feat(dnn/cuda): add impl for fusing warp perspective and dimshuffle

GitOrigin-RevId: 51e025973f

dnn/scripts/opr_param_defs.py

@@ -43,6 +43,12 @@ pdef('Axis').add_fields('int32', 'axis', 0)
           Doc('NCHW4_NCHW32', 'NCHW4_NCHW32 means input tensors are nchw4 layout, output tensor is nchw32 layout'), 
           Doc('NCHW32_NCHW4', 'NCHW32_NCHW4 means input tensors are nchw32 layout, output tensor is nchw4 layout'), 
           Doc('NCHW4_NCHW', 'NCHW4_NCHW means input tensors are nchw4 layout, output tensor is nchw layout'), 
+          Doc('NHWC_NCHW', 'NHWC_NCHW means input tensors are nhwc layout, '
+              'output tensor is nchw layout'),
+          Doc('NHWC_NCHW4_IC_SMALL', 'NHWC_NCHW4_IC_SMALL means input tensors are nhwc(c < 4) layout, '
+              'output tensor is nchw4 layout, padding c=4'),
+          Doc('NCHW_NCHW4_IC_SMALL', 'NCHW_NCHW4_IC_SMALL means input tensors are nchw(c < 4) layout, '
+              'output tensor is nchw4 layout, padding c=4'),
           Doc('CHWN4', 'CHWN4 is currently only used on Nvidia platform for fast implementation '
               'of convolution using CUDA/SASS. The channels are splitted to groups of 4 channels.'))
  )

dnn/src/common/warp_perspective.cpp

@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 #include "megdnn/oprs.h"
 
@@ -14,20 +15,17 @@
 
 namespace megdnn {
 
-void WarpPerspectiveBase::check_layout_fwd(const TensorLayout &src,
-        const TensorLayout &mat,
-        const TensorLayout &mat_idx,
-        const TensorLayout &dst)
-{
+void WarpPerspectiveBase::check_layout_fwd(const TensorLayout& src,
+                                           const TensorLayout& mat,
+                                           const TensorLayout& mat_idx,
+                                           const TensorLayout& dst) {
     megdnn_assert_contiguous(mat);
     megdnn_assert_contiguous(src);
     megdnn_assert_contiguous(dst);
     auto errmsg = [&]() {
-        return megdnn_layout_msg(src) + ", " +
-            megdnn_layout_msg(mat) + ", " +
-            megdnn_layout_msg(mat_idx) + ", " +
-            megdnn_layout_msg(dst) + ", " +
-            param_msg();
+        return megdnn_layout_msg(src) + ", " + megdnn_layout_msg(mat) + ", " +
+               megdnn_layout_msg(mat_idx) + ", " + megdnn_layout_msg(dst) +
+               ", " + param_msg();
     };
     MEGDNN_MARK_USED_VAR(errmsg);
     if (param().format == param::WarpPerspective::Format::NHWCD4 ||
@@ -35,9 +33,17 @@ void WarpPerspectiveBase::check_layout_fwd(const TensorLayout &src,
         megdnn_assert(src.ndim == 5_z, "%s", errmsg().c_str());
         megdnn_assert(dst.ndim == 5_z, "%s", errmsg().c_str());
 
+    } else if (param().format ==
+                       param::WarpPerspective::Format::NHWC_NCHW4_IC_SMALL ||
+               param().format ==
+                       param::WarpPerspective::Format::NCHW_NCHW4_IC_SMALL) {
+        megdnn_assert(src.ndim == 4_z, "%s", errmsg().c_str());
+        megdnn_assert(dst.ndim == 5_z, "%s", errmsg().c_str());
     } else {
         megdnn_assert(param().format == param::WarpPerspective::Format::NHWC ||
-                      param().format == param::WarpPerspective::Format::NCHW);
+                      param().format == param::WarpPerspective::Format::NCHW ||
+                      param().format ==
+                              param::WarpPerspective::Format::NHWC_NCHW);
         megdnn_assert(src.ndim == 4_z, "%s", errmsg().c_str());
         megdnn_assert(dst.ndim == 4_z, "%s", errmsg().c_str());
     }
@@ -45,7 +51,7 @@ void WarpPerspectiveBase::check_layout_fwd(const TensorLayout &src,
     megdnn_assert(dst.shape[0] == mat.shape[0], "%s", errmsg().c_str());
     if (mat_idx.ndim) {
         megdnn_assert(mat_idx.dtype == dtype::Int32() && mat_idx.ndim == 1,
-                "%s", errmsg().c_str());
+                      "%s", errmsg().c_str());
         megdnn_assert(mat.shape[0] == mat_idx.shape[0], "%s", errmsg().c_str());
         megdnn_assert_contiguous(mat_idx);
     } else {
@@ -54,35 +60,103 @@ void WarpPerspectiveBase::check_layout_fwd(const TensorLayout &src,
     megdnn_assert(mat.shape[1] == 3_z, "%s", errmsg().c_str());
     megdnn_assert(mat.shape[2] == 3_z, "%s", errmsg().c_str());
 
-    if (param().format == param::WarpPerspective::Format::NCHW) {
-        megdnn_assert(
-                src.dtype.enumv() == DTypeEnum::Float32 ||
-                        MEGDNN_FLOAT16_SELECT(
-                                (src.dtype.enumv() == DTypeEnum::Float16 ||
-                                 src.dtype.enumv() == DTypeEnum::BFloat16),
-                                false) ||
-                        src.dtype.enumv() == DTypeEnum::Int8 ||
-                        src.dtype.enumv() == DTypeEnum::Uint8 ||
-                        (src.dtype.enumv() == DTypeEnum::QuantizedS8 ||
-                         src.dtype.enumv() == DTypeEnum::Quantized8Asymm),
-                "WarpPerspective NCHW input dtype should be "
-                "Float32/Int8/Uint8/QInt8/QUint8" MEGDNN_FLOAT16_SELECT(
-                        "/Float16/BFloat16", "") ".");
-        megdnn_assert(
-                (src.dtype.category() == DTypeCategory::FLOAT &&
-                 (src.dtype == mat.dtype ||
-                  mat.dtype.enumv() == DTypeEnum::Float32)) ||
-                        ((src.dtype.category() == DTypeCategory::INT ||
-                          src.dtype.category() == DTypeCategory::QUANTIZED) &&
-                         mat.dtype.enumv() == DTypeEnum::Float32),
-                "The input to WarpPerspective is in NCHW format, in this "
-                "case, if the input dtype is floating point, the "
-                "transformation matrix should have same dtype as the "
-                "input, otherwise, it should be in Float32, %s given.",
-                mat.dtype.name());
+    if (src.format == dst.format && dst.dtype == src.dtype) {
+        if (param().format == param::WarpPerspective::Format::NCHW) {
+            megdnn_assert(
+                    src.dtype.enumv() == DTypeEnum::Float32 ||
+                            MEGDNN_FLOAT16_SELECT(
+                                    (src.dtype.enumv() == DTypeEnum::Float16 ||
+                                     src.dtype.enumv() == DTypeEnum::BFloat16),
+                                    false) ||
+                            src.dtype.enumv() == DTypeEnum::Int8 ||
+                            src.dtype.enumv() == DTypeEnum::Uint8 ||
+                            (src.dtype.enumv() == DTypeEnum::QuantizedS8 ||
+                             src.dtype.enumv() == DTypeEnum::Quantized8Asymm),
+                    "WarpPerspective NCHW input dtype should be "
+                    "Float32/Int8/Uint8/QInt8/QUint8" MEGDNN_FLOAT16_SELECT(
+                            "/Float16/BFloat16", "") ".");
+            megdnn_assert(
+                    (src.dtype.category() == DTypeCategory::FLOAT &&
+                     (src.dtype == mat.dtype ||
+                      mat.dtype.enumv() == DTypeEnum::Float32)) ||
+                            ((src.dtype.category() == DTypeCategory::INT ||
+                              src.dtype.category() ==
+                                      DTypeCategory::QUANTIZED) &&
+                             mat.dtype.enumv() == DTypeEnum::Float32),
+                    "The input to WarpPerspective is in NCHW format, in this "
+                    "case, if the input dtype is floating point, the "
+                    "transformation matrix should have same dtype as the "
+                    "input, otherwise, it should be in Float32, %s given.",
+                    mat.dtype.name());
+
+            megdnn_assert(src.shape[1] == dst.shape[1], "%s", errmsg().c_str());
 
