提交 09dab387 编写于 作者: M Megvii Engine Team 提交者: “wenjuan”

feat(cuda): support int1 simplewq conv

GitOrigin-RevId: 9c37c41bc7e450f3df81e6059603101de3f14416
上级 6554e262
......@@ -561,7 +561,8 @@ void ConvolutionBase<Parameter>::check_or_deduce_dtype_fwd(
src.enumv() == DTypeEnum::QuantizedS8 ||
src.enumv() == DTypeEnum::Quantized8Asymm ||
src.enumv() == DTypeEnum::QuantizedS4 ||
src.enumv() == DTypeEnum::Quantized4Asymm) {
src.enumv() == DTypeEnum::Quantized4Asymm ||
src.enumv() == DTypeEnum::QuantizedS1) {
supported_dst_dtype.push_back(dtype::QuantizedS32(mul_scale(src, filter)));
bool cond_dst = dst.valid() && (dst.enumv() == src.enumv() ||
((dst.enumv() == DTypeEnum::QuantizedS4 ||
......
......@@ -25,7 +25,7 @@ ConvBiasForwardImpl::AlgoPack::AlgoPack() {
non_cudnn_algos.push_back(&matmul);
non_cudnn_algos.push_back(&matmul8x8x32);
non_cudnn_algos.push_back(&batched_matmul);
non_cudnn_algos.push_back(&int1_simple);
fill_cudnn_algos();
for (auto&& algo : cudnn_conv_bias_activations) {
all_algos.push_back(&algo);
......@@ -45,6 +45,7 @@ ConvBiasForwardImpl::AlgoPack::AlgoPack() {
conv_algos.push_back(&matmul8x8x32);
conv_algos.push_back(&batched_matmul);
conv_algos.push_back(&group);
conv_algos.push_back(&int1_simple);
for (auto&& algo : conv_algos) {
all_algos.push_back(algo);
......
......@@ -87,6 +87,7 @@ public:
CUDA_FALLBACK_NCHW_INT4,
CUDA_IMPLICIT_BATCHED_GEMM_FMA_NCHW_F32,
CUDA_IMPLICIT_BATCHED_GEMM_HMMA_NCHW_F16,
CUDA_SIMPLE_INT1,
};
using Mapper = std::unordered_map<AlgorithmDesc, AlgoBase*>;
......@@ -1089,6 +1090,24 @@ private:
WorkspaceBundle get_workspace_bundle(void* ptr, const SizeArgs& args) const;
};
class ConvBiasForwardImpl::AlgoSimpleInt1 final : public AlgoBase {
public:
bool is_available(const SizeArgs& args) const override;
size_t get_workspace_in_bytes(const SizeArgs& args) const override;
void exec(const ExecArgs& args) const override;
std::vector<SearchItem> get_subopr_list(
const TensorLayoutArray& layouts, const OperatorBase* opr) const override;
const char* name() const override { return "CONVBIAS_SIMPLE_INT1"; }
AlgoAttribute attribute() const override { return AlgoAttribute::REPRODUCIBLE; }
MEGDNN_DECL_ALGO_TYPE(CUDA_SIMPLE_INT1)
private:
WorkspaceBundle get_workspace_bundle(void* ptr, const SizeArgs& args) const;
};
class ConvBiasForwardImpl::AlgoPack : NonCopyableObj {
private:
AlgoBase::Mapper m_all_algos_map;
......@@ -1132,6 +1151,7 @@ public:
std::vector<AlgoFloat16NCHWHMMAImplicitBatchedGemm> f16_implicit_bmm;
AlgoGroupConvGeneral group;
AlgoBFloat16 bfloat16;
AlgoSimpleInt1 int1_simple;
AlgoBase* cudnn_conv_bias_act_from_enum(cudnnConvolutionFwdAlgo_t algo);
......
......@@ -30,6 +30,8 @@ bool ConvBiasForwardImpl::AlgoCUDNNConvBiasActivation::is_available(
return false;
}
}
if (args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS1)
return false;
if ((args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS4 ||
args.src_layout->dtype.enumv() == DTypeEnum::Quantized4Asymm) &&
args.filter_layout->dtype.enumv() == DTypeEnum::QuantizedS4)
......
......@@ -134,6 +134,9 @@ void ConvBiasDesc::set_conv(
namespace conv_bias {
bool is_cudnn_supported(const BiasForwardSizeArgs& args) {
if (args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS1)
return false;
if ((args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS4 ||
args.src_layout->dtype.enumv() == DTypeEnum::Quantized4Asymm) &&
args.filter_layout->dtype.enumv() == DTypeEnum::QuantizedS4)
......
......@@ -221,6 +221,11 @@ ConvBiasForward::Algorithm* ConvBiasForwardImpl::get_algorithm_heuristic(
return &sm_algo_pack.fallback_nchw_qs8;
}
if (sm_algo_pack.int1_simple.is_available_attribute(
args, positive_attr, negative_attr, workspace_limit_in_bytes)) {
return &sm_algo_pack.int1_simple;
}
if (args.src_layout->dtype.enumv() != DTypeTrait<dtype::BFloat16>::enumv) {
return megdnn::get_algo_match_attribute<ConvBiasForwardImpl>(
sm_algo_pack.non_cudnn_algos, args, workspace_limit_in_bytes,
......
