refactor(mgb): make convolution3D handle noncontiguous tensors

GitOrigin-RevId: 3d3c31b02161532637948ba9aec42d161ec05e92

dnn/src/common/convolution3d.cpp

@@ -122,8 +122,6 @@ Convolution3DBase::CanonizedFilterMeta Convolution3DBase::deduce_layout_fwd(
         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 >= 5_z, "%s", errmsg().c_str());
     megdnn_assert(src.dtype == filter.dtype, "%s", errmsg().c_str());
     if (param().data_type == Param::DataType::FLOAT) {
@@ -170,6 +168,8 @@ Convolution3DBase::CanonizedFilterMeta Convolution3DBase::deduce_layout_fwd(
 Convolution3DBase::CanonizedFilterMeta Convolution3DBase::check_layout_fwd(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst) const {
+    megdnn_assert_contiguous(src);
+    megdnn_assert_contiguous(filter);
     TensorLayout dst_expected;
     auto ret = deduce_layout_fwd(src, filter, dst_expected);
     megdnn_assert_eq_layout(dst_expected, dst);
@@ -185,7 +185,12 @@ void Convolution3DForward::deduce_layout(const TensorLayout& src,
 Convolution3DBase::CanonizedFilterMeta Convolution3DForward::check_exec(
         const TensorLayout& src, const TensorLayout& filter,
         const TensorLayout& dst, size_t workspace_in_bytes) {
-    auto ret = check_layout_fwd(src, filter, dst);
+    auto src_fwd = src;
+    auto dst_fwd = dst;
+    src_fwd.init_contiguous_stride();
+    dst_fwd.init_contiguous_stride();
+
+    auto ret = check_layout_fwd(src_fwd, filter, dst_fwd);
     auto required_workspace_in_bytes = get_workspace_in_bytes(src, filter, dst);
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
     return ret;
@@ -196,7 +201,12 @@ Convolution3DBase::CanonizedFilterMeta Convolution3DBackwardData::check_exec(
         const TensorLayout& grad, size_t workspace_in_bytes) {
     megdnn_assert(param().data_type == Param::DataType::FLOAT,
                   "only float type is supported for conv backward");
-    auto ret = check_layout_fwd(grad, filter, diff);
+    auto diff_fwd = diff;
+    auto grad_fwd = grad;
+    diff_fwd.init_contiguous_stride();
+    grad_fwd.init_contiguous_stride();
+
+    auto ret = check_layout_fwd(grad_fwd, filter, diff_fwd);
     auto required_workspace_in_bytes =
             get_workspace_in_bytes(filter, diff, grad);
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
@@ -244,7 +254,12 @@ Convolution3DBase::CanonizedFilterMeta Convolution3DBackwardFilter::check_exec(
         const TensorLayout& grad, size_t workspace_in_bytes) {
     megdnn_assert(param().data_type == Param::DataType::FLOAT,
                   "only float type is supported for conv backward");
-    auto ret = check_layout_fwd(src, grad, diff);
+    auto src_fwd = src;
+    auto diff_fwd = diff;
+    src_fwd.init_contiguous_stride();
+    diff_fwd.init_contiguous_stride();
+
+    auto ret = check_layout_fwd(src_fwd, grad, diff_fwd);
     auto required_workspace_in_bytes = get_workspace_in_bytes(src, diff, grad);
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
     return ret;

dnn/src/cuda/convolution/backward_data/group_conv.cpp

@@ -44,6 +44,8 @@ bool ConvolutionBackwardDataImpl::AlgoGroupConvGeneral::is_available(
          args.diff_layout->dtype == dtype::QuantizedS8())) {
         return false;
     }
+    if (args.filter_meta.group <= 1)
+        return false;
     auto sub_args = args;
     TensorLayout diff_pg, grad_pg;
     modify_size_args(sub_args, diff_pg, grad_pg);

dnn/src/cuda/convolution/backward_filter/chanwise.cpp

@@ -19,7 +19,7 @@ using namespace convolution;
 
 bool ConvolutionBackwardFilterImpl::AlgoChanwise::is_available(
         const SizeArgs &args) const {
-    if (!args.grad_layout->is_contiguous() ||
+    if (!args.src_layout->is_contiguous() ||
         !args.diff_layout->is_contiguous()) {
         return false;
     }