-        megdnn_assert(dst.dtype == src.dtype);
-        megdnn_assert(src.shape[1] == dst.shape[1], "%s", errmsg().c_str());
+            megdnn_assert(param().imode ==
+                          param::WarpPerspective::InterpolationMode::LINEAR);
+            megdnn_assert(param().bmode !=
+                          param::WarpPerspective::BorderMode::TRANSPARENT);
+            megdnn_assert(param().bmode !=
+                          param::WarpPerspective::BorderMode::ISOLATED);
+        } else if (param().format == param::WarpPerspective::Format::NHWC) {
+            megdnn_assert(src.shape[3] == dst.shape[3], "%s", errmsg().c_str());
+        } else if (param().format == param::WarpPerspective::Format::NCHW4) {
+            megdnn_assert(src.dtype.enumv() == DTypeEnum::QuantizedS8,
+                          "src expected QuantizedS8, but got %s",
+                          src.dtype.name());
+            megdnn_assert(mat.dtype == dtype::Float32(),
+                          "matrix dtype expected float, got %s",
+                          mat.dtype.name());
+            megdnn_assert(src.shape[4] == 4 && dst.shape[4] == 4);
+            megdnn_assert(src.shape[1] == dst.shape[1], "%s", errmsg().c_str());
+
+            megdnn_assert(param().imode ==
+                          param::WarpPerspective::InterpolationMode::LINEAR);
+            megdnn_assert(param().bmode !=
+                          param::WarpPerspective::BorderMode::TRANSPARENT);
+            megdnn_assert(param().bmode !=
+                          param::WarpPerspective::BorderMode::ISOLATED);
+        } else {
+            megdnn_assert(param().format ==
+                          param::WarpPerspective::Format::NHWCD4);
+            megdnn_assert(
+                    src.dtype == dtype::Float32() ||
+                            MEGDNN_FLOAT16_SELECT(
+                                    (src.dtype == dtype::Float16() ||
+                                     src.dtype == dtype::BFloat16()),
+                                    false) ||
+                            src.dtype.enumv() == DTypeEnum::QuantizedS8 ||
+                            src.dtype.enumv() == DTypeEnum::Quantized8Asymm,
+                    "WarpPerspective NHWCD4 input dtype should be "
+                    "Float32" MEGDNN_FLOAT16_SELECT(
+                            "/Float16/BFloat16",
+                            "") ",QunatizedS8, Quantized8Asymm.");
+            megdnn_assert(
+                    (src.dtype == mat.dtype || mat.dtype == dtype::Float32()),
+                    "The input to WarpPerspective is in NHWCD4 format, in this "
+                    "case, if the input dtype is floating point, the "
+                    "transformation matrix should have same dtype as the "
+                    "input, %s given.",
+                    mat.dtype.name());
+            //! number of channels is same
+            megdnn_assert(src.shape[2] == dst.shape[2], "%s", errmsg().c_str());
+            megdnn_assert(param().imode ==
+                          param::WarpPerspective::InterpolationMode::LINEAR);
+            megdnn_assert(param().bmode !=
+                          param::WarpPerspective::BorderMode::TRANSPARENT);
+            megdnn_assert(param().bmode !=
+                          param::WarpPerspective::BorderMode::ISOLATED);
+        }
+    } else if (param().format ==
+                       param::WarpPerspective::Format::NHWC_NCHW4_IC_SMALL ||
+               param().format ==
+                       param::WarpPerspective::Format::NCHW_NCHW4_IC_SMALL) {
+        megdnn_assert((src.dtype.enumv() == DTypeEnum::Quantized8Asymm ||
+                       src.dtype.enumv() == DTypeEnum::Uint8),
+                      "src expected Quantized8Asymm or Uint8, but got %s",
+                      src.dtype.name());
+        megdnn_assert(mat.dtype == dtype::Float32(),
+                      "matrix dtype expected float, got %s", mat.dtype.name());
+        megdnn_assert(dst.shape[4] == 4);
 
         megdnn_assert(param().imode ==
                       param::WarpPerspective::InterpolationMode::LINEAR);
@@ -90,16 +164,14 @@ void WarpPerspectiveBase::check_layout_fwd(const TensorLayout &src,
                       param::WarpPerspective::BorderMode::TRANSPARENT);
         megdnn_assert(param().bmode !=
                       param::WarpPerspective::BorderMode::ISOLATED);
-    } else if (param().format == param::WarpPerspective::Format::NHWC) {
-        megdnn_assert(src.shape[3] == dst.shape[3], "%s", errmsg().c_str());
-    } else if (param().format == param::WarpPerspective::Format::NCHW4) {
-        megdnn_assert(dst.dtype == src.dtype);
-        megdnn_assert(src.dtype.enumv() == DTypeEnum::QuantizedS8,
-                "src expected QuantizedS8, but got %s", src.dtype.name());
+    } else if (param().format == param::WarpPerspective::Format::NHWC_NCHW) {
+        megdnn_assert((src.dtype.enumv() == DTypeEnum::Quantized8Asymm ||
+                       src.dtype.enumv() == DTypeEnum::Uint8),
+                      "src expected Quantized8Asymm or Uint8, but got %s",
+                      src.dtype.name());
         megdnn_assert(mat.dtype == dtype::Float32(),
                       "matrix dtype expected float, got %s", mat.dtype.name());
-        megdnn_assert(src.shape[4] == 4 && dst.shape[4] == 4);
-        megdnn_assert(src.shape[1] == dst.shape[1], "%s", errmsg().c_str());
+        megdnn_assert(src.shape[3] == dst.shape[1], "%s", errmsg().c_str());
 
         megdnn_assert(param().imode ==
                       param::WarpPerspective::InterpolationMode::LINEAR);
@@ -108,40 +180,14 @@ void WarpPerspectiveBase::check_layout_fwd(const TensorLayout &src,
         megdnn_assert(param().bmode !=
                       param::WarpPerspective::BorderMode::ISOLATED);
     } else {
-        megdnn_assert(param().format == param::WarpPerspective::Format::NHWCD4);
-        megdnn_assert(
-                src.dtype == dtype::Float32() ||
-                        MEGDNN_FLOAT16_SELECT((src.dtype == dtype::Float16() ||
-                                               src.dtype == dtype::BFloat16()),
-                                              false) ||
-                        src.dtype.enumv() == DTypeEnum::QuantizedS8 ||
-                        src.dtype.enumv() == DTypeEnum::Quantized8Asymm,
-                "WarpPerspective NHWCD4 input dtype should be "
-                "Float32" MEGDNN_FLOAT16_SELECT(
-                        "/Float16/BFloat16",
-                        "") ",QunatizedS8, Quantized8Asymm.");
-        megdnn_assert(
-                (src.dtype == mat.dtype || mat.dtype == dtype::Float32()),
-                "The input to WarpPerspective is in NHWCD4 format, in this "
-                "case, if the input dtype is floating point, the "
-                "transformation matrix should have same dtype as the "
-                "input, %s given.",
-                mat.dtype.name());
-        megdnn_assert(dst.dtype == src.dtype);
-        //! number of channels is same
-        megdnn_assert(src.shape[2] == dst.shape[2], "%s", errmsg().c_str());
-        megdnn_assert(param().imode ==
-                      param::WarpPerspective::InterpolationMode::LINEAR);
-        megdnn_assert(param().bmode !=
-                      param::WarpPerspective::BorderMode::TRANSPARENT);
-        megdnn_assert(param().bmode !=
-                      param::WarpPerspective::BorderMode::ISOLATED);
+        megdnn_assert(param().format == param::WarpPerspective::Format::NCHW);
+        megdnn_assert((src.dtype.enumv() == DTypeEnum::Quantized8Asymm ||
+                       src.dtype.enumv() == DTypeEnum::Uint8) &&
+                      dst.dtype.enumv() == DTypeEnum::Float32);
     }
-    megdnn_assert(src.format == dst.format);
 }
 
-std::string WarpPerspectiveBase::param_msg() const
-{
+std::string WarpPerspectiveBase::param_msg() const {
     std::string res;
     res.append(megdnn_mangle("imode="));
     switch (param().imode) {
@@ -191,31 +237,25 @@ std::string WarpPerspectiveBase::param_msg() const
     return res;
 }
 
-int WarpPerspectiveBase::get_real_coord(int p, int len)
-{
+int WarpPerspectiveBase::get_real_coord(int p, int len) {
     auto bmode = param().bmode;
-    if( (unsigned)p < (unsigned)len )
+    if ((unsigned)p < (unsigned)len)
         ;
-    else if( bmode == BorderMode::REPLICATE )
+    else if (bmode == BorderMode::REPLICATE)
         p = p < 0 ? 0 : len - 1;
-    else if( bmode == BorderMode::REFLECT || bmode == BorderMode::REFLECT_101 )
-    {
+    else if (bmode == BorderMode::REFLECT || bmode == BorderMode::REFLECT_101) {
         int delta = (bmode == BorderMode::REFLECT_101);
-        if( len == 1 )
+        if (len == 1)
             return 0;
-        do
-        {
-            if( p < 0 )
+        do {
+            if (p < 0)
                 p = -p - 1 + delta;
             else
                 p = len - 1 - (p - len) - delta;
-        }
-        while( (unsigned)p >= (unsigned)len );
-    }
-    else if( bmode == BorderMode::WRAP )
-    {
-        if( p < 0 )
-            p -= ((p-len+1)/len)*len;
+        } while ((unsigned)p >= (unsigned)len);
+    } else if (bmode == BorderMode::WRAP) {
+        if (p < 0)
+            p -= ((p - len + 1) / len) * len;
         /*
         if( p >= len )
             p %= len;
@@ -223,18 +263,16 @@ int WarpPerspectiveBase::get_real_coord(int p, int len)
         while (p >= len) {
             p -= len;
         }
-    }
-    else if( bmode == BorderMode::CONSTANT )
+    } else if (bmode == BorderMode::CONSTANT)
         p = -1;
     return p;
 }
 
