sglang_v0.5.2/pytorch_2.8.0/third_party/XNNPACK/bench/binary.cc

398 lines
15 KiB
C++

// Copyright 2023 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.
#include <algorithm>
#include <array>
#include <cassert>
#include <cstddef>
#include <cstdint>
#include <limits>
#include <memory>
#include <random>
#include <vector>
#include "utils.h"
#include "xnnpack.h"
#include "xnnpack/datatype.h"
#include "xnnpack/buffer.h"
#include "xnnpack/math.h"
#include <benchmark/benchmark.h>
#ifdef BENCHMARK_TENSORFLOW_LITE
#include "flatbuffers/include/flatbuffers/flatbuffers.h"
#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/kernels/test_util.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/schema/schema_generated.h"
#include "tensorflow/lite/version.h"
#endif // BENCHMARK_TENSORFLOW_LITE
void init_params(xnn_binary_operator op_type, xnn_datatype datatype,
xnn_binary_params& params,
xnn_quantization_params& input_quantization,
xnn_quantization_params& output_quantization) {
switch (op_type) {
case xnn_datatype_qint8:
input_quantization = {0, 1.0f / 128.0f};
output_quantization = {128, 1.0f / 128.0f};
break;
case xnn_datatype_quint8:
input_quantization = {128, 1.0f / 128.0f};
output_quantization = {0, 1.0f / 256.0f};
break;
default:
input_quantization = {0, 1.0f};
output_quantization = {0, 1.0f};
break;
}
}
template <typename T>
static void benchmark_binary_operator(benchmark::State& state,
xnn_binary_operator op_type) {
const size_t batch_size = state.range(0);
state.SetItemsProcessed(batch_size);
state.SetBytesProcessed(batch_size * (sizeof(T) * 3));
xnn_binary_params params;
xnn_quantization_params input_quantization;
xnn_quantization_params output_quantization;
init_params(op_type, xnn_datatype_of<T>(), params, input_quantization,
output_quantization);
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32dist = std::uniform_real_distribution<float>(
std::max<float>(std::numeric_limits<T>::lowest(), -128.0f),
std::min<float>(std::numeric_limits<T>::max(), 127.0f));
xnnpack::Buffer<T> input1(batch_size + XNN_EXTRA_BYTES / sizeof(T));
xnnpack::Buffer<T> input2(batch_size + XNN_EXTRA_BYTES / sizeof(T));
xnnpack::Buffer<T> output(batch_size);
std::generate(input1.begin(), input1.end(), [&]() { return f32dist(rng); });
std::generate(input2.begin(), input2.end(), [&]() { return f32dist(rng); });
xnn_status status = xnn_initialize(nullptr /* allocator */);
if (status != xnn_status_success) {
state.SkipWithError("failed to initialize XNNPACK");
return;
}
xnn_operator_t op = nullptr;
status = xnn_create_binary_elementwise_nd(
op_type, xnn_datatype_of<T>(), &input_quantization, &input_quantization,
&output_quantization, /*flags*/ 0, &op);
if (status != xnn_status_success || op == nullptr) {
state.SkipWithError("failed to create operator");
return;
}
const size_t input_shape[] = {batch_size};
status =
xnn_reshape_binary_elementwise_nd(op, /*num_input1_dims*/ 1, input_shape,
/*num_input2_dims*/ 1, input_shape,
/*threadpool=*/nullptr);
if (status != xnn_status_success) {
state.SkipWithError("failed to reshape operator");
return;
}
status = xnn_setup_binary_elementwise_nd(op, input1.data(), input2.data(),
output.data());
if (status != xnn_status_success) {
state.SkipWithError("failed to setup operator");
return;
}
for (auto _ : state) {
