diff --git a/mlir/include/mlir/ExecutionEngine/Float16bits.h b/mlir/include/mlir/ExecutionEngine/Float16bits.h new file mode 100644 --- /dev/null +++ b/mlir/include/mlir/ExecutionEngine/Float16bits.h @@ -0,0 +1,39 @@ +//===--- Float16bits.h - supports 2-byte floats ---------------------------===// +// +// 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 f16 and bf16 to support the compilation and execution +// of programs using these types. +// +//===----------------------------------------------------------------------===// + +#ifndef MLIR_EXECUTIONENGINE_FLOAT16BITS_H_ +#define MLIR_EXECUTIONENGINE_FLOAT16BITS_H_ + +#include +#include + +// Implements half precision and bfloat with f16 and bf16, using the MLIR type +// names. These data types are also used for c-interface runtime routines. +extern "C" { +struct f16 { + f16(float f = 0); + uint16_t bits; +}; + +struct bf16 { + bf16(float f = 0); + uint16_t bits; +}; +} + +// Outputs a half precision value. +std::ostream &operator<<(std::ostream &os, const f16 &f); +// Outputs a bfloat value. +std::ostream &operator<<(std::ostream &os, const bf16 &d); + +#endif // MLIR_EXECUTIONENGINE_FLOAT16BITS_H_ diff --git a/mlir/include/mlir/ExecutionEngine/SparseTensorUtils.h b/mlir/include/mlir/ExecutionEngine/SparseTensorUtils.h --- a/mlir/include/mlir/ExecutionEngine/SparseTensorUtils.h +++ b/mlir/include/mlir/ExecutionEngine/SparseTensorUtils.h @@ -15,6 +15,7 @@ #define MLIR_EXECUTIONENGINE_SPARSETENSORUTILS_H_ #include "mlir/ExecutionEngine/CRunnerUtils.h" +#include "mlir/ExecutionEngine/Float16bits.h" #include #include @@ -77,12 +78,14 @@ enum class PrimaryType : uint32_t { kF64 = 1, kF32 = 2, - kI64 = 3, - kI32 = 4, - kI16 = 5, - kI8 = 6, - kC64 = 7, - kC32 = 8 + kF16 = 3, + kBF16 = 4, + kI64 = 5, + kI32 = 6, + kI16 = 7, + kI8 = 8, + kC64 = 9, + kC32 = 10 }; // This x-macro only specifies the non-complex `V` types, because the ABI @@ -97,6 +100,8 @@ #define FOREVERY_SIMPLEX_V(DO) \ DO(F64, double) \ DO(F32, float) \ + DO(F16, f16) \ + DO(BF16, bf16) \ DO(I64, int64_t) \ DO(I32, int32_t) \ DO(I16, int16_t) \ diff --git a/mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp b/mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp --- a/mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp +++ b/mlir/lib/Dialect/SparseTensor/Transforms/CodegenUtils.cpp @@ -104,6 +104,10 @@ return PrimaryType::kF64; if (elemTp.isF32()) return PrimaryType::kF32; + if (elemTp.isF16()) + return PrimaryType::kF16; + if (elemTp.isBF16()) + return PrimaryType::kBF16; if (elemTp.isInteger(64)) return PrimaryType::kI64; if (elemTp.isInteger(32)) diff --git a/mlir/lib/ExecutionEngine/CMakeLists.txt b/mlir/lib/ExecutionEngine/CMakeLists.txt --- a/mlir/lib/ExecutionEngine/CMakeLists.txt +++ b/mlir/lib/ExecutionEngine/CMakeLists.txt @@ -7,6 +7,7 @@ CudaRuntimeWrappers.cpp SparseTensorUtils.cpp ExecutionEngine.cpp + Float16bits.cpp RocmRuntimeWrappers.cpp RunnerUtils.cpp OptUtils.cpp @@ -121,6 +122,7 @@ add_mlir_library(mlir_c_runner_utils SHARED CRunnerUtils.cpp + Float16bits.cpp SparseTensorUtils.cpp EXCLUDE_FROM_LIBMLIR diff --git a/mlir/lib/ExecutionEngine/Float16bits.cpp b/mlir/lib/ExecutionEngine/Float16bits.cpp new file mode 100644 --- /dev/null +++ b/mlir/lib/ExecutionEngine/Float16bits.cpp @@ -0,0 +1,143 @@ +//===--- Float16bits.cpp - supports 2-byte floats ------------------------===// +// +// 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 f16 and