refactor(mgb): make conv handle noncontiguous tensors

GitOrigin-RevId: 86282709b3013ae44e13f74639a1c45a0dad97b0

dnn/include/megdnn/oprs/nn.h

@@ -511,6 +511,12 @@ protected:
             const TensorLayout& bias, const TensorLayout& z,
             const TensorLayout& dst, size_t workspace_in_bytes,
             const PreprocessedFilter* preprocessed_filter);
+
+    CanonizedFilterMeta check_exec_allow_noncontiguous(
+            const TensorLayout& src, const TensorLayout& filter,
+            const TensorLayout& bias, const TensorLayout& z,
+            const TensorLayout& dst, size_t workspace_in_bytes,
+            const PreprocessedFilter* preprocessed_filter);
 };
 using ConvBias = ConvBiasForward;
 

dnn/src/common/conv_bias.cpp

@@ -11,30 +11,18 @@
  */
 
 #include "src/common/conv_bias.h"
-#include "megdnn/oprs/nn.h"
 #include "src/common/utils.h"
 #include "src/common/opr_delegate.h"
 
 namespace megdnn {
+namespace {
 
-void ConvBiasForward::deduce_dtype(DType src, DType filter, DType /* bias */,
-                                   DType /* z */, DType& dst) {
-    check_or_deduce_dtype_fwd(src, filter, dst);
-}
-
-void ConvBiasForward::deduce_layout(const TensorLayout& src,
-                                    const TensorLayout& filter,
-                                    const TensorLayout& /* bias */,
-                                    const TensorLayout& /* z */,
-                                    TensorLayout& dst) {
-    deduce_layout_fwd(src, filter, dst);
-}
-
-ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
-        const TensorLayout& src, const TensorLayout& filter,
-        const TensorLayout& bias, const TensorLayout& z,
-        const TensorLayout& dst, size_t workspace_in_bytes,
-        const PreprocessedFilter* preprocessed_filter) {
+void do_check_exec_common(
+        ConvBiasForward* opr, const TensorLayout& src,
+        const TensorLayout& filter, const TensorLayout& bias,
+        const TensorLayout& z, const TensorLayout& dst,
+        size_t workspace_in_bytes,
+        const ConvBiasForward::PreprocessedFilter* preprocessed_filter) {
     megdnn_assert((src.dtype.enumv() == filter.dtype.enumv()) ||
                   (src.dtype.enumv() == DTypeEnum::Quantized4Asymm &&
                    filter.dtype.enumv() == DTypeEnum::QuantizedS4));
@@ -52,9 +40,8 @@ ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
         }
     }
 