status = xnn_run_operator(op, /*threadpool=*/nullptr);
if (status != xnn_status_success) {
state.SkipWithError("failed to run operator");
return;
}
}
status = xnn_delete_operator(op);
if (status != xnn_status_success) {
state.SkipWithError("failed to delete operator");
return;
}
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
}
#ifdef BENCHMARK_TENSORFLOW_LITE
tflite::BuiltinOperator xnn_binary_operator_to_tflite(xnn_binary_operator op) {
switch (op) {
case xnn_binary_add:
return tflite::BuiltinOperator_STABLEHLO_ADD;
case xnn_binary_subtract:
return tflite::BuiltinOperator_STABLEHLO_SUBTRACT;
case xnn_binary_multiply:
return tflite::BuiltinOperator_STABLEHLO_MULTIPLY;
case xnn_binary_divide:
return tflite::BuiltinOperator_STABLEHLO_DIVIDE;
case xnn_binary_maximum:
return tflite::BuiltinOperator_STABLEHLO_MAXIMUM;
case xnn_binary_minimum:
return tflite::BuiltinOperator_STABLEHLO_MINIMUM;
case xnn_binary_copysign:
return tflite::BuiltinOperator_CUSTOM; // No corresponding TFlite op
case xnn_binary_squared_difference:
return tflite::BuiltinOperator_SQUARED_DIFFERENCE;
case xnn_binary_prelu:
return tflite::BuiltinOperator_PRELU;
case xnn_binary_modulus:
return tflite::BuiltinOperator_CUSTOM; // No corresponding TFlite op
case xnn_binary_atan2:
return tflite::BuiltinOperator_ATAN2;
case xnn_binary_pow:
return tflite::BuiltinOperator_STABLEHLO_POWER;
case xnn_binary_bitwise_and:
return tflite::BuiltinOperator_STABLEHLO_AND;
case xnn_binary_bitwise_or:
return tflite::BuiltinOperator_STABLEHLO_OR;
case xnn_binary_bitwise_xor:
return tflite::BuiltinOperator_BITWISE_XOR;
case xnn_binary_shift_left:
return tflite::BuiltinOperator_STABLEHLO_SHIFT_LEFT;
case xnn_binary_shift_right_logical:
// TODO: BuiltinOperator_RIGHT_SHIFT is logical only for unsigned types
return tflite::BuiltinOperator_RIGHT_SHIFT;
case xnn_binary_shift_right_arithmetic:
// TODO: BuiltinOperator_RIGHT_SHIFT is arithmetic only for unsigned types
return tflite::BuiltinOperator_RIGHT_SHIFT;
default:
XNN_UNREACHABLE;
return tflite::BuiltinOperator_CUSTOM;
}
}
template <typename T>
struct TypeToTfliteType {
using type = T;
static constexpr auto tensor_type =
tflite::TensorTypeFor<typename xnnpack::unwrap_quantized<T>::type>::value;
};
template <>
struct TypeToTfliteType<xnn_float16> {
using type = TfLiteFloat16;
static constexpr auto tensor_type = tflite::TensorType_FLOAT16;
};
template <>
struct TypeToTfliteType<xnn_bfloat16> {
using type = TfLiteBFloat16;
static constexpr auto tensor_type = tflite::TensorType_BFLOAT16;
};
template <typename T, class BuildInQuantization, class BuildOutQuantization>
static void benchmark_tflite_binary_operator(
benchmark::State& state, BuildInQuantization in_quantization,
BuildOutQuantization out_quantization, tflite::BuiltinOperator op_code) {
const size_t batch_size = state.range(0);
flatbuffers::FlatBufferBuilder builder;
const flatbuffers::Offset<tflite::OperatorCode> operator_code =
CreateOperatorCode(builder, op_code);
const std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers{{
tflite::CreateBuffer(builder, builder.CreateVector({})),
}};
const std::array<int32_t, 1> shape{{static_cast<int32_t>(batch_size)}};
const std::array<flatbuffers::Offset<tflite::Tensor>, 3> tensors{{
tflite::CreateTensor(
