//===- PolynomialApproximation.cpp - Approximate math operations ----------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This file implements expansion of math operations to fast approximations // that do not rely on any of the library functions. // //===----------------------------------------------------------------------===// #include #include #include #include "mlir/Dialect/Arith/IR/Arith.h" #include "mlir/Dialect/Math/IR/Math.h" #include "mlir/Dialect/Math/Transforms/Approximation.h" #include "mlir/Dialect/Math/Transforms/Passes.h" #include "mlir/Dialect/Utils/IndexingUtils.h" #include "mlir/Dialect/Vector/IR/VectorOps.h" #include "mlir/Dialect/Vector/Utils/VectorUtils.h" #include "mlir/Dialect/X86Vector/X86VectorDialect.h" #include "mlir/IR/Builders.h" #include "mlir/IR/BuiltinTypes.h" #include "mlir/IR/ImplicitLocOpBuilder.h" #include "mlir/IR/OpDefinition.h" #include "mlir/IR/PatternMatch.h" #include "mlir/IR/TypeUtilities.h" #include "mlir/Transforms/DialectConversion.h" #include "mlir/Transforms/GreedyPatternRewriteDriver.h" #include "llvm/ADT/ArrayRef.h" #include "llvm/ADT/STLExtras.h" #include "llvm/Support/MathExtras.h" using namespace mlir; using namespace mlir::math; using namespace mlir::vector; // Returns vector shape if the type is a vector. Returns an empty shape if it is // not a vector. static ArrayRef vectorShape(Type type) { auto vectorType = dyn_cast(type); return vectorType ? vectorType.getShape() : ArrayRef(); } static ArrayRef vectorShape(Value value) { return vectorShape(value.getType()); } //----------------------------------------------------------------------------// // Broadcast scalar types and values into vector types and values. //----------------------------------------------------------------------------// // Broadcasts scalar type into vector type (iff shape is non-scalar). static Type broadcast(Type type, ArrayRef shape) { assert(!isa(type) && "must be scalar type"); return !shape.empty() ? VectorType::get(shape, type) : type; } // Broadcasts scalar value into vector (iff shape is non-scalar). static Value broadcast(ImplicitLocOpBuilder &builder, Value value, ArrayRef shape) { assert(!isa(value.getType()) && "must be scalar value"); auto type = broadcast(value.getType(), shape); return !shape.empty() ? builder.create(type, value) : value; } //----------------------------------------------------------------------------// // Helper function to handle n-D vectors with 1-D operations. //----------------------------------------------------------------------------// // Expands and unrolls n-D vector operands into multiple fixed size 1-D vectors // and calls the compute function with 1-D vector operands. Stitches back all // results into the original n-D vector result. // // Examples: vectorWidth = 8 // - vector<4x8xf32> unrolled 4 times // - vector<16xf32> expanded to vector<2x8xf32> and unrolled 2 times // - vector<4x16xf32> expanded to vector<4x2x8xf32> and unrolled 4*2 times // // Some math approximations rely on ISA-specific operations that only accept // fixed size 1-D vectors (e.g. AVX expects vectors of width 8). // // It is the caller's responsibility to verify that the inner dimension is // divisible by the vectorWidth, and that all operands have the same vector // shape. static Value handleMultidimensionalVectors(ImplicitLocOpBuilder &builder, ValueRange operands, int64_t vectorWidth, llvm::function_ref compute) { assert(!operands.empty() && "operands must be not empty"); assert(vectorWidth > 0 && "vector width must be larger than 0"); VectorType inputType = cast(operands[0].getType()); ArrayRef inputShape = inputType.getShape(); // If input shape matches target vector width, we can just call the // user-provided compute function with the operands. if (inputShape == llvm::ArrayRef(vectorWidth)) return compute(operands); // Check if the inner dimension has to be expanded, or we can directly iterate // over the outer dimensions of the vector. int64_t innerDim = inputShape.back(); int64_t expansionDim = innerDim / vectorWidth; assert((innerDim % vectorWidth == 0) && "invalid inner dimension size"); // Maybe expand operands to the higher rank vector shape that we'll use to // iterate over and extract one dimensional vectors. SmallVector expandedShape(inputShape.begin(), inputShape.end()); SmallVector expandedOperands(operands); if (expansionDim > 1) { // Expand shape from [..., innerDim] to [..., expansionDim, vectorWidth]. expandedShape.insert(expandedShape.end() - 1, expansionDim); expandedShape.back() = vectorWidth; for (unsigned i = 0; i < operands.size(); ++i) { auto operand = operands[i]; auto eltType = cast(operand.getType()).getElementType(); auto expandedType = VectorType::get(expandedShape, eltType); expandedOperands[i] = builder.create(expandedType, operand); } } // Iterate over all outer dimensions of the compute shape vector type. auto iterationDims = ArrayRef(expandedShape).drop_back(); int64_t maxIndex = computeMaxLinearIndex(iterationDims); auto strides = computeStrides(iterationDims); // Compute results for each one dimensional vector. SmallVector results(maxIndex); for (int64_t i = 0; i < maxIndex; ++i) { auto offsets = delinearize(i, strides); SmallVector extracted(expandedOperands.size()); for (const auto &tuple : llvm::enumerate(expandedOperands)) extracted[tuple.index()] = builder.create(tuple.value(), offsets); results[i] = compute(extracted); } // Stitch results together into one large vector. Type resultEltType = cast(results[0].getType()).getElementType(); Type resultExpandedType = VectorType::get(expandedShape, resultEltType); Value result = builder.create( resultExpandedType, builder.getZeroAttr(resultExpandedType)); for (int64_t i = 0; i < maxIndex; ++i) result = builder.create(results[i], result, delinearize(i, strides)); // Reshape back to the original vector shape. return builder.create( VectorType::get(inputShape, resultEltType), result); } //----------------------------------------------------------------------------// // Helper functions to create constants. //----------------------------------------------------------------------------// static Value floatCst(ImplicitLocOpBuilder &builder, float value, Type elementType) { assert((elementType.isF16() || elementType.isF32()) && "x must be f16 