dnn/src/cuda/convolution/backward_filter/group_conv.cpp

@@ -42,6 +42,8 @@ bool ConvolutionBackwardFilterImpl::AlgoGroupConvGeneral::is_available(
         args.diff_layout->dtype == dtype::BFloat16()) {
         return false;
     }
+    if (args.grad_filter_meta.group <= 1)
+        return false;
     auto sub_args = args;
     TensorLayout src_pg, diff_pg;
     modify_size_args(sub_args, src_pg, diff_pg);

dnn/src/cuda/convolution3d/backward_data/algo.cpp

@@ -64,7 +64,7 @@ Convolution3DBackwardDataImpl::AlgoBase::SizeArgs::SizeArgs(
         Convolution3DBackwardDataImpl *o,
         const TensorLayout &filter, const TensorLayout &diff,
         const TensorLayout &grad):
-    SizeArgs(o, o->check_layout_fwd(grad, filter, diff), diff, grad)
+    SizeArgs(o, o->make_canonized_filter_meta(grad.ndim, filter), diff, grad)
 {
 }
 

dnn/src/cuda/convolution3d/backward_data/chanwise.cpp

@@ -19,6 +19,10 @@ using namespace convolution3d;
 
 bool Convolution3DBackwardDataImpl::AlgoChanwise::is_available(
         const SizeArgs &args) const {
+    if (!args.grad_layout->is_contiguous() ||
+        !args.diff_layout->is_contiguous()) {
+        return false;
+    }
     auto &&fm = args.filter_meta;
     return args.filter_meta.format == Param::Format::NCDHW &&
         args.diff_layout->dtype.category() == DTypeCategory::FLOAT &&

dnn/src/cuda/convolution3d/backward_data/group_conv.cpp

@@ -38,6 +38,8 @@ Convolution3DBackwardDataImpl::AlgoGroupConvGeneral::AlgoGroupConvGeneral(
 
 bool Convolution3DBackwardDataImpl::AlgoGroupConvGeneral::is_available(
         const SizeArgs &args) const {
+    if (args.filter_meta.group <= 1)
+        return false;
     auto sub_args = args;
     TensorLayout diff_pg, grad_pg;
     modify_size_args(sub_args, diff_pg, grad_pg);

dnn/src/cuda/convolution3d/backward_filter/algo.cpp

@@ -67,7 +67,7 @@ Convolution3DBackwardFilterImpl::AlgoBase::SizeArgs::SizeArgs(
         Convolution3DBackwardFilterImpl *o,
         const TensorLayout &src, const TensorLayout &diff,
         const TensorLayout &grad):
-    SizeArgs(o, src, diff, o->check_layout_fwd(src, grad, diff))
+    SizeArgs(o, src, diff, o->make_canonized_filter_meta(src.ndim, grad))
 {
 }
 

dnn/src/cuda/convolution3d/backward_filter/chanwise.cpp

@@ -19,6 +19,10 @@ using namespace convolution3d;
 
 bool Convolution3DBackwardFilterImpl::AlgoChanwise::is_available(
         const SizeArgs &args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.diff_layout->is_contiguous()) {
+        return false;
+    }
     auto &&fm = args.grad_filter_meta;
     return fm.format == Param::Format::NCDHW &&
         args.diff_layout->dtype.category() == DTypeCategory::FLOAT &&

dnn/src/cuda/convolution3d/backward_filter/group_conv.cpp

@@ -38,6 +38,8 @@ Convolution3DBackwardFilterImpl::AlgoGroupConvGeneral::AlgoGroupConvGeneral(
 
 bool Convolution3DBackwardFilterImpl::AlgoGroupConvGeneral::is_available(
         const SizeArgs &args) const {
+    if (args.grad_filter_meta.group <= 1)
+        return false;
     auto sub_args = args;
     TensorLayout src_pg, diff_pg;
     modify_size_args(sub_args, src_pg, diff_pg);

dnn/src/cuda/convolution3d/backward_filter/inplace_matmul.cpp

@@ -17,6 +17,10 @@ using namespace cuda;
 
 bool Convolution3DBackwardFilterImpl::AlgoInplaceMatmul::is_available(
         const SizeArgs &args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.diff_layout->is_contiguous()) {
+        return false;
+    }
     auto &&fm = args.grad_filter_meta;
     return args.grad_filter_meta.format == Param::Format::NCDHW &&
         args.src_layout->dtype == dtype::Float32() &&

dnn/src/cuda/convolution3d/forward/algo.cpp

@@ -69,7 +69,7 @@ Convolution3DForwardImpl::AlgoBase::SizeArgs::SizeArgs(
         Convolution3DForwardImpl *o,
         const TensorLayout &src, const TensorLayout &filter,
         const TensorLayout &dst):
-    SizeArgs(o, src, o->check_layout_fwd(src, filter, dst), dst)
+    SizeArgs(o, src, o->make_canonized_filter_meta(src.ndim, filter), dst)
 {
 }
 

dnn/src/cuda/convolution3d/forward/chanwise.cpp

@@ -19,6 +19,10 @@ using namespace convolution3d;
 
 bool Convolution3DForwardImpl::AlgoChanwise::is_available(
         const SizeArgs &args) const {
+    if (!args.src_layout->is_contiguous() ||
+        !args.dst_layout->is_contiguous()) {
+        return false;
+    }
     auto &&fm = args.filter_meta;
     return args.filter_meta.format == Param::Format::NCDHW &&
         args.src_layout->dtype.category() == DTypeCategory::FLOAT &&

dnn/src/cuda/convolution3d/forward/group_conv.cpp

@@ -45,6 +45,8 @@ Convolution3DForwardImpl::AlgoGroupConvGeneral::AlgoGroupConvGeneral(
 
 bool Convolution3DForwardImpl::AlgoGroupConvGeneral::is_available(
         const SizeArgs &args) const {
+    if (args.filter_meta.group <= 1)
+        return false;
     auto sub_args = args;
     TensorLayout src_pg, dst_pg;
     modify_size_args(sub_args, src_pg, dst_pg);

dnn/src/naive/convolution3d/helper.h

@@ -215,7 +215,6 @@ void backward_data(_megdnn_tensor_in filter,
         _megdnn_tensor_in diff,
         _megdnn_tensor_out grad,
         const Convolution3D::CanonizedFilterMeta &filter_meta) {
-    megdnn_assert(grad.layout.is_contiguous());
     memset(grad.raw_ptr, 0, grad.layout.span().dist_byte());
     megdnn_assert(filter_meta.spatial_ndim == 3);
     compute3d<gtype, ftype, dtype, StrategyBwdData>(
@@ -227,7 +226,6 @@ void backward_filter(_megdnn_tensor_in src,
         _megdnn_tensor_in diff,
         _megdnn_tensor_out grad,
         const Convolution3D::CanonizedFilterMeta &filter_meta) {
-    megdnn_assert(grad.layout.is_contiguous());
     memset(grad.raw_ptr, 0, grad.layout.span().dist_byte());
     megdnn_assert(filter_meta.spatial_ndim == 3);
     compute3d<stype, gtype, dtype, StrategyBwdFlt>(

dnn/test/cuda/convolution.cpp

@@ -384,16 +384,6 @@ TEST_F(CUDA, CONVOLUTION_BACKWARD_DATA_INT8_NCHW_DP4A) {
         }
         checker.set_rng(0, &rng).set_rng(1, &rng).set_param(arg.param).exec(
                 TensorLayoutArray{filter, dst, src});
-        //! noncontiguous case
-        {
-            param::Convolution param;
-            param.pad_h = param.pad_w = 1;
-            checker.set_param(param).execl(TensorLayoutArray{
-                    {{16, 16, 3, 3}, {144, 9, 3, 1}, dtype::QuantizedS8{1.3f}},
-                    {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::QuantizedS8{1.2f}},
-                    {{2, 16, 7, 7}, {1568, 49, 7, 1}, dtype::QuantizedS8{1.2f}}
-            });
-        }
     }
 }
 