-    auto ret = check_layout_fwd(src, filter, dst);
     megdnn_assert_contiguous(bias);
-    auto required_workspace_in_bytes = get_workspace_in_bytes(
+    auto required_workspace_in_bytes = opr->get_workspace_in_bytes(
             src, filter, bias, z, dst, preprocessed_filter);
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes,
                   "worksapce have size of %zu, but need %zu",
@@ -68,55 +55,58 @@ ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
                 return bias.eq_layout(dst);
             }
         };
-        if (check_eq(bias, dst))
-            return ret;
-        if (param().format == param::ConvBias::Format::NCHW ||
-            param().format == param::ConvBias::Format::NCHW4_NCHW) {
+        if (check_eq(bias, dst)) {
+            return;
+        }
+        if (opr->param().format == param::ConvBias::Format::NCHW ||
+            opr->param().format == param::ConvBias::Format::NCHW4_NCHW) {
             megdnn_assert(bias.shape[0] == 1);
             megdnn_assert(bias.shape[1] == dst.shape[1], "bias:%s, dst:%s",
                           bias.to_string().c_str(), dst.to_string().c_str());
             megdnn_assert(bias.shape[2] == 1);
             megdnn_assert(bias.shape[3] == 1);
-        } else if (param().format == param::ConvBias::Format::NHWC) {
+        } else if (opr->param().format == param::ConvBias::Format::NHWC) {
             megdnn_assert(bias.shape[0] == 1);
             megdnn_assert(bias.shape[1] == 1);
             megdnn_assert(bias.shape[2] == 1);
             megdnn_assert(bias.shape[3] == dst.shape[3], "bias:%s, dst:%s",
                           bias.to_string().c_str(), dst.to_string().c_str());
-        } else if (param().format == param::ConvBias::Format::NCHW4 ||
-                   param().format == param::ConvBias::Format::NCHW44 ||
-                   param().format == param::ConvBias::Format::NCHW44_DOT ||
-                   param().format == param::ConvBias::Format::NCHW32_NCHW4) {
+        } else if (opr->param().format == param::ConvBias::Format::NCHW4 ||
+                   opr->param().format == param::ConvBias::Format::NCHW44 ||
+                   opr->param().format == param::ConvBias::Format::NCHW44_DOT ||
+                   opr->param().format ==
+                           param::ConvBias::Format::NCHW32_NCHW4) {
             megdnn_assert(bias.shape[0] == 1);
             megdnn_assert(bias.shape[1] == dst.shape[1], "bias:%s, dst:%s",
                           bias.to_string().c_str(), dst.to_string().c_str());
             megdnn_assert(bias.shape[2] == 1);
             megdnn_assert(bias.shape[3] == 1);
             megdnn_assert(bias.shape[4] == 4);
-        } else if (param().format == param::ConvBias::Format::NCHW8 ||
-                   param().format == param::ConvBias::Format::NCHW88 ) {
+        } else if (opr->param().format == param::ConvBias::Format::NCHW8 ||
+                   opr->param().format == param::ConvBias::Format::NCHW88) {
             megdnn_assert(bias.shape[0] == 1);
             megdnn_assert(bias.shape[1] == dst.shape[1], "bias:%s, dst:%s",
                           bias.to_string().c_str(), dst.to_string().c_str());
             megdnn_assert(bias.shape[2] == 1);
             megdnn_assert(bias.shape[3] == 1);
             megdnn_assert(bias.shape[4] == 8);
-        } else if (param().format == param::ConvBias::Format::NCHW32 ||
-                   param().format == param::ConvBias::Format::NCHW4_NCHW32) {
+        } else if (opr->param().format == param::ConvBias::Format::NCHW32 ||
+                   opr->param().format ==
+                           param::ConvBias::Format::NCHW4_NCHW32) {
             megdnn_assert(bias.shape[0] == 1);
             megdnn_assert(bias.shape[1] == dst.shape[1], "bias:%s, dst:%s",
                           bias.to_string().c_str(), dst.to_string().c_str());
             megdnn_assert(bias.shape[2] == 1);
             megdnn_assert(bias.shape[3] == 1);
             megdnn_assert(bias.shape[4] == 32);
-        } else if (param().format == param::ConvBias::Format::CHWN4) {
+        } else if (opr->param().format == param::ConvBias::Format::CHWN4) {
             megdnn_assert(bias.shape[0] == dst.shape[0], "bias:%s, dst:%s",
                           bias.to_string().c_str(), dst.to_string().c_str());
             megdnn_assert(bias.shape[1] == 1);
             megdnn_assert(bias.shape[2] == 1);
             megdnn_assert(bias.shape[3] == 1);
             megdnn_assert(bias.shape[4] == 4);
-        } else if (param().format == param::ConvBias::Format::NCHW64) {
+        } else if (opr->param().format == param::ConvBias::Format::NCHW64) {
             megdnn_assert(bias.shape[0] == 1);
             megdnn_assert(bias.shape[1] == dst.shape[1], "bias:%s, dst:%s",
                           bias.to_string().c_str(), dst.to_string().c_str());
@@ -124,7 +114,8 @@ ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
             megdnn_assert(bias.shape[3] == 1);
             megdnn_assert(bias.shape[4] == 64);
         } else {
-            megdnn_assert(param().format == param::ConvBias::Format::NHWCD4);
+            megdnn_assert(opr->param().format ==
+                          param::ConvBias::Format::NHWCD4);
             megdnn_assert(bias.shape[0] == 1);
             megdnn_assert(bias.shape[1] == 1);
             megdnn_assert(bias.shape[2] == dst.shape[2], "bias:%s, dst:%s",
@@ -135,11 +126,53 @@ ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
     }
 