or f32 type."); return builder.create( builder.getFloatAttr(elementType, value)); } static Value f32Cst(ImplicitLocOpBuilder &builder, double value) { return builder.create(builder.getF32FloatAttr(value)); } static Value i32Cst(ImplicitLocOpBuilder &builder, int32_t value) { return builder.create(builder.getI32IntegerAttr(value)); } static Value f32FromBits(ImplicitLocOpBuilder &builder, uint32_t bits) { Value i32Value = i32Cst(builder, static_cast(bits)); return builder.create(builder.getF32Type(), i32Value); } //----------------------------------------------------------------------------// // Helper functions to build math functions approximations. //----------------------------------------------------------------------------// // Return the minimum of the two values or NaN if value is NaN static Value min(ImplicitLocOpBuilder &builder, Value value, Value bound) { return builder.create( builder.create(arith::CmpFPredicate::ULT, value, bound), value, bound); } // Return the maximum of the two values or NaN if value is NaN static Value max(ImplicitLocOpBuilder &builder, Value value, Value bound) { return builder.create( builder.create(arith::CmpFPredicate::UGT, value, bound), value, bound); } // Return the clamped value or NaN if value is NaN static Value clamp(ImplicitLocOpBuilder &builder, Value value, Value lowerBound, Value upperBound) { return max(builder, min(builder, value, upperBound), lowerBound); } // Decomposes given floating point value `arg` into a normalized fraction and // an integral power of two (see std::frexp). Returned values have float type. static std::pair frexp(ImplicitLocOpBuilder &builder, Value arg, bool isPositive = false) { assert(getElementTypeOrSelf(arg).isF32() && "arg must be f32 type"); ArrayRef shape = vectorShape(arg); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; auto i32 = builder.getIntegerType(32); auto i32Vec = broadcast(i32, shape); auto f32Vec = broadcast(builder.getF32Type(), shape); Value cst126f = f32Cst(builder, 126.0f); Value cstHalf = f32Cst(builder, 0.5f); Value cstInvMantMask = f32FromBits(builder, ~0x7f800000u); // Bitcast to i32 for bitwise operations. Value i32Half = builder.create(i32, cstHalf); Value i32InvMantMask = builder.create(i32, cstInvMantMask); Value i32Arg = builder.create(i32Vec, arg); // Compute normalized fraction. Value tmp0 = builder.create(i32Arg, bcast(i32InvMantMask)); Value tmp1 = builder.create(tmp0, bcast(i32Half)); Value normalizedFraction = builder.create(f32Vec, tmp1); // Compute exponent. Value arg0 = isPositive ? arg : builder.create(arg); Value biasedExponentBits = builder.create( builder.create(i32Vec, arg0), bcast(i32Cst(builder, 23))); Value biasedExponent = builder.create(f32Vec, biasedExponentBits); Value exponent = builder.create(biasedExponent, bcast(cst126f)); return {normalizedFraction, exponent}; } // Computes exp2 for an i32 argument. static Value exp2I32(ImplicitLocOpBuilder &builder, Value arg) { assert(getElementTypeOrSelf(arg).isInteger(32) && "arg must be i32 type"); ArrayRef shape = vectorShape(arg); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; auto f32Vec = broadcast(builder.getF32Type(), shape); // The exponent of f32 located at 23-bit. auto exponetBitLocation = bcast(i32Cst(builder, 23)); // Set the exponent bias to zero. auto bias = bcast(i32Cst(builder, 127)); Value biasedArg = builder.create(arg, bias); Value exp2ValueInt = builder.create(biasedArg, exponetBitLocation); Value exp2ValueF32 = builder.create(f32Vec, exp2ValueInt); return exp2ValueF32; } namespace { Value makePolynomialCalculation(ImplicitLocOpBuilder &builder, llvm::ArrayRef coeffs, Value x) { Type elementType = getElementTypeOrSelf(x); assert((elementType.isF32() || elementType.isF16()) && "x must be f32 or f16 type"); ArrayRef shape = vectorShape(x); if (coeffs.empty()) return broadcast(builder, floatCst(builder, 0.0f, elementType), shape); if (coeffs.size() == 1) return coeffs[0]; Value res = builder.create(x, coeffs[coeffs.size() - 1], coeffs[coeffs.size() - 2]); for (auto i = ptrdiff_t(coeffs.size()) - 3; i >= 0; --i) { res = builder.create(x, res, coeffs[i]); } return res; } } // namespace //----------------------------------------------------------------------------// // Helper function/pattern to insert casts for reusing F32 bit expansion. //----------------------------------------------------------------------------// template LogicalResult insertCasts(Operation *op, PatternRewriter &rewriter) { // Conservatively only allow where the operand and result types are exactly 1. Type origType = op->getResultTypes().front(); for (Type t : llvm::drop_begin(op->getResultTypes())) if (origType != t) return rewriter.notifyMatchFailure(op, "required all types to match"); for (Type t : op->getOperandTypes()) if (origType != t) return rewriter.notifyMatchFailure(op, "required all types to match"); // Skip if already F32 or larger than 32 bits. if (getElementTypeOrSelf(origType).isF32() || getElementTypeOrSelf(origType).getIntOrFloatBitWidth() > 32) return failure(); // Create F32 equivalent type. Type newType; if (auto shaped = dyn_cast(origType)) { newType = shaped.clone(rewriter.getF32Type()); } else if (isa(origType)) { newType = rewriter.getF32Type(); } else { return rewriter.notifyMatchFailure(op, "unable to find F32 equivalent type"); } Location loc = op->getLoc(); SmallVector operands; for (auto operand : op->getOperands()) operands.push_back(rewriter.create(loc, newType, operand)); auto result = rewriter.create(loc, TypeRange{newType}, operands, op->getAttrs()); rewriter.replaceOpWithNewOp(op, origType, result); return success(); } namespace { // Pattern to cast to F32 to reuse F32 expansion as fallback for single-result // op. // TODO: Consider revising to avoid adding multiple casts for a subgraph that is // all in lower precision. Currently this is only fallback support and performs // simplistic casting. template struct ReuseF32Expansion : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(T op, PatternRewriter &rewriter) const final { static_assert( T::template hasTrait(), "requires same operands and result types"); return insertCasts(op, rewriter); } }; } // namespace //----------------------------------------------------------------------------// // AtanOp approximation. //----------------------------------------------------------------------------// namespace { struct AtanApproximation : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::AtanOp op, PatternRewriter &rewriter) const final; }; } // namespace LogicalResult AtanApproximation::matchAndRewrite(math::AtanOp op, PatternRewriter &rewriter) const { auto operand = op.getOperand(); if (!getElementTypeOrSelf(operand).