......@@ -72,6 +72,7 @@ public:
class AlgoInt4Int4NHWCIMMAImplicitGemm;
class AlgoUInt4Int4NHWCIMMAImplicitGemm;
class AlgoBFloat16;
class AlgoSimpleInt1;
// The following algorithms are suitable for channel wise convolution
class AlgoFloat32NCHWFMAImplicitBatchedGemm;
class AlgoFloat16NCHWHMMAImplicitBatchedGemm;
......
/**
* \file dnn/src/cuda/conv_bias/simple_int1.cpp
* MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
*
* Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
* implied.
*/
#include "src/common/algo_base.h"
#include "src/cuda/conv_bias/algo.h"
#include "src/cuda/handle.h"
#include "src/cuda/utils.cuh"
#include "src/cuda/utils.h"
using namespace megdnn;
using namespace cuda;
using namespace conv_bias;
namespace {
std::pair<TensorLayoutArray, ConvBiasForwardImpl::Param> sub_opr_config(
const TensorLayoutArray& layouts, const ConvBiasForwardImpl* opr) {
megdnn_assert(layouts.size() >= 3);
std::pair<TensorLayoutArray, ConvBiasForwardImpl::Param> ret;
ret.first = layouts;
auto change_dtype = [](TensorLayout& layout) {
if (layout.dtype.enumv() == DTypeEnum::QuantizedS1 ||
layout.dtype.enumv() == DTypeEnum::QuantizedS32) {
layout.dtype = dtype::Float32();
}
};
change_dtype(ret.first[0]);
change_dtype(ret.first[1]);
change_dtype(ret.first[2]);
change_dtype(ret.first[3]);
change_dtype(ret.first[4]);
ret.second = opr->param();
ret.second.compute_mode = ConvBiasForwardImpl::Param::ComputeMode::DEFAULT;
return ret;
}
std::pair<TensorLayoutArray, std::unique_ptr<ConvBiasForward>> prepare_sub_opr(
const ConvBiasForwardImpl::AlgoBase::SizeArgs& args) {
auto convbias_opr = args.handle->create_operator<ConvBias>();
auto&& config = sub_opr_config(
{*args.src_layout, *args.filter_layout, *args.bias_layout, *args.z_layout,
*args.dst_layout},
args.opr);
convbias_opr->param() = config.second;
return {config.first, std::move(convbias_opr)};
}
} // namespace
std::vector<Algorithm::SearchItem> ConvBiasForwardImpl::AlgoSimpleInt1::get_subopr_list(
const TensorLayoutArray& layouts, const OperatorBase* opr) const {
auto&& config =
sub_opr_config(layouts, static_cast<const ConvBiasForwardImpl*>(opr));
std::string param_str;
Algorithm::serialize_write_pod(config.second, param_str);
return {{Algorithm::OprType::CONVBIAS_FORWARD, param_str, config.first}};
}
bool ConvBiasForwardImpl::AlgoSimpleInt1::is_available(const SizeArgs& args) const {
if (args.src_layout->dtype.valid() && args.filter_layout->dtype.valid() &&
args.bias_layout->dtype.valid() && args.z_layout->dtype.valid() &&
args.dst_layout->dtype.valid()) {
auto config = prepare_sub_opr(args);
return args.src_layout->dtype.enumv() == args.filter_layout->dtype.enumv() &&
args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS1 &&
get_algorithm(
static_cast<ConvBiasForwardImpl*>(config.second.get()),
config.first[0], config.first[1], config.first[2],
config.first[3], config.first[4]);
} else {
return false;
}
}
WorkspaceBundle ConvBiasForwardImpl::AlgoSimpleInt1::get_workspace_bundle(
void* ptr, const SizeArgs& args) const {
auto config = prepare_sub_opr(args);
SmallVector<size_t> sizes;
auto get_workspace = [&sizes](const TensorLayout& src, const TensorLayout& dst) {
if (src.dtype != dst.dtype) {
sizes.push_back(dst.span().dist_byte());
}
};
get_workspace(*args.src_layout, config.first[0]);
get_workspace(*args.filter_layout, config.first[1]);
get_workspace(*args.bias_layout, config.first[2]);
get_workspace(*args.z_layout, config.first[3]);
get_workspace(*args.dst_layout, config.first[4]);
sizes.push_back(config.second->get_workspace_in_bytes(
config.first[0], config.first[1], config.first[2], config.first[3],
config.first[4], nullptr));
return {ptr, std::move(sizes)};
}