     if (z.ndim != 0) {
-        megdnn_assert(param().format != param::ConvBias::Format::NCHW4_NCHW32);
-        megdnn_assert(param().format != param::ConvBias::Format::NCHW32_NCHW4);
+        megdnn_assert(opr->param().format !=
+                      param::ConvBias::Format::NCHW4_NCHW32);
+        megdnn_assert(opr->param().format !=
+                      param::ConvBias::Format::NCHW32_NCHW4);
         megdnn_assert(z.dtype.enumv() == dst.dtype.enumv());
         megdnn_assert(z.eq_shape(dst));
     }
+}
+
+}  // namespace
+
+void ConvBiasForward::deduce_dtype(DType src, DType filter, DType /* bias */,
+                                   DType /* z */, DType& dst) {
+    check_or_deduce_dtype_fwd(src, filter, dst);
+}
+
+void ConvBiasForward::deduce_layout(const TensorLayout& src,
+                                    const TensorLayout& filter,
+                                    const TensorLayout& /* bias */,
+                                    const TensorLayout& /* z */,
+                                    TensorLayout& dst) {
+    deduce_layout_fwd(src, filter, dst);
+}
+
+ConvBiasForward::CanonizedFilterMeta ConvBiasForward::check_exec(
+        const TensorLayout& src, const TensorLayout& filter,
+        const TensorLayout& bias, const TensorLayout& z,
+        const TensorLayout& dst, size_t workspace_in_bytes,
+        const PreprocessedFilter* preprocessed_filter) {
+    do_check_exec_common(this, src, filter, bias, z, dst, workspace_in_bytes,
+                         preprocessed_filter);
+    auto ret = check_layout_fwd(src, filter, dst);
+    return ret;
+}
+
+ConvBiasForward::CanonizedFilterMeta
+ConvBiasForward::check_exec_allow_noncontiguous(
+        const TensorLayout& src, const TensorLayout& filter,
+        const TensorLayout& bias, const TensorLayout& z,
+        const TensorLayout& dst, size_t workspace_in_bytes,
+        const PreprocessedFilter* preprocessed_filter) {
+    do_check_exec_common(this, src, filter, bias, z, dst, workspace_in_bytes,
+                         preprocessed_filter);
+    TensorLayout dst_expected;
+    dst_expected.dtype = dst.dtype;
+    auto ret = deduce_layout_fwd(src, filter, dst_expected);
+    megdnn_assert_eq_shape(dst_expected, dst);
     return ret;
 }
 

dnn/src/common/conv_bias.h

@@ -12,6 +12,7 @@
 #include "megdnn/handle.h"
 #include "megdnn/opr_param_defs.h"
 #include "megdnn/oprs/general.h"
+#include "megdnn/oprs/nn.h"
 #include "megdnn/oprs/nn_int.h"
 #include "src/common/utils.h"
 

dnn/src/common/convolution.cpp

@@ -595,8 +595,6 @@ ConvolutionBase<Parameter>::deduce_layout_fwd(const TensorLayout& src,
                                               TensorLayout& dst) const {
     auto errmsg = [&]() { return get_errmsg(src, filter, dst, param()); };
     MEGDNN_MARK_USED_VAR(errmsg);
-    megdnn_assert_contiguous(src);
-    megdnn_assert_contiguous(filter);
     megdnn_assert(src.ndim >= 3_z, "%s", errmsg().c_str());
     megdnn_assert(((src.dtype.enumv() == filter.dtype.enumv()) ||
                    (src.dtype.enumv() == DTypeEnum::Quantized4Asymm &&
@@ -976,6 +974,8 @@ ConvolutionBase<param::Convolution>::CanonizedFilterMeta
 ConvolutionBase<param::Convolution>::check_layout_fwd(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst) const {
+    megdnn_assert_contiguous(src);
+    megdnn_assert_contiguous(filter);
     TensorLayout dst_expected;
     dst_expected.dtype = dst.dtype;
 
