bench/utils.h

// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#pragma once

#include <cstddef>
#include <cstdint>

#include <xnnpack.h>
#include <xnnpack/allocator.h>

#include <benchmark/benchmark.h>

namespace benchmark {
namespace utils {

uint32_t WipeCache();
uint32_t PrefetchToL1(const void* ptr, size_t size);

// Disable support for denormalized numbers in floating-point units.
void DisableDenormals();

// Return clock rate, in Hz, for the currently used logical processor.
uint64_t GetCurrentCpuFrequency();

// Return maximum (across all cores/clusters/sockets) last level cache size.
// Can overestimate, but not underestimate LLC size.
size_t GetMaxCacheSize();

// Set number of elements for a unary elementwise microkernel such that:
// - It is divisible by 2, 3, 4, 5, 6.
// - It is divisible by AVX512 width.
// - Total memory footprint does not exceed the characteristic cache size for 
//   the architecture.
template<class InType, class OutType>
void UnaryElementwiseParameters(benchmark::internal::Benchmark* benchmark) {
  benchmark->ArgName("N");

  size_t characteristic_l1 = 32 * 1024;
  size_t characteristic_l2 = 256 * 1024;
#if XNN_ARCH_ARM
  characteristic_l1 = 16 * 1024;
  characteristic_l2 = 128 * 1024;
#endif  // XNN_ARCH_ARM

  const size_t elementwise_size = sizeof(InType) + sizeof(OutType);
  benchmark->Arg(characteristic_l1 / elementwise_size / 960 * 960);
  benchmark->Arg(characteristic_l2 / elementwise_size / 960 * 960);
}

// Set number of elements for a binary elementwise microkernel such that:
// - It is divisible by 2, 3, 4, 5, 6.
// - It is divisible by AVX512 width.
// - Total memory footprint does not exceed the characteristic cache size for 
//   the architecture.
template<class InType, class OutType>
void BinaryElementwiseParameters(benchmark::internal::Benchmark* benchmark) {
  benchmark->ArgName("N");

  size_t characteristic_l1 = 32 * 1024;
  size_t characteristic_l2 = 256 * 1024;
#if XNN_ARCH_ARM
  characteristic_l1 = 16 * 1024;
  characteristic_l2 = 128 * 1024;
#endif  // XNN_ARCH_ARM

  const size_t elementwise_size = 2 * sizeof(InType) + sizeof(OutType);
  benchmark->Arg(characteristic_l1 / elementwise_size / 960 * 960);
  benchmark->Arg(characteristic_l2 / elementwise_size / 960 * 960);
}

// Set multi-threading parameters appropriate for the processor.
void MultiThreadingParameters(benchmark::internal::Benchmark* benchmark);

typedef bool (*IsaCheckFunction)(benchmark::State& state);

// Check if either ARM VFPv2 or VFPv3 extension is supported.
// If VFP is unsupported, report error in benchmark state, and return false.
bool CheckVFP(benchmark::State& state);

// Check if ARM NEON extension is supported.
// If NEON is unsupported, report error in benchmark state, and return false.
bool CheckNEON(benchmark::State& state);

// Check if ARM NEON-FP16 extension is supported.
// If NEON-FP16 is unsupported, report error in benchmark state, and return false.
bool CheckNEONFP16(benchmark::State& state);

// Check if ARM NEON-FMA extension is supported.
// If NEON-FMA is unsupported, report error in benchmark state, and return false.
bool CheckNEONFMA(benchmark::State& state);

// Check if ARMv8 NEON instructions are supported.
// If ARMv8 NEON is unsupported, report error in benchmark state, and return false.
bool CheckNEONV8(benchmark::State& state);

// Check if ARM NEON-FP16-ARITH extension is supported.
// If NEON-FP16-ARITH is unsupported, report error in benchmark state, and return false.
bool CheckNEONFP16ARITH(benchmark::State& state);

// Check if ARM DOT extension is supported.
// If DOT is unsupported, report error in benchmark state, and return false.
bool CheckNEONDOT(benchmark::State& state);

// Check if x86 SSSE3 extension is supported.
// If SSSE3 is unsupported, report error in benchmark state, and return false.
bool CheckSSSE3(benchmark::State& state);

// Check if x86 SSE4.1 extension is supported.
// If SSE4.1 is unsupported, report error in benchmark state, and return false.
bool CheckSSE41(benchmark::State& state);

// Check if x86 AVX extension is supported.
// If AVX is unsupported, report error in benchmark state, and return false.
bool CheckAVX(benchmark::State& state);

// Check if x86 F16C extension is supported.
// If F16C is unsupported, report error in benchmark state, and return false.
bool CheckF16C(benchmark::State& state);

// Check if x86 XOP extension is supported.
// If XOP is unsupported, report error in benchmark state, and return false.
bool CheckXOP(benchmark::State& state);

// Check if x86 FMA3 extension is supported.
// If FMA3 is unsupported, report error in benchmark state, and return false.
bool CheckFMA3(benchmark::State& state);

// Check if x86 AVX2 extension is supported.
// If AVX2 is unsupported, report error in benchmark state, and return false.
bool CheckAVX2(benchmark::State& state);

// Check if x86 AVX512F extension is supported.
// If AVX512F is unsupported, report error in benchmark state, and return false.
bool CheckAVX512F(benchmark::State& state);

// Check if x86 SKX-level AVX512 extensions (AVX512F, AVX512CD, AVX512BW, AVX512DQ, and AVX512VL) are supported.
// If SKX-level AVX512 extensions are unsupported, report error in benchmark state, and return false.
bool CheckAVX512SKX(benchmark::State& state);

template <class T>
inline T DivideRoundUp(T x, T q) {
  return x / q + T(x % q != 0);
}

template <class T>
inline T RoundUp(T x, T q) {
  return q * DivideRoundUp(x, q);
}

template <class T>
inline T Doz(T a, T b) {
  return a >= b ? a - b : T(0);
}

// A struct that uses RAII pattern to allocate and release code memory.
struct CodeMemoryHelper {
  CodeMemoryHelper();
  ~CodeMemoryHelper();

  xnn_code_buffer buffer;
  xnn_status status;
};

}  // namespace utils
}  // namespace benchmark