fix(dnn/cuda): fix AlgoFallbackNCHWQS8 to support Float32 dst

GitOrigin-RevId: 06f90f5cf384bc4ddb2f97860e4f530ee9a85705

dnn/src/cuda/conv_bias/conv_nchwqs8.cpp

@@ -50,15 +50,23 @@ void ConvBiasForwardImpl::AlgoFallbackNCHWQS8::make_inner_layout(
     deduce_reformat_layout(relayout_src, *args.filter_layout,
                            inner_weight_layout,
                            RelayoutFormat::Param::Mode::NCHW_NCHW4_WEIGHT);
-    deduce_reformat_layout(relayout_src, *args.dst_layout, inner_dst_layout,
-                           RelayoutFormat::Param::Mode::NCHW_NCHW4, 0,
-                           args.filter_meta.group);
-    deduce_reformat_layout(relayout_src, *args.bias_layout, inner_bias_layout,
-                           RelayoutFormat::Param::Mode::NCHW_NCHW4, 0,
-                           args.filter_meta.group);
-    deduce_reformat_layout(relayout_src, *args.z_layout, inner_z_layout,
-                           RelayoutFormat::Param::Mode::NCHW_NCHW4, 0,
-                           args.filter_meta.group);
+    bool dst_float = args.dst_layout->dtype.enumv() == DTypeEnum::Float32;
+    if (dst_float) {
+        inner_dst_layout = *args.dst_layout;
+        inner_bias_layout = *args.bias_layout;
+        inner_z_layout = *args.z_layout;
+    } else {
+        deduce_reformat_layout(relayout_src, *args.dst_layout, inner_dst_layout,
+                               RelayoutFormat::Param::Mode::NCHW_NCHW4, 0,
+                               args.filter_meta.group);
+        deduce_reformat_layout(relayout_src, *args.bias_layout,
+                               inner_bias_layout,
+                               RelayoutFormat::Param::Mode::NCHW_NCHW4, 0,
+                               args.filter_meta.group);
+        deduce_reformat_layout(relayout_src, *args.z_layout, inner_z_layout,
+                               RelayoutFormat::Param::Mode::NCHW_NCHW4, 0,
+                               args.filter_meta.group);
+    }
 };
 
 bool ConvBiasForwardImpl::AlgoFallbackNCHWQS8::is_available(
@@ -70,8 +78,7 @@ bool ConvBiasForwardImpl::AlgoFallbackNCHWQS8::is_available(
     auto&& param = args.opr->param();
     bool is_format_ok = param.format == param::ConvBias::Format::NCHW;
     bool is_version_ok = CUDNN_VERSION >= 7500;
-    bool is_dtype_ok =
-            args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS8;
+    bool is_dtype_ok = args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS8;
     bool is_bias_ok =
             args.bias_layout->ndim == 0 ||
             (args.bias_layout->ndim == 4 && args.bias_layout->shape[0] == 1 &&
@@ -90,17 +97,23 @@ WorkspaceBundle ConvBiasForwardImpl::AlgoFallbackNCHWQS8::get_workspace_bundle(
     TensorLayout inner_z_layout;
     make_inner_layout(args, inner_src_layout, inner_weight_layout,
                       inner_dst_layout, inner_bias_layout, inner_z_layout);
-    auto opr = args.handle->create_operator<ConvBiasForward>();
     Param inner_conv_param = args.opr->param();
-    inner_conv_param.format = Param::Format::NCHW4;
+    size_t ws_dst = 0, ws_bias = 0, ws_z = 0;
+    if (args.dst_layout->dtype.enumv() == DTypeEnum::Float32) {
+        inner_conv_param.format = Param::Format::NCHW4_NCHW;
+    } else {
+        inner_conv_param.format = Param::Format::NCHW4;
+        ws_dst = inner_dst_layout.span().dist_byte();
+        ws_bias = inner_bias_layout.span().dist_byte();
+        ws_z = inner_z_layout.span().dist_byte();
+    }
+    auto opr = args.handle->create_operator<ConvBiasForward>();
     opr->param() = inner_conv_param;
-    return WorkspaceBundle(ptr, {inner_src_layout.span().dist_byte(),
-                                 inner_weight_layout.span().dist_byte(),
-                                 inner_dst_layout.span().dist_byte(),
-                                 inner_bias_layout.span().dist_byte(),
-                                 inner_z_layout.span().dist_byte(),
-                                 opr->get_workspace_in_bytes(
-                                         inner_src_layout, inner_weight_layout,
+    return WorkspaceBundle(
+            ptr,
+            {inner_src_layout.span().dist_byte(),
+             inner_weight_layout.span().dist_byte(), ws_dst, ws_bias, ws_z,
+             opr->get_workspace_in_bytes(inner_src_layout, inner_weight_layout,
                                          inner_bias_layout, inner_z_layout,
                                          inner_dst_layout, nullptr)});
 }
@@ -145,22 +158,33 @@ void ConvBiasForwardImpl::AlgoFallbackNCHWQS8::exec(
     TensorND inner_bias(bundle.get(3), inner_bias_layout);
     TensorND inner_z(bundle.get(4), inner_z_layout);
 