-void WarpPerspectiveForward::check_exec(const TensorLayout &src,
-        const TensorLayout &mat,
-        const TensorLayout &mat_idx,
-        const TensorLayout &dst,
-        size_t workspace_in_bytes)
-{
+void WarpPerspectiveForward::check_exec(const TensorLayout& src,
+                                        const TensorLayout& mat,
+                                        const TensorLayout& mat_idx,
+                                        const TensorLayout& dst,
+                                        size_t workspace_in_bytes) {
     check_exec_allow_nhwc_mat_idx(src, mat, mat_idx, dst, workspace_in_bytes);
 }
 
@@ -248,7 +286,10 @@ void WarpPerspectiveForward::check_exec_allow_nhwc_mat_idx(
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
     if (param().format != Param::Format::NHWC &&
         param().format != Param::Format::NCHW &&
-        param().format != Param::Format::NCHW4) {
+        param().format != Param::Format::NCHW4 &&
+        param().format != Param::Format::NHWC_NCHW &&
+        param().format != Param::Format::NHWC_NCHW4_IC_SMALL &&
+        param().format != Param::Format::NCHW_NCHW4_IC_SMALL) {
         megdnn_assert(!mat_idx.ndim,
                       "mat_idx not supported for current format");
     }
@@ -263,7 +304,8 @@ void WarpPerspectiveBackwardData::check_exec(const TensorLayout& mat,
     megdnn_assert(grad.dtype == dtype::Float32() MEGDNN_INC_FLOAT16(
                                         || grad.dtype == dtype::BFloat16()),
                   "Backward WarpPerspective only supports Float32/BFloat16.");
-    auto required_workspace_in_bytes = get_workspace_in_bytes(mat, mat_idx, diff, grad);
+    auto required_workspace_in_bytes =
+            get_workspace_in_bytes(mat, mat_idx, diff, grad);
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
 }
 
@@ -283,6 +325,6 @@ void WarpPerspectiveBackwardMat::check_exec(const TensorLayout& src,
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
 }
 
-} // namespace megdnn
+}  // namespace megdnn
 
 // vim: syntax=cpp.doxygen

dnn/src/cuda/warp_perspective/common.h

@@ -12,6 +12,7 @@
 #pragma once
 #include <cuda_runtime_api.h>
 #include "src/common/cv/enums.h"
+#include "src/cuda/utils.cuh"
 #include "megcore_cdefs.h"
 
 namespace megdnn {
@@ -34,6 +35,22 @@ void forward_proxy_nchw4(const ctype* src, const float* mat, const int* mat_idx,
                          megcore::AsyncErrorInfo* error_info,
                          void* error_tracker, cudaStream_t stream);
 
+template <typename src_dtype, typename src_ctype, typename dst_ctype>
+void forward_proxy_quint8_dimshuffle_typecvt_nchw4(
+        bool is_nhwc, const src_ctype* src, const float* mat,
+        const int* mat_idx, dst_ctype* dst, int N_SRC, int N_MAT, int C, int IH,
+        int IW, int OH, int OW, src_ctype bval, DTypeParamImpl<src_dtype> param,
+        BorderMode bmode, megcore::AsyncErrorInfo* error_info,
+        void* error_tracker, cudaStream_t stream);
+
+template <typename src_dtype, typename src_ctype, typename dst_ctype>
+void forward_proxy_quint8_dimshuffle_typecvt_nchw(
+        bool is_nhwc, const src_ctype* src, const float* mat,
+        const int* mat_idx, dst_ctype* dst, int N_SRC, int N_MAT, int C, int IH,
+        int IW, int OH, int OW, src_ctype bval, DTypeParamImpl<src_dtype> param,
+        BorderMode bmode, megcore::AsyncErrorInfo* error_info,
+        void* error_tracker, cudaStream_t stream);
+
 void backward_data_proxy(const float* mat, const int* midx, const float* diff,
                          float* grad, float* workspace, int N, int N_SRC, int C,
                          int IH, int IW, int OH, int OW, float bval,

