Skip to content

MegEngine
MegEngine

2020fengziyang / MegEngine
与 Fork 源项目一致

Fork自 MegEngine 天元 / MegEngine

通知 1
Star 1
Fork 0
MegEngine
MegEngine

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

提交 edd7e167 编写于 作者: M Megvii Engine Team 提交者： Xu Xinran
浏览文件
操作

feat(dnn/fallback): add im2col filterpreprocess function 

GitOrigin-RevId: 61c54ad258a42301711d3efdae0caef47d7b0584
上级 9e9e8ca0
展开全部 隐藏空白更改
内联 并排
Showing with 1207 addition and 252 deletion
dnn/src/fallback/conv_bias/im2col/algos.h
浏览文件 @ edd7e167
......@@ -22,27 +22,6 @@ namespace megdnn {
namespace fallback {
class ConvBiasImpl::AlgoIm2col final : public AlgoBase {
//! calculate m_oc_tile_size in choice_ohw_oc_block() fucntion,
//! when m_oc_tile_size < this value m_oc_tile_size = ohw
static constexpr size_t DEFAULT_OHW_MIN_TILE_SIZE = 32;
//! when nr_threads > 1 and round(ohw,nr_threads)>nr_threads,
//! m_oc_tile_size = DEFAULT_OC_TILE_SIZE
static constexpr size_t DEFAULT_OC_TILE_SIZE = 512;
//! when m_oc_tile_size > this value m_oc_tile_size =
//! DEFAULT_OC_MAX_TILE_SIZE
static constexpr size_t DEFAULT_OC_MAX_TILE_SIZE = 1024;
//! when m_oc_tile_size < this value m_oc_tile_size =
//! DEFAULT_OC_MIN_TILE_SIZE the purpose is aligning the calculation
static constexpr size_t DEFAULT_OC_MIN_TILE_SIZE = 128;
fallback::MatrixMulImpl::KernSizeParam get_matmul_kern_param(
const NCBKernSizeParam& param, size_t ohw_tile_size,
size_t oc_tile_size) const;
WorkspaceBundle get_bundle(const NCBKernSizeParam& param) const;
void choice_ohw_oc_block(
const NCBKernSizeParam& param, size_t& oc_tile_size,
size_t& ohw_tile_size, size_t block_m, size_t block_n,
fallback::MatrixMulImpl::AlgoBase::PackMode pack_mode) const;
public:
AlgoIm2col(MatrixMulImpl::AlgoBase* matmul_algo, size_t ohw_tile_size)
: m_matmul_algo(matmul_algo),
......@@ -59,10 +38,16 @@ public:
bool usable(const NCBKernSizeParam& param,
AlgoSelectionStrategy algo_selection_strategy) const override;
size_t get_workspace(const NCBKernSizeParam& param) const override;
SmallVector<NCBKern> dispatch_kerns(
SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam& param) const override;
SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
const NCBKernSizeParam& param) const override;
size_t get_preprocess_workspace(
const NCBKernSizeParam& /*param*/) const override {
return 0;
}
SmallVector<NCBKern> dispatch_preprocess_kerns(
const NCBKernSizeParam& param) const override;
bool is_preferred(
const NCBKernSizeParam& param) const override {
bool is_preferred(const NCBKernSizeParam& param) const override {
if (param.src_type.category() == DTypeCategory::QUANTIZED) {
static CpuOprDelegationStorage<1> storage;
auto conv_bias_opr = storage.get<ConvBias, 0>();
......