+    bool dst_float = args.dst_layout->dtype.enumv() == DTypeEnum::Float32;
+
     Param inner_conv_param = args.opr->param();
-    inner_conv_param.format = Param::Format::NCHW4;
+    inner_conv_param.format =
+            dst_float ? Param::Format::NCHW4_NCHW : Param::Format::NCHW4;
     auto inner_opr = args.handle->create_operator<ConvBiasForward>();
     inner_opr->param() = inner_conv_param;
 
     relayout_nchw_nchw4->exec(*args.src_tensor, inner_src, {});
     relayout_weight->exec(*args.filter_tensor, inner_weight, {});
-    if (inner_bias_layout.ndim > 0) {
-        relayout_nchw_nchw4->exec(*args.bias_tensor, inner_bias, {});
-    }
-    if (inner_z_layout.ndim > 0) {
-        relayout_nchw_nchw4->exec(*args.z_tensor, inner_z, {});
+
+    if (dst_float) {
+        inner_opr->exec(inner_src, inner_weight, *args.bias_tensor,
+                        *args.z_tensor, *args.dst_tensor, nullptr,
+                        Workspace((dt_byte*)bundle.get(5), bundle.get_size(5)));
+    } else {
+        if (inner_bias_layout.ndim > 0) {
+            relayout_nchw_nchw4->exec(*args.bias_tensor, inner_bias, {});
+        }
+        if (inner_z_layout.ndim > 0) {
+            relayout_nchw_nchw4->exec(*args.z_tensor, inner_z, {});
+        }
+        inner_opr->exec(inner_src, inner_weight, inner_bias, inner_z, inner_dst,
+                        nullptr,
+                        Workspace((dt_byte*)bundle.get(5), bundle.get_size(5)));
+        relayout_nchw4_nchw->exec(inner_dst, *args.dst_tensor, {});
     }
-    inner_opr->exec(inner_src, inner_weight, inner_bias, inner_z, inner_dst,
-                    nullptr, Workspace((dt_byte*)bundle.get(5), bundle.get_size(5)));
-    relayout_nchw4_nchw->exec(inner_dst, *args.dst_tensor, {});
 }
 
 // vim: syntax=cpp.doxygen

dnn/src/cuda/conv_bias/cudnn_conv_bias_activation.cpp

@@ -192,8 +192,7 @@ void ConvBiasForwardImpl::AlgoCUDNNConvBiasActivation::exec(
          dst_dtype = args.dst_layout->dtype;
     megdnn_assert(
             (src_dtype.category() == dst_dtype.category()) ||
-            (args.opr->param().format == param::ConvBias::Format::NCHW4_NCHW &&
-             src_dtype.enumv() == DTypeEnum::QuantizedS8 &&
+            (src_dtype.enumv() == DTypeEnum::QuantizedS8 &&
              dst_dtype.enumv() == DTypeEnum::Float32));
     megdnn_assert(src_dtype.category() == filter_dtype.category());
 

dnn/test/cuda/conv_bias_int8.cpp

@@ -28,6 +28,15 @@ namespace megdnn {
 namespace test {
 namespace conv{
 
+TEST_F(CUDA, CONV_BIAS_INT8_NCHW4_CUDNN_CONVOLUTION) {
+    require_compute_capability(7, 5);
+    conv_bias::check_conv_bias(
+            dtype::QuantizedS8{1.2f}, dtype::QuantizedS8{1.3f},
+            dtype::QuantizedS32{1.2f * 1.3f}, dtype::QuantizedS8{1.3f},
+            handle_cuda(), "DEFAULT:CUDNN:ConvBiasActivation:",
+            param::ConvBias::Format::NCHW4);
+}
+
 TEST_F(CUDA, CONV_BIAS_INT8_NCHW4_1x1) {
     require_compute_capability(6, 1);
     conv_bias::check_conv_bias(
@@ -689,6 +698,82 @@ TEST_F(CUDA, CONV_BIAS_INT8_CHWN4_UNROLL_WIDTH_TENSORCORE_1x1_ALGO_2) {
 }
 
