jefflim-google/highway

Performance-portable SIMD with runtime dispatch

Efficient and performance-portable SIMD wrapper

This library provides type-safe and source-code portable wrappers over existing platform-specific intrinsics. Its design aims for simplicity, reliable efficiency across platforms, and immediate usability with current compilers.

Current status

Implemented for scalar/SSE4/AVX2/AVX-512/ARMv8 targets, each with unit tests.

Tested on GCC 8.3.0 and Clang 6.0.1.

A quick-reference page briefly lists all operations and their parameters.

Design philosophy

• Performance is important but not the sole consideration. Anyone who goes to the trouble of using SIMD clearly cares about speed. However, portability, maintainability and readability also matter, otherwise we would write in assembly. We aim for performance within 10-20% of a hand-written assembly implementation on the development platform.

• The guiding principles of C++ are "pay only for what you use" and "leave no room for a lower-level language below C++". We apply these by defining a SIMD API that ensures operation costs are visible, predictable and minimal.

• Performance portability is important, i.e. the API should be efficient on all target platforms. Unfortunately, common idioms for one platform can be inefficient on others. For example: summing lanes horizontally versus shuffling. Documenting which operations are expensive does not prevent their use, as evidenced by widespread use of HADDPS. Performance acceptance tests may detect large regressions, but do not help choose the approach during initial development. Analysis tools can warn about some potential inefficiencies, but likely not all. We instead provide a carefully chosen set of vector types and operations that are efficient on all target platforms (PPC8, SSE4/AVX2+, ARMv8), plus some useful but less performance-portable operations in an ext namespace to make their cost visible.

• Future SIMD hardware features are difficult to predict. For example, AVX2 came with surprising semantics (almost no interaction between 128-bit blocks) and AVX-512 added two kinds of predicates (writemask and zeromask). To ensure the API reflects hardware realities, we suggest a flexible approach that adds new operations as they become commonly available.

• Masking is not yet widely supported on current CPUs. It is difficult to define an interface that provides access to all platform features while retaining performance portability. The P0214R5 proposal lacks support for AVX-512/ARM SVE zeromasks. We suggest limiting usage of masks to the IfThen[Zero]Else[Zero] functions until the community has gained more experience with them.

• "Width-agnostic" SIMD is more future-proof than user-specified fixed sizes. For example, valarray-like code can iterate over a 1D array with a library-specified vector width. This will result in better code when vector sizes increase, and matches the direction taken by ARM SVE and RiscV hardware as well as Agner Fog's ForwardCom instruction set proposal. However, some applications may require fixed sizes, so we also guarantee support for 128-bit vectors in each instruction set.

• The API and its implementation should be usable and efficient with commonly used compilers. Some of our open-source users cannot upgrade, so we need to support ~4 year old compilers. For example, we write ShiftLeft<3>(v) instead of v << 3 because MSVC 2017 (ARM64) does not propagate the literal (https://godbolt.org/g/rKx5Ga). However, we do require function-specific target attributes, supported by GCC 4.9 / Clang 3.9 / MSVC 2015.

• Efficient and safe runtime dispatch is important. Modules such as image or video codecs are typically embedded into larger applications such as browsers, so they cannot require separate binaries for each CPU. Libraries also cannot predict whether the application already uses AVX2 (and pays the frequency throttling cost), so this decision must be left to the application. Using only the lowest-common denominator instructions sacrifices too much performance. Therefore, we provide code paths for multiple instruction sets and choose the most suitable at runtime. To reduce overhead, dispatch should be hoisted to higher layers instead of checking inside every low-level function. Generating each code path from the same source reduces implementation- and debugging cost.

• Not every CPU need be supported. For example, pre-SSE4.1 CPUs are increasingly rare and the AVX instruction set is limited to floating-point operations. To reduce code size and compile time, we provide specializations for SSE4, AVX2 and AVX-512 instruction sets on x86.

• Access to platform-specific intrinsics is necessary for acceptance in performance-critical projects. We provide conversions to and from intrinsics to allow utilizing specialized platform-specific functionality, and simplify incremental porting of existing code.

• The core API should be compact and easy to learn. We provide only the few dozen operations which are necessary and sufficient for most of the 150+ SIMD applications we examined.

Differences versus P0214R5 proposal

1. Adding widely used and portable operations such as AndNot, AverageRound, bit-shift by immediates and IfThenElse.

2. Designing the API to avoid or minimize overhead on AVX2/AVX-512 caused by crossing 128-bit 'block' boundaries.

3. Avoiding the need for non-native vectors. By contrast, P0214R5's simd_cast returns fixed_size<> vectors which are more expensive to access because they reside on the stack. We can avoid this plus additional overhead on ARM/AVX2 by defining width-expanding operations as functions of a vector part, e.g. promoting half a vector of uint8_t lanes to one full vector of uint16_t, or demoting full vectors to half vectors with half-width lanes.