builder, builder.CreateVector<int32_t>(shape.data(), shape.size()),
TypeToTfliteType<T>::tensor_type,
/*buffer=*/0, /*name=*/0, in_quantization(builder)),
tflite::CreateTensor(
builder, builder.CreateVector<int32_t>(shape.data(), shape.size()),
TypeToTfliteType<T>::tensor_type,
/*buffer=*/0, /*name=*/0, in_quantization(builder)),
tflite::CreateTensor(
builder, builder.CreateVector<int32_t>(shape.data(), shape.size()),
TypeToTfliteType<T>::tensor_type,
/*buffer=*/0, /*name=*/0, out_quantization(builder)),
}};
const std::array<int32_t, 2> op_inputs{{0, 1}};
const std::array<int32_t, 1> op_outputs{{2}};
flatbuffers::Offset<tflite::Operator> op = tflite::CreateOperator(
builder, 0 /* opcode_index */,
builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()),
builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size()));
const std::array<int32_t, 2> graph_inputs{{0, 1}};
const std::array<int32_t, 1> graph_outputs{{2}};
const flatbuffers::Offset<tflite::SubGraph> subgraph = tflite::CreateSubGraph(
builder, builder.CreateVector(tensors.data(), tensors.size()),
builder.CreateVector<int32_t>(graph_inputs.data(), graph_inputs.size()),
builder.CreateVector<int32_t>(graph_outputs.data(), graph_outputs.size()),
builder.CreateVector(&op, 1));
const flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(
builder, TFLITE_SCHEMA_VERSION, builder.CreateVector(&operator_code, 1),
builder.CreateVector(&subgraph, 1), builder.CreateString("Binary model"),
builder.CreateVector(buffers.data(), buffers.size()));
builder.Finish(model_buffer);
const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer());
tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates resolver;
tflite::InterpreterBuilder interpreterBuilder(model, resolver);
std::unique_ptr<tflite::Interpreter> interpreter;
if (interpreterBuilder(&interpreter) != kTfLiteOk || interpreter == nullptr) {
state.SkipWithError("failed to create TFLite interpreter");
return;
}
interpreter->SetNumThreads(1);
if (interpreter->AllocateTensors() != kTfLiteOk) {
state.SkipWithError("failed to allocate tensors");
return;
}
std::random_device random_device;
auto rng = std::mt19937(random_device());
auto f32dist = std::uniform_real_distribution<float>(
std::max<float>(std::numeric_limits<T>::lowest(), -128.0f),
std::min<float>(std::numeric_limits<T>::max(), 127.0f));
T* input_ptr = reinterpret_cast<T*>(interpreter->tensor(0)->data.raw);
std::generate(input_ptr, input_ptr + batch_size,
[&]() { return f32dist(rng); });
for (auto _ : state) {
if (interpreter->Invoke() != kTfLiteOk) {
state.SkipWithError("failed to invoke TFLite interpreter");
return;
}
}
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
if (cpu_frequency != 0) {
state.counters["cpufreq"] = cpu_frequency;
}
state.counters["elements"] = benchmark::Counter(
uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate);
const size_t bytes_per_iteration = 2 * batch_size * sizeof(float);
state.counters["bytes"] =
benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration,
benchmark::Counter::kIsRate);
interpreter.reset();
}
static flatbuffers::Offset<tflite::QuantizationParameters> no_quantization(
flatbuffers::FlatBufferBuilder& builder) {
return 0;
}
static auto CreateTfLiteQuantizationParameters(
flatbuffers::FlatBufferBuilder& builder,