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ArrayRef shape = vectorShape(op.getOperand()); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); Value abs = builder.create(operand); auto one = broadcast(builder, f32Cst(builder, 1.0), shape); // When 0.66 < x <= 2.41 we do (x-1) / (x+1): auto twoThirds = broadcast(builder, f32Cst(builder, 0.66), shape); Value cmp2 = builder.create(arith::CmpFPredicate::OGT, abs, twoThirds); Value addone = builder.create(abs, one); Value subone = builder.create(abs, one); Value xnum = builder.create(cmp2, subone, abs); Value xden = builder.create(cmp2, addone, one); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; // Break into the <= 0.66 or > 2.41 we do x or 1/x: auto tan3pio8 = bcast(f32Cst(builder, 2.41421356237309504880)); Value cmp1 = builder.create(arith::CmpFPredicate::OGT, abs, tan3pio8); xnum = builder.create(cmp1, one, xnum); xden = builder.create(cmp1, abs, xden); Value x = builder.create(xnum, xden); Value xx = builder.create(x, x); // Perform the Taylor series approximation for atan over the range // [0.0, 0.66]. auto p0 = bcast(f32Cst(builder, -8.750608600031904122785e-01)); auto p1 = bcast(f32Cst(builder, -1.615753718733365076637e+01)); auto p2 = bcast(f32Cst(builder, -7.500855792314704667340e+01)); auto p3 = bcast(f32Cst(builder, -1.228866684490136173410e+02)); auto p4 = bcast(f32Cst(builder, -6.485021904942025371773e+01)); auto q0 = bcast(f32Cst(builder, +2.485846490142306297962e+01)); auto q1 = bcast(f32Cst(builder, +1.650270098316988542046e+02)); auto q2 = bcast(f32Cst(builder, +4.328810604912902668951e+02)); auto q3 = bcast(f32Cst(builder, +4.853903996359136964868e+02)); auto q4 = bcast(f32Cst(builder, +1.945506571482613964425e+02)); // Apply the polynomial approximation for the numerator: Value n = p0; n = builder.create(xx, n, p1); n = builder.create(xx, n, p2); n = builder.create(xx, n, p3); n = builder.create(xx, n, p4); n = builder.create(n, xx); // Apply the polynomial approximation for the denominator: Value d = q0; d = builder.create(xx, d, q1); d = builder.create(xx, d, q2); d = builder.create(xx, d, q3); d = builder.create(xx, d, q4); // Compute approximation of theta: Value ans0 = builder.create(n, d); ans0 = builder.create(ans0, x, x); // Correct for the input mapping's angles: Value mpi4 = bcast(f32Cst(builder, llvm::numbers::pi / 4)); Value ans2 = builder.create(mpi4, ans0); Value ans = builder.create(cmp2, ans2, ans0); Value mpi2 = bcast(f32Cst(builder, llvm::numbers::pi / 2)); Value ans1 = builder.create(mpi2, ans0); ans = builder.create(cmp1, ans1, ans); // Correct for signing of the input. rewriter.replaceOpWithNewOp(op, ans, operand); return success(); } //----------------------------------------------------------------------------// // AtanOp approximation. //----------------------------------------------------------------------------// namespace { struct Atan2Approximation : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::Atan2Op op, PatternRewriter &rewriter) const final; }; } // namespace LogicalResult Atan2Approximation::matchAndRewrite(math::Atan2Op op, PatternRewriter &rewriter) const { auto y = op.getOperand(0); auto x = op.getOperand(1); if (!getElementTypeOrSelf(x).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); ArrayRef shape = vectorShape(op.getResult()); // Compute atan in the valid range. auto div = builder.create(y, x); auto atan = builder.create(div); // Determine what the atan would be for a 180 degree rotation. auto zero = broadcast(builder, f32Cst(builder, 0.0f), shape); auto pi = broadcast(builder, f32Cst(builder, 3.14159265359f), shape); auto addPi = builder.create(atan, pi); auto subPi = builder.create(atan, pi); auto atanGt = builder.create(arith::CmpFPredicate::OGT, atan, zero); auto flippedAtan = builder.create(atanGt, subPi, addPi); // Determine whether to directly use atan or use the 180 degree flip auto xGt = builder.create(arith::CmpFPredicate::OGT, x, zero); Value result = builder.create(xGt, atan, flippedAtan); // Handle x = 0, y > 0 Value xZero = builder.create(arith::CmpFPredicate::OEQ, x, zero); Value yGt = builder.create(arith::CmpFPredicate::OGT, y, zero); Value isHalfPi = builder.create(xZero, yGt); auto halfPi = broadcast(builder, f32Cst(builder, 1.57079632679f), shape); result = builder.create(isHalfPi, halfPi, result); // Handle x = 0, y < 0 Value yLt = builder.create(arith::CmpFPredicate::OLT, y, zero); Value isNegativeHalfPiPi = builder.create(xZero, yLt); auto negativeHalfPiPi = broadcast(builder, f32Cst(builder, -1.57079632679f), shape); result = builder.create(isNegativeHalfPiPi, negativeHalfPiPi, result); // Handle x = 0, y = 0; Value yZero = builder.create(arith::CmpFPredicate::OEQ, y, zero); Value isNan = builder.create(xZero, yZero); Value cstNan = broadcast(builder, f32FromBits(builder, 0x7fc00000), shape); result = builder.create(isNan, cstNan, result); rewriter.replaceOp(op, result); return success(); } //----------------------------------------------------------------------------// // TanhOp approximation. //----------------------------------------------------------------------------// namespace { struct TanhApproximation : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::TanhOp op, PatternRewriter &rewriter) const final; }; } // namespace LogicalResult TanhApproximation::matchAndRewrite(math::TanhOp op, PatternRewriter &rewriter) const { if (!getElementTypeOrSelf(op.getOperand()).