[FP16] Implement conversion operations. (WebAssembly#6974)

Note: FP16 is a little different from F32/F64 since it can't represent the full 2^16 integer range. 65504 is the max whole integer. This leads to some slightly strange behavior when converting integers greater than 65504 since they become infinity. Specified at https://github.com/WebAssembly/half-precision/blob/main/proposals/half-precision/Overview.md
aheejin · Sep 26, 2024 · c3a71ff · c3a71ff
1 parent 3856a2d
commit c3a71ff
Show file tree

Hide file tree

Showing 19 changed files with 303 additions and 6 deletions.
diff --git a/scripts/gen-s-parser.py b/scripts/gen-s-parser.py
@@ -540,6 +540,10 @@
     ("i32x4.trunc_sat_f64x2_u_zero",  "makeUnary(UnaryOp::TruncSatZeroUVecF64x2ToVecI32x4)"),
     ("f32x4.demote_f64x2_zero",       "makeUnary(UnaryOp::DemoteZeroVecF64x2ToVecF32x4)"),
     ("f64x2.promote_low_f32x4",       "makeUnary(UnaryOp::PromoteLowVecF32x4ToVecF64x2)"),
+    ("i16x8.trunc_sat_f16x8_s",  "makeUnary(UnaryOp::TruncSatSVecF16x8ToVecI16x8)"),
+    ("i16x8.trunc_sat_f16x8_u",  "makeUnary(UnaryOp::TruncSatUVecF16x8ToVecI16x8)"),
+    ("f16x8.convert_i16x8_s",    "makeUnary(UnaryOp::ConvertSVecI16x8ToVecF16x8)"),
+    ("f16x8.convert_i16x8_u",    "makeUnary(UnaryOp::ConvertUVecI16x8ToVecF16x8)"),
 
     # relaxed SIMD ops
     ("i8x16.relaxed_swizzle", "makeBinary(BinaryOp::RelaxedSwizzleVecI8x16)"),

diff --git a/src/gen-s-parser.inc b/src/gen-s-parser.inc
@@ -326,12 +326,34 @@ switch (buf[0]) {
               default: goto parse_error;
             }
           }
-          case 'c':
-            if (op == "f16x8.ceil"sv) {
-              CHECK_ERR(makeUnary(ctx, pos, annotations, UnaryOp::CeilVecF16x8));
-              return Ok{};
+          case 'c': {
+            switch (buf[7]) {
+              case 'e':
+                if (op == "f16x8.ceil"sv) {
+                  CHECK_ERR(makeUnary(ctx, pos, annotations, UnaryOp::CeilVecF16x8));
+                  return Ok{};
+                }
+                goto parse_error;
+              case 'o': {
+                switch (buf[20]) {
+                  case 's':
+                    if (op == "f16x8.convert_i16x8_s"sv) {
+                      CHECK_ERR(makeUnary(ctx, pos, annotations, UnaryOp::ConvertSVecI16x8ToVecF16x8));
+                      return Ok{};
+                    }
+                    goto parse_error;
+                  case 'u':
+                    if (op == "f16x8.convert_i16x8_u"sv) {
+                      CHECK_ERR(makeUnary(ctx, pos, annotations, UnaryOp::ConvertUVecI16x8ToVecF16x8));
+                      return Ok{};
+                    }
+                    goto parse_error;
+                  default: goto parse_error;
+                }
+              }
+              default: goto parse_error;
             }
-            goto parse_error;
+          }
           case 'd':
             if (op == "f16x8.div"sv) {
               CHECK_ERR(makeBinary(ctx, pos, annotations, BinaryOp::DivVecF16x8));
@@ -2038,6 +2060,23 @@ switch (buf[0]) {
               default: goto parse_error;
             }
           }
+          case 't': {
+            switch (buf[22]) {
+              case 's':
+                if (op == "i16x8.trunc_sat_f16x8_s"sv) {
+                  CHECK_ERR(makeUnary(ctx, pos, annotations, UnaryOp::TruncSatSVecF16x8ToVecI16x8));
+                  return Ok{};
+                }
+                goto parse_error;
+              case 'u':
+                if (op == "i16x8.trunc_sat_f16x8_u"sv) {
+                  CHECK_ERR(makeUnary(ctx, pos, annotations, UnaryOp::TruncSatUVecF16x8ToVecI16x8));
+                  return Ok{};
+                }
+                goto parse_error;
+              default: goto parse_error;
+            }
+          }
           default: goto parse_error;
         }
       }