dnn/src/cuda/warp_perspective/forward.cpp

@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 #include "src/cuda/warp_perspective/opr_impl.h"
 #include "src/cuda/warp_perspective/warp_perspective_cv.cuh"
@@ -166,6 +167,30 @@ void WarpPerspectiveForwardImpl::exec(_megdnn_tensor_in ssrc,
                 IW = src.layout.shape[3];
                 OH = dst.layout.shape[2];
                 OW = dst.layout.shape[3];
+            } else if (param().format == Param::Format::NHWC_NCHW) {
+                C = src.layout.shape[3];
+                IH = src.layout.shape[1];
+                IW = src.layout.shape[2];
+                OH = dst.layout.shape[2];
+                OW = dst.layout.shape[3];
+            } else if (param().format == Param::Format::NHWC_NCHW4_IC_SMALL) {
+                C = src.layout.shape[3];
+                IH = src.layout.shape[1];
+                IW = src.layout.shape[2];
+                OH = dst.layout.shape[2];
+                OW = dst.layout.shape[3];
+                megdnn_assert(
+                        (C == 1) || (C == 3),
+                        "NHWC_NCHW4_IC_SMALL only support C == 1 or C == 3");
+            } else if (param().format == Param::Format::NCHW_NCHW4_IC_SMALL) {
+                C = src.layout.shape[1];
+                IH = src.layout.shape[2];
+                IW = src.layout.shape[3];
+                OH = dst.layout.shape[2];
+                OW = dst.layout.shape[3];
+                megdnn_assert(
+                        (C == 1) || (C == 3),
+                        "NCHW_NCHW4_IC_SMALL only support C == 1 or C == 3");
             } else {
                 megdnn_assert(
                         param().format == param::WarpPerspective::Format::NCHW,
@@ -180,55 +205,123 @@ void WarpPerspectiveForwardImpl::exec(_megdnn_tensor_in ssrc,
                           "unsupported interpolation mode for NCHW format");
             auto bval = param().border_val;
             auto bmode = warp_perspective::get_bmode(param().bmode);
-
-            if (src.layout.dtype == dtype::Float32{}) {
-                warp_perspective::forward_proxy(
-                        is_nhwc, src.ptr<dt_float32>(), mat.ptr<dt_float32>(),
-                        mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
-                        dst.ptr<dt_float32>(), src.layout[0], mat.layout[0], C,
-                        IH, IW, OH, OW, bval, bmode, async_error_info(handle()),
-                        m_error_tracker, stream);
-            } else if (MEGDNN_FLOAT16_SELECT(
-                               src.layout.dtype == dtype::Float16(), false)) {
+            if (src.layout.dtype == dst.layout.dtype) {
+                if (src.layout.dtype == dtype::Float32{}) {
+                    warp_perspective::forward_proxy(
+                            is_nhwc, src.ptr<dt_float32>(),
+                            mat.ptr<dt_float32>(),
+                            mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
+                            dst.ptr<dt_float32>(), src.layout[0], mat.layout[0],
+                            C, IH, IW, OH, OW, bval, bmode,
+                            async_error_info(handle()), m_error_tracker,
+                            stream);
+                } else if (MEGDNN_FLOAT16_SELECT(
+                                   src.layout.dtype == dtype::Float16(),
+                                   false)) {
 #ifndef MEGDNN_DISABLE_FLOAT16
-                warp_perspective::forward_proxy(
-                        is_nhwc, src.ptr<dt_float16>(), mat.ptr<dt_float32>(),
-                        mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
-                        dst.ptr<dt_float16>(), src.layout[0], mat.layout[0], C,
-                        IH, IW, OH, OW, static_cast<dt_float16>(bval), bmode,
-                        async_error_info(handle()), m_error_tracker, stream);
+                    warp_perspective::forward_proxy(
+                            is_nhwc, src.ptr<dt_float16>(),
+                            mat.ptr<dt_float32>(),
+                            mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
+                            dst.ptr<dt_float16>(), src.layout[0], mat.layout[0],
+                            C, IH, IW, OH, OW, static_cast<dt_float16>(bval),
+                            bmode, async_error_info(handle()), m_error_tracker,
+                            stream);
 #endif
-            } else if (src.layout.dtype == dtype::Uint8()) {
-                warp_perspective::forward_proxy<dt_uint8>(
-                        is_nhwc, src.ptr<dt_uint8>(), mat.ptr<dt_float32>(),
-                        mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
-                        dst.ptr<dt_uint8>(), src.layout[0], mat.layout[0], C,
-                        IH, IW, OH, OW, bval, bmode, async_error_info(handle()),
-                        m_error_tracker, stream);
-            } else if (src.layout.dtype == dtype::Int8()) {
-                megdnn_assert(
-                        !is_nhwc,
-                        "WarpPerspective on CUDA does not support NHWC + Int8");
-                warp_perspective::forward_proxy<dt_int8>(
-                        false, src.ptr<dt_int8>(), mat.ptr<dt_float32>(),
-                        mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
-                        dst.ptr<dt_int8>(), src.layout[0], mat.layout[0], C, IH,
-                        IW, OH, OW,
-                        bval /* implicit float -> int8 conversion, should be
-                                safe */
-                        ,
-                        bmode, async_error_info(handle()), m_error_tracker,
-                        stream);
-            } else if (src.layout.dtype.enumv() == DTypeEnum::QuantizedS8) {
-                megdnn_assert(param().format == Param::Format::NCHW4,
-                              "WarpPerspective on CUDA supports NCHW4 + "
-                              "QuantizedS8 only");
-                warp_perspective::forward_proxy_nchw4<dt_int8>(
-                        src.compatible_ptr<dt_int8>(), mat.ptr<dt_float32>(),
-                        mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
-                        dst.compatible_ptr<dt_int8>(), src.layout[0],
-                        mat.layout[0], C, IH, IW, OH, OW, bval, bmode,
-                        async_error_info(handle()), m_error_tracker, stream);
+                } else if (src.layout.dtype == dtype::Uint8()) {
+                    warp_perspective::forward_proxy<dt_uint8>(
+                            is_nhwc, src.ptr<dt_uint8>(), mat.ptr<dt_float32>(),
+                            mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
+                            dst.ptr<dt_uint8>(), src.layout[0], mat.layout[0],
+                            C, IH, IW, OH, OW, bval, bmode,
+                            async_error_info(handle()), m_error_tracker,
+                            stream);
+                } else if (src.layout.dtype == dtype::Int8()) {
+                    megdnn_assert(!is_nhwc,
+                                  "WarpPerspective on CUDA does not support "
+                                  "NHWC + Int8");
+                    warp_perspective::forward_proxy<dt_int8>(
+                            false, src.ptr<dt_int8>(), mat.ptr<dt_float32>(),
+                            mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
+                            dst.ptr<dt_int8>(), src.layout[0], mat.layout[0], C,
+                            IH, IW, OH, OW,
+                            bval /* implicit float -> int8 conversion,
+                                    should be safe */
+                            ,
+                            bmode, async_error_info(handle()), m_error_tracker,
+                            stream);
+                } else if (src.layout.dtype.enumv() == DTypeEnum::QuantizedS8) {
+                    megdnn_assert(param().format == Param::Format::NCHW4,
+                                  "WarpPerspective on CUDA supports NCHW4 + "
+                                  "QuantizedS8 only");
+                    warp_perspective::forward_proxy_nchw4<dt_int8>(
+                            src.compatible_ptr<dt_int8>(),
+                            mat.ptr<dt_float32>(),
+                            mat_idx.raw_ptr ? mat_idx.ptr<int>() : nullptr,
+                            dst.compatible_ptr<dt_int8>(), src.layout[0],
+                            mat.layout[0], C, IH, IW, OH, OW, bval, bmode,
+                            async_error_info(handle()), m_error_tracker,
+                            stream);
+                }
+            } else if ((src.layout.dtype.enumv() ==
+                                DTypeEnum::Quantized8Asymm ||
+                        src.layout.dtype.enumv() == DTypeEnum::Uint8)) {
+                uint8_t zero_point = 0;
+                float scale = 1.f;
+                if (src.layout.dtype.enumv() == DTypeEnum::Quantized8Asymm) {
+                    zero_point =
+                            src.layout.dtype.param<dtype::Quantized8Asymm>()
+                                    .zero_point;
+                    scale = src.layout.dtype.param<dtype::Quantized8Asymm>()
+                                    .scale;
+                } else if (src.layout.dtype.enumv() == DTypeEnum::Uint8 &&
+                           dst.layout.dtype.enumv() == DTypeEnum::QuantizedS8) {
+                    zero_point = 128;
+                    scale = 1.f;
+                }
+                DTypeParamImpl<dt_quint8> src_dtype_param(scale, zero_point);
+
+                if ((dst.layout.dtype.enumv() == DTypeEnum::QuantizedS8 &&
+                     dst.layout.dtype.param<dtype::QuantizedS8>().scale ==
+                             scale) &&
+                    ((param().format == Param::Format::NCHW_NCHW4_IC_SMALL) ||
+                     (param().format == Param::Format::NHWC_NCHW4_IC_SMALL))) {
+                    bool is_nhwc_ic_small =
+                            (param().format ==
+                             Param::Format::NHWC_NCHW4_IC_SMALL);
+                    warp_perspective::
+                            forward_proxy_quint8_dimshuffle_typecvt_nchw4<
+                                    dt_quint8, dt_uint8, dt_int8>(
+                                    is_nhwc_ic_small,
+                                    src.compatible_ptr<dt_uint8>(),
+                                    mat.ptr<dt_float32>(),
+                                    mat_idx.raw_ptr ? mat_idx.ptr<int>()
+                                                    : nullptr,
+                                    dst.compatible_ptr<dt_int8>(),
+                                    src.layout[0], mat.layout[0], C, IH, IW, OH,
+                                    OW, bval, src_dtype_param, bmode,
+                                    async_error_info(handle()), m_error_tracker,
+                                    stream);
+                } else {
+                    megdnn_assert(
+                            ((dst.layout.dtype.enumv() == DTypeEnum::Float32) &&
+                             ((param().format == Param::Format::NCHW) ||
+                              (param().format == Param::Format::NHWC_NCHW))),
+                            "invalid format for Quantized8Asymm input");
+                    bool is_nhwc = (param().format == Param::Format::NHWC_NCHW);
+                    warp_perspective::
+                            forward_proxy_quint8_dimshuffle_typecvt_nchw<
+                                    dt_quint8, dt_uint8, dt_float32>(
+                                    is_nhwc, src.compatible_ptr<dt_uint8>(),
+                                    mat.ptr<dt_float32>(),
+                                    mat_idx.raw_ptr ? mat_idx.ptr<int>()
+                                                    : nullptr,
+                                    dst.compatible_ptr<dt_float32>(),
+                                    src.layout[0], mat.layout[0], C, IH, IW, OH,
+                                    OW, bval, src_dtype_param, bmode,
+                                    async_error_info(handle()), m_error_tracker,
+                                    stream);
+                }
             } else {
                 megdnn_throw(ssprintf("unsupported dtype: %s",
                                       src.layout.dtype.name()));

dnn/src/naive/warp_perspective/opr_impl.cpp

@@ -249,6 +249,162 @@ void WarpPerspectiveForwardImpl::kern_naive_nhwcd4(
     MIDOUT_END();
 }
 