dnn/src/fallback/conv_bias/im2col/strategy_base.h
浏览文件 @ edd7e167
......@@ -40,9 +40,11 @@ struct StrategyParam {
size_t block_n;
size_t block_k;
size_t pack_oc_size;
size_t packA_group_size;
bool skip_copy_dst;
bool is_dst_8bit;
bool is_ohw_size_bigger;
bool enable_filter_preprocess;
};
class StrategyBase {
......@@ -62,7 +64,7 @@ public:
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::MatmulDescription&
matmul_desec,
size_t pack_size) = 0;
const StrategyParam& sparam) = 0;
virtual void exec_im2col(
const WorkspaceBundle& bundle, const WorkspaceBundle& bundle_thread,
......@@ -296,7 +298,7 @@ public:
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::MatmulDescription&
matmul_desc,
size_t pack_size) override;
const StrategyParam& sparam) override;
virtual void exec_im2col(
const WorkspaceBundle& bundle, const WorkspaceBundle& bundle_thread,
const StrategyParam& sparam,
......@@ -375,7 +377,7 @@ public:
const fallback::MatrixMulImpl::AlgoBase* matmul_algo,
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::MatmulDescription& MDsec,
size_t pack_size) override;
const StrategyParam& sparam) override;
void exec_matmul(const fallback::ConvBiasImpl::NCBKernParam& param,
const StrategyParam& sparam, const WorkspaceBundle& bundle,
......@@ -431,7 +433,7 @@ public:
const fallback::MatrixMulImpl::AlgoBase* matmul_algo,
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::MatmulDescription& MDsec,
size_t pack_size) override;
const StrategyParam& sparam) override;
void exec_im2col(
const WorkspaceBundle& bundle, const WorkspaceBundle& bundle_thread,
......
dnn/src/fallback/conv_bias/im2col/strategy_default.cpp
浏览文件 @ edd7e167
......@@ -25,19 +25,23 @@ void Strategy<src_ctype, bias_ctype, dst_ctype, op_ctype, op_dtype,
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::MatmulDescription&
matmul_desc,
size_t) {
const StrategyParam& sparam) {
fallback::MatrixMulImpl::KernParam matmul_param;
size_t group_id = ncb_index.ndrange_id[0];
static_cast<fallback::MatrixMulImpl::KernSizeParam&>(matmul_param) =
matmulparam;
size_t packA_group_size = matmul_algo->get_bundle(matmul_param).get_size(0);
size_t packed_per_oc_block_size =
round_up(matmul_param.K, matmul_desc.innerblocksize.k) *
matmul_desc.innerblocksize.m * matmul_desc.packa_type_size;
size_t a_panel_offset = ncb_index.ndrange_id[1] * packed_per_oc_block_size;
int8_t* a_panel = static_cast<int8_t*>(bundle.get(BUNDLE_PACKA_INDEX)) +
group_id * packA_group_size + a_panel_offset;
int8_t* tmp_ptr =
sparam.enable_filter_preprocess
? static_cast<int8_t*>(
param.preprocessed_filter->tensors[0].raw_ptr)
: static_cast<int8_t*>(bundle.get(BUNDLE_PACKA_INDEX));
int8_t* a_panel =
tmp_ptr + group_id * sparam.packA_group_size + a_panel_offset;
matmul_param.A_ptr =
const_cast<src_ctype*>(param.filter<src_ctype>(group_id));
matmul_algo->pack_A(matmul_param, a_panel, ncb_index.ndrange_id[1],
......@@ -149,15 +153,20 @@ void Strategy<src_ctype, bias_ctype, dst_ctype, op_ctype, op_dtype,
size_t packA_per_oc_block_size =
round_up(matmul_param.K, matmul_desc.innerblocksize.k) *
sparam.oc_tile_size * matmul_desc.packa_type_size;
size_t packA_group_size = matmul_algo->get_bundle(matmul_param).get_size(0);
size_t packA_group_size = sparam.packA_group_size;
size_t a_panel_offset = ncb_index.ndrange_id[1] * packA_group_size +
ncb_index.ndrange_id[3] * packA_per_oc_block_size;
void* matmul_dst = get_matmul_dst_ptr(param, bundle_thread, sparam);
src_ctype* a_panel = reinterpret_cast<src_ctype*>(
reinterpret_cast<uintptr_t>(bundle.get(BUNDLE_PACKA_INDEX)) +
a_panel_offset);
int8_t* tmp_ptr =
sparam.enable_filter_preprocess
? static_cast<int8_t*>(
param.preprocessed_filter->tensors[0].raw_ptr)
: static_cast<int8_t*>(bundle.get(BUNDLE_PACKA_INDEX));
src_ctype* a_panel =
reinterpret_cast<src_ctype*>(tmp_ptr + a_panel_offset);
src_ctype* b_panel =
reinterpret_cast<src_ctype*>(reinterpret_cast<uintptr_t>(
bundle_thread.get(THREAD_BUNDLE_PACKB_INDEX)));
......
dnn/src/fallback/conv_bias/im2col/strategy_nopack.cpp
浏览文件 @ edd7e167
......@@ -26,7 +26,7 @@ void Strategy<src_ctype, bias_ctype, dst_ctype, op_ctype, op_dtype,
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::
MatmulDescription& /*matmul_dsec*/,
size_t) {
const StrategyParam&) {
MEGDNN_MARK_USED_VAR(bundle);
MEGDNN_MARK_USED_VAR(param);
MEGDNN_MARK_USED_VAR(matmulparam);
......
dnn/src/fallback/conv_bias/im2col/strategy_onlypacka.cpp
浏览文件 @ edd7e167
......@@ -26,7 +26,7 @@ void Strategy<src_ctype, bias_ctype, dst_ctype, op_ctype, op_dtype,
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::
MatmulDescription& /*matmul_desc*/,
size_t) {
const StrategyParam& sparam) {
fallback::MatrixMulImpl::KernParam matmul_param;
static_cast<fallback::MatrixMulImpl::KernSizeParam&>(matmul_param) =
matmulparam;
......@@ -36,12 +36,17 @@ void Strategy<src_ctype, bias_ctype, dst_ctype, op_ctype, op_dtype,
size_t output_block_oc_size =
std::min(oc_tile_size, OC - ncb_index.ndrange_id[1] * oc_tile_size);
size_t oc_cur_index = ncb_index.ndrange_id[1] * oc_tile_size;
size_t packA_group_size =
bundle.get_size(BUNDLE_PACKA_INDEX) / param.filter_meta.group;
size_t a_panel_offset = ncb_index.ndrange_id[1] *
matmul_algo->get_bundle(matmul_param).get_size(0);
int8_t* a_panel = static_cast<int8_t*>(bundle.get(BUNDLE_PACKA_INDEX)) +
group_id * packA_group_size + a_panel_offset;
int8_t* tmp_ptr =
sparam.enable_filter_preprocess
? static_cast<int8_t*>(
param.preprocessed_filter->tensors[0].raw_ptr)
: static_cast<int8_t*>(bundle.get(BUNDLE_PACKA_INDEX));
int8_t* a_panel = tmp_ptr +
group_id * sparam.packA_group_size + a_panel_offset;
matmul_param.A_ptr =
const_cast<src_ctype*>(param.filter<src_ctype>(group_id)) +
oc_cur_index * matmul_param.K;
......@@ -60,20 +65,22 @@ void Strategy<src_ctype, bias_ctype, dst_ctype, op_ctype, op_dtype,
fallback::MatrixMulImpl::KernParam matmul_param,
const fallback::MatrixMulImpl::AlgoBase* matmul_algo,
const fallback::ConvBiasImpl::NCBKernIndex& ncb_index,
const fallback::MatrixMulImpl::AlgoBase::
MatmulDescription& /*matmul_desc*/
) {
size_t packA_group_size =
bundle.get_size(BUNDLE_PACKA_INDEX) / param.filter_meta.group;
const fallback::MatrixMulImpl::AlgoBase::MatmulDescription&
/*matmul_desc*/) {
size_t a_panel_offset = ncb_index.ndrange_id[3] *
matmul_algo->get_bundle(matmul_param).get_size(0);
a_panel_offset = sparam.group_id * packA_group_size + a_panel_offset;
a_panel_offset =
sparam.group_id * sparam.packA_group_size + a_panel_offset;
void* matmul_dst = get_matmul_dst_ptr(param, bundle_thread, sparam);
src_ctype* a_panel = reinterpret_cast<src_ctype*>(
reinterpret_cast<uintptr_t>(bundle.get(BUNDLE_PACKA_INDEX)) +
a_panel_offset);
int8_t* tmp_ptr =
sparam.enable_filter_preprocess
? static_cast<int8_t*>(
param.preprocessed_filter->tensors[0].raw_ptr)
: static_cast<int8_t*>(bundle.get(BUNDLE_PACKA_INDEX));
src_ctype* a_panel = reinterpret_cast<src_ctype*>(tmp_ptr + a_panel_offset);
src_ctype* b_panel = nullptr;
src_ctype* im2col_dst = static_cast<src_ctype*>(
......