@@ -989,6 +989,8 @@ ConvolutionBase<param::ConvBias>::CanonizedFilterMeta
 ConvolutionBase<param::ConvBias>::check_layout_fwd(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst) const {
+    megdnn_assert_contiguous(src);
+    megdnn_assert_contiguous(filter);
     TensorLayout dst_expected;
     dst_expected.dtype = dst.dtype;
 
@@ -1002,6 +1004,8 @@ ConvolutionBase<param::BatchConvBias>::CanonizedFilterMeta
 ConvolutionBase<param::BatchConvBias>::check_layout_fwd(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst) const {
+    megdnn_assert_contiguous(src);
+    megdnn_assert_contiguous(filter);
     TensorLayout dst_expected;
     dst_expected.dtype = dst.dtype;
 

dnn/src/cuda/conv_bias/algo.cpp

@@ -116,8 +116,9 @@ ConvBiasForwardImpl::AlgoBase::SizeArgs::SizeArgs(
         const TensorLayout& filter, const TensorLayout& bias,
         const TensorLayout& z, const TensorLayout& dst,
         const PreprocessedFilter* preprocessed_filter)
-        : SizeArgs(o, src, filter, o->check_layout_fwd(src, filter, dst), bias,
-                   z, dst, preprocessed_filter) {}
+        : SizeArgs(o, src, filter,
+                   o->make_canonized_filter_meta(src.ndim, filter), bias, z,
+                   dst, preprocessed_filter) {}
 
 ConvBiasForwardImpl::AlgoBase::SizeArgs::SizeArgs(
         ConvBiasForwardImpl* o, const TensorLayout& src,

dnn/src/cuda/conv_bias/batched_matmul.cpp

@@ -75,8 +75,8 @@ ConvBiasForwardImpl::AlgoBatchedMatmul::get_subopr_list(
         const TensorLayoutArray& layouts, const OperatorBase* opr) const {
     const ConvBiasForwardImpl* conv_bias_opr =
             static_cast<const ConvBiasForwardImpl*>(opr);
-    CanonizedFilterMeta fm =
-            conv_bias_opr->check_layout_fwd(layouts[0], layouts[1], layouts[4]);
+    CanonizedFilterMeta fm = conv_bias_opr->make_canonized_filter_meta(
+            layouts[0].ndim, layouts[1]);
     auto&& config = sub_opr_config(fm, layouts[0], layouts[1], layouts[4],
                                    conv_bias_opr);
 

dnn/src/cuda/conv_bias/chanwise.cpp

@@ -20,6 +20,10 @@ using namespace conv_bias;
 
 bool ConvBiasForwardImpl::AlgoChanwise::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.src_layout->dtype == args.filter_layout->dtype &&
         args.src_layout->dtype == dtype::BFloat16()) {
         return false;

dnn/src/cuda/conv_bias/chanwise_8x8x32.cpp

@@ -21,6 +21,10 @@ using namespace conv_bias;
 
 bool ConvBiasForwardImpl::AlgoChanwise8x8x32::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.z_layout->ndim > 0)
         return false;
     using NonlineMode = param::ConvBias::NonlineMode;

dnn/src/cuda/conv_bias/chanwise_small.cpp

@@ -30,6 +30,10 @@ inline bool is_available_small(const chanwise::Param& param) {
 
 bool ConvBiasForwardImpl::AlgoChanwiseSmall::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.src_layout->dtype == args.filter_layout->dtype &&
         args.src_layout->dtype == dtype::BFloat16()) {
         return false;