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ArrayRef shape = vectorShape(op.getOperand()); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; // Clamp operand into [plusClamp, minusClamp] range. Value minusClamp = bcast(f32Cst(builder, -7.99881172180175781f)); Value plusClamp = bcast(f32Cst(builder, 7.99881172180175781f)); Value x = clamp(builder, op.getOperand(), minusClamp, plusClamp); // Mask for tiny values that are approximated with `operand`. Value tiny = bcast(f32Cst(builder, 0.0004f)); Value tinyMask = builder.create( arith::CmpFPredicate::OLT, builder.create(op.getOperand()), tiny); // The monomial coefficients of the numerator polynomial (odd). Value alpha1 = bcast(f32Cst(builder, 4.89352455891786e-03f)); Value alpha3 = bcast(f32Cst(builder, 6.37261928875436e-04f)); Value alpha5 = bcast(f32Cst(builder, 1.48572235717979e-05f)); Value alpha7 = bcast(f32Cst(builder, 5.12229709037114e-08f)); Value alpha9 = bcast(f32Cst(builder, -8.60467152213735e-11f)); Value alpha11 = bcast(f32Cst(builder, 2.00018790482477e-13f)); Value alpha13 = bcast(f32Cst(builder, -2.76076847742355e-16f)); // The monomial coefficients of the denominator polynomial (even). Value beta0 = bcast(f32Cst(builder, 4.89352518554385e-03f)); Value beta2 = bcast(f32Cst(builder, 2.26843463243900e-03f)); Value beta4 = bcast(f32Cst(builder, 1.18534705686654e-04f)); Value beta6 = bcast(f32Cst(builder, 1.19825839466702e-06f)); // Since the polynomials are odd/even, we need x^2. Value x2 = builder.create(x, x); // Evaluate the numerator polynomial p. Value p = builder.create(x2, alpha13, alpha11); p = builder.create(x2, p, alpha9); p = builder.create(x2, p, alpha7); p = builder.create(x2, p, alpha5); p = builder.create(x2, p, alpha3); p = builder.create(x2, p, alpha1); p = builder.create(x, p); // Evaluate the denominator polynomial q. Value q = builder.create(x2, beta6, beta4); q = builder.create(x2, q, beta2); q = builder.create(x2, q, beta0); // Divide the numerator by the denominator. Value res = builder.create( tinyMask, x, builder.create(p, q)); rewriter.replaceOp(op, res); return success(); } #define LN2_VALUE \ 0.693147180559945309417232121458176568075500134360255254120680009493393621L #define LOG2E_VALUE \ 1.442695040888963407359924681001892137426645954152985934135449406931109219L //----------------------------------------------------------------------------// // LogOp and Log2Op approximation. //----------------------------------------------------------------------------// namespace { template struct LogApproximationBase : public OpRewritePattern { using OpRewritePattern::OpRewritePattern; /// Base 2 if 'base2' is set; natural logarithm (base e) otherwise. LogicalResult logMatchAndRewrite(Op op, PatternRewriter &rewriter, bool base2) const; }; } // namespace // This approximation comes from Julien Pommier's SSE math library. // Link: http://gruntthepeon.free.fr/ssemath template LogicalResult LogApproximationBase::logMatchAndRewrite(Op op, PatternRewriter &rewriter, bool base2) const { if (!getElementTypeOrSelf(op.getOperand()).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ArrayRef shape = vectorShape(op.getOperand()); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; Value cstZero = bcast(f32Cst(builder, 0.0f)); Value cstOne = bcast(f32Cst(builder, 1.0f)); Value cstNegHalf = bcast(f32Cst(builder, -0.5f)); // The smallest non denormalized float number. Value cstMinNormPos = bcast(f32FromBits(builder, 0x00800000u)); Value cstMinusInf = bcast(f32FromBits(builder, 0xff800000u)); Value cstPosInf = bcast(f32FromBits(builder, 0x7f800000u)); Value cstNan = bcast(f32FromBits(builder, 0x7fc00000)); // Polynomial coefficients. Value cstCephesSQRTHF = bcast(f32Cst(builder, 0.707106781186547524f)); Value cstCephesLogP0 = bcast(f32Cst(builder, 7.0376836292E-2f)); Value cstCephesLogP1 = bcast(f32Cst(builder, -1.1514610310E-1f)); Value cstCephesLogP2 = bcast(f32Cst(builder, 1.1676998740E-1f)); Value cstCephesLogP3 = bcast(f32Cst(builder, -1.2420140846E-1f)); Value cstCephesLogP4 = bcast(f32Cst(builder, +1.4249322787E-1f)); Value cstCephesLogP5 = bcast(f32Cst(builder, -1.6668057665E-1f)); Value cstCephesLogP6 = bcast(f32Cst(builder, +2.0000714765E-1f)); Value cstCephesLogP7 = bcast(f32Cst(builder, -2.4999993993E-1f)); Value cstCephesLogP8 = bcast(f32Cst(builder, +3.3333331174E-1f)); Value x = op.getOperand(); // Truncate input values to the minimum positive normal. x = max(builder, x, cstMinNormPos); // Extract significant in the range [0.5,1) and exponent. std::pair pair = frexp(builder, x, /*isPositive=*/true); x = pair.first; Value e = pair.second; // Shift the inputs from the range [0.5,1) to [sqrt(1/2), sqrt(2)) and shift // by -1.0. The values are then centered around 0, which improves the // stability of the polynomial evaluation: // // if( x < SQRTHF ) { // e -= 1; // x = x + x - 1.0; // } else { x = x - 1.0; } Value mask = builder.create(arith::CmpFPredicate::OLT, x, cstCephesSQRTHF); Value tmp = builder.create(mask, x, cstZero); x = builder.create(x, cstOne); e = builder.create( e, builder.create(mask, cstOne, cstZero)); x = builder.create(x, tmp); Value x2 = builder.create(x, x); Value x3 = builder.create(x2, x); // Evaluate the polynomial approximant of degree 8 in three parts. Value y0, y1, y2; y0 = builder.create(cstCephesLogP0, x, cstCephesLogP1); y1 = builder.create(cstCephesLogP3, x, cstCephesLogP4); y2 = builder.create(cstCephesLogP6, x, cstCephesLogP7); y0 = builder.create(y0, x, cstCephesLogP2); y1 = builder.create(y1, x, cstCephesLogP5); y2 = builder.create(y2, x, cstCephesLogP8); y0 = builder.create(y0, x3, y1); y0 = builder.create(y0, x3, y2); y0 = builder.create(y0, x3); y0 = builder.create(cstNegHalf, x2, y0); x = builder.create(x, y0); if (base2) { Value cstLog2e = bcast(f32Cst(builder, static_cast(LOG2E_VALUE))); x = builder.create(x, cstLog2e, e); } else { Value cstLn2 = bcast(f32Cst(builder, static_cast(LN2_VALUE))); x = builder.create(e, cstLn2, x); } Value invalidMask = builder.create(arith::CmpFPredicate::ULT, op.getOperand(), cstZero); Value zeroMask = builder.create(arith::CmpFPredicate::OEQ, op.getOperand(), cstZero); Value posInfMask = builder.create(arith::CmpFPredicate::OEQ, op.getOperand(), cstPosInf); // Filter out invalid values: // • x == 0 -> -INF // • x < 0 -> NAN // • x == +INF -> +INF Value aproximation = builder.create( zeroMask, cstMinusInf, builder.create( invalidMask, cstNan, builder.create(posInfMask, cstPosInf, x))); rewriter.replaceOp(op, aproximation); return success(); } namespace { struct LogApproximation : public LogApproximationBase { using LogApproximationBase::LogApproximationBase; LogicalResult matchAndRewrite(math::LogOp op, PatternRewriter &rewriter) const final { return logMatchAndRewrite(op, rewriter, /*base2=*/false); } }; } // namespace namespace { struct Log2Approximation : public LogApproximationBase { using LogApproximationBase::LogApproximationBase; LogicalResult matchAndRewrite(math::Log2Op op, PatternRewriter &rewriter) const final { return logMatchAndRewrite(op, rewriter, /*base2=*/true); } }; } // namespace //----------------------------------------------------------------------------// // Log1p approximation. //----------------------------------------------------------------------------// namespace { struct Log1pApproximation : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::Log1pOp op, PatternRewriter &rewriter) const final; }; } // namespace // Approximate log(1+x). LogicalResult Log1pApproximation::matchAndRewrite(math::Log1pOp op, PatternRewriter &rewriter) const { if (!getElementTypeOrSelf(op.getOperand()).