diff --git a/src/ir/child-typer.h b/src/ir/child-typer.h
@@ -423,6 +423,10 @@ template<typename Subtype> struct ChildTyper : OverriddenVisitor<Subtype> {
       case RelaxedTruncUVecF32x4ToVecI32x4:
       case RelaxedTruncZeroSVecF64x2ToVecI32x4:
       case RelaxedTruncZeroUVecF64x2ToVecI32x4:
+      case TruncSatSVecF16x8ToVecI16x8:
+      case TruncSatUVecF16x8ToVecI16x8:
+      case ConvertSVecI16x8ToVecF16x8:
+      case ConvertUVecI16x8ToVecF16x8:
       case AnyTrueVec128:
       case AllTrueVecI8x16:
       case AllTrueVecI16x8:

diff --git a/src/ir/cost.h b/src/ir/cost.h
@@ -257,6 +257,10 @@ struct CostAnalyzer : public OverriddenVisitor<CostAnalyzer, CostType> {
       case RelaxedTruncUVecF32x4ToVecI32x4:
       case RelaxedTruncZeroSVecF64x2ToVecI32x4:
       case RelaxedTruncZeroUVecF64x2ToVecI32x4:
+      case TruncSatSVecF16x8ToVecI16x8:
+      case TruncSatUVecF16x8ToVecI16x8:
+      case ConvertSVecI16x8ToVecF16x8:
+      case ConvertUVecI16x8ToVecF16x8:
         ret = 1;
         break;
       case InvalidUnary:

diff --git a/src/literal.h b/src/literal.h
@@ -377,14 +377,18 @@ class Literal {
   Literal extendS32() const;
   Literal wrapToI32() const;
 
+  Literal convertSIToF16() const;
+  Literal convertUIToF16() const;
   Literal convertSIToF32() const;
   Literal convertUIToF32() const;
   Literal convertSIToF64() const;
   Literal convertUIToF64() const;
   Literal convertF32ToF16() const;
 
+  Literal truncSatToSI16() const;
   Literal truncSatToSI32() const;
   Literal truncSatToSI64() const;
+  Literal truncSatToUI16() const;
   Literal truncSatToUI32() const;
   Literal truncSatToUI64() const;
 
@@ -693,6 +697,10 @@ class Literal {
   Literal truncSatZeroUToI32x4() const;
   Literal demoteZeroToF32x4() const;
   Literal promoteLowToF64x2() const;
+  Literal truncSatToSI16x8() const;
+  Literal truncSatToUI16x8() const;
+  Literal convertSToF16x8() const;
+  Literal convertUToF16x8() const;
   Literal swizzleI8x16(const Literal& other) const;
   Literal relaxedMaddF16x8(const Literal& left, const Literal& right) const;
   Literal relaxedNmaddF16x8(const Literal& left, const Literal& right) const;

diff --git a/src/passes/Print.cpp b/src/passes/Print.cpp
@@ -1368,6 +1368,18 @@ struct PrintExpressionContents
       case RelaxedTruncZeroUVecF64x2ToVecI32x4:
         o << "i32x4.relaxed_trunc_f64x2_u_zero";
         break;
+      case TruncSatSVecF16x8ToVecI16x8:
+        o << "i16x8.trunc_sat_f16x8_s";
+        break;
+      case TruncSatUVecF16x8ToVecI16x8:
+        o << "i16x8.trunc_sat_f16x8_u";
+        break;
+      case ConvertSVecI16x8ToVecF16x8:
+        o << "f16x8.convert_i16x8_s";
+        break;
+      case ConvertUVecI16x8ToVecF16x8:
+        o << "f16x8.convert_i16x8_u";
+        break;
       case InvalidUnary:
         WASM_UNREACHABLE("unvalid unary operator");
     }