dnn/src/cuda/conv_bias/conv_nchwqs8.cpp

@@ -63,6 +63,10 @@ void ConvBiasForwardImpl::AlgoFallbackNCHWQS8::make_inner_layout(
 
 bool ConvBiasForwardImpl::AlgoFallbackNCHWQS8::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     auto&& param = args.opr->param();
     bool is_format_ok = param.format == param::ConvBias::Format::NCHW;
     bool is_version_ok = CUDNN_VERSION >= 7500;

dnn/src/cuda/conv_bias/cudnn_conv_bias_activation.cpp

@@ -24,6 +24,10 @@ using namespace conv_bias;
 
 bool ConvBiasForwardImpl::AlgoCUDNNConvBiasActivation::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if ((args.src_layout->dtype.enumv() == DTypeEnum::QuantizedS4 ||
          args.src_layout->dtype.enumv() == DTypeEnum::Quantized4Asymm) &&
         args.filter_layout->dtype.enumv() == DTypeEnum::QuantizedS4)

dnn/src/cuda/conv_bias/implicit_gemm_int8_chwn4_dp4a.cpp

@@ -74,6 +74,10 @@ void dispatch_kernel(const int8_t* d_src, const int8_t* d_filter,
 
 bool ConvBiasForwardImpl::AlgoInt8CHWN4DotProdImplicitGemm::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.bias_layout->ndim <= 0)
         return false;
 

dnn/src/cuda/conv_bias/implicit_gemm_int8_chwn4_imma.cpp

@@ -62,6 +62,10 @@ void dispatch_kernel(const int8_t* d_src, const int8_t* d_filter,
 
 bool ConvBiasForwardImpl::AlgoInt8CHWN4IMMAImplicitGemm::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.bias_layout->ndim <= 0)
         return false;
 

dnn/src/cuda/conv_bias/implicit_gemm_int8_chwn4_imma_reorder_filter.cpp

@@ -109,6 +109,10 @@ INST(PerChannelBiasVisitor);
 
 bool ConvBiasForwardImpl::AlgoInt8CHWN4IMMAImplicitGemmReorderFilter::
         is_available(const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.bias_layout->ndim <= 0)
         return false;
 

dnn/src/cuda/conv_bias/implicit_gemm_int8_chwn4_imma_unroll_width.cpp

@@ -109,6 +109,10 @@ INST(PerChannelBiasVisitor);
 
 bool ConvBiasForwardImpl::AlgoInt8CHWN4IMMAImplicitGemmUnrollWidth::
         is_available(const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.bias_layout->ndim <= 0)
         return false;
 

dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw32_imma.cpp

@@ -23,6 +23,10 @@ using namespace convolution;
 #if CUDA_VERSION >= 10020
 bool ConvBiasForwardImpl::AlgoInt8NCHW32IMMAImplicitGemm::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.bias_layout->ndim <= 0)
         return false;
 

dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw4_dp4a.cpp

@@ -20,6 +20,10 @@ using namespace cuda;
 
 bool ConvBiasForwardImpl::AlgoInt8NCHW4DotProdImplicitGemm::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.bias_layout->ndim <= 0)
         return false;
 

dnn/src/cuda/conv_bias/implicit_gemm_int8_nchw4_imma.cpp

@@ -20,6 +20,10 @@ using namespace cuda;
 #if CUDA_VERSION >= 10000
 bool ConvBiasForwardImpl::AlgoInt8NCHW4IMMAImplicitGemm::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.bias_layout->ndim <= 0)
         return false;
 

dnn/src/cuda/conv_bias/matmul.cpp

@@ -61,8 +61,8 @@ ConvBiasForwardImpl::AlgoMatmul::get_subopr_list(
         const TensorLayoutArray& layouts, const OperatorBase* opr) const {
     const ConvBiasForwardImpl* conv_bias_opr =
             static_cast<const ConvBiasForwardImpl*>(opr);
-    CanonizedFilterMeta fm =
-            conv_bias_opr->check_layout_fwd(layouts[0], layouts[1], layouts[4]);
+    CanonizedFilterMeta fm = conv_bias_opr->make_canonized_filter_meta(
+            layouts[0].ndim, layouts[1]);
     auto&& config = sub_opr_config(fm, layouts[0], layouts[1], layouts[4],
                                    conv_bias_opr);
 

dnn/src/cuda/conv_bias/opr_impl.cpp

@@ -16,6 +16,7 @@
 #include "src/cuda/handle.h"
 #include "src/cuda/utils.h"
 
