/arm-neon-dot-product.c
@@ -0,0 +1,76 @@
+// RUN: %clang_cc1 -triple armv8-linux-gnueabihf -target-cpu cortex-a75 -target-feature +dotprod \
+// RUN: -disable-O0-optnone  -emit-llvm -o - %s | opt -S -instcombine | FileCheck %s
+
+// REQUIRES: arm-registered-target
+
+// Test ARM v8.2-A dot product intrinsics
+
+#include <arm_neon.h>
+
+uint32x2_t test_vdot_u32(uint32x2_t a, uint8x8_t b, uint8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_u32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.arm.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_u32(a, b, c);
+}
+
+uint32x4_t test_vdotq_u32(uint32x4_t a, uint8x16_t b, uint8x16_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_u32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.arm.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_u32(a, b, c);
+}
+
+int32x2_t test_vdot_s32(int32x2_t a, int8x8_t b, int8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_s32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.arm.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_s32(a, b, c);
+}
+
+int32x4_t test_vdotq_s32(int32x4_t a, int8x16_t b, int8x16_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_s32(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.arm.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> %c)
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_s32(a, b, c);
+}
+
+uint32x2_t test_vdot_lane_u32(uint32x2_t a, uint8x8_t b, uint8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_lane_u32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <2 x i32> <i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.arm.neon.udot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> [[CAST2]])
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_lane_u32(a, b, c, 1);
+}
+
+uint32x4_t test_vdotq_lane_u32(uint32x4_t a, uint8x16_t b, uint8x8_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_lane_u32(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.arm.neon.udot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> [[CAST2]])
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_lane_u32(a, b, c, 1);
+}
+
+int32x2_t test_vdot_lane_s32(int32x2_t a, int8x8_t b, int8x8_t c) {
+// CHECK-LABEL: define <2 x i32> @test_vdot_lane_s32(<2 x i32> %a, <8 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <2 x i32> <i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <2 x i32> [[SHUFFLE]] to <8 x i8>
+// CHECK: [[RESULT:%.*]] = call <2 x i32> @llvm.arm.neon.sdot.v2i32.v8i8(<2 x i32> %a, <8 x i8> %b, <8 x i8> [[CAST2]])
+// CHECK: ret <2 x i32> [[RESULT]]
+  return vdot_lane_s32(a, b, c, 1);
+}
+
+int32x4_t test_vdotq_lane_s32(int32x4_t a, int8x16_t b, int8x8_t c) {
+// CHECK-LABEL: define <4 x i32> @test_vdotq_lane_s32(<4 x i32> %a, <16 x i8> %b, <8 x i8> %c)
+// CHECK: [[CAST1:%.*]] = bitcast <8 x i8> %c to <2 x i32>
+// CHECK: [[SHUFFLE:%.*]] = shufflevector <2 x i32> [[CAST1]], <2 x i32> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
+// CHECK: [[CAST2:%.*]] = bitcast <4 x i32> [[SHUFFLE]] to <16 x i8>
+// CHECK: [[RESULT:%.*]] = call <4 x i32> @llvm.arm.neon.sdot.v4i32.v16i8(<4 x i32> %a, <16 x i8> %b, <16 x i8> [[CAST2]])
+// CHECK: ret <4 x i32> [[RESULT]]
+  return vdotq_lane_s32(a, b, c, 1);
+}
diff --git a/utils/TableGen/NeonEmitter.cpp b/utils/TableGen/NeonEmitter.cpp
@@ -995,6 +995,19 @@ void Type::applyModifier(char Mod) {
     if (!AppliedQuad)
       Bitwidth *= 2;
     break;
+  case '7':
+    if (AppliedQuad)
+      Bitwidth /= 2;
+    ElementBitwidth = 8;
+    break;
+  case '8':
+    ElementBitwidth = 8;
+    break;
+  case '9':
+    if (!AppliedQuad)
+      Bitwidth *= 2;
+    ElementBitwidth = 8;
+    break;
   default:
     llvm_unreachable("Unhandled character!");
   }