dnn/test/cuda/convolution3d.cpp

@@ -150,6 +150,77 @@ TEST_F(CUDA, CONVOLUTION3D_MATMUL_FORWARD) {
     }
 }
 
+TEST_F(CUDA, CONVOLUTION3D_FORWARD_NONCONTIG_CUDNN) {
+    using namespace convolution3d;
+    Checker<Convolution3DForward> checker(handle_cuda());
+    checker.set_before_exec_callback(AlgoChecker<Convolution3DForward>(
+            "CUDNN"));
+    param::Convolution3D param;
+    param.pad_d = param.pad_h = param.pad_w = 1;
+    checker.set_dtype(0, dtype::Float32())
+                .set_dtype(1, dtype::Float32())
+                .set_epsilon(1e-3);
+
+    //! noncontiguous case
+    {
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()},
+                {{5, 5, 3, 3, 3}, {135, 27, 9, 3, 1}, dtype::Float32()},
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()}});
+    }
+}
+
+TEST_F(CUDA, CONVOLUTION3D_FORWARD_NONCONTIG_INPLACE_MATMUL) {
+    using namespace convolution3d;
+    Checker<Convolution3DForward> checker(handle_cuda());
+    checker.set_before_exec_callback(AlgoChecker<Convolution3DForward>(
+            "INPLACE_MATMUL"));
+    param::Convolution3D param;
+    param.pad_d = param.pad_h = param.pad_w = 1;
+    checker.set_dtype(0, dtype::Float32())
+                .set_dtype(1, dtype::Float32())
+                .set_epsilon(1e-3);
+
+    //! noncontiguous case
+    {
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()},
+                {{5, 5, 3, 3, 3}, {135, 27, 9, 3, 1}, dtype::Float32()},
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()}});
+    }
+}
+
+TEST_F(CUDA, CONVOLUTION3D_FORWARD_NONCONTIG_1x1x1) {
+    using namespace convolution3d;
+    Checker<Convolution3DForward> checker(handle_cuda());
+    checker.set_before_exec_callback(AlgoChecker<Convolution3DForward>(
+            "1x1x1"));
+    param::Convolution3D param;
+    checker.set_dtype(0, dtype::Float32())
+                .set_dtype(1, dtype::Float32())
+                .set_epsilon(1e-3);
+
+    //! noncontiguous case
+    {
+        checker.set_param(param).execl(TensorLayoutArray{
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()},
+                {{5, 5, 1, 1, 1}, {5, 1, 1, 1, 1}, dtype::Float32()},
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()}});
+    }
+}
+
 #if MEGDNN_WITH_BENCHMARK
 TEST_F(CUDA, BENCHMARK_CONVOLUTION3D_MATMUL_BACKWARD_FILTER) {
     using namespace convolution3d;
@@ -343,6 +414,60 @@ TEST_F(CUDA, CONVOLUTION3D_MATMUL_BACKWARD_FILTER) {
     }
 }
 
+TEST_F(CUDA, CONVOLUTION3D_BACKWARD_DATA_NONCONTIG_CUDNN) {
+    using namespace convolution3d;
+    Checker<Convolution3DBackwardData> checker(handle_cuda());
+    checker.set_before_exec_callback(AlgoChecker<Convolution3DBackwardData>(
+            "CUDNN"));
+    Convolution3DBackwardData::Param param;
+    param.pad_d = param.pad_h = param.pad_w = 1;
+    NormalRNG default_rng;
+    checker.set_dtype(0, dtype::Float32())
+            .set_dtype(1, dtype::Float32())
+            .set_rng(0, &default_rng)
+            .set_rng(1, &default_rng)
+            .set_epsilon(1e-3)
+            .set_param(param);
+    //! noncontiguous case
+    {
+        checker.execl(TensorLayoutArray{
+                {{5, 5, 3, 3, 3}, {135, 27, 9, 3, 1}, dtype::Float32()},
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()},
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()}});
+    }
+}
+
+TEST_F(CUDA, CONVOLUTION3D_BACKWARD_FILTER_NONCONTIG_CUDNN) {
+    using namespace convolution3d;
+    Checker<Convolution3DBackwardFilter> checker(handle_cuda());
+    checker.set_before_exec_callback(AlgoChecker<Convolution3DBackwardFilter>(
+            "CUDNN"));
+    Convolution3DBackwardFilter::Param param;
+    param.pad_d = param.pad_h = param.pad_w = 1;
+    NormalRNG default_rng;
+    checker.set_dtype(0, dtype::Float32())
+            .set_dtype(1, dtype::Float32())
+            .set_rng(0, &default_rng)
+            .set_rng(1, &default_rng)
+            .set_epsilon(1e-3)
+            .set_param(param);
+    //! noncontiguous case
+    {
+        checker.execl(TensorLayoutArray{
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()},
+                {{4, 5, 16, 16, 16},
+                 {40960, 4096, 256, 16, 1},
+                 dtype::Float32()},
+                {{5, 5, 3, 3, 3}, {135, 27, 9, 3, 1}, dtype::Float32()}});
+    }
+}
+
 /*
 TEST_F(CUDA, CONV_CONFIG_COMBINATIONS) {
     auto eps_getter = [](bool f16, int stage, const char *name) -> float {