bf16 to support the compilation and execution +// of programs using these types. +// +//===----------------------------------------------------------------------===// + +#include "mlir/ExecutionEngine/Float16bits.h" + +namespace { + +// Union used to make the int/float aliasing explicit so we can access the raw +// bits. +union Float32Bits { + uint32_t u; + float f; +}; + +const uint32_t kF32MantiBits = 23; +const uint32_t kF32HalfMantiBitDiff = 13; +const uint32_t kF32HalfBitDiff = 16; +const Float32Bits kF32Magic = {113 << kF32MantiBits}; +const uint32_t kF32HalfExpAdjust = (127 - 15) << kF32MantiBits; + +// Constructs the 16 bit representation for a half precision value from a float +// value. This implementation is adapted from Eigen. +uint16_t float2half(float floatValue) { + const Float32Bits inf = {255 << kF32MantiBits}; + const Float32Bits f16max = {(127 + 16) << kF32MantiBits}; + const Float32Bits denormMagic = {((127 - 15) + (kF32MantiBits - 10) + 1) + << kF32MantiBits}; + uint32_t signMask = 0x80000000u; + uint16_t halfValue = static_cast(0x0u); + Float32Bits f; + f.f = floatValue; + uint32_t sign = f.u & signMask; + f.u ^= sign; + + if (f.u >= f16max.u) { + const uint32_t halfQnan = 0x7e00; + const uint32_t halfInf = 0x7c00; + // Inf or NaN (all exponent bits set). + halfValue = (f.u > inf.u) ? halfQnan : halfInf; // NaN->qNaN and Inf->Inf + } else { + // (De)normalized number or zero. + if (f.u < kF32Magic.u) { + // The resulting FP16 is subnormal or zero. + // + // Use a magic value to align our 10 mantissa bits at the bottom of the + // float. As long as FP addition is round-to-nearest-even this works. + f.f += denormMagic.f; + + halfValue = static_cast(f.u - denormMagic.u); + } else { + uint32_t mantOdd = + (f.u >> kF32HalfMantiBitDiff) & 1; // Resulting mantissa is odd. + + // Update exponent, rounding bias part 1. The following expressions are + // equivalent to `f.u += ((unsigned int)(15 - 127) << kF32MantiBits) + + // 0xfff`, but without arithmetic overflow. + f.u += 0xc8000fffU; + // Rounding bias part 2. + f.u += mantOdd; + halfValue = static_cast(f.u >> kF32HalfMantiBitDiff); + } + } + + halfValue |= static_cast(sign >> kF32HalfBitDiff); + return halfValue; +} + +// Converts the 16 bit representation of a half precision value to a float +// value. This implementation is adapted from Eigen. +float half2float(uint16_t halfValue) { + const uint32_t shiftedExp = + 0x7c00 << kF32HalfMantiBitDiff; // Exponent mask after shift. + + // Initialize the float representation with the exponent/mantissa bits. + Float32Bits f = { + static_cast((halfValue & 0x7fff) << kF32HalfMantiBitDiff)}; + const uint32_t exp = shiftedExp & f.u; + f.u += kF32HalfExpAdjust; // Adjust the exponent + + // Handle exponent special cases. + if (exp == shiftedExp) { + // Inf/NaN + f.u += kF32HalfExpAdjust; + } else if (exp == 0) { + // Zero/Denormal? + f.u += 1 << kF32MantiBits; + f.f -= kF32Magic.f; + } + + f.u |= (halfValue & 0x8000) << kF32HalfBitDiff; // Sign bit. + return f.f; +} + +const uint32_t kF32BfMantiBitDiff = 16; + +// Constructs the 16 bit representation for a bfloat value from a float value. +// This implementation is adapted from Eigen. +uint16_t float2bfloat(float floatValue) { + Float32Bits floatBits; + floatBits.f = floatValue; + uint16_t bfloatBits; + + // Least significant bit of resulting bfloat. + uint32_t lsb = (floatBits.u >> kF32BfMantiBitDiff) & 1; + uint32_t rounding_bias = 0x7fff + lsb; + floatBits.u += rounding_bias; + bfloatBits = static_cast(floatBits.u >> kF32BfMantiBitDiff); + return bfloatBits; +} + +// Converts the 16 bit representation of a bfloat value to a float value. This +// implementation is adapted from Eigen. +float bfloat2float(uint16_t bfloatBits) { + Float32Bits floatBits; + floatBits.u = static_cast(bfloatBits) << kF32BfMantiBitDiff; + return floatBits.f; +} + +} // namespace + +f16::f16(float f) : bits(float2half(f)) {} + +bf16::bf16(float f) : bits(float2bfloat(f)) {} + +std::ostream &operator<<(std::ostream &os, const f16 &f) { + os << half2float(f.bits); + return os; +} + +std::ostream &operator<<(std::ostream &os, const bf16 &d) { + os << bfloat2float(d.bits); + return os; +} diff --git a/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp b/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp --- a/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp +++ b/mlir/lib/ExecutionEngine/SparseTensorUtils.cpp @@ -1567,6 +1567,16 @@ CASE(OverheadType::kU8, OverheadType::kU8, PrimaryType::kF32, uint8_t, uint8_t, float); + // Two-byte floats with both overheads of the same type. + CASE_SECSAME(OverheadType::kU64, PrimaryType::kF16, uint64_t, f16); + CASE_SECSAME(OverheadType::kU64, PrimaryType::kBF16, uint64_t, bf16); + CASE_SECSAME(OverheadType::kU32, PrimaryType::kF16, uint32_t, f16); + CASE_SECSAME(OverheadType::kU32, PrimaryType::kBF16, uint32_t, bf16); + CASE_SECSAME(OverheadType::kU16, PrimaryType::kF16, uint16_t, f16); + CASE_SECSAME(OverheadType::kU16, PrimaryType::kBF16, uint16_t, bf16); + CASE_SECSAME(OverheadType::kU8, PrimaryType::kF16, uint8_t, f16); + CASE_SECSAME(OverheadType::kU8, PrimaryType::kBF16, uint8_t, bf16); + // Integral matrices with both overheads of the same type. CASE_SECSAME(OverheadType::kU64, PrimaryType::kI64, uint64_t, int64_t); CASE_SECSAME(OverheadType::kU64, PrimaryType::kI32, uint64_t, int32_t); diff --git a/mlir/test/Dialect/SparseTensor/conversion_sparse2dense.mlir b/mlir/test/Dialect/SparseTensor/conversion_sparse2dense.mlir --- a/mlir/test/Dialect/SparseTensor/conversion_sparse2dense.mlir +++ b/mlir/test/Dialect/SparseTensor/conversion_sparse2dense.mlir @@ -28,9 +28,8 @@ // CHECK-DAG: %[[PermD:.*]] = memref.cast %[[PermS]] : memref<1xindex> to memref // CHECK-DAG: memref.store %[[I0]], %[[PermS]][%[[I0]]] : memref<1xindex> // CHECK-DAG: %[[zeroI32:.*]] = arith.constant 0 : i32 -// CHECK-DAG: %[[ElemTp:.*]] = arith.constant 4 : i32 -// CHECK-DAG: %[[ActionToIter:.*]] = arith.constant 6 : i32 -// CHECK-DAG: %[[Iter:.*]] = call @newSparseTensor(%[[AttrsD]], %[[SizesD]], %[[PermD]], %[[zeroI32]], %[[zeroI32]], %[[ElemTp]], %[[ActionToIter]], %[[Arg]]) : (memref, memref, memref, i32, i32, i32, i32, !llvm.ptr) -> !llvm.ptr +// CHECK-DAG: %[[ElemTpActionToIter:.*]] = arith.constant 6 : i32 +// CHECK-DAG: %[[Iter:.*]] = call @newSparseTensor(%[[AttrsD]], %[[SizesD]], %[[PermD]], %[[zeroI32]], %[[zeroI32]], %[[ElemTpActionToIter]], %[[ElemTpActionToIter]], %[[Arg]]) : (memref, memref, memref, i32, i32, i32, i32, !llvm.ptr) -> !llvm.ptr // CHECK-DAG: %[[IndS:.*]] = memref.alloca() : memref<1xindex> // CHECK-DAG: %[[IndD:.*]] = memref.cast %[[IndS]] : memref<1xindex> to memref // CHECK-DAG: %[[ElemBuffer:.*]] = memref.alloca() : memref @@ -67,9 +66,8 @@ // CHECK-DAG: %[[PermD:.*]] = memref.cast %[[PermS]] : memref<1xindex> to memref // CHECK-DAG: memref.store %[[I0]], %[[PermS]][%[[I0]]] : memref<1xindex> // CHECK-DAG: %[[zeroI32:.