const xnn_quantization_params& params) {
return tflite::CreateQuantizationParameters(
builder, 0 /*min*/, 0 /*max*/,
builder.CreateVector<float>({params.scale}),
builder.CreateVector<int64_t>({params.zero_point}));
}
template <typename T>
static void benchmark_tflite_binary_operator(benchmark::State& state,
xnn_binary_operator op) {
xnn_binary_params params;
xnn_quantization_params input_quantization;
xnn_quantization_params output_quantization;
init_params(op, xnn_datatype_of<T>(), params, input_quantization,
output_quantization);
auto in_quantization = [=](flatbuffers::FlatBufferBuilder& builder) {
return CreateTfLiteQuantizationParameters(builder, input_quantization);
};
auto out_quantization = [=](flatbuffers::FlatBufferBuilder& builder) {
return CreateTfLiteQuantizationParameters(builder, output_quantization);
};
tflite::BuiltinOperator op_code = xnn_binary_operator_to_tflite(op);
if (op_code == tflite::BuiltinOperator_CUSTOM) {
state.SkipWithMessage("no corresponding TFLite operator");
return;
}
if (!xnnpack::is_quantized<T>::value) {
return benchmark_tflite_binary_operator<T>(state, no_quantization,
no_quantization, op_code);
} else {
return benchmark_tflite_binary_operator<T>(state, in_quantization,
out_quantization, op_code);
}
}
#define BENCHMARK_OP_TYPE(op, type_name, type) \
void xnnpack_##op##_##type_name(benchmark::State& state) { \
benchmark_binary_operator<type>(state, xnn_binary_##op); \
} \
void tflite_##op##_##type_name(benchmark::State& state) { \
benchmark_tflite_binary_operator<type>(state, xnn_binary_##op); \
} \
BENCHMARK(xnnpack_##op##_##type_name) \
->Apply(benchmark::utils::BinaryElementwiseParameters<type, type>) \
->UseRealTime(); \
BENCHMARK(tflite_##op##_##type_name) \
->Apply(benchmark::utils::BinaryElementwiseParameters<type, type>) \
->UseRealTime();
#else // BENCHMARK_TENSORFLOW_LITE
#define BENCHMARK_OP_TYPE(op, type_name, type) \
void xnnpack_##op##_##type_name(benchmark::State& state) { \
benchmark_binary_operator<type>(state, xnn_binary_##op); \
} \
BENCHMARK(xnnpack_##op##_##type_name) \
->Apply(benchmark::utils::BinaryElementwiseParameters<type, type>) \
->UseRealTime();
#endif // BENCHMARK_TENSORFLOW_LITE
#define BENCHMARK_OP_REAL(op) \
BENCHMARK_OP_TYPE(op, f32, float) \
BENCHMARK_OP_TYPE(op, f16, xnn_float16) \
BENCHMARK_OP_TYPE(op, bf16, xnn_bfloat16) \
BENCHMARK_OP_TYPE(op, qs8, xnnpack::quantized<int8_t>) \
BENCHMARK_OP_TYPE(op, qu8, xnnpack::quantized<uint8_t>)
#define BENCHMARK_OP_INTEGRAL(op) BENCHMARK_OP_TYPE(op, s32, int32_t)
#define BENCHMARK_OP_ALL(op) \
BENCHMARK_OP_REAL(op) \
BENCHMARK_OP_INTEGRAL(op)
BENCHMARK_OP_ALL(add);
BENCHMARK_OP_ALL(subtract);
BENCHMARK_OP_ALL(multiply);
BENCHMARK_OP_ALL(divide);
BENCHMARK_OP_ALL(maximum);
BENCHMARK_OP_ALL(minimum);
// BENCHMARK_OP_ALL(copysign); // Missing in TFLite
BENCHMARK_OP_REAL(squared_difference);
BENCHMARK_OP_REAL(prelu);
// BENCHMARK_OP_ALL(modulus); // Missing in TFLite
BENCHMARK_OP_REAL(atan2);
BENCHMARK_OP_ALL(pow);
BENCHMARK_OP_INTEGRAL(bitwise_and);
BENCHMARK_OP_INTEGRAL(bitwise_or);
BENCHMARK_OP_INTEGRAL(bitwise_xor);
BENCHMARK_OP_INTEGRAL(shift_left);
BENCHMARK_OP_INTEGRAL(shift_right_logical);
BENCHMARK_OP_INTEGRAL(shift_right_arithmetic);
#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif