feat(mgb/gopt): global layout transform support cuda fp16

GitOrigin-RevId: 1449c54215

src/gopt/impl/global_layout_transform/layout_transform_context.cpp

@@ -78,13 +78,13 @@ std::unique_ptr<LayoutTransformContext> make_cuda_ctx(
                      OprFormatConfigID::NHWC})
             .add_opr_config(
                     opr::PoolingForward::typeinfo(),
-                    {OprFormatConfigID::NCHW4, OprFormatConfigID::NCHW32,
-                     OprFormatConfigID::NHWC, OprFormatConfigID::NCHW64,
-                     OprFormatConfigID::CHWN4})
+                    {OprFormatConfigID::NCHW, OprFormatConfigID::NCHW4,
+                     OprFormatConfigID::NCHW32, OprFormatConfigID::NHWC,
+                     OprFormatConfigID::NCHW64, OprFormatConfigID::CHWN4})
             .add_opr_config(
                     opr::WarpPerspectiveForward::typeinfo(),
-                    {OprFormatConfigID::NHWC, OprFormatConfigID::NCHW4,
-                     OprFormatConfigID::NCHW64});
+                    {OprFormatConfigID::NCHW, OprFormatConfigID::NHWC,
+                     OprFormatConfigID::NCHW4, OprFormatConfigID::NCHW64});
     return ctx;
 }
 

src/gopt/impl/global_layout_transform/opr_tensor_formats_config.cpp

@@ -191,8 +191,11 @@ struct OprSingleInOutTensorFormatsDispatcherImpl<OprFormatConfigID::NHWC> {
         config.typeinfo = opr->dyn_typeinfo();
         config.opr_format = OprFormat::NHWC;
         config.config_id = OprFormatConfigID::NHWC;
-        bool available = opr->input(0)->dtype().enumv() == DTypeEnum::Quantized4Asymm ||
+        bool f16_config = DNN_FLOAT16_SELECT(
+                (opr->input(0)->dtype().enumv() == DTypeEnum::Float16), true);
+        bool i4_config = opr->input(0)->dtype().enumv() == DTypeEnum::Quantized4Asymm ||
                          opr->input(0)->dtype().enumv() == DTypeEnum::QuantizedS4;
+        bool available = f16_config || i4_config;
         config.input_dtypes = {opr->input(0)->dtype().enumv()};
         config.input_tensor_types = {TensorType::FEATURE};
         available &= opr->output(0)->dtype().enumv() == opr->input(0)->dtype().enumv();
@@ -275,16 +278,22 @@ struct ConvTensorFormatsDispatcherImpl<Opr, OprFormatConfigID::NHWC> {
         config.opr_format = OprFormat::NHWC;
         config.config_id = OprFormatConfigID::NHWC;
         auto check_dtype = [](const DType& dt) {
+            bool f16_config =
+                    DNN_FLOAT16_SELECT((dt.enumv() == DTypeEnum::Float16), true);
             bool i4_config = dt.enumv() == DTypeEnum::Quantized4Asymm ||
                              dt.enumv() == DTypeEnum::QuantizedS4;
             bool i8_config = dt.enumv() == DTypeEnum::QuantizedS8;
-            return i4_config || i8_config;
+            return f16_config || i4_config || i8_config;
         };
         bool available = true;
         for (size_t i = 0; i < opr->input().size(); ++i) {
-            if (i == 2)
-                available &= opr->input(i)->dtype().enumv() == DTypeEnum::QuantizedS32;
-            else {
+            if (i == 2) {
+                available &=
+                        opr->input(i)->dtype().enumv() == DTypeEnum::QuantizedS32 ||
+                        DNN_FLOAT16_SELECT(
+                                opr->input(i)->dtype().enumv() == DTypeEnum::Float16,
+                                true);
+            } else {
                 available &= check_dtype(opr->input(i)->dtype());
             }
             config.input_dtypes.emplace_back(opr->input(i)->dtype().enumv());
@@ -866,12 +875,18 @@ struct ConvTensorFormatsDispatcherImpl<
         config.config_id = OprFormatConfigID::NHWC;
         bool available = true;
         for (size_t i = 0; i < opr->input().size(); ++i) {
-            available &= opr->input(i)->dtype().enumv() == DTypeEnum::QuantizedS8;
+            available &=
+                    opr->input(i)->dtype().enumv() == DTypeEnum::QuantizedS8 ||
+                    DNN_FLOAT16_SELECT(
+                            opr->input(i)->dtype().enumv() == DTypeEnum::Float16, true);
             config.input_dtypes.emplace_back(opr->input(i)->dtype().enumv());
             TensorType tensor_type = i == 0 ? TensorType::WEIGHT : TensorType::FEATURE;
             config.input_tensor_types.emplace_back(tensor_type);
         }
-        available &= opr->output(0)->dtype().enumv() == DTypeEnum::QuantizedS8;
+        available &=
+                opr->output(0)->dtype().enumv() == DTypeEnum::QuantizedS8 ||
+                DNN_FLOAT16_SELECT(
+                        opr->output(0)->dtype().enumv() == DTypeEnum::Float16, true);
         config.output_dtypes.emplace_back(opr->output(0)->dtype().enumv());
         available &= conv.param().sparse == opr::ConvBias::Param::Sparse::DENSE;
         config.input_tensor_formats = {
@@ -934,6 +949,7 @@ StaticData::StaticData() {
     OPR_TENSOR_FORMATS_CONFIG_REG(ConvBias, NCHW44_DOT_HYBRID);
 
     OPR_TENSOR_FORMATS_CONFIG_REG(ConvolutionForward, NCHW);
+    OPR_TENSOR_FORMATS_CONFIG_REG(ConvolutionForward, NHWC);
     OPR_TENSOR_FORMATS_CONFIG_REG(ConvolutionForward, NCHW4);
     OPR_TENSOR_FORMATS_CONFIG_REG(ConvolutionForward, NCHW44);
     OPR_TENSOR_FORMATS_CONFIG_REG(ConvolutionForward, NCHW88);

src/gopt/test/layout_transform_pass.cpp

@@ -29,6 +29,8 @@
 #include "./cache_data.h"
 #endif
 
+#include "megbrain/plugin/opr_io_dump.h"
+
 using namespace mgb;
 using namespace gopt;
 using namespace serialization;
@@ -748,6 +750,95 @@ TEST(TestLayoutTransform, CanonicalizeLayoutTransform) {
     MGB_ASSERT_TENSOR_EQ(t1, t2);
 }
 
+#if MGB_CUDA
+TEST(TestLayoutTransform, Resnet18_F16) {
+    REQUIRE_GPU(1);
+    auto cn = CompNode::load("gpu0");
+    auto&& prop = CompNodeEnv::from_comp_node(cn).cuda_env().device_prop;
+    auto sm_ver = prop.major * 10 + prop.minor;
+    if (sm_ver < 70) {
+        printf("This testcast ignored due to insufficient cuda cap(got: %d, "
+               "expected: %d)\n",
+               sm_ver, 70);
+        return;
+    }
+
+    Network network(cn);
+    auto output = make_resnet18(network, 16);
+    using S = opr::mixin::AlgoChooserHelper::ExecutionPolicy::Strategy;
+    S strategy = S::PROFILE;
+    gopt::modify_opr_algo_strategy_inplace({{output}}, strategy);
+
+    HostTensorND t1;
+    auto func1 = network.graph->compile({make_callback_copy(output, t1)});
+    func1->execute();
+
+    using OprFormatConfigID = LayoutTransformContext::OprFormatConfigID;
+    using OprList = LayoutTransformContext::OprList;
+    using Attribute = LayoutTransformContext::Attribute;
+    using Target = LayoutTransformContext::Target;
+    using ReformatAttribute = LayoutTransformContext::ReformatAttribute;
+    OprList opr_list = {
+            opr::ConvBiasForward::typeinfo(), opr::ElemwiseMultiType::typeinfo(),
+            opr::Elemwise::typeinfo(),        opr::TypeCvt::typeinfo(),
+            opr::PoolingForward::typeinfo(),  opr::WarpPerspectiveForward::typeinfo(),
+    };
+    SmallVector<TensorFormats> available_tensor_formats = {
+            TensorFormats::NCHW, TensorFormats::NHWC};
+    Attribute attribute = {
+            OprFormatConfigID::NCHW, TensorFormats::NCHW, Target::UNSPEC,
+            ReformatAttribute::AUTO_PADDING_NHWC};
+    auto ctx = std::make_unique<LayoutTransformContext>(
+            std::move(opr_list), std::move(available_tensor_formats), attribute);
+    ctx->add_opr_config(
+               opr::ConvBiasForward::typeinfo(),
+               {OprFormatConfigID::NCHW, OprFormatConfigID::NHWC})
+            .add_opr_config(
+                    opr::PoolingForward::typeinfo(),
+                    {OprFormatConfigID::NCHW, OprFormatConfigID::NHWC});
+#if MGB_WITH_CACHED_TEST
+    auto profiler = std::make_unique<ProfilerMock>(
+            static_cast<const uint8_t*>(TestLayoutTransform_Resnet18_F16.data()),
+            TestLayoutTransform_Resnet18_F16.size());
+#else
+    auto profiler = ProfilerBase::make_cached_profiler(
+            "TestLayoutTransform.Resnet18_F16.cache");
+#endif
+    std::unique_ptr<SolverBase> solver{
+            new DynamicProgrammingSolver(std::move(profiler))};
+    auto new_output =
+            gopt::GraphOptimizer{}
+                    .add_pass(ConvertF32ToF16Pass::make(false))
+                    .add_pass<FuseConvBiasNonlinPass>()
+                    .add_pass<FuseConvBiasZPass>()
+                    .add_pass<LayoutTransformPass>(std::move(ctx), std::move(solver))
+                    .add_pass<ShuffleShuffleRemovePass>()
+                    .add_pass(FuseNCHW4Int8Preprocess::make())
+                    .add_pass<FoldingConvBiasDimshufflePass>()
+                    .add_pass<ParamFusePass>()
+                    .add_pass<ParamMergePass>()
+                    .apply({{output}})
+                    .endpoint_vars();
+    auto new_out_var = new_output[0];
+    /// check global layout transform pass
+    auto nr_dimshuffle = find_opr_num<opr::Dimshuffle>(new_out_var);
+    ASSERT_EQ(nr_dimshuffle, 4u);
+    /// check pass fuse conv bias with z
+    auto nr_elemwise = find_opr_num<opr::Elemwise>(new_out_var);
+    ASSERT_EQ(nr_elemwise, 4u);
+    /// 21 convolutions, 21 weights and 21 bias, total 42 parameters
+    const auto& param_merge = find_opr<opr::MultipleDeviceTensorHolder>(new_out_var);
+    ASSERT_EQ(param_merge.output().size(), 42u);
+
+    GraphProfiler gprof{network.graph.get()};
+    HostTensorND t2;
+    auto func2 = network.graph->compile({make_callback_copy(new_out_var, t2)});
+    func2->execute();
+    gprof.to_json_full(func2.get())->writeto_fpath(output_file("resnet18_f16.json"));
+    MGB_ASSERT_TENSOR_NEAR(t1, t2, 1e-3);
+}
+#endif
+
 TEST(TestLayoutTransform, Resnet18_F32) {
     auto cn = CompNode::load("cpu0");
 
@@ -1115,4 +1206,5 @@ TEST(TestLayoutTransform, MobileNetV2_NCHW44_DOT) {
     /// check correct
     MGB_ASSERT_TENSOR_EQ(t1, t2);
 }
+
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/gopt/test/network.cpp

@@ -38,8 +38,13 @@ SymbolVar Network::add_conv(
         param.nonlineMode = opr::ConvBias::Param::NonlineMode::IDENTITY;
     }
 
-    auto conv = opr::ConvBias::make(
-            f, weight, bias, param, {}, OperatorNodeConfig{out_dtype});
+    SymbolVar conv;
+    if (out_dtype.category() == DTypeCategory::QUANTIZED) {
+        conv = opr::ConvBias::make(
+                f, weight, bias, param, {}, OperatorNodeConfig{out_dtype});
+    } else {
+        conv = opr::ConvBias::make(f, weight, bias, param, {});
+    }
     weight_idx++;
     bias_idx++;
     return conv;