diff --git a/src/support/safe_integer.cpp b/src/support/safe_integer.cpp
@@ -98,6 +98,11 @@ int64_t wasm::toSInteger64(double x) {
  * 1 11111111 111...11 => 0xffffffff => -nan(0x7fffff)
  */
 
+bool wasm::isInRangeI16TruncS(int32_t i) {
+  uint32_t u = i;
+  return (u < 0x47000000U) || (u >= 0x80000000U && u <= 0xc7000000U);
+}
+
 bool wasm::isInRangeI32TruncS(int32_t i) {
   uint32_t u = i;
   return (u < 0x4f000000U) || (u >= 0x80000000U && u <= 0xcf000000U);
@@ -108,6 +113,11 @@ bool wasm::isInRangeI64TruncS(int32_t i) {
   return (u < 0x5f000000U) || (u >= 0x80000000U && u <= 0xdf000000U);
 }
 
+bool wasm::isInRangeI16TruncU(int32_t i) {
+  uint32_t u = i;
+  return (u < 0x47800000) || (u >= 0x80000000U && u < 0xbf800000U);
+}
+
 bool wasm::isInRangeI32TruncU(int32_t i) {
   uint32_t u = i;
   return (u < 0x4f800000U) || (u >= 0x80000000U && u < 0xbf800000U);

diff --git a/src/support/safe_integer.h b/src/support/safe_integer.h
@@ -32,8 +32,10 @@ uint64_t toUInteger64(double x);
 int64_t toSInteger64(double x);
 // The isInRange* functions all expect to be passed the binary representation
 // of a float or double.
+bool isInRangeI16TruncS(int32_t i);
 bool isInRangeI32TruncS(int32_t i);
 bool isInRangeI64TruncS(int32_t i);
+bool isInRangeI16TruncU(int32_t i);
 bool isInRangeI32TruncU(int32_t i);
 bool isInRangeI64TruncU(int32_t i);
 bool isInRangeI32TruncS(int64_t i);

diff --git a/src/tools/fuzzing/fuzzing.cpp b/src/tools/fuzzing/fuzzing.cpp
@@ -3143,7 +3143,11 @@ Expression* TranslateToFuzzReader::makeUnary(Type type) {
                          CeilVecF16x8,
                          FloorVecF16x8,
                          TruncVecF16x8,
-                         NearestVecF16x8)),
+                         NearestVecF16x8,
+                         TruncSatSVecF16x8ToVecI16x8,
+                         TruncSatUVecF16x8ToVecI16x8,
+                         ConvertSVecI16x8ToVecF16x8,
+                         ConvertUVecI16x8ToVecF16x8)),
              make(Type::v128)});
       }
       WASM_UNREACHABLE("invalid value");

diff --git a/src/wasm-binary.h b/src/wasm-binary.h
@@ -1039,6 +1039,10 @@ enum ASTNodes {
   F16x8Max = 0x142,
   F16x8Pmin = 0x143,
   F16x8Pmax = 0x144,
+  I16x8TruncSatF16x8S = 0x145,
+  I16x8TruncSatF16x8U = 0x146,
+  F16x8ConvertI16x8S = 0x147,
+  F16x8ConvertI16x8U = 0x148,
 