size_t ConvBiasForwardImpl::AlgoSimpleInt1::get_workspace_in_bytes(
const SizeArgs& args) const {
return get_workspace_bundle(nullptr, args).total_size_in_bytes();
}
void ConvBiasForwardImpl::AlgoSimpleInt1::exec(const ExecArgs& args) const {
TensorND fsrc_tensor = *args.src_tensor;
TensorND ffilter_tensor = *args.filter_tensor;
TensorND fbias_tensor = *args.bias_tensor;
TensorND fz_tensor = *args.z_tensor;
TensorND fdst_tensor = *args.dst_tensor;
auto config = prepare_sub_opr(args);
auto bundle = get_workspace_bundle(args.workspace.raw_ptr, args);
CompTypeCvter<dtype::QuantizedS1, dtype::Float32> cvter(args.handle, &bundle);
{
cvter.src_to_comp_type(*args.src_tensor, fsrc_tensor)
.src_to_comp_type(*args.filter_tensor, ffilter_tensor);
}
WorkspaceBundle dst_bundle = {
bundle.get(2),
{bundle.get_size(2), bundle.get_size(3), bundle.get_size(4),
bundle.get_size(5)}};
CompTypeCvter<dtype::QuantizedS32, dtype::Float32> dst_cvter(
args.handle, &dst_bundle);
{
dst_cvter.src_to_comp_type(*args.bias_tensor, fbias_tensor)
.src_to_comp_type(*args.z_tensor, fz_tensor)
.src_to_comp_type(*args.dst_tensor, fdst_tensor);
}
config.second->exec(
fsrc_tensor, ffilter_tensor, fbias_tensor, fz_tensor, fdst_tensor, nullptr,
dst_cvter.workspace());
{ dst_cvter.comp_to_dst_type(fdst_tensor, *args.dst_tensor); }
}
// vim: syntax=cpp.doxygen
......@@ -44,6 +44,10 @@ std::pair<TensorLayoutArray, ConvBiasForward::Param> sub_opr_config(
src.dtype.param<dtype::Quantized4Asymm>().scale *
filter.dtype.param<dtype::Quantized4Asymm>().scale);
} else if (src.dtype.enumv() == DTypeEnum::QuantizedS1) {
bias_type = dtype::QuantizedS32(
src.dtype.param<dtype::QuantizedS1>().scale *
filter.dtype.param<dtype::QuantizedS1>().scale);
} else {
megdnn_assert(src.dtype.category() == DTypeCategory::FLOAT);
bias_type = src.dtype;
......
......@@ -278,6 +278,9 @@ void ConvBiasForwardImpl::exec(
DISPATCH_RAW(
Quantized4Asymm, QuantizedS4, QuantizedS32, QuantizedS32, DEFAULT,
(convolution::forward_bias<dt_quint4, dt_qint4, dt_qint32, dt_qint32>))
DISPATCH_RAW(
QuantizedS1, QuantizedS1, QuantizedS32, QuantizedS32, FLOAT32,
(convolution::forward_bias<dt_qint1, dt_qint1, dt_qint32, dt_qint32>))
#if !MEGDNN_DISABLE_FLOAT16
DISPATCH(Float16, Float16)
DISPATCH_RAW(
......
......@@ -84,6 +84,15 @@ inline void StrategyFwd::on(
d += cast(s) * cast(f);
}
template <>
inline void StrategyFwd::on(
dt_qint1& s, dt_qint1& f, dt_qint32& d, DType, DType, DType) {
auto cast = [](const dt_qint1& val) {
return dt_qint32(static_cast<int32_t>(val.as_int8()));
};
d += cast(s) * cast(f);
}
struct StrategyBwdData {
template <typename st, typename ft, typename dt>
static void on(st& s, ft& f, dt& d, DType, DType, DType) {
......
......@@ -133,6 +133,32 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_BF16) {
}
}
TEST_F(CUDA, CONV_BIAS_FORWARD_QS1) {
require_compute_capability(6, 1);
UniformIntRNG int_rng{1, 1};
Checker<ConvBiasForward> checker(handle_cuda());
checker.set_before_exec_callback(AlgoChecker<ConvBiasForward>(
ExecutionPolicyAlgoName{"CONVBIAS_SIMPLE_INT1", {{"MATMUL", {}}}}));
ConvBias::Param param;
param.format = ConvBias::Param::Format::NCHW;
param.compute_mode = param::Convolution::ComputeMode::FLOAT32;
{
auto src_shape = TensorShape{20, 2, 224, 224};
auto filter_shape = TensorShape{20, 2, 3, 3};
checker.set_dtype(0, dtype::QuantizedS1(1.0f))
.set_dtype(1, dtype::QuantizedS1(1.0f))
.set_dtype(2, dtype::QuantizedS32(1.0f))
.set_dtype(3, dtype::QuantizedS32(1.0f))
.set_dtype(4, dtype::QuantizedS32(1.0f))
.set_rng(0, &int_rng)
.set_rng(1, &int_rng)
.set_param(param)
.execs({src_shape, filter_shape, {}, {}, {}});
}
}
TEST_F(CUDA, CONV_BIAS_FORWARD_QS8) {
require_compute_capability(6, 1);
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册