+template <typename ctype, typename dst_ctype, typename mtype>
+void WarpPerspectiveForwardImpl::kern_naive_dimshuffle_typecvt(
+        const KernParam<ctype, mtype>& kern_param, size_t task_id) {
+    MEGDNN_MARK_USED_VAR(kern_param);
+    MIDOUT_BEGIN(megdnn_naive_warpperspective, ctype, mtype, midout_iv(2)) {
+        UNPACK_WARP_PERSPECTIVE_FWD_KERN_PARAM(kern_param);
+        MEGDNN_MARK_USED_VAR(N_MAT);
+        //! strides of C, H, W on src and dst
+        size_t sstrd[3], dstrd[3];
+        auto set_sstrd = [&](size_t s0, size_t s1, size_t s2) {
+            sstrd[0] = s0;
+            sstrd[1] = s1;
+            sstrd[2] = s2;
+        };
+        auto set_dstrd = [&](size_t s0, size_t s1, size_t s2) {
+            dstrd[0] = s0;
+            dstrd[1] = s1;
+            dstrd[2] = s2;
+        };
+        switch (kern_param.format) {
+            case Format::NCHW:
+            case Format::NCHW_NCHW4_IC_SMALL:
+                set_sstrd(IH * IW, IW, 1);
+                set_dstrd(OH * OW, OW, 1);
+                break;
+            case Format::NHWC_NCHW:
+            case Format::NHWC_NCHW4_IC_SMALL:
+                set_sstrd(1, IW * C, C);
+                set_dstrd(OH * OW, OW, 1);
+                break;
+            default:
+                megdnn_throw("bad format");
+        }
+
+        uint8_t zero_point = 0;
+        float scale = 1.f;
+
+        bool is_dst_float = kern_param.dst_dtype.enumv() == DTypeEnum::Float32;
+        if (kern_param.src_dtype.enumv() ==
+            DTypeTrait<dtype::Quantized8Asymm>::enumv) {
+            auto dtype_param =
+                    kern_param.src_dtype
+                            .template param<dtype::Quantized8Asymm>();
+            zero_point = dtype_param.zero_point;
+            scale = dtype_param.scale;
+        } else if (kern_param.src_dtype.enumv() == DTypeEnum::Uint8) {
+            zero_point =
+                    (kern_param.dst_dtype.enumv() == DTypeEnum::QuantizedS8)
+                            ? 128
+                            : 0;
+            scale = 1.f;
+        }
+
+        dst_ctype* dst_ptr = reinterpret_cast<dst_ctype*>(dptr);
+
+        bool is_dst_nchw4 =
+                (kern_param.format == Format::NCHW_NCHW4_IC_SMALL) ||
+                (kern_param.format == Format::NHWC_NCHW4_IC_SMALL);
+        auto visit_src = [&sptr, sstrd](size_t c, int h, int w) -> float {
+            return sptr[sstrd[0] * c + sstrd[1] * h + sstrd[2] * w];
+        };
+        auto visit_src_bd = [&sptr, sstrd, border_val](size_t c, int h,
+                                                       int w) -> float {
+            if (h != -1 && w != -1) {
+                return sptr[sstrd[0] * c + sstrd[1] * h + sstrd[2] * w];
+            } else
+                return border_val;
+        };
+        auto visit_dst = [&dst_ptr, dstrd, is_dst_nchw4](size_t c, int h,
+                                                         int w) -> dst_ctype& {
+            if (!is_dst_nchw4)
+                return dst_ptr[dstrd[0] * c + dstrd[1] * h + dstrd[2] * w];
+            else
+                return dst_ptr[((dstrd[0] * (c >> 2) + dstrd[1] * h +
+                                 dstrd[2] * w)
+                                << 2) +
+                               (c & 0b11)];
+        };
+
+        rounding::RoundingConverter<dst_ctype> output_converter;
+        auto orig_sptr = sptr;
+        size_t n = task_id / OH;
+        size_t oh = task_id % OH;
+        mptr = mptr + n * 3 * 3;
+        dst_ptr = is_dst_nchw4 ? (dst_ptr + n * OH * OW * 4)
+                               : (dst_ptr + n * C * OH * OW);
+        if (midx_ptr) {
+            size_t idx = midx_ptr[n];
+            megdnn_assert(
+                    idx < N_SRC,
+                    "mat_idx out of bound: mat_idx[%zu]=%zu src_batch=%zu", n,
+                    idx, N_SRC);
+            sptr = orig_sptr + idx * (C * IH * IW);
+        } else if (n) {
+            sptr += n * C * IH * IW;
+        }
+        rep(ow, OW) {
+            float numeratorw = mptr[0] * ow + mptr[1] * oh + mptr[2];
+            float numeratorh = mptr[3] * ow + mptr[4] * oh + mptr[5];
+            float denominator = mptr[6] * ow + mptr[7] * oh + mptr[8];
+            float alphaw = numeratorw / denominator;
+            float alphah = numeratorh / denominator;
+
+            int iw0 = get_real_coord(std::floor(alphaw) + 0, IW);
+            int iw1 = get_real_coord(std::floor(alphaw) + 1, IW);
+            int ih0 = get_real_coord(std::floor(alphah) + 0, IH);
+            int ih1 = get_real_coord(std::floor(alphah) + 1, IH);
+
+            alphaw -= floor(alphaw);
+            alphah -= floor(alphah);
+            if (bmode != BorderMode::CONSTANT) {
+                rep(c, C) {
+                    auto val =
+                            visit_src(c, ih0, iw0) * (1.0f - alphaw) *
+                                    (1.0f - alphah) +
+                            visit_src(c, ih0, iw1) * alphaw * (1.0f - alphah) +
+                            visit_src(c, ih1, iw0) * (1.0f - alphaw) * alphah +
+                            visit_src(c, ih1, iw1) * alphaw * alphah;
+                    val = is_dst_float ? (val - zero_point) * scale
+                                       : val - zero_point;
+                    visit_dst(c, oh, ow) = output_converter(val);
+                }
+            } else {
+                rep(c, C) {
+                    auto val = visit_src_bd(c, ih0, iw0) * (1.0f - alphaw) *
+                                       (1.0f - alphah) +
+                               visit_src_bd(c, ih0, iw1) * alphaw *
+                                       (1.0f - alphah) +
+                               visit_src_bd(c, ih1, iw0) * (1.0f - alphaw) *
+                                       alphah +
+                               visit_src_bd(c, ih1, iw1) * alphaw * alphah;
+                    val = std::isfinite(val) ? val : border_val;
+                    val = is_dst_float ? (val - zero_point) * scale
+                                       : val - zero_point;
+                    visit_dst(c, oh, ow) = output_converter(val);
+                }
+            }
+            if (is_dst_nchw4) {
+                for (auto c = C; c < 4; ++c) {
+                    visit_dst(c, oh, ow) = 0;
+                }
+            }
+        }
+    }
+    MIDOUT_END();
+}
+
+#define INST(ctype, drc_ctype, mtype)                                        \
+    template void WarpPerspectiveForwardImpl::kern_naive_dimshuffle_typecvt< \
+            ctype, drc_ctype, mtype>(const KernParam<ctype, mtype>&, size_t);
+
+INST(uint8_t, int8_t, float);
+INST(uint8_t, float, float);
+
+#undef INST
+
 void WarpPerspectiveForwardImpl::exec(_megdnn_tensor_in src,
                                       _megdnn_tensor_in mat,
                                       _megdnn_tensor_in mat_idx,
@@ -320,6 +476,65 @@ void WarpPerspectiveForwardImpl::exec(_megdnn_tensor_in src,
                               src.layout.dtype.name())
                              .c_str());
     }
+
+    bool is_fusion_dtype = src.layout.dtype.enumv() != dst.layout.dtype.enumv();
+    bool is_u8_or_qu8_in =
+            src.layout.dtype.enumv() == DTypeTrait<dtype::Uint8>::enumv ||
+            src.layout.dtype.enumv() ==
+                    DTypeTrait<dtype::Quantized8Asymm>::enumv;
+
+    if (is_fusion_dtype && is_u8_or_qu8_in &&
+        ((param().format == Format::NCHW_NCHW4_IC_SMALL) ||
+         (param().format == Format::NHWC_NCHW4_IC_SMALL) ||
+         (param().format == Format::NHWC_NCHW) ||
+         (param().format == Format::NCHW))) {
+        if (src.layout.dtype.enumv() ==
+                    DTypeTrait<dtype::Quantized8Asymm>::enumv ||
+            src.layout.dtype.enumv() == DTypeTrait<dtype::Uint8>::enumv) {
+            float scale = 1.f;
+
+            if (src.layout.dtype.enumv() ==
+                DTypeTrait<dtype::Quantized8Asymm>::enumv) {
+                scale = src.layout.dtype.param<dtype::Quantized8Asymm>().scale;
+            }
+
+            auto kparam = KernParam<uint8_t, float>::from_tensors(
+                    param().format, param().bmode, param().border_val, src, mat,
+                    mat_idx, dst, workspace);
+
+            if (dst.layout.dtype.enumv() == DTypeTrait<dtype::Float32>::enumv) {
+                auto run = [kparam, this](size_t index, size_t) {
+                    kern_naive_dimshuffle_typecvt<uint8_t, float, float>(kparam,
+                                                                         index);
+                };
+                MEGDNN_DISPATCH_MULTI_THREAD_CPU_KERN_OPR(run,
+                                                          kparam.oh * batch);
+                return;
+            } else if ((dst.layout.dtype.enumv() ==
+                        DTypeTrait<dtype::QuantizedS8>::enumv) &&
+                       (dst.layout.dtype.param<dtype::QuantizedS8>().scale ==
+                        scale)) {
+                auto run = [kparam, this](size_t index, size_t) {
+                    kern_naive_dimshuffle_typecvt<uint8_t, int8_t, float>(
+                            kparam, index);
+                };
+                MEGDNN_DISPATCH_MULTI_THREAD_CPU_KERN_OPR(run,
+                                                          kparam.oh * batch);
+                return;
+            } else {
+                megdnn_throw(ssprintf("Unsupported DType in "
+                                      "WarpPerspective Dimshuffle Typecvt: %s",
+                                      src.layout.dtype.name())
+                                     .c_str());
+            }
+        }
+
+        megdnn_throw(ssprintf("Unsupported input DType in "
+                              "WarpPerspective: %s",
+                              src.layout.dtype.name())
+                             .c_str());
+    }
+
     if (warp::is_cv_available(src.layout, mat.layout, dst.layout, param().imode,
                               param().format)) {
         MIDOUT_BEGIN(megdnn_naive_warpperspective, void) {
@@ -331,12 +546,12 @@ void WarpPerspectiveForwardImpl::exec(_megdnn_tensor_in src,
         megdnn_assert(warp::is_dnn_available(src.layout, mat.layout, dst.layout,
                                              param().imode, param().format));
         /*!
-         * We currently use floating point for all WarpPerspective computation,
-         * so even if the input ctype is one of the integer type, mtype should
-         * always be float32.
+         * We currently use floating point for all WarpPerspective
+         * computation, so even if the input ctype is one of the integer
+         * type, mtype should always be float32.
          *
-         * \warning It's different with \c WarpAffine, with mtype be float16 if
-         * input type is float16.
+         * \warning It's different with \c WarpAffine, with mtype be float16
+         * if input type is float16.
          */
 