4. Guaranteeing access to the underlying intrinsic vector type. This ensures all platform-specific capabilities can be used. P0214R5 instead only 'encourages' implementations to provide access.

5. Enabling safe runtime dispatch and inlining in the same binary. P0214R5 is based on the Vc library, which does not provide assistance for linking multiple instruction sets into the same binary. The Vc documentation suggests compiling separate executables for each instruction set or using GCC's ifunc (indirect functions). The latter is compiler-specific and risks crashes due to ODR violations when compiling the same function with different compiler flags. We solve this problem via target-specific attributes (see HOWTO section below). We also permit a mix of static target selection and runtime dispatch for hotspots that may benefit from newer instruction sets if available.

6. Using built-in PPC vector types without a wrapper class. This leads to much better code generation with GCC 6.3: https://godbolt.org/z/pd2PNP. By contrast, P0214R5 requires a wrapper. We avoid this by using only the member operators provided by the PPC vectors; all other functions and typedefs are non-members. 2019-04 update: Clang power64le does not have this issue, so we simplified get_part(d, v) to GetLane(v).

• Omitting inefficient or non-performance-portable operations such as hmax, operator[], and unsupported integer comparisons. Applications can often replace these operations at lower cost than emulating that exact behavior.

• Omitting long double types: these are not commonly available in hardware.

• Ensuring signed integer overflow has well-defined semantics (wraparound).

• Simple header-only implementation and less than a tenth of the size of the Vc library from which P0214 was derived (98,000 lines in https://github.com/VcDevel/Vc according to the gloc Chrome extension).

• Avoiding hidden performance costs. P0214R5 allows implicit conversions from integer to float, which costs 3-4 cycles on x86. We make these conversions explicit to ensure their cost is visible.

Prior API designs

The author has been writing SIMD code since 2002: first via assembly language, then intrinsics, later Intel's F32vec4 wrapper, followed by three generations of custom vector classes. The first used macros to generate the classes, which reduces duplication but also readability. The second used templates instead. The third (used in highwayhash and PIK) added support for AVX2 and runtime dispatch. The current design (used in JPEG XL) enables code generation for multiple platforms and/or instruction sets from the same source, and improves runtime dispatch.

Other related work

• Neat SIMD adopts a similar approach with interchangeable vector/scalar types and a compact interface. It allows access to the underlying intrinsics, but does not appear to be designed for other platforms than x86.

• UME::SIMD (code, paper) also adopts an explicit vectorization model with vector classes. However, it exposes the union of all platform capabilities, which makes the API harder to learn (209-page spec) and implement (the estimated LOC count is 500K). The API is less performance-portable because it allows applications to use operations that are inefficient on other platforms.

• Inastemp (code, paper) is a vector library for scientific computing with some innovative features: automatic FLOPS counting, and "if/else branches" using lambda functions. It supports IBM Power8, but only provides float and double types.

Overloaded function API

Most C++ vector APIs rely on class templates. However, two PPC compilers including GCC 6.3 generate inefficient code for classes with a SIMD vector member: an extra load/store for every function argument/return value. To avoid this overhead, we use built-in vector types on PPC. These provide overloaded arithmetic operators but do not allow member functions/typedefs such as size() or value_type. We instead rely on overloaded functions.

Because vectors of various lengths are the same type on PPC, we need an additional tag argument for disambiguation. Any function template with multiple return types uses a descriptor argument to specify the return type. For example, the return type of Zero(Desc<T, N>) is VT<Desc<T, N>>. For brevity, Desc is abbreviated to D for template arguments and d in lvalues.

It may seem preferable to write Zero<D>() rather than Zero(D()), but there are technical difficulties. We prefer generic implementations where possible rather than overloading for every single T. Because C++ does not allow partial specialization of function templates, we need multiple overloads: one primary template per target. Thus, functions cannot be invoked using template syntax. Can we instead add a wrapper function template that calls the appropriate overload? Unfortunately, the compiler mechanism for avoiding dangerous per-file -mavx2 requires per-function annotations, and these attributes are not generic. Thus, a wrapper into which SIMD functions are inlined cannot be a function, because it would also need a target-specific attribute. A macro ZERO(D) could work, but this is hardly more clear than a normal function with arguments. Note that descriptors occur often, so user code can define a const HWY_FULL(float) d; and then write Zero(d).

Masks

AVX-512 introduced a major change to the SIMD interface: special mask registers (one bit per lane) that serve as predicates. It would be expensive to force AVX-512 implementations to conform to the prior model of full vectors with lanes set to all one or all zero bits. We instead provide a Mask type that emulates a subset of this functionality on other platforms at zero cost.