   // bulk memory opcodes
 

diff --git a/src/wasm-interpreter.h b/src/wasm-interpreter.h
@@ -632,6 +632,14 @@ class ExpressionRunner : public OverriddenVisitor<SubType, Flow> {
         return value.demoteZeroToF32x4();
       case PromoteLowVecF32x4ToVecF64x2:
         return value.promoteLowToF64x2();
+      case TruncSatSVecF16x8ToVecI16x8:
+        return value.truncSatToSI16x8();
+      case TruncSatUVecF16x8ToVecI16x8:
+        return value.truncSatToUI16x8();
+      case ConvertSVecI16x8ToVecF16x8:
+        return value.convertSToF16x8();
+      case ConvertUVecI16x8ToVecF16x8:
+        return value.convertUToF16x8();
       case InvalidUnary:
         WASM_UNREACHABLE("invalid unary op");
     }

diff --git a/src/wasm.h b/src/wasm.h
@@ -238,6 +238,10 @@ enum UnaryOp {
 
   // Half precision SIMD
   SplatVecF16x8,
+  TruncSatSVecF16x8ToVecI16x8,
+  TruncSatUVecF16x8ToVecI16x8,
+  ConvertSVecI16x8ToVecF16x8,
+  ConvertUVecI16x8ToVecF16x8,
 
   InvalidUnary
 };

diff --git a/src/wasm/literal.cpp b/src/wasm/literal.cpp
@@ -798,6 +798,20 @@ Literal Literal::wrapToI32() const {
   return Literal((int32_t)i64);
 }
 
+Literal Literal::convertSIToF16() const {
+  if (type == Type::i32) {
+    return Literal(fp16_ieee_from_fp32_value(float(i32)));
+  }
+  WASM_UNREACHABLE("invalid type");
+}
+
+Literal Literal::convertUIToF16() const {
+  if (type == Type::i32) {
+    return Literal(fp16_ieee_from_fp32_value(float(uint16_t(i32))));
+  }
+  WASM_UNREACHABLE("invalid type");
+}
+
 Literal Literal::convertSIToF32() const {
   if (type == Type::i32) {
     return Literal(float(i32));
@@ -861,6 +875,14 @@ static Literal saturating_trunc(typename AsInt<F>::type val) {
   return Literal(I(std::trunc(bit_cast<F>(val))));
 }
 
+Literal Literal::truncSatToSI16() const {
+  if (type == Type::f32) {
+    return saturating_trunc<float, int16_t, isInRangeI16TruncS>(
+      Literal(*this).castToI32().geti32());
+  }
+  WASM_UNREACHABLE("invalid type");
+}
+
 Literal Literal::truncSatToSI32() const {
   if (type == Type::f32) {
     return saturating_trunc<float, int32_t, isInRangeI32TruncS>(
@@ -885,6 +907,14 @@ Literal Literal::truncSatToSI64() const {
   WASM_UNREACHABLE("invalid type");
 }
 
+Literal Literal::truncSatToUI16() const {
+  if (type == Type::f32) {
+    return saturating_trunc<float, uint16_t, isInRangeI16TruncU>(
+      Literal(*this).castToI32().geti32());
+  }
+  WASM_UNREACHABLE("invalid type");
+}
+
 Literal Literal::truncSatToUI32() const {
   if (type == Type::f32) {
     return saturating_trunc<float, uint32_t, isInRangeI32TruncU>(
@@ -1997,6 +2027,19 @@ Literal Literal::convertUToF32x4() const {
   return unary<4, &Literal::getLanesI32x4, &Literal::convertUIToF32>(*this);
 }
 
+Literal Literal::truncSatToSI16x8() const {
+  return unary<8, &Literal::getLanesF16x8, &Literal::truncSatToSI16>(*this);
+}
+Literal Literal::truncSatToUI16x8() const {
+  return unary<8, &Literal::getLanesF16x8, &Literal::truncSatToUI16>(*this);
+}
+Literal Literal::convertSToF16x8() const {
+  return unary<8, &Literal::getLanesSI16x8, &Literal::convertSIToF16>(*this);
+}
+Literal Literal::convertUToF16x8() const {
+  return unary<8, &Literal::getLanesSI16x8, &Literal::convertUIToF16>(*this);
+}
+
 Literal Literal::anyTrueV128() const {
   auto lanes = getLanesI32x4();
   for (size_t i = 0; i < 4; ++i) {