         DISPATCH_ST(dtype::Float32, float, float, KERN);

dnn/src/naive/warp_perspective/opr_impl.h

@@ -26,6 +26,7 @@ protected:
         float border_val;
         size_t n_src, n_mat, c, ih, iw, oh, ow;
         ctype *sptr, *dptr;
+        DType src_dtype, dst_dtype;
         mtype* mptr;
         int* midx_ptr;  //!< can be null
         Workspace workspace;
@@ -41,6 +42,8 @@ protected:
             ret.bmode = bmode;
             ret.border_val = border_val;
             ret.n_src = src.layout.shape[0];
+            ret.src_dtype = src.layout.dtype;
+            ret.dst_dtype = dst.layout.dtype;
             if (mat_idx.raw_ptr) {
                 megdnn_assert(mat_idx.layout.ndim == 1);
                 ret.n_mat = mat_idx.layout.shape[0];
@@ -50,7 +53,8 @@ protected:
                 ret.n_mat = ret.n_src;
                 ret.midx_ptr = nullptr;
             }
-            if (format == Format::NCHW) {
+            if (format == Format::NCHW ||
+                format == Format::NCHW_NCHW4_IC_SMALL) {
                 ret.c = src.layout.shape[1];
                 ret.ih = src.layout.shape[2];
                 ret.iw = src.layout.shape[3];
@@ -62,6 +66,13 @@ protected:
                 ret.iw = src.layout.shape[2];
                 ret.oh = dst.layout.shape[1];
                 ret.ow = dst.layout.shape[2];
+            } else if (format == Format::NHWC_NCHW ||
+                       format == Format::NHWC_NCHW4_IC_SMALL) {
+                ret.c = src.layout.shape[3];
+                ret.ih = src.layout.shape[1];
+                ret.iw = src.layout.shape[2];
+                ret.oh = dst.layout.shape[2];
+                ret.ow = dst.layout.shape[3];
             } else if (format == Format::NCHW4) {
                 ret.c = src.layout.shape[1] * 4;
                 ret.ih = src.layout.shape[2];
@@ -76,15 +87,16 @@ protected:
                 ret.oh = dst.layout.shape[1];
                 ret.ow = dst.layout.shape[3];
             }
-            if (src.layout.dtype.enumv() == DTypeEnum::Float32 ||
-                MEGDNN_FLOAT16_SELECT(
-                        (src.layout.dtype.enumv() == DTypeEnum::Float16 ||
-                         src.layout.dtype.enumv() == DTypeEnum::BFloat16),
-                        false) ||
-                src.layout.dtype.enumv() == DTypeEnum::Int8 ||
-                src.layout.dtype.enumv() == DTypeEnum::Uint8 ||
-                src.layout.dtype.enumv() == DTypeEnum::QuantizedS8 ||
-                src.layout.dtype.enumv() == DTypeEnum::Quantized8Asymm) {
+            if ((src.layout.dtype.enumv() == DTypeEnum::Float32 ||
+                 MEGDNN_FLOAT16_SELECT(
+                         (src.layout.dtype.enumv() == DTypeEnum::Float16 ||
+                          src.layout.dtype.enumv() == DTypeEnum::BFloat16),
+                         false) ||
+                 src.layout.dtype.enumv() == DTypeEnum::Int8 ||
+                 src.layout.dtype.enumv() == DTypeEnum::Uint8 ||
+                 src.layout.dtype.enumv() == DTypeEnum::QuantizedS8 ||
+                 src.layout.dtype.enumv() == DTypeEnum::Quantized8Asymm) &&
+                (src.layout.dtype == dst.layout.dtype)) {
                 ret.sptr = src.compatible_ptr<ctype>();
                 ret.mptr = mat.ptr<mtype>();
                 ret.dptr = dst.compatible_ptr<ctype>();
@@ -92,6 +104,13 @@ protected:
                 ret.sptr = src.compatible_ptr<ctype>();
                 ret.mptr = mat.ptr<mtype>();
                 ret.dptr = dst.compatible_ptr<ctype>();
+            } else if ((src.layout.dtype.enumv() == DTypeEnum::Uint8 ||
+                        src.layout.dtype.enumv() ==
+                                DTypeEnum::Quantized8Asymm) &&
+                       src.layout.dtype.enumv() != dst.layout.dtype.enumv()) {
+                ret.sptr = src.compatible_ptr<ctype>();
+                ret.mptr = mat.ptr<mtype>();
+                ret.dptr = reinterpret_cast<ctype*>(dst.raw_ptr);
             } else {
                 ret.sptr = nullptr;
                 ret.mptr = nullptr;
@@ -122,6 +141,9 @@ private:
     template <typename ctype, typename mtype>
     void kern_naive_nhwcd4(const KernParam<ctype, mtype>& kern_param,
                            size_t task_id);
+    template <typename ctype, typename dst_ctype, typename mtype>
+    void kern_naive_dimshuffle_typecvt(
+            const KernParam<ctype, mtype>& kern_param, size_t task_id);
 };
 
 class WarpPerspectiveBackwardDataImpl : public WarpPerspectiveBackwardData {

dnn/test/cuda/warp_perspective.cpp

@@ -23,8 +23,7 @@ using namespace megdnn;
 using namespace test;
 
 class NanMatRNG : public RNG {
-    void gen(const TensorND& tensor_) override
-    {
+    void gen(const TensorND& tensor_) override {
         auto& gen = RandomState::generator();
         std::uniform_real_distribution<dt_float32> pdist3(1.9f, 2.1f);
         std::uniform_real_distribution<dt_float32> pdist(0.9f, 1.1f);
@@ -335,6 +334,144 @@ TEST_F(CUDA, WARP_PERSPECTIVE_NCHW4) {
     }
 }
 