Masks are returned by comparisons and TestBit; they serve as the input to IfThen*. We provide conversions between masks and vector lanes. For clarity and safety, we use FF..FF as the definition of true. To also benefit from x86 instructions that only require the sign bit of floating-point inputs to be set, we provide a special ZeroIfNegative function.

Use cases and HOWTO

Whenever possible, use full-width descriptors for maximum performance portability: HWY_FULL(float). When necessary, applications may also rely on 128-bit vectors: Desc<float, 4>. There is also the option of a vector of up to N lanes: HWY_CAPPED(float, N) (the length is always a power of two).

• Single instruction set per platform: use normal C++ functions with HWY_ATTR annotation and #include "static_targets.h". Note that this can co-exist with runtime dispatch in the same translation unit.

• Runtime dispatch means calling the best available implementation:

  #include "jxl/hwy/runtime_dispatch.h"
  Dispatch(TargetBitfield().Best(), Module()/*, args*/);

The module is declared in a header as a non-const member function:

  struct Module {
    HWY_DECLARE(void, (args))
  };

The module.cc is automatically compiled multiple times, once per target CPU:

  #ifndef HWY_TARGET_INCLUDE
  #define HWY_TARGET_INCLUDE "path/module.cc"
  #include "path/module.h"
  #endif  // end of non-SIMD code

  #include "jxl/hwy/foreach_target.h"

  namespace HWY_NAMESPACE { namespace {
    HWY_ATTR void Implementation() {}
  }}
  HWY_ATTR void Module::HWY_FUNC(/*args*/) {
    HWY_NAMESPACE::Implementation();
  }

Example source code

void FloorLog2(const uint8_t* HWY_RESTRICT values,
               uint8_t* HWY_RESTRICT log2) {
  // Descriptors for all required data types:
  const HWY_FULL(int32_t) d32;
  const HWY_FULL(float) df;
  const HWY_CAPPED(uint8_t, d32.N) d8;

  const auto u8 = Load(d8, values);
  const auto bits = BitCast(d32, ConvertTo(df, ConvertTo(d32, u8)));
  const auto exponent = ShiftRight<23>(bits) - Set(d32, 127);
  Store(ConvertTo(d8, exponent), d8, log2);
}

This generates the following SSE4 and AVX2 code, as shown by IACA:

 p0  p1  p5
|   |   | 1 | CP | pmovzxbd xmm1, dword [rsp+0x25c]
|   | 1 |   |    | cvtdq2ps xmm1, xmm1
| 1 |   |   |    | psrad xmm1, 0x17
|   | 1 |   |    | paddd xmm1, xmm0
|   |   | 1 | CP | packusdw xmm1, xmm0
|   |   | 1 | CP | packuswb xmm1, xmm0
|   |   |   |    | movd [rsp+0x45c], xmm1

|   |   | 1 | CP | vpmovzxbd ymm1, qword [rsp+0x228]
|   | 1 |   |    | vcvtdq2ps ymm1, ymm1
| 1 |   |   |    | vpsrad ymm1, ymm1, 0x17
|   | 1 |   |    | vpaddd ymm1, ymm1, ymm0
|   |   | 1 | CP | vpackusdw ymm1, ymm1, ymm0
|   |   | 1 | CP | vpermq ymm1, ymm1, 0xe8
|   |   | 1 | CP | vpackuswb xmm1, xmm1, xmm0
|   |   |   |    | vmovq [rsp+0x448], xmm1

void Copy(const uint8_t* HWY_RESTRICT from, const size_t size,
          uint8_t* HWY_RESTRICT to) {
  // Width-agnostic (library-specified N)
  const HWY_FULL(uint8_t) d;
  const Scalar<uint8_t> ds;
  size_t i = 0;
  for (; i + d.N <= size; i += d.N) {
    const auto bytes = Load(d, from + i);
    Store(bytes, d, to + i);
  }

  for (; i < size; ++i) {
    // (Same loop body as above, could factor into a shared template)
    const auto bytes = Load(ds, from + i);
    Store(bytes, ds, to + i);
  }
}

void MulAdd(const T* HWY_RESTRICT mul_array, const T* HWY_RESTRICT add_array,
            const size_t size, T* HWY_RESTRICT x_array) {
  // Type-agnostic (caller-specified lane type) and width-agnostic (uses
  // best available instruction set).
  const HWY_FULL(T) d;
  for (size_t i = 0; i < size; i += d.N) {
    const auto mul = Load(d, mul_array + i);
    const auto add = Load(d, add_array + i);
    auto x = Load(d, x_array + i);
    x = MulAdd(mul, x, add);
    Store(x, d, x_array + i);
  }
}

Additional resources

• Overview of instructions per operation on different architectures

This is not an officially supported Google product. Contact: [email protected]