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ArrayRef shape = vectorShape(op.getOperand()); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; // Approximate log(1+x) using the following, due to W. Kahan: // u = x + 1.0; // if (u == 1.0 || u == inf) return x; // return x * log(u) / (u - 1.0); // ^^^^^^^^^^^^^^^^^^^^^^ // "logLarge" below. Value cstOne = bcast(f32Cst(builder, 1.0f)); Value x = op.getOperand(); Value u = builder.create(x, cstOne); Value uSmall = builder.create(arith::CmpFPredicate::OEQ, u, cstOne); Value logU = builder.create(u); Value uInf = builder.create(arith::CmpFPredicate::OEQ, u, logU); Value logLarge = builder.create( x, builder.create( logU, builder.create(u, cstOne))); Value approximation = builder.create( builder.create(uSmall, uInf), x, logLarge); rewriter.replaceOp(op, approximation); return success(); } //----------------------------------------------------------------------------// // Erf approximation. //----------------------------------------------------------------------------// // Approximates erf(x) with // a - P(x)/Q(x) // where P and Q are polynomials of degree 4. // Different coefficients are chosen based on the value of x. // The approximation error is ~2.5e-07. // Boost's minimax tool that utilizes the Remez method was used to find the // coefficients. LogicalResult ErfPolynomialApproximation::matchAndRewrite(math::ErfOp op, PatternRewriter &rewriter) const { Value operand = op.getOperand(); Type elementType = getElementTypeOrSelf(operand); if (!(elementType.isF32() || elementType.isF16())) return rewriter.notifyMatchFailure(op, "only f32 and f16 type is supported."); ArrayRef shape = vectorShape(operand); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; const int intervalsCount = 3; const int polyDegree = 4; Value zero = bcast(floatCst(builder, 0, elementType)); Value one = bcast(floatCst(builder, 1, elementType)); Value pp[intervalsCount][polyDegree + 1]; pp[0][0] = bcast(floatCst(builder, +0.00000000000000000e+00f, elementType)); pp[0][1] = bcast(floatCst(builder, +1.12837916222975858e+00f, elementType)); pp[0][2] = bcast(floatCst(builder, -5.23018562988006470e-01f, elementType)); pp[0][3] = bcast(floatCst(builder, +2.09741709609267072e-01f, elementType)); pp[0][4] = bcast(floatCst(builder, +2.58146801602987875e-02f, elementType)); pp[1][0] = bcast(floatCst(builder, +0.00000000000000000e+00f, elementType)); pp[1][1] = bcast(floatCst(builder, +1.12750687816789140e+00f, elementType)); pp[1][2] = bcast(floatCst(builder, -3.64721408487825775e-01f, elementType)); pp[1][3] = bcast(floatCst(builder, +1.18407396425136952e-01f, elementType)); pp[1][4] = bcast(floatCst(builder, +3.70645533056476558e-02f, elementType)); pp[2][0] = bcast(floatCst(builder, -3.30093071049483172e-03f, elementType)); pp[2][1] = bcast(floatCst(builder, +3.51961938357697011e-03f, elementType)); pp[2][2] = bcast(floatCst(builder, -1.41373622814988039e-03f, elementType)); pp[2][3] = bcast(floatCst(builder, +2.53447094961941348e-04f, elementType)); pp[2][4] = bcast(floatCst(builder, -1.71048029455037401e-05f, elementType)); Value qq[intervalsCount][polyDegree + 1]; qq[0][0] = bcast(floatCst(builder, +1.000000000000000000e+00f, elementType)); qq[0][1] = bcast(floatCst(builder, -4.635138185962547255e-01f, elementType)); qq[0][2] = bcast(floatCst(builder, +5.192301327279782447e-01f, elementType)); qq[0][3] = bcast(floatCst(builder, -1.318089722204810087e-01f, elementType)); qq[0][4] = bcast(floatCst(builder, +7.397964654672315005e-02f, elementType)); qq[1][0] = bcast(floatCst(builder, +1.00000000000000000e+00f, elementType)); qq[1][1] = bcast(floatCst(builder, -3.27607011824493086e-01f, elementType)); qq[1][2] = bcast(floatCst(builder, +4.48369090658821977e-01f, elementType)); qq[1][3] = bcast(floatCst(builder, -8.83462621207857930e-02f, elementType)); qq[1][4] = bcast(floatCst(builder, +5.72442770283176093e-02f, elementType)); qq[2][0] = bcast(floatCst(builder, +1.00000000000000000e+00f, elementType)); qq[2][1] = bcast(floatCst(builder, -2.06069165953913769e+00f, elementType)); qq[2][2] = bcast(floatCst(builder, +1.62705939945477759e+00f, elementType)); qq[2][3] = bcast(floatCst(builder, -5.83389859211130017e-01f, elementType)); qq[2][4] = bcast(floatCst(builder, +8.21908939856640930e-02f, elementType)); Value offsets[intervalsCount]; offsets[0] = bcast(floatCst(builder, 0.0f, elementType)); offsets[1] = bcast(floatCst(builder, 0.0f, elementType)); offsets[2] = bcast(floatCst(builder, 1.0f, elementType)); Value bounds[intervalsCount]; bounds[0] = bcast(floatCst(builder, 0.8f, elementType)); bounds[1] = bcast(floatCst(builder, 2.0f, elementType)); bounds[2] = bcast(floatCst(builder, 3.75f, elementType)); Value isNegativeArg = builder.create(arith::CmpFPredicate::OLT, operand, zero); Value negArg = builder.create(operand); Value x = builder.create(isNegativeArg, negArg, operand); Value offset = offsets[0]; Value p[polyDegree + 1]; Value q[polyDegree + 1]; for (int i = 0; i <= polyDegree; ++i) { p[i] = pp[0][i]; q[i] = qq[0][i]; } // TODO: maybe use vector stacking to reduce the number of selects. Value isLessThanBound[intervalsCount]; for (int j = 0; j < intervalsCount - 1; ++j) { isLessThanBound[j] = builder.create(arith::CmpFPredicate::OLT, x, bounds[j]); for (int i = 0; i <= polyDegree; ++i) { p[i] = builder.create(isLessThanBound[j], p[i], pp[j + 1][i]); q[i] = builder.create(isLessThanBound[j], q[i], qq[j + 1][i]); } offset = builder.create(isLessThanBound[j], offset, offsets[j + 1]); } isLessThanBound[intervalsCount - 1] = builder.create( arith::CmpFPredicate::ULT, x, bounds[intervalsCount - 1]); Value pPoly = makePolynomialCalculation(builder, p, x); Value qPoly = makePolynomialCalculation(builder, q, x); Value rationalPoly = builder.create(pPoly, qPoly); Value formula = builder.create(offset, rationalPoly); formula = builder.create(isLessThanBound[intervalsCount - 1], formula, one); // erf is odd function: erf(x) = -erf(-x). Value negFormula = builder.create(formula); Value res = builder.create(isNegativeArg, negFormula, formula); rewriter.replaceOp(op, res); return success(); } //----------------------------------------------------------------------------// // Exp approximation. //----------------------------------------------------------------------------// namespace { Value clampWithNormals(ImplicitLocOpBuilder &builder, const llvm::ArrayRef shape, Value value, float lowerBound, float upperBound) { assert(!std::isnan(lowerBound)); assert(!std::isnan(upperBound)); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; auto selectCmp = [&builder](auto pred, Value value, Value bound) { return builder.create( builder.create(pred, value, bound), value, bound); }; // Note: prefer UGE/ULE vs. UGT/ULT, since they generate vmaxps/vminps vs. // vcmpleps+vmovaps on x86_64. The latter outcome is also obtained with // arith::{Max,Min}FOp. value = selectCmp(arith::CmpFPredicate::UGE, value, bcast(f32Cst(builder, lowerBound))); value = selectCmp(arith::CmpFPredicate::ULE, value, bcast(f32Cst(builder, upperBound))); return value; } struct ExpApproximation : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::ExpOp op, PatternRewriter &rewriter) const final; }; LogicalResult ExpApproximation::matchAndRewrite(math::ExpOp op, PatternRewriter &rewriter) const { auto shape = vectorShape(op.getOperand().getType()); auto elementTy = getElementTypeOrSelf(op.getType()); if (!elementTy.isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto add = [&](Value a, Value b) -> Value { return builder.create(a, b); }; auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; auto floor = [&](Value a) { return builder.create(a); }; auto fmla = [&](Value a, Value b, Value c) { return builder.create(a, b, c); }; auto mul = [&](Value a, Value b) -> Value { return builder.create(a, b); }; // Polynomial approximation from Cephes. // // To compute e^x, we re-express it as // // e^x = e^(a + b) // = e^(a + n log(2)) // = e^a * 2^n. // // We choose n = round(x / log(2)), restricting the value of `a` to // (-log(2)/2, log(2)/2). We then use a polynomial to compute e^a. The // relative error between our approximation and the true value of e^a is less // than 2^-22.5 for all values of `a` within this range. // Restrict input to a small range, including some values that evaluate to // +/- inf. Note that for our lower bound, we choose log(2^-126) instead of // log(F32_EPSILON). We do so because this routine always flushes denormal // floating points to 0. Therefore, we only need to worry about exponentiating // up to the smallest representable non-denormal floating point, which is // 2^-126. // Constants. Value cstHalf = bcast(f32Cst(builder, 0.5f)); Value cstOne = bcast(f32Cst(builder, 1.0f)); // 1/log(2) Value cstLog2ef = bcast(f32Cst(builder, 1.44269504088896341f)); Value cstExpC1 = bcast(f32Cst(builder, -0.693359375f)); Value cstExpC2 = bcast(f32Cst(builder, 2.12194440e-4f)); Value cstExpP0 = bcast(f32Cst(builder, 1.9875691500E-4f)); Value cstExpP1 = bcast(f32Cst(builder, 1.3981999507E-3f)); Value cstExpP2 = bcast(f32Cst(builder, 8.3334519073E-3f)); Value cstExpP3 = bcast(f32Cst(builder, 4.1665795894E-2f)); Value cstExpP4 = bcast(f32Cst(builder, 1.6666665459E-1f)); Value cstExpP5 = bcast(f32Cst(builder, 5.0000001201E-1f)); // Our computations below aren't particularly sensitive to the exact choices // here, so we choose values a bit larger/smaller than // // log(F32_MAX) = 88.723... // log(2^-126) = -87.337... Value x = op.getOperand(); x = clampWithNormals(builder, shape, x, -87.8f, 88.8f); Value n = floor(fmla(x, cstLog2ef, cstHalf)); // When we eventually do the multiplication in e^a * 2^n, we need to handle // the case when n > 127, the max fp32 exponent (so 2^n == inf) but e^a < 1 // (so e^a * 2^n != inf). There's a similar problem for n < -126, the // smallest fp32 exponent. // // A straightforward solution would be to detect n out of range and split it // up, doing // // e^a * 2^n = e^a * 2^(n1 + n2) // = (2^n1 * e^a) * 2^n2. // // But it turns out this approach is quite slow, probably because it // manipulates subnormal values. // // The approach we use instead is to clamp n to [-127, 127]. Let n' be the // value of n clamped to [-127, 127]. In the case where n' = 127, `a` can grow // up to as large as 88.8 - 127 * log(2) which is about 0.7703. Even though // this value of `a` is outside our previously specified range, e^a will still // only have a relative error of approximately 2^-16 at worse. In practice // this seems to work well enough; it passes our exhaustive tests, breaking // only one result, and by one ulp (we return exp(88.7228394) = max-float but // we should return inf). // // In the case where n' = -127, the original input value of x is so small that // e^x, our final answer, is less than 2^-126. Since 2^-126 is the smallest // normal floating point, and since we flush denormals, we simply return 0. We // do this in a branchless way by observing that our code for constructing 2^n // produces 0 if n = -127. // // The proof that n' = -127 implies e^x < 2^-126 is as follows: // // n' = -127 implies n <= -127 // implies round(x / log(2)) <= -127 // implies x/log(2) < -126.5 // implies x < -126.5 * log(2) // implies e^x < e^(-126.5 * log(2)) // implies e^x < 2^-126.5 < 2^-126 // // This proves that n' = -127 implies e^x < 2^-126. n = clampWithNormals(builder, shape, n, -127.0f, 127.0f); // Computes x = x - n' * log(2), the value for `a` x = fmla(cstExpC1, n, x); x = fmla(cstExpC2, n, x); // Polynomial to compute z = e^a, accurate for a in (-0.5, 0.5). Value z = fmla(x, cstExpP0, cstExpP1); z = fmla(z, x, cstExpP2); z = fmla(z, x, cstExpP3); z = fmla(z, x, cstExpP4); z = fmla(z, x, cstExpP5); z = fmla(z, mul(x, x), x); z = add(cstOne, z); // Convert n' to an i32. This is safe because we clamped it above. auto i32Vec = broadcast(builder.getI32Type(), shape); Value nI32 = builder.create(i32Vec, n); // Creates the value 2^n' if -126 <= n' <= 127 and 0 if n' = -127. Value pow2 = exp2I32(builder, nI32); // Return z * 2^n' if -126 <= n' <= 127 and 0 if n = -127. Value ret = mul(z, pow2); rewriter.replaceOp(op, ret); return mlir::success(); } } // namespace //----------------------------------------------------------------------------// // ExpM1 approximation. //----------------------------------------------------------------------------// namespace { struct ExpM1Approximation : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::ExpM1Op op, PatternRewriter &rewriter) const final; }; } // namespace LogicalResult ExpM1Approximation::matchAndRewrite(math::ExpM1Op op, PatternRewriter &rewriter) const { if (!getElementTypeOrSelf(op.getOperand()).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ArrayRef shape = vectorShape(op.getOperand()); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; // expm1(x) = exp(x) - 1 = u - 1. // We have to handle it carefully when x is near 0, i.e. u ~= 1, // and when the input is ~= -inf, i.e. u - 1 ~= -1. Value cstOne = bcast(f32Cst(builder, 1.0f)); Value cstNegOne = bcast(f32Cst(builder, -1.0f)); Value x = op.getOperand(); Value u = builder.create(x); Value uEqOneOrNaN = builder.create(arith::CmpFPredicate::UEQ, u, cstOne); Value uMinusOne = builder.create(u, cstOne); Value uMinusOneEqNegOne = builder.create( arith::CmpFPredicate::OEQ, uMinusOne, cstNegOne); // logU = log(u) ~= x Value logU = builder.create(u); // Detect exp(x) = +inf; written this way to avoid having to form +inf. Value isInf = builder.create(arith::CmpFPredicate::OEQ, logU, u); // (u - 1) * (x / ~x) Value expm1 = builder.create( uMinusOne, builder.create(x, logU)); expm1 = builder.create(isInf, u, expm1); Value approximation = builder.create( uEqOneOrNaN, x, builder.create(uMinusOneEqNegOne, cstNegOne, expm1)); rewriter.replaceOp(op, approximation); return success(); } //----------------------------------------------------------------------------// // Sin and Cos approximation. //----------------------------------------------------------------------------// namespace { template struct SinAndCosApproximation : public OpRewritePattern { public: using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(OpTy op, PatternRewriter &rewriter) const final; }; } // namespace #define TWO_OVER_PI \ 0.6366197723675813430755350534900574481378385829618257949906693762L #define PI_OVER_2 \ 1.5707963267948966192313216916397514420985846996875529104874722961L // Approximates sin(x) or cos(x) by finding the best approximation polynomial in // the reduced range [0, pi/2] for both sin(x) and cos(x). Then given y in the // reduced range sin(x) will be computed as sin(y), -sin(y), cos(y) or -cos(y). template LogicalResult SinAndCosApproximation::matchAndRewrite( OpTy op, PatternRewriter &rewriter) const { static_assert( llvm::is_one_of::value, "SinAndCosApproximation pattern expects math::SinOp or math::CosOp"); if (!getElementTypeOrSelf(op.getOperand()).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ArrayRef shape = vectorShape(op.getOperand()); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; auto mul = [&](Value a, Value b) -> Value { return builder.create(a, b); }; auto sub = [&](Value a, Value b) -> Value { return builder.create(a, b); }; auto floor = [&](Value a) { return builder.create(a); }; auto i32Vec = broadcast(builder.getI32Type(), shape); auto fPToSingedInteger = [&](Value a) -> Value { return builder.create(i32Vec, a); }; auto modulo4 = [&](Value a) -> Value { return builder.create(a, bcast(i32Cst(builder, 3))); }; auto isEqualTo = [&](Value a, Value b) -> Value { return builder.create(arith::CmpIPredicate::eq, a, b); }; auto isGreaterThan = [&](Value a, Value b) -> Value { return builder.create(arith::CmpIPredicate::sgt, a, b); }; auto select = [&](Value cond, Value t, Value f) -> Value { return builder.create(cond, t, f); }; auto fmla = [&](Value a, Value b, Value c) { return builder.create(a, b, c); }; auto bitwiseOr = [&](Value a, Value b) { return builder.create(a, b); }; Value twoOverPi = bcast(f32Cst(builder, (float)TWO_OVER_PI)); Value piOverTwo = bcast(f32Cst(builder, (float)PI_OVER_2)); Value x = op.getOperand(); Value k = floor(mul(x, twoOverPi)); Value y = sub(x, mul(k, piOverTwo)); Value cstOne = bcast(f32Cst(builder, 1.0)); Value cstNegativeOne = bcast(f32Cst(builder, -1.0)); Value cstSC2 = bcast(f32Cst(builder, -0.16666667163372039794921875f)); Value cstSC4 = bcast(f32Cst(builder, 8.333347737789154052734375e-3f)); Value cstSC6 = bcast(f32Cst(builder, -1.9842604524455964565277099609375e-4f)); Value cstSC8 = bcast(f32Cst(builder, 2.760012648650445044040679931640625e-6f)); Value cstSC10 = bcast(f32Cst(builder, -2.50293279435709337121807038784027099609375e-8f)); Value cstCC2 = bcast(f32Cst(builder, -0.5f)); Value cstCC4 = bcast(f32Cst(builder, 4.166664183139801025390625e-2f)); Value cstCC6 = bcast(f32Cst(builder, -1.388833043165504932403564453125e-3f)); Value cstCC8 = bcast(f32Cst(builder, 2.47562347794882953166961669921875e-5f)); Value cstCC10 = bcast(f32Cst(builder, -2.59630184018533327616751194000244140625e-7f)); Value kMod4 = modulo4(fPToSingedInteger(k)); Value kR0 = isEqualTo(kMod4, bcast(i32Cst(builder, 0))); Value kR1 = isEqualTo(kMod4, bcast(i32Cst(builder, 1))); Value kR2 = isEqualTo(kMod4, bcast(i32Cst(builder, 2))); Value kR3 = isEqualTo(kMod4, bcast(i32Cst(builder, 3))); Value sinuseCos = isSine ? bitwiseOr(kR1, kR3) : bitwiseOr(kR0, kR2); Value negativeRange = isSine ? isGreaterThan(kMod4, bcast(i32Cst(builder, 1))) : bitwiseOr(kR1, kR2); Value y2 = mul(y, y); Value base = select(sinuseCos, cstOne, y); Value cstC2 = select(sinuseCos, cstCC2, cstSC2); Value cstC4 = select(sinuseCos, cstCC4, cstSC4); Value cstC6 = select(sinuseCos, cstCC6, cstSC6); Value cstC8 = select(sinuseCos, cstCC8, cstSC8); Value cstC10 = select(sinuseCos, cstCC10, cstSC10); Value v1 = fmla(y2, cstC10, cstC8); Value v2 = fmla(y2, v1, cstC6); Value v3 = fmla(y2, v2, cstC4); Value v4 = fmla(y2, v3, cstC2); Value v5 = fmla(y2, v4, cstOne); Value v6 = mul(base, v5); Value approximation = select(negativeRange, mul(cstNegativeOne, v6), v6); rewriter.replaceOp(op, approximation); return success(); } //----------------------------------------------------------------------------// // Cbrt approximation. //----------------------------------------------------------------------------// namespace { struct CbrtApproximation : public OpRewritePattern { using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::CbrtOp op, PatternRewriter &rewriter) const final; }; } // namespace // Estimation of cube-root using an algorithm defined in // Hacker's Delight 2nd Edition. LogicalResult CbrtApproximation::matchAndRewrite(math::CbrtOp op, PatternRewriter &rewriter) const { auto operand = op.getOperand(); if (!getElementTypeOrSelf(operand).