+TEST_F(CUDA, WARP_PERSPECTIVE_NCHW_NCHW4_IC_SMALL) {
+    using Param = WarpPerspective::Param;
+    WarpPerspective::Param param;
+    Checker<WarpPerspectiveForward> checker(handle_cuda());
+    WarpPerspectiveMatRNG rng;
+
+    param.format = Param::Format::NCHW_NCHW4_IC_SMALL;
+
+    checker.set_rng(1, &rng);
+    checker.set_dtype(0, dtype::Quantized8Asymm(0.1f, 128));
+    checker.set_dtype(2, dtype::QuantizedS8(0.1f));
+    for (auto bmode : {WarpPerspective::BorderMode::WRAP,
+                       WarpPerspective::BorderMode::REFLECT,
+                       WarpPerspective::BorderMode::REPLICATE,
+                       WarpPerspective::BorderMode::CONSTANT}) {
+        param.border_val = 0.3f;
+        param.bmode = bmode;
+        param.imode = Param::InterpolationMode::LINEAR;
+
+        checker.set_param(param);
+        checker.set_epsilon(1 + 1e-3);
+        checker.execs({{2, 3, 10, 11}, {2, 3, 3}, {2, 1, 11, 12, 4}});
+        checker.execs({{1, 3, 25, 510}, {1, 3, 3}, {1, 1, 25, 25, 4}});
+        checker.execs({{1, 3, 25, 25}, {1, 3, 3}, {1, 1, 51, 51, 4}});
+        checker.execs({{1, 3, 51, 51}, {1, 3, 3}, {1, 1, 25, 25, 4}});
+    }
+    {
+        Checker<WarpPerspective, WarpPerspectiveMatIdxProxy> checker(
+                handle_cuda());
+        constexpr int N_SRC = 5;
+        UniformIntRNG mat_idx_rng{0, N_SRC - 1};
+        checker.set_dtype(0, dtype::Quantized8Asymm(0.1f, 128));
+        checker.set_rng(1, &rng);
+        checker.set_dtype(2, dtype::Int32());
+        checker.set_rng(2, &mat_idx_rng);
+        checker.set_dtype(3, dtype::QuantizedS8(0.1f));
+        param.bmode = WarpPerspective::Param::BorderMode::REFLECT;
+        param.imode = param::WarpPerspective::InterpolationMode::LINEAR;
+        checker.set_param(param);
+        checker.set_epsilon(1 + 1e-3);
+        checker.execs({{N_SRC, 3, 10, 11}, {2, 3, 3}, {2}, {2, 1, 11, 12, 4}});
+        checker.execs(
+                {{N_SRC, 3, 17, 13}, {123, 3, 3}, {123}, {123, 1, 16, 15, 4}});
+    }
+}
+
+TEST_F(CUDA, WARP_PERSPECTIVE_NHWC_NCHW4_IC_SMALL) {
+    using Param = WarpPerspective::Param;
+    WarpPerspective::Param param;
+    Checker<WarpPerspectiveForward> checker(handle_cuda());
+    WarpPerspectiveMatRNG rng;
+
+    param.format = Param::Format::NHWC_NCHW4_IC_SMALL;
+
+    checker.set_rng(1, &rng);
+    checker.set_dtype(0, dtype::Uint8());
+    checker.set_dtype(2, dtype::QuantizedS8(1.f));
+    for (auto bmode : {WarpPerspective::BorderMode::WRAP,
+                       WarpPerspective::BorderMode::REFLECT,
+                       WarpPerspective::BorderMode::REPLICATE,
+                       WarpPerspective::BorderMode::CONSTANT}) {
+        param.border_val = 0.3f;
+        param.bmode = bmode;
+        param.imode = Param::InterpolationMode::LINEAR;
+
+        checker.set_param(param);
+        checker.set_epsilon(1 + 1e-3);
+        checker.execs({{2, 10, 11, 3}, {2, 3, 3}, {2, 1, 11, 12, 4}});
+        checker.execs({{1, 25, 510, 3}, {1, 3, 3}, {1, 1, 25, 25, 4}});
+        checker.execs({{1, 25, 25, 3}, {1, 3, 3}, {1, 1, 51, 51, 4}});
+        checker.execs({{1, 51, 51, 3}, {1, 3, 3}, {1, 1, 25, 25, 4}});
+    }
+    {
+        Checker<WarpPerspective, WarpPerspectiveMatIdxProxy> checker(
+                handle_cuda());
+        constexpr int N_SRC = 5;
+        UniformIntRNG mat_idx_rng{0, N_SRC - 1};
+        checker.set_dtype(0, dtype::Uint8());
+        checker.set_rng(1, &rng);
+        checker.set_dtype(2, dtype::Int32());
+        checker.set_rng(2, &mat_idx_rng);
+        checker.set_dtype(3, dtype::QuantizedS8(1.f));
+        param.bmode = WarpPerspective::Param::BorderMode::REFLECT;
+        param.imode = param::WarpPerspective::InterpolationMode::LINEAR;
+        checker.set_param(param);
+        checker.set_epsilon(1 + 1e-3);
+        checker.execs({{N_SRC, 10, 11, 3}, {2, 3, 3}, {2}, {2, 1, 11, 12, 4}});
+        checker.execs(
+                {{N_SRC, 17, 13, 3}, {123, 3, 3}, {123}, {123, 1, 16, 15, 4}});
+    }
+}
+
+TEST_F(CUDA, WARP_PERSPECTIVE_NHWC_NCHW) {
+    using Param = WarpPerspective::Param;
+    WarpPerspective::Param param;
+    Checker<WarpPerspectiveForward> checker(handle_cuda());
+    WarpPerspectiveMatRNG rng;
+
+    param.format = Param::Format::NHWC_NCHW;
+
+    checker.set_rng(1, &rng);
+    checker.set_dtype(0, dtype::Uint8());
+    checker.set_dtype(2, dtype::Float32());
+    for (auto bmode : {WarpPerspective::BorderMode::WRAP,
+                       WarpPerspective::BorderMode::REFLECT,
+                       WarpPerspective::BorderMode::REPLICATE,
+                       WarpPerspective::BorderMode::CONSTANT}) {
+        param.border_val = 0.3f;
+        param.bmode = bmode;
+        param.imode = Param::InterpolationMode::LINEAR;
+
+        checker.set_param(param);
+        checker.set_epsilon(1 + 1e-3);
+        checker.execs({{2, 10, 11, 3}, {2, 3, 3}, {2, 3, 11, 12}});
+        checker.execs({{1, 25, 510, 3}, {1, 3, 3}, {1, 3, 25, 25}});
+        checker.execs({{1, 25, 25, 3}, {1, 3, 3}, {1, 3, 51, 51}});
+        checker.execs({{1, 51, 51, 3}, {1, 3, 3}, {1, 3, 25, 25}});
+    }
+    {
+        Checker<WarpPerspective, WarpPerspectiveMatIdxProxy> checker(
+                handle_cuda());
+        constexpr int N_SRC = 5;
+        UniformIntRNG mat_idx_rng{0, N_SRC - 1};
+        checker.set_dtype(0, dtype::Uint8());
+        checker.set_rng(1, &rng);
+        checker.set_dtype(2, dtype::Int32());
+        checker.set_rng(2, &mat_idx_rng);
+        checker.set_dtype(3, dtype::Float32());
+        param.bmode = WarpPerspective::Param::BorderMode::REFLECT;
+        param.imode = param::WarpPerspective::InterpolationMode::LINEAR;
+        checker.set_param(param);
+        checker.set_epsilon(1 + 1e-3);
+        checker.execs({{N_SRC, 10, 11, 3}, {2, 3, 3}, {2}, {2, 3, 11, 12}});
+        checker.execs(
+                {{N_SRC, 17, 13, 3}, {123, 3, 3}, {123}, {123, 3, 16, 15}});
+    }
+}
+
 TEST_F(CUDA, WARP_PERSPECTIVE_FORWARD_NCHW_INT8) {
     warp_perspective::run_int8_test(handle_cuda());
 }

src/gopt/impl/framework.cpp

@@ -6,7 +6,8 @@
  *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
- * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
  */
 
 #include "megbrain/gopt/framework.h"
@@ -35,13 +36,13 @@ using namespace gopt;
 /* ================ SubGraph ================ */
 
 OperatorNodeBase* SubGraph::Rewriter::auto_replace_outputs(
-        OperatorNodeBase *opr) {
-    auto &&new_inp = m_opr_new_inp_cache;
+        OperatorNodeBase* opr) {
+    auto&& new_inp = m_opr_new_inp_cache;
     new_inp.clear();
     new_inp.reserve(opr->input().size());
     bool has_replaced_inp = false;
 
-    for (auto i: opr->input()) {
+    for (auto i : opr->input()) {
         auto new_var = get_var(i);
         if (new_var != i) {
             has_replaced_inp = true;
@@ -52,14 +53,14 @@ OperatorNodeBase* SubGraph::Rewriter::auto_replace_outputs(
     }
 
     if (has_replaced_inp) {
-        auto new_opr = serialization::copy_opr_shallow(
-                *opr, new_inp, opr->config());
+        auto new_opr =
+                serialization::copy_opr_shallow(*opr, new_inp, opr->config());
         auto &&out0 = opr->output(), &&out1 = new_opr->output();
         size_t i = 0;
         auto err_msg = [opr, new_opr] {
-            return ssprintf("bad opr copy: src=%s{%s} dst=%s{%s}",
-                    opr->cname(), opr->dyn_typeinfo()->name,
-                    new_opr->cname(), new_opr->dyn_typeinfo()->name);
+            return ssprintf("bad opr copy: src=%s{%s} dst=%s{%s}", opr->cname(),
+                            opr->dyn_typeinfo()->name, new_opr->cname(),
+                            new_opr->dyn_typeinfo()->name);
         };
         MGB_MARK_USED_VAR(err_msg);
         // opr output size mismatch may be caused by:
@@ -67,33 +68,33 @@ OperatorNodeBase* SubGraph::Rewriter::auto_replace_outputs(
         //     1) other post-insert optimization (e.g. const folding)
         // we can't handle only usable_output here, since some output var with
         // volatile flag could be the graph's endpoint (e.g. RemoteSend)
-        for (; i < std::min(out0.size(), out1.size()); ++ i) {
+        for (; i < std::min(out0.size(), out1.size()); ++i) {
             bool v0 = out0[i]->contain_flag(VarNode::Flag::VOLATILE_CONTENT),
                  v1 = out1[i]->contain_flag(VarNode::Flag::VOLATILE_CONTENT);
             mgb_assert(v0 == v1, "%s", err_msg().c_str());
 
-            auto &&ins = m_varmap.insert({out0[i], {true, nullptr}});
+            auto&& ins = m_varmap.insert({out0[i], {true, nullptr}});
             mgb_assert(ins.second || ins.first->second.first,
                        "opr output already replaced");
             // handle repeated call on the same opr
             ins.first->second.second = out1[i];
             on_var_replaced(out0[i], out1[i], nullptr);
         }
-        for (; i < out0.size(); ++ i) {
+        for (; i < out0.size(); ++i) {
             mgb_assert(out0[i]->contain_flag(VarNode::Flag::VOLATILE_CONTENT),
-                    "%s", err_msg().c_str());
+                       "%s", err_msg().c_str());
         }
-        for (; i < out1.size(); ++ i) {
+        for (; i < out1.size(); ++i) {
             mgb_assert(out1[i]->contain_flag(VarNode::Flag::VOLATILE_CONTENT),
-                    "%s", err_msg().c_str());
+                       "%s", err_msg().c_str());
         }
         return new_opr;
     }
     return opr;
 }
 