dnn/src/fallback/conv_bias/opr_impl.cpp
浏览文件 @ edd7e167
......@@ -154,7 +154,8 @@ void ConvBiasImpl::exec_preprocess(const TensorLayout& src_layout,
bias{nullptr, bias_layout};
auto fparam = make_ncb_kern_param(src, filter, bias, dst, workspace,
preprocessed_filter);
ConvolutionImpl::Algorithm* algo = get_algorithm(fparam, workspace.size);
//! should not pass workspace_size limit otherwise can not find match algo
ConvBiasImpl::Algorithm* algo = get_algorithm(fparam);
if (!is_naive_algo(algo) && NCB_ALGO_FUNC(get_preprocess_workspace, algo,
fparam) <= workspace.size) {
exec_preprocess_with_ncb_kern(fparam, algo);
......
dnn/src/fallback/conv_bias/opr_impl.h
浏览文件 @ edd7e167
......@@ -299,6 +299,11 @@ private:
const PreprocessedFilter* preprocessed_filter);
};
inline bool is_enable_filter_preprocess(
const ConvBiasImpl::NCBKernSizeParam& param) {
return param.preprocessed_filter &&
param.preprocessed_filter->tensors.size() >= 1;
}
} // namespace fallback
} // namespace megdnn
......
dnn/src/fallback/convolution/opr_impl.cpp
浏览文件 @ edd7e167
......@@ -109,7 +109,9 @@ void ConvolutionImpl::exec_preprocess(const TensorLayout& src_layout,
TensorND src{nullptr, src_layout}, dst{nullptr, dst_layout};
auto fparam = make_ncb_kern_param(src, filter, dst, preprocessed_filter,
workspace);
ConvolutionImpl::Algorithm* algo = get_algorithm(fparam, workspace.size);
//! should not pass workspace_size limit otherwise can not find match algo
ConvolutionImpl::Algorithm* algo = get_algorithm(fparam);
if (!is_naive_algo(algo) && NCB_ALGO_FUNC(get_preprocess_workspace, algo,
fparam) <= workspace.size) {
exec_preprocess_with_ncb_kern(fparam, algo);
......
dnn/test/common/conv_bias.cpp
浏览文件 @ edd7e167
......@@ -1118,6 +1118,30 @@ void checker_conv_bias_int8x8x16(std::vector<conv_bias::TestArg> args,
}
}
void check_conv_bias_preprocess(std::vector<conv_bias::TestArg> args,
Handle* handle, RNG* rng, float epsilon,
DType type0, DType type1, DType type2,
DType type3, const char* algo_name) {
using namespace conv_bias;
Checker<ConvBiasForward, OprWeightPreprocessProxy<ConvBiasForward>> checker(
handle);
checker.set_dtype(0, type0);
checker.set_dtype(1, type1);
checker.set_dtype(2, type2);
checker.set_dtype(4, type3);
checker.set_epsilon(epsilon);
if (NULL != rng) {
checker.set_rng(0, rng).set_rng(1, rng).set_rng(2, rng).set_rng(3, rng);
}
checker.set_before_exec_callback(
conv_bias::ConvBiasAlgoChecker<ConvBias>(algo_name));
for (auto&& arg : args) {
checker.set_param(arg.param).execs(
{arg.src, arg.filter, arg.bias, {}, {}});
}
}
void winograd_algo_extra_impl(const TensorNDArray& tensors, uint32_t m,
param::ConvBias param, Handle* handle,
......
dnn/test/common/conv_bias.h
浏览文件 @ edd7e167
......@@ -58,7 +58,10 @@ std::vector<TestArg> get_int8_chwn4_tensorcore_args(size_t kernel_size);
std::vector<TestArg> get_int8_nchw44_args(size_t kernel_size, size_t pack_size,
bool compute_float32 = false,
bool group_mode = false);
void check_conv_bias_preprocess(std::vector<conv_bias::TestArg> args,
Handle* handle, RNG* rng, float epsilon,
DType type0, DType type1, DType type2,
DType type3, const char* algo_name);
template <typename Opr>
using ConvBiasAlgoChecker = AlgoChecker<Opr>;
......