+#include "src/common/conv_bias.h"
 #include "src/common/algo_chooser.h"
 
 #include "src/cuda/cudnn_with_check.h"
@@ -28,8 +29,9 @@ void ConvBiasForwardImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                                _megdnn_tensor_out dst,
                                const PreprocessedFilter* preprocessed_filter,
                                _megdnn_workspace workspace) {
-    check_exec(src.layout, filter.layout, bias.layout, z.layout, dst.layout,
-               workspace.size, preprocessed_filter);
+    check_exec_allow_noncontiguous(src.layout, filter.layout, bias.layout,
+                                   z.layout, dst.layout, workspace.size,
+                                   preprocessed_filter);
     AlgoBase::ExecArgs args(this, src, filter, bias, z, dst, workspace,
                             preprocessed_filter);
     auto algo = get_algorithm(this, src.layout, filter.layout, bias.layout,

dnn/src/cuda/conv_bias/opr_impl.h

@@ -87,6 +87,7 @@ public:
             const AlgoAttribute& negative_attr) override;
 
 private:
+
     static AlgoPack sm_algo_pack;
 };
 

dnn/src/cuda/conv_bias/quint4x4x32_wmma.cpp

@@ -25,6 +25,10 @@ using namespace activation_u4;
 #if CUDA_VERSION >= 10000
 bool ConvBiasForwardImpl::AlgoQUInt4x4x32WMMA::is_available(
         const SizeArgs& args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     if (args.z_layout->ndim > 0)
         return false;
 

dnn/src/naive/conv_bias/opr_impl.cpp

@@ -233,9 +233,9 @@ void ConvBiasForwardImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
         dt_byte* workspace_ptr = workspace.raw_ptr;
         // ============================w * f + b================================
 
-        auto filter_meta =
-                check_exec(src.layout, filter.layout, bias.layout, z.layout,
-                           dst.layout, workspace.size, preprocessed_filter);
+        auto filter_meta = check_exec_allow_noncontiguous(
+                src.layout, filter.layout, bias.layout, z.layout, dst.layout,
+                workspace.size, preprocessed_filter);
         auto sfb = dst;
         if (bias.layout.dtype.enumv() != dst.layout.dtype.enumv()) {
             // intermediate result

dnn/test/cuda/conv_bias.cpp

@@ -749,6 +749,18 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_CUDNN_CONVOLUTION) {
                 .set_param(arg.param)
                 .execs({arg.src, arg.filter, arg.bias, {}, {}});
     }
+    //! noncontiguous case
+    {
+        param::ConvBias param;
+        param.pad_h = param.pad_w = 1;
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::Float32()},
+                {{16, 16, 3, 3}, {144, 9, 3, 1}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::Float32()},
+        });
+    }
 }
 
 TEST_F(CUDA, CONV_BIAS_FORWARD_INPLACE_MATMUL) {
@@ -791,6 +803,18 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_INPLACE_MATMUL) {
                 .execs({{2, 3, 3, 16}, {5, 3, 3, 3}, {1, 5, 1, 1}, {}, {}})
                 .execs({{2, 2, 8, 3}, {3, 2, 3, 3}, {1, 3, 1, 1}, {}, {}});
     }
+    //! noncontiguous case
+    {
+        param::ConvBias param;
+        param.pad_h = param.pad_w = 1;
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::Float32()},
+                {{16, 16, 3, 3}, {144, 9, 3, 1}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::Float32()},
+        });
+    }
 }
 