*]] = arith.constant 0 : i32 -// CHECK-DAG: %[[ElemTp:.*]] = arith.constant 4 : i32 -// CHECK-DAG: %[[ActionToIter:.*]] = arith.constant 6 : i32 -// CHECK-DAG: %[[Iter:.*]] = call @newSparseTensor(%[[AttrsD]], %[[SizesD]], %[[PermD]], %[[zeroI32]], %[[zeroI32]], %[[ElemTp]], %[[ActionToIter]], %[[Arg]]) : (memref, memref, memref, i32, i32, i32, i32, !llvm.ptr) -> !llvm.ptr +// CHECK-DAG: %[[ElemTpActionToIter:.*]] = arith.constant 6 : i32 +// CHECK-DAG: %[[Iter:.*]] = call @newSparseTensor(%[[AttrsD]], %[[SizesD]], %[[PermD]], %[[zeroI32]], %[[zeroI32]], %[[ElemTpActionToIter]], %[[ElemTpActionToIter]], %[[Arg]]) : (memref, memref, memref, i32, i32, i32, i32, !llvm.ptr) -> !llvm.ptr // CHECK-DAG: %[[IndS:.*]] = memref.alloca() : memref<1xindex> // CHECK-DAG: %[[IndD:.*]] = memref.cast %[[IndS]] : memref<1xindex> to memref // CHECK-DAG: %[[ElemBuffer:.*]] = memref.alloca() : memref diff --git a/mlir/test/Integration/Dialect/SparseTensor/CPU/dense_output_f16.mlir b/mlir/test/Integration/Dialect/SparseTensor/CPU/dense_output_f16.mlir new file mode 100644 --- /dev/null +++ b/mlir/test/Integration/Dialect/SparseTensor/CPU/dense_output_f16.mlir @@ -0,0 +1,90 @@ +// RUN: mlir-opt %s --sparse-compiler | \ +// RUN: mlir-cpu-runner \ +// RUN: -e entry -entry-point-result=void \ +// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext | \ +// RUN: FileCheck %s + +#SparseVector = #sparse_tensor.encoding<{dimLevelType = ["compressed"]}> +#DenseVector = #sparse_tensor.encoding<{dimLevelType = ["dense"]}> + +#trait_vec_op = { + indexing_maps = [ + affine_map<(i) -> (i)>, // a (in) + affine_map<(i) -> (i)>, // b (in) + affine_map<(i) -> (i)> // x (out) + ], + iterator_types = ["parallel"] +} + +module { + // Creates a dense vector using the minimum values from two input sparse vectors. + // When there is no overlap, include the present value in the output. + func.func @vector_min(%arga: tensor, + %argb: tensor) -> tensor { + %c = arith.constant 0 : index + %d = tensor.dim %arga, %c : tensor + %xv = bufferization.alloc_tensor (%d) : tensor + %0 = linalg.generic #trait_vec_op + ins(%arga, %argb: tensor, tensor) + outs(%xv: tensor) { + ^bb(%a: f16, %b: f16, %x: f16): + %1 = sparse_tensor.binary %a, %b : f16, f16 to f16 + overlap={ + ^bb0(%a0: f16, %b0: f16): + %cmp = arith.cmpf "olt", %a0, %b0 : f16 + %2 = arith.select %cmp, %a0, %b0: f16 + sparse_tensor.yield %2 : f16 + } + left=identity + right=identity + linalg.yield %1 : f16 + } -> tensor + return %0 : tensor + } + + // Dumps a dense vector of type f16. + func.func @dump_vec(%arg0: tensor) { + // Dump the values array to verify only sparse contents are stored. + %c0 = arith.constant 0 : index + %d0 = arith.constant -1.0 : f16 + %0 = sparse_tensor.values %arg0 : tensor to memref + %1 = vector.transfer_read %0[%c0], %d0: memref, vector<32xf16> + %f1 = arith.extf %1: vector<32xf16> to vector<32xf32> + vector.print %f1 : vector<32xf32> + return + } + + // Driver method to call and verify the kernel. + func.func @entry() { + %c0 = arith.constant 0 : index + + // Setup sparse vectors. + %v1 = arith.constant sparse< + [ [0], [3], [11], [17], [20], [21], [28], [29], [31] ], + [ 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0 ] + > : tensor<32xf16> + %v2 = arith.constant sparse< + [ [1], [3], [4], [10], [16], [18], [21], [28], [29], [31] ], + [11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0, 20.0 ] + > : tensor<32xf16> + %sv1 = sparse_tensor.convert %v1 : tensor<32xf16> to tensor + %sv2 = sparse_tensor.convert %v2 : tensor<32xf16> to tensor + + // Call the sparse vector kernel. + %0 = call @vector_min(%sv1, %sv2) + : (tensor, + tensor) -> tensor + + // + // Verify the result. + // + // CHECK: ( 1, 11, 0, 2, 13, 0, 0, 0, 0, 0, 14, 3, 0, 0, 0, 0, 15, 4, 16, 0, 5, 6, 0, 0, 0, 0, 0, 0, 7, 8, 0, 9 ) + call @dump_vec(%0) : (tensor) -> () + + // Release the resources. + sparse_tensor.release %sv1 : tensor + sparse_tensor.release %sv2 : tensor + sparse_tensor.release %0 : tensor + return + } +} diff --git a/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_sum_f16.mlir b/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_sum_f16.mlir new file mode 100644 --- /dev/null +++ b/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_sum_f16.mlir @@ -0,0 +1,78 @@ + // RUN: mlir-opt %s --sparse-compiler | \ +// RUN: mlir-cpu-runner \ +// RUN: -e entry -entry-point-result=void \ +// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext | \ +// RUN: FileCheck %s + +!Filename = !llvm.ptr + +#SparseMatrix = #sparse_tensor.encoding<{ + dimLevelType = [ "compressed", "compressed" ] +}> + +#trait_sum_reduce = { + indexing_maps = [ + affine_map<(i,j) -> (i,j)>, // A + affine_map<(i,j) -> ()> // x (out) + ], + iterator_types = ["reduction", "reduction"], + doc = "x += A(i,j)" +} + +module { + // + // A kernel that sum-reduces a matrix to a single scalar. + // + func.func @kernel_sum_reduce(%arga: tensor, + %argx: tensor {linalg.inplaceable = true}) -> tensor { + %0 = linalg.generic #trait_sum_reduce + ins(%arga: tensor) + outs(%argx: tensor) { + ^bb(%a: f16, %x: f16): + %0 = arith.addf %x, %a : f16 + linalg.yield %0 : f16 + } -> tensor + return %0 : tensor + } + + func.func private @getTensorFilename(index) -> (!Filename) + + // + // Main driver that reads matrix from file and calls the sparse kernel. + // + func.func @entry() { + // Setup input sparse matrix from compressed constant. + %d = arith.constant dense <[ + [ 1.1, 1.2, 0.0, 1.4 ], + [ 0.0, 0.0, 0.0, 0.0 ], + [ 3.1, 0.0, 3.3, 3.4 ] + ]> : tensor<3x4xf16> + %a = sparse_tensor.convert %d : tensor<3x4xf16> to tensor + + %d0 = arith.constant 0.0 : f16 + // Setup memory for a single reduction scalar, + // initialized to zero. + %xdata = memref.alloc() : memref + memref.store %d0, %xdata[] : memref + %x = bufferization.to_tensor %xdata : memref + + // Call the kernel. + %0 = call @kernel_sum_reduce(%a, %x) + : (tensor, tensor) -> tensor + + // Print the result for verification. + // + // CHECK: 13.5 + // + %m = bufferization.to_memref %0 : memref + %v = memref.load %m[] : memref + %vf = arith.extf %v: f16 to f32 + vector.print %vf : f32 + + // Release the resources. + memref.dealloc %xdata : memref + sparse_tensor.release %a : tensor + + return + } +} diff --git a/utils/bazel/llvm-project-overlay/mlir/BUILD.bazel b/utils/bazel/llvm-project-overlay/mlir/BUILD.bazel --- a/utils/bazel/llvm-project-overlay/mlir/BUILD.bazel +++ b/utils/bazel/llvm-project-overlay/mlir/BUILD.bazel @@ -6415,10 +6415,12 @@ name = "mlir_c_runner_utils", srcs = [ "lib/ExecutionEngine/CRunnerUtils.cpp", + "lib/ExecutionEngine/Float16bits.cpp", "lib/ExecutionEngine/SparseTensorUtils.cpp", ], hdrs = [ "include/mlir/ExecutionEngine/CRunnerUtils.h", + "include/mlir/ExecutionEngine/Float16bits.h", "include/mlir/ExecutionEngine/Msan.h", "include/mlir/ExecutionEngine/SparseTensorUtils.h", ],