dnn/test/x86/conv_bias.cpp
浏览文件 @ edd7e167
......@@ -752,7 +752,7 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_DIRECT_STRIDE2) {
}
}
TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_INT8X8X) {
TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_INT8X8X32) {
using namespace conv_bias;
std::vector<TestArg> args;
......@@ -842,6 +842,98 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_INT8X8X) {
#undef cb2
}
TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_INT8X8X32_FILTER_PREPROCESS) {
using namespace conv_bias;
std::vector<TestArg> args;
auto run = [&](size_t oc, size_t ic, size_t w, size_t h, size_t kernel,
size_t p, NonlineMode nonline_mode) {
if (w + 2 * p < kernel || h + 2 * p < kernel)
return;
param::ConvBias param;
param.stride_h = 1;
param.stride_w = 1;
param.pad_h = p;
param.pad_w = p;
param.nonlineMode = nonline_mode;
//! no bias
args.emplace_back(param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel}, TensorShape{});
};
for (size_t kernel : {2, 3, 4, 5, 6, 7})
for (size_t ic : {1, 4, 8, 16})
for (size_t oc : {1, 4, 8})
for (size_t p : {0, 2})
for (size_t size : {20, 21, 24})
for (NonlineMode nonline_mode :
{NonlineMode::IDENTITY}) {
run(oc, ic, size, size, kernel, p, nonline_mode);
}
//! test OC block
run(2046, 1, 8, 8, 2, 0, NonlineMode::IDENTITY);
Checker<ConvBiasForward, OprWeightPreprocessProxy<ConvBiasForward>> checker(
handle());
UniformIntRNG rng{-50, 50};
#define cb(algo_name) \
checker.set_before_exec_callback( \
conv_bias::ConvBiasAlgoChecker<ConvBias>(algo_name)); \
checker.set_dtype(0, dtype::Int8()); \
checker.set_dtype(1, dtype::Int8()); \
checker.set_dtype(2, dtype::Int32()); \
checker.set_dtype(4, dtype::Int32()); \
for (auto&& arg : args) { \
checker.set_param(arg.param).execs({arg.src, arg.filter, {}, {}, {}}); \
} \
for (auto&& arg : args) { \
checker.set_dtype(0, dtype::QuantizedS8(2.5f)) \
.set_dtype(1, dtype::QuantizedS8(2.5f)) \
.set_dtype(2, dtype::QuantizedS32(6.25f)) \
.set_dtype(4, {}) \
.set_rng(0, &rng) \
.set_rng(1, &rng) \
.set_rng(2, &rng) \
.set_param(arg.param) \
.execs({arg.src, arg.filter, {}, {}, {}}); \
}
#define cb2(algo_name) \
checker.set_before_exec_callback( \
conv_bias::ConvBiasAlgoChecker<ConvBias>(algo_name)); \
checker.set_dtype(0, dtype::Int8()); \
checker.set_dtype(1, dtype::Int8()); \
checker.set_dtype(2, dtype::Int16()); \
checker.set_dtype(4, dtype::Int16()); \
for (auto&& arg : args) { \
checker.set_param(arg.param).execs({arg.src, arg.filter, {}, {}, {}}); \
}
#if MEGDNN_X86_WITH_MKL_DNN
if (megdnn::x86::is_supported(x86::SIMDType::VNNI)) {
cb("IM2COLMATMUL:X86_INT8X8X32_MKLDNN");
}
#endif
#if MEGDNN_X86_WITH_VNNI
if (megdnn::x86::is_supported(x86::SIMDType::VNNI)) {
cb("IM2COLMATMUL:X86_INT8X8X32_VNNI");
}
#endif
if (megdnn::x86::is_supported(x86::SIMDType::AVX2)) {
cb("IM2COLMATMUL:X86_INT8X8X32_AVX2_2X4X16");
cb("IM2COLMATMUL:X86_INT8X8X32_AVX2_4X16X2");
cb2("IM2COLMATMUL:X86_INT8X8X16_AVX2");
}
if (::megdnn::x86::is_supported(::megdnn::x86::SIMDType::SSE4_2)) {
cb("IM2COLMATMUL:X86_INT8X8X32_SSE_4X8X2");
cb2("IM2COLMATMUL:X86_INT8X8X16_SSE");
}
#undef cb
#undef cb2
}
TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_FP32) {
using namespace conv_bias;
std::vector<TestArg> args;
......@@ -950,6 +1042,61 @@ TEST_F(X86, CONV_BIAS_IM2COLMATMUL_FP32) {
#undef cb
}
TEST_F(X86, CONV_BIAS_IM2COLMATMUL_FP32_NOPACK_PREPROCESS) {
using namespace conv_bias;
std::vector<TestArg> args;
auto run = [&](size_t oc, size_t ic, size_t w, size_t h, size_t kernel,
size_t p, NonlineMode nonline_mode) {
if (w + 2 * p < kernel || h + 2 * p < kernel)
return;
param::ConvBias param;
param.stride_h = 1;
param.stride_w = 1;
param.pad_h = p;
param.pad_w = p;
param.nonlineMode = nonline_mode;
//! no bias
args.emplace_back(param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel}, TensorShape{});
args.emplace_back(param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel},
TensorShape{1, oc, 1, 1});
args.emplace_back(
param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel},
TensorShape{1, oc, (h + 2 * p - kernel) / param.stride_h + 1,
(w + 2 * p - kernel) / param.stride_w + 1});
};
for (size_t kernel : {2, 3, 4, 5, 6, 7})
for (size_t ic : {1, 4, 8, 16})
for (size_t oc : {1, 4, 8, 16, 300})
for (size_t p : {0, 2})
for (size_t size : {8, 24})
for (NonlineMode nonline_mode :
{NonlineMode::IDENTITY, NonlineMode::RELU}) {
run(oc, ic, size, size, kernel, p, nonline_mode);
}
run(2046, 8, 20, 20, 3, 1, NonlineMode::IDENTITY);
Checker<ConvBiasForward, OprWeightPreprocessProxy<ConvBiasForward>> checker(
handle());
#define cb(algo_name) \
checker.set_before_exec_callback( \
conv_bias::ConvBiasAlgoChecker<ConvBias>(algo_name)); \
for (auto&& arg : args) { \
checker.set_param(arg.param).execs( \
{arg.src, arg.filter, arg.bias, {}, {}}); \
}
cb("IM2COLMATMUL:X86_F32_BLAS");
#undef cb
}
#endif
......@@ -1020,6 +1167,73 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_FP32_PACKA) {
#undef cb
}
TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_FP32_PACKA_FILTER_PREPROCESS) {
using namespace conv_bias;
std::vector<TestArg> args;
auto run = [&](size_t oc, size_t ic, size_t w, size_t h, size_t kernel,
size_t p, NonlineMode nonline_mode) {
if (w + 2 * p < kernel || h + 2 * p < kernel)
return;
param::ConvBias param;
param.stride_h = 1;
param.stride_w = 1;
param.pad_h = p;
param.pad_w = p;
param.nonlineMode = nonline_mode;
//! no bias
args.emplace_back(param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel}, TensorShape{});
args.emplace_back(param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel},
TensorShape{1, oc, 1, 1});
args.emplace_back(
param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel},
TensorShape{1, oc, (h + 2 * p - kernel) / param.stride_h + 1,
(w + 2 * p - kernel) / param.stride_w + 1});
param.sparse = param::ConvBias::Sparse::GROUP;
args.emplace_back(param, TensorShape{1, 2 * ic, h, w},
TensorShape{2, oc, ic, kernel, kernel},
TensorShape{});
args.emplace_back(param, TensorShape{1, 2 * ic, h, w},
TensorShape{2, oc, ic, kernel, kernel},
TensorShape{1, oc * 2, 1, 1});
args.emplace_back(
param, TensorShape{1, 2 * ic, h, w},
TensorShape{2, oc, ic, kernel, kernel},
TensorShape{1, 2 * oc, (h + 2 * param.pad_h - kernel) / 1 + 1,
(w + 2 * param.pad_w - kernel) / 1 + 1});
};
for (size_t kernel : {2, 3, 4, 5, 6, 7})
for (size_t ic : {1, 4, 8, 16})
for (size_t oc : {1, 4, 8, 16})
for (size_t p : {0, 1})
for (size_t size : {8, 24})
for (NonlineMode nonline_mode :
{NonlineMode::IDENTITY, NonlineMode::RELU}) {
run(oc, ic, size, size, kernel, p, nonline_mode);
}
run(2046, 8, 20, 20, 3, 1, NonlineMode::IDENTITY);
Checker<ConvBiasForward, OprWeightPreprocessProxy<ConvBiasForward>> checker(
handle());
#define cb(algo_name) \
checker.set_before_exec_callback( \
conv_bias::ConvBiasAlgoChecker<ConvBias>(algo_name)); \
for (auto&& arg : args) { \
checker.set_param(arg.param).execs( \
{arg.src, arg.filter, arg.bias, {}, {}}); \
}
cb("IM2COLMATMUL:X86_F32_MKL_PACKA:192");
#undef cb
}
/**************************** Conv1x1 PackA *************************/
namespace {
void checker_conv_bias(std::vector<conv_bias::TestArg> args, Handle* handle,
......@@ -1169,6 +1383,77 @@ TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_QINT8) {
#undef cb
}
TEST_F(X86_MULTI_THREADS, CONV_BIAS_IM2COLMATMUL_QINT8_FILTER_PREPROCESS) {
using namespace conv_bias;
std::vector<TestArg> args;
auto run = [&](size_t oc, size_t ic, size_t w, size_t h, size_t kernel,
size_t p, NonlineMode nonline_mode) {
if (w + 2 * p < kernel || h + 2 * p < kernel)
return;
param::ConvBias param;
param.stride_h = 1;
param.stride_w = 1;
param.pad_h = p;
param.pad_w = p;
param.nonlineMode = nonline_mode;
//! no bias
args.emplace_back(param, TensorShape{1, ic, h, w},
TensorShape{oc, ic, kernel, kernel}, TensorShape{});
//! bias channel
args.emplace_back(param, TensorShape{2, ic, h, w},
TensorShape{oc, ic, kernel, kernel},
TensorShape{1, oc, 1, 1});
};
for (size_t kernel : {2, 3, 4, 5, 6, 7})
for (size_t ic : {1, 4, 8, 16})
for (size_t oc : {1, 4, 8})
for (size_t p : {0, 2})
for (size_t size : {20, 21, 24})
for (NonlineMode nonline_mode :
{NonlineMode::IDENTITY, NonlineMode::RELU,
NonlineMode::H_SWISH}) {
run(oc, ic, size, size, kernel, p, nonline_mode);
}
run(2046, 8, 20, 20, 3, 1, NonlineMode::IDENTITY);
Checker<ConvBiasForward, OprWeightPreprocessProxy<ConvBiasForward>> checker(
handle());
#define cb(algo_name) \
checker.set_before_exec_callback( \
conv_bias::ConvBiasAlgoChecker<ConvBias>(algo_name)); \
UniformIntRNG rng{-50, 50}; \
for (auto&& arg : args) { \
checker.set_dtype(0, dtype::QuantizedS8(2.5f)) \
.set_dtype(1, dtype::QuantizedS8(2.5f)) \
.set_dtype(2, dtype::QuantizedS32(6.25f)) \
.set_dtype(4, dtype::QuantizedS8(60.25)) \
.set_rng(0, &rng) \
.set_rng(1, &rng) \
.set_rng(2, &rng) \
.set_param(arg.param) \
.execs({arg.src, arg.filter, {}, {}, {}}); \
}
#if MEGDNN_X86_WITH_MKL_DNN
if (x86::is_supported(x86::SIMDType::VNNI)) {
cb("IM2COLMATMUL:X86_INT8X8X32_MKLDNN");
}
#endif
#if MEGDNN_X86_WITH_VNNI
if (x86::is_supported(x86::SIMDType::VNNI)) {
cb("IM2COLMATMUL:X86_INT8X8X32_VNNI");
}
#endif
if (x86::is_supported(x86::SIMDType::AVX2)) {
cb("IM2COLMATMUL:X86_INT8X8X32_AVX2_2X4X16");
}
#undef cb
}
TEST_F(X86, CONV_BIAS_MATMUL) {
using namespace conv_bias;
std::vector<TestArg> args;
......
Markdown is supported
0% 或 .
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 或