 
+TEST_F(CUDA, FALLBACK_CONV_QS8) {
+    require_compute_capability_eq(7, 5);
+    Checker<ConvBiasForward> checker(handle_cuda());
+    auto check = [&checker](const std::string&& algo) {
+        checker.set_before_exec_callback(
+                conv_bias::ConvBiasAlgoChecker<ConvBiasForward>(algo.c_str()));
+        UniformIntRNG rng{-3, 3};
+        UniformIntRNG bias_rng{-50, 50};
+        checker.set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .set_rng(2, &bias_rng)
+                .set_rng(3, &rng)
+                .set_dtype(0, dtype::QuantizedS8{1.2f})
+                .set_dtype(1, dtype::QuantizedS8{1.3f})
+                .set_dtype(2, dtype::QuantizedS32{1.2f * 1.3f})
+                .set_dtype(3, dtype::QuantizedS8{19.990229f})
+                .set_dtype(4, dtype::QuantizedS8{19.990228f})
+                .set_epsilon(1e-3)
+                .set_max_avg_error(1e-1)
+                .set_max_avg_biased_error(1e-3);
+        param::ConvBias param;
+        param.pad_h = param.pad_w = 1;
+        param.stride_h = param.stride_w = 2;
+        param.format = param::ConvBias::Format::NCHW;
+        checker.set_param(param).execs({{16, 15, 14, 14},
+                                        {28, 15, 3, 3},
+                                        {1, 28, 1, 1},
+                                        {16, 28, 7, 7},
+                                        {}});
+        checker.set_param(param).execs({{16, 32, 14, 14},
+                                        {32, 32, 3, 3},
+                                        {1, 32, 1, 1},
+                                        {},
+                                        {}});
+    };
+    check("FALLBACK_CONV_NCHW_QS8");
+}
+
+TEST_F(CUDA, FALLBACK_CONV_QS8_F32) {
+    require_compute_capability_eq(7, 5);
+    Checker<ConvBiasForward> checker(handle_cuda());
+    auto check = [&checker](const std::string&& algo) {
+        checker.set_before_exec_callback(
+                conv_bias::ConvBiasAlgoChecker<ConvBiasForward>(algo.c_str()));
+        UniformIntRNG rng{-3, 3};
+        UniformFloatRNG bias_rng{-50.f, 50.f};
+        checker.set_rng(0, &rng)
+                .set_rng(1, &rng)
+                .set_rng(2, &bias_rng)
+                .set_rng(3, &rng)
+                .set_dtype(0, dtype::QuantizedS8{1.2f})
+                .set_dtype(1, dtype::QuantizedS8{1.3f})
+                .set_dtype(2, dtype::Float32{})
+                .set_dtype(3, dtype::Float32{})
+                .set_dtype(4, dtype::Float32{})
+                .set_epsilon(1e-3)
+                .set_max_avg_error(1e-1)
+                .set_max_avg_biased_error(1e-3);
+        param::ConvBias param;
+        param.pad_h = param.pad_w = 1;
+        param.stride_h = param.stride_w = 2;
+        param.format = param::ConvBias::Format::NCHW;
+        checker.set_param(param).execs({{16, 15, 14, 14},
+                                        {28, 15, 3, 3},
+                                        {1, 28, 1, 1},
+                                        {16, 28, 7, 7},
+                                        {}});
+        checker.set_param(param).execs({{16, 32, 14, 14},
+                                        {32, 32, 3, 3},
+                                        {1, 32, 1, 1},
+                                        {},
+                                        {}});
+    };
+    check("FALLBACK_CONV_NCHW_QS8");
+}
+
 TEST_F(CUDA, CUTLASS_CONV_BIAS_INT8_WEIGHT_PREPROCESS) {
     require_compute_capability(6, 1);
     Checker<ConvBiasForward, OprWeightPreprocessProxy<ConvBiasForward>> checker(