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ImplicitLocOpBuilder b(op->getLoc(), rewriter); ArrayRef shape = vectorShape(operand); Type floatTy = getElementTypeOrSelf(operand.getType()); Type intTy = b.getIntegerType(floatTy.getIntOrFloatBitWidth()); // Convert to vector types if necessary. floatTy = broadcast(floatTy, shape); intTy = broadcast(intTy, shape); auto bconst = [&](TypedAttr attr) -> Value { Value value = b.create(attr); return broadcast(b, value, shape); }; // Declare the initial values: Value intTwo = bconst(b.getI32IntegerAttr(2)); Value intFour = bconst(b.getI32IntegerAttr(4)); Value intEight = bconst(b.getI32IntegerAttr(8)); Value intMagic = bconst(b.getI32IntegerAttr(0x2a5137a0)); Value fpThird = bconst(b.getF32FloatAttr(0.33333333f)); Value fpTwo = bconst(b.getF32FloatAttr(2.0f)); Value fpZero = bconst(b.getF32FloatAttr(0.0f)); // Compute an approximation of one third: // union {int ix; float x;}; // x = x0; // ix = ix/4 + ix/16; Value absValue = b.create(operand); Value intValue = b.create(intTy, absValue); Value divideBy4 = b.create(intValue, intTwo); Value divideBy16 = b.create(intValue, intFour); intValue = b.create(divideBy4, divideBy16); // ix = ix + ix/16; divideBy16 = b.create(intValue, intFour); intValue = b.create(intValue, divideBy16); // ix = ix + ix/256; Value divideBy256 = b.create(intValue, intEight); intValue = b.create(intValue, divideBy256); // ix = 0x2a5137a0 + ix; intValue = b.create(intValue, intMagic); // Perform one newtons step: // x = 0.33333333f*(2.0f*x + x0/(x*x)); Value floatValue = b.create(floatTy, intValue); Value squared = b.create(floatValue, floatValue); Value mulTwo = b.create(floatValue, fpTwo); Value divSquared = b.create(absValue, squared); floatValue = b.create(mulTwo, divSquared); floatValue = b.create(floatValue, fpThird); // x = 0.33333333f*(2.0f*x + x0/(x*x)); squared = b.create(floatValue, floatValue); mulTwo = b.create(floatValue, fpTwo); divSquared = b.create(absValue, squared); floatValue = b.create(mulTwo, divSquared); floatValue = b.create(floatValue, fpThird); // Check for zero and restore sign. Value isZero = b.create(arith::CmpFPredicate::OEQ, absValue, fpZero); floatValue = b.create(isZero, fpZero, floatValue); floatValue = b.create(floatValue, operand); rewriter.replaceOp(op, floatValue); return success(); } //----------------------------------------------------------------------------// // Rsqrt approximation. //----------------------------------------------------------------------------// namespace { struct RsqrtApproximation : public OpRewritePattern { using OpRewritePattern::OpRewritePattern; LogicalResult matchAndRewrite(math::RsqrtOp op, PatternRewriter &rewriter) const final; }; } // namespace LogicalResult RsqrtApproximation::matchAndRewrite(math::RsqrtOp op, PatternRewriter &rewriter) const { if (!getElementTypeOrSelf(op.getOperand()).isF32()) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ArrayRef shape = vectorShape(op.getOperand()); // Only support already-vectorized rsqrt's. if (shape.empty() || shape.back() % 8 != 0) return rewriter.notifyMatchFailure(op, "unsupported operand type"); ImplicitLocOpBuilder builder(op->getLoc(), rewriter); auto bcast = [&](Value value) -> Value { return broadcast(builder, value, shape); }; Value cstPosInf = bcast(f32FromBits(builder, 0x7f800000u)); Value cstOnePointFive = bcast(f32Cst(builder, 1.5f)); Value cstNegHalf = bcast(f32Cst(builder, -0.5f)); Value cstMinNormPos = bcast(f32FromBits(builder, 0x00800000u)); Value negHalf = builder.create(op.getOperand(), cstNegHalf); // Select only the inverse sqrt of positive normals (denormals are // flushed to zero). Value ltMinMask = builder.create( arith::CmpFPredicate::OLT, op.getOperand(), cstMinNormPos); Value infMask = builder.create(arith::CmpFPredicate::OEQ, op.getOperand(), cstPosInf); Value notNormalFiniteMask = builder.create(ltMinMask, infMask); // Compute an approximate result. Value yApprox = handleMultidimensionalVectors( builder, op->getOperands(), 8, [&builder](ValueRange operands) -> Value { return builder.create(operands); }); // Do a single step of Newton-Raphson iteration to improve the approximation. // This uses the formula y_{n+1} = y_n * (1.5 - y_n * (0.5 * x) * y_n). // It is essential to evaluate the inner term like this because forming // y_n^2 may over- or underflow. Value inner = builder.create(negHalf, yApprox); Value fma = builder.create(yApprox, inner, cstOnePointFive); Value yNewton = builder.create(yApprox, fma); // Select the result of the Newton-Raphson step for positive normal arguments. // For other arguments, choose the output of the intrinsic. This will // return rsqrt(+inf) = 0, rsqrt(x) = NaN if x < 0, and rsqrt(x) = +inf if // x is zero or a positive denormalized float (equivalent to flushing positive // denormalized inputs to zero). Value res = builder.create(notNormalFiniteMask, yApprox, yNewton); rewriter.replaceOp(op, res); return success(); } //----------------------------------------------------------------------------// void mlir::populateMathPolynomialApproximationPatterns( RewritePatternSet &patterns, const MathPolynomialApproximationOptions &options) { // Patterns for leveraging existing f32 lowerings on other data types. patterns .add, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion, ReuseF32Expansion>( patterns.getContext()); patterns.add, SinAndCosApproximation>( patterns.getContext()); if (options.enableAvx2) { patterns.add>( patterns.getContext()); } }