diff --git a/src/wasm/wasm-binary.cpp b/src/wasm/wasm-binary.cpp
@@ -6522,6 +6522,22 @@ bool WasmBinaryReader::maybeVisitSIMDUnary(Expression*& out, uint32_t code) {
       curr = allocator.alloc<Unary>();
       curr->op = RelaxedTruncZeroUVecF64x2ToVecI32x4;
       break;
+    case BinaryConsts::I16x8TruncSatF16x8S:
+      curr = allocator.alloc<Unary>();
+      curr->op = TruncSatSVecF16x8ToVecI16x8;
+      break;
+    case BinaryConsts::I16x8TruncSatF16x8U:
+      curr = allocator.alloc<Unary>();
+      curr->op = TruncSatUVecF16x8ToVecI16x8;
+      break;
+    case BinaryConsts::F16x8ConvertI16x8S:
+      curr = allocator.alloc<Unary>();
+      curr->op = ConvertSVecI16x8ToVecF16x8;
+      break;
+    case BinaryConsts::F16x8ConvertI16x8U:
+      curr = allocator.alloc<Unary>();
+      curr->op = ConvertUVecI16x8ToVecF16x8;
+      break;
     default:
       return false;
   }

diff --git a/src/wasm/wasm-stack.cpp b/src/wasm/wasm-stack.cpp
@@ -1324,6 +1324,22 @@ void BinaryInstWriter::visitUnary(Unary* curr) {
       o << int8_t(BinaryConsts::SIMDPrefix)
         << U32LEB(BinaryConsts::I32x4RelaxedTruncF64x2UZero);
       break;
+    case TruncSatSVecF16x8ToVecI16x8:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I16x8TruncSatF16x8S);
+      break;
+    case TruncSatUVecF16x8ToVecI16x8:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::I16x8TruncSatF16x8U);
+      break;
+    case ConvertSVecI16x8ToVecF16x8:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::F16x8ConvertI16x8S);
+      break;
+    case ConvertUVecI16x8ToVecF16x8:
+      o << int8_t(BinaryConsts::SIMDPrefix)
+        << U32LEB(BinaryConsts::F16x8ConvertI16x8U);
+      break;
     case InvalidUnary:
       WASM_UNREACHABLE("invalid unary op");
   }

diff --git a/src/wasm/wasm-validator.cpp b/src/wasm/wasm-validator.cpp
@@ -2160,6 +2160,10 @@ void FunctionValidator::visitUnary(Unary* curr) {
     case RelaxedTruncUVecF32x4ToVecI32x4:
     case RelaxedTruncZeroSVecF64x2ToVecI32x4:
     case RelaxedTruncZeroUVecF64x2ToVecI32x4:
+    case TruncSatSVecF16x8ToVecI16x8:
+    case TruncSatUVecF16x8ToVecI16x8:
+    case ConvertSVecI16x8ToVecF16x8:
+    case ConvertUVecI16x8ToVecF16x8:
       shouldBeEqual(curr->type, Type(Type::v128), curr, "expected v128 type");
       shouldBeEqual(
         curr->value->type, Type(Type::v128), curr, "expected v128 operand");

diff --git a/src/wasm/wasm.cpp b/src/wasm/wasm.cpp
@@ -703,6 +703,10 @@ void Unary::finalize() {
     case RelaxedTruncUVecF32x4ToVecI32x4:
     case RelaxedTruncZeroSVecF64x2ToVecI32x4:
     case RelaxedTruncZeroUVecF64x2ToVecI32x4:
+    case TruncSatSVecF16x8ToVecI16x8:
+    case TruncSatUVecF16x8ToVecI16x8:
+    case ConvertSVecI16x8ToVecF16x8:
+    case ConvertUVecI16x8ToVecF16x8:
       type = Type::v128;
       break;
     case AnyTrueVec128: