From 40e778fbd1e175b2a722b93fc280705cbcad3cde Mon Sep 17 00:00:00 2001
From: Megvii Engine Team <megengine@megvii.com>
Date: Tue, 8 Sep 2020 19:35:19 +0800
Subject: [PATCH] fix(mgb): fix-deconv-io16c32-pass-fail

GitOrigin-RevId: 4b44cb48cb15206c4f5735f33927200407dbd283
---
 src/gopt/impl/inference.cpp | 24 ++++++++++++++++++++++++
 src/gopt/test/inference.cpp | 27 +++++++++++++++++++++++++++
 2 files changed, 51 insertions(+)

diff --git a/src/gopt/impl/inference.cpp b/src/gopt/impl/inference.cpp
index f26b81b8b..8115f61c5 100644
--- a/src/gopt/impl/inference.cpp
+++ b/src/gopt/impl/inference.cpp
@@ -771,6 +771,29 @@ std::unique_ptr<ConvertF32ToF16Pass> ConvertF32ToF16Pass::make(
         return new_conv_opr.node()->owner_opr();
     };
 
+    auto replace_deconv_opr = [use_f32_comp](OperatorNodeBase* opr,
+                                           const VarNodeArray& new_inp) {
+        mgb_assert(opr->input().size() == new_inp.size());
+        auto& deconv_opr = opr->cast_final_safe<opr::ConvolutionBackwardData>();
+        auto new_param = deconv_opr.param();
+        if (use_f32_comp) {
+            new_param.compute_mode =
+                    megdnn::param::Convolution::ComputeMode::FLOAT32;
+        }
+        mgb_assert(new_inp[0]->dtype() == dtype::Float16(),
+                   "inp %s:%s, owner_opr:%s", new_inp[0]->dtype().name(),
+                   new_inp[0]->name().c_str(),
+                   new_inp[0]->owner_opr()->name().c_str());
+        mgb_assert(new_inp[1]->dtype() == dtype::Float16(),
+                   "inp %s:%s, owner_opr:%s", new_inp[1]->dtype().name(),
+                   new_inp[1]->name().c_str(),
+                   new_inp[1]->owner_opr()->name().c_str());
+        auto new_deconv_opr = opr::ConvolutionBackwardData::make(
+                new_inp[0], new_inp[1], new_param, deconv_opr.execution_policy(),
+                deconv_opr.config());
+        return new_deconv_opr.node()->owner_opr();
+    };
+
     auto replace_convbias_opr = [use_f32_comp](OperatorNodeBase* opr,
                                                const VarNodeArray& new_inp) {
         auto& convbias_opr = opr->cast_final_safe<opr::ConvBiasForward>();
@@ -941,6 +964,7 @@ std::unique_ptr<ConvertF32ToF16Pass> ConvertF32ToF16Pass::make(
     replace_func[opr::Host2DeviceCopy::typeinfo()] = replace_h2d_opr;
     replace_func[opr::SharedDeviceTensor::typeinfo()] = replace_sdt_opr;
     replace_func[opr::Convolution::typeinfo()] = replace_conv_opr;
+    replace_func[opr::ConvolutionBackwardData::typeinfo()] = replace_deconv_opr;
     replace_func[opr::ConvBias::typeinfo()] = replace_convbias_opr;
     replace_func[opr::MatrixMul::typeinfo()] = replace_matmul_opr;
     replace_func[opr::Reduce::typeinfo()] = replace_reduce_opr;
diff --git a/src/gopt/test/inference.cpp b/src/gopt/test/inference.cpp
index 6faa9ac62..cb2cd183b 100644
--- a/src/gopt/test/inference.cpp
+++ b/src/gopt/test/inference.cpp
@@ -710,6 +710,33 @@ TEST(TestGoptInference, Float16IOFloat32Compute) {
     MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-3);
 }
 
+TEST(TestGoptInference, Float16IOFloat32ComputeDeConv) {
+    constexpr size_t INP_H = 10, INP_W = 10;
+    HostTensorGenerator<> gen;
+    auto graph = ComputingGraph::make();
+    auto mkvar = [&](const char* name, const TensorShape& shp) {
+        return opr::Host2DeviceCopy::make(*graph, gen(shp)).rename(name);
+    };
+    graph->options().graph_opt_level = 0;
+
+    auto s0 = mkvar("s0", {5, 5, 3, 3}),
+         s1 = mkvar("s1", {1, 5, INP_H, INP_W});
+    auto y = opr::ConvolutionBackwardData::make(s0, s1, {}, {});
+    SymbolVar y_opt;
+    auto options = gopt::OptimizeForInferenceOptions{};
+    options.enable_f16_io_f32_comp();
+    unpack_vector(gopt::optimize_for_inference({y}, options), y_opt);
+    ASSERT_EQ(find_opr<opr::ConvolutionBackwardData>(y_opt).param().compute_mode,
+              opr::ConvBias::Param::ConvBias::ComputeMode::FLOAT32);
+    ASSERT_EQ(y_opt.dtype(), dtype::Float32());
+
+    HostTensorND host_y, host_y_opt;
+    auto func = graph->compile({make_callback_copy(y, host_y),
+                                make_callback_copy(y_opt, host_y_opt)});
+    func->execute();
+    MGB_ASSERT_TENSOR_NEAR(host_y, host_y_opt, 1e-2);
+}
+
 TEST(TestGoptInference, Float16IOFloat32ComputeWarpPerspective) {
     constexpr size_t INP_H = 10, INP_W = 10, N = 2;
     HostTensorGenerator<> gen;
-- 
GitLab