 TEST_F(CUDA, CONV_BIAS_FORWARD_MATMUL) {
@@ -835,6 +859,18 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_MATMUL) {
                 .execs({{2, 3, 3, 16}, {5, 3, 3, 3}, {1, 5, 1, 1}, {}, {}})
                 .execs({{2, 2, 8, 3}, {3, 2, 3, 3}, {1, 3, 1, 1}, {}, {}});
     }
+    //! noncontiguous case
+    {
+        param::ConvBias param;
+        param.pad_h = param.pad_w = 1;
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::Float32()},
+                {{16, 16, 3, 3}, {144, 9, 3, 1}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::Float32()},
+        });
+    }
 }
 
 TEST_F(CUDA, CONV_BIAS_FORWARD_MATMUL_8x8x32) {
@@ -880,6 +916,21 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_MATMUL_8x8x32) {
                 .execs({{2, 3, 16, 3}, {5, 3, 3, 3}, {1, 1, 1, 5}, {}, {}})
                 .execs({{2, 8, 3, 2}, {3, 3, 3, 2}, {1, 1, 1, 3}, {}, {}});
     }
+    //! noncontiguous case
+    {
+        param::ConvBias param;
+        param.pad_h = param.pad_w = 1;
+        param.format = param::ConvBias::Format::NHWC;
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{2, 7, 7, 16}, {1568, 224, 32, 1}, dtype::QuantizedS8{1.2f}},
+                {{16, 3, 3, 16}, {144, 48, 16, 1}, dtype::QuantizedS8{1.3f}},
+                {{}, {}, dtype::QuantizedS32{1.2f * 1.3f}},
+                {{}, {}, dtype::QuantizedS8{1.1f}},
+                {{2, 7, 7, 16},
+                 {1568, 224, 32, 1},
+                 dtype::QuantizedS32{1.2f * 1.3f}},
+        });
+    }
 }
 
 TEST_F(CUDA, CONV_BIAS_FORWARD_MATMUL_NCHW4) {
@@ -913,6 +964,21 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_MATMUL_NCHW4) {
     checker.exec({{1, 4, 2, 2, 4}, {16, 4, 3, 3, 4}, {1, 4, 1, 1, 4}, {}, {}});
     checker.exec(
             {{8, 64, 12, 12, 4}, {256, 64, 3, 3, 4}, {1, 64, 1, 1, 4}, {}, {}});
+    //! noncontiguous case
+    {
+        param::ConvBias param;
+        param.pad_h = param.pad_w = 1;
+        param.format = ConvBias::Param::Format::NCHW4;
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{2, 4, 7, 7, 4}, {1568, 196, 28, 4, 1}, dtype::QuantizedS8{1.2f}},
+                {{16, 4, 3, 3, 4}, {144, 36, 12, 4, 1}, dtype::QuantizedS8{1.3f}},
+                {{}, {}, dtype::QuantizedS32{1.2f * 1.3f}},
+                {{}, {}, dtype::QuantizedS8{1.1f}},
+                {{2, 4, 7, 7, 4},
+                 {1568, 196, 28, 4, 1},
+                 dtype::QuantizedS32{1.2f * 1.3f}},
+        });
+    }
 }
 
 TEST_F(CUDA, CONV_BIAS_FORWARD_BATCHED_MATMUL) {
@@ -939,6 +1005,17 @@ TEST_F(CUDA, CONV_BIAS_FORWARD_BATCHED_MATMUL) {
         checker.set_param(arg.param);
         checker.execs({arg.src, arg.filter, arg.bias, {}, {}});
     }
+    //! noncontiguous case
+    {
+        param::ConvBias param;
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::Float32()},
+                {{16, 16, 1, 1}, {16, 1, 1, 1}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{}, {}, dtype::Float32()},
+                {{2, 16, 7, 7}, {784, 49, 7, 1}, dtype::Float32()},
+        });
+    }
 }
 
 TEST_F(CUDA, CONV_BIAS_FORWARD_GROUP) {