-void SubGraph::Rewriter::replace_var(
-        VarNode *src, VarNode *dst, const char *msg) {
+void SubGraph::Rewriter::replace_var(VarNode* src, VarNode* dst,
+                                     const char* msg) {
     if (src == dst)
         return;
 
@@ -103,19 +104,19 @@ void SubGraph::Rewriter::replace_var(
             "dst %s maps back to src %s in SubGraph::Rewriter::replace_var",
             dst->cname(), src->cname());
 
-    auto &&ins = m_varmap.insert({src, {false, dst}});
+    auto&& ins = m_varmap.insert({src, {false, dst}});
     if (!ins.second) {
-        auto &&old_rep = ins.first->second;
+        auto&& old_rep = ins.first->second;
         mgb_assert(old_rep.first || old_rep.second == dst,
-                "can not replace a var twice");
+                   "can not replace a var twice");
         old_rep.first = false;
         old_rep.second = dst;
     }
     on_var_replaced(src, dst, msg);
 }
 
-void SubGraph::Rewriter::on_var_replaced(
-        VarNode* src, VarNode* dst, const char* msg) {
+void SubGraph::Rewriter::on_var_replaced(VarNode* src, VarNode* dst,
+                                         const char* msg) {
     if (auto state = m_owner_graph->owner_opt_state()) {
         state->on_var_replaced(src, dst, msg);
     }
@@ -124,7 +125,7 @@ void SubGraph::Rewriter::on_var_replaced(
 void SubGraph::Rewriter::apply_inplace() const {
     m_owner_graph->m_endpoint_oprs.clear();
     m_owner_graph->m_endpoint_vars_set.clear();
-    for (auto &&var: m_owner_graph->m_endpoint_vars) {
+    for (auto&& var : m_owner_graph->m_endpoint_vars) {
         var = get_var(var.node());
         m_owner_graph->m_endpoint_oprs.insert(var.node()->owner_opr());
         m_owner_graph->m_endpoint_vars_set.insert(var.node());
@@ -150,33 +151,30 @@ std::pair<bool, VarNode*> SubGraph::Rewriter::get_var_internal(VarNode* var) {
     return it->second = {it_next->second.first & it->second.first, next.second};
 }
 
-SubGraph::SubGraph(const SymbolVarArray &endpoint_vars):
-    m_endpoint_vars(endpoint_vars)
-{
+SubGraph::SubGraph(const SymbolVarArray& endpoint_vars)
+        : m_endpoint_vars(endpoint_vars) {
     mgb_assert(!endpoint_vars.empty(), "endpoints can not be empty");
     m_comp_graph = endpoint_vars[0].node()->owner_graph();
-    for (auto i: endpoint_vars) {
+    for (auto i : endpoint_vars) {
         m_endpoint_oprs.insert(i.node()->owner_opr());
         m_endpoint_vars_set.insert(i.node());
         mgb_assert(m_comp_graph == i.node()->owner_graph(),
-                "endpoints belong to different computing graphs");
+                   "endpoints belong to different computing graphs");
     }
 }
 
-void SubGraph::iter(
-        const Callback& cb,
-        std::shared_ptr<ExtraDep> extra_dep) const {
+void SubGraph::iter(const Callback& cb,
+                    std::shared_ptr<ExtraDep> extra_dep) const {
     Callback on_opr;
 
     if (m_owner_opt_state) {
-        on_opr = [state=m_owner_opt_state, &cb](OperatorNodeBase *opr) {
+        on_opr = [state = m_owner_opt_state, &cb](OperatorNodeBase* opr) {
             state->m_opr_property_flag = OprPropertyFlag::ALL;
             state->m_cur_iter_src_opr = cg::get_opr_root_source_opr(opr);
             state->m_cur_iter_opr_priority =
-                opr->node_prop().attribute().priority;
+                    opr->node_prop().attribute().priority;
             state->m_cur_iter_opr_stream_prop_type =
-                state->m_comp_node_opt.stream_prop_type(
-                        opr->output(0));
+                    state->m_comp_node_opt.stream_prop_type(opr->output(0));
             mgb_assert(state->m_oprs_inserted.empty());
             cb(opr);
             state->m_opr_property_flag = OprPropertyFlag::NONE;
@@ -188,19 +186,19 @@ void SubGraph::iter(
     }
 
     cg::DepOprIter dep_iter{on_opr, std::move(extra_dep)};
-    for (auto i: m_endpoint_oprs)
+    for (auto i : m_endpoint_oprs)
         dep_iter.add(i);
 }
 
 ThinHashMap<VarNode*, size_t> SubGraph::get_var2nr_val_dep_oprs() const {
     ThinHashMap<VarNode*, size_t> ret;
-    auto cb = [&](OperatorNodeBase *opr) {
-        for (auto &&i: opr->node_prop().dep_map()) {
+    auto cb = [&](OperatorNodeBase* opr) {
+        for (auto&& i : opr->node_prop().dep_map()) {
             if (OperatorNodeBase::NodeProp::is_device_value_dep(i.second)) {
-                ++ ret.at(i.first);
+                ++ret.at(i.first);
             }
         }
-        for (auto i: opr->output()) {
+        for (auto i : opr->output()) {
             if (!i->contain_flag(VarNode::Flag::VOLATILE_CONTENT)) {
                 auto ins = ret.insert({i, 0});
                 mgb_assert(ins.second);
@@ -208,13 +206,13 @@ ThinHashMap<VarNode*, size_t> SubGraph::get_var2nr_val_dep_oprs() const {
         }
     };
     iter(cb);
-    for (auto i: m_endpoint_vars_set) {
+    for (auto i : m_endpoint_vars_set) {
         auto iter = ret.find(i);
         if (iter == ret.end()) {
             mgb_assert(i->contain_flag(VarNode::Flag::VOLATILE_CONTENT));
             ret[i] = 1;
         } else {
-            ++ ret.at(i);
+            ++ret.at(i);
         }
     }
     return ret;
@@ -222,10 +220,8 @@ ThinHashMap<VarNode*, size_t> SubGraph::get_var2nr_val_dep_oprs() const {
 
 /* ================ UniqReaderCheck ================ */
 
-UniqReaderCheck::UniqReaderCheck(const SubGraph &graph):
-    m_var2nr_val_dep{graph.get_var2nr_val_dep_oprs()}
-{
-}
+UniqReaderCheck::UniqReaderCheck(const SubGraph& graph)
+        : m_var2nr_val_dep{graph.get_var2nr_val_dep_oprs()} {}
 
 void UniqReaderCheck::update_on_opr_auto_replace(OperatorNodeBase* opr,
                                                  OperatorNodeBase* repl_opr) {
@@ -253,32 +249,30 @@ void UniqReaderCheck::update_on_opr_auto_replace(OperatorNodeBase* opr,
 
 /* ================ OptState ================ */
 
-OptState::OptState(
-        const GraphOptimizer *owner_optimizer, const SubGraph& graph):
-    m_owner_optimizer{owner_optimizer},
-    m_var_replace_map{
-        const_cast<ThinHashMap<VarNode*, VarNode*>*>(
-                &GraphOptimizer::var_replace_map(*graph.comp_graph()))},
-    m_comp_node_opt{graph.comp_graph()->seq_comp_node_optimizer()},
-    m_graph{graph}
-{
+OptState::OptState(const GraphOptimizer* owner_optimizer, const SubGraph& graph)
+        : m_owner_optimizer{owner_optimizer},
+          m_var_replace_map{const_cast<ThinHashMap<VarNode*, VarNode*>*>(
+                  &GraphOptimizer::var_replace_map(*graph.comp_graph()))},
+          m_comp_node_opt{graph.comp_graph()->seq_comp_node_optimizer()},
+          m_graph{graph} {
     mgb_assert(!m_graph.m_owner_opt_state);
     m_var_replace_map->clear();
     m_graph.m_owner_opt_state = this;
     m_oprs_inserted.clear();
 
-    auto on_opr_insert = [this](const cg::event::OprInserted &ev) {
+    auto on_opr_insert = [this](const cg::event::OprInserted& ev) {
         auto need_src_opr = m_opr_property_flag & OprPropertyFlag::SOURCE_OPR,
              need_priority = m_opr_property_flag & OprPropertyFlag::PRIORITY;
         if (need_src_opr)
-            mgb_assert(m_cur_iter_src_opr, "opr %s{%s} created outside from "
-                    "SubGraph::iter",
-                    ev.opr->cname(), ev.opr->dyn_typeinfo()->name);
+            mgb_assert(m_cur_iter_src_opr,
+                       "opr %s{%s} created outside from "
+                       "SubGraph::iter",
+                       ev.opr->cname(), ev.opr->dyn_typeinfo()->name);
         if (ev.exc || ev.is_dedup)
             return;
 
-        auto &&new_attr = ev.opr->node_prop().attribute();
-        auto &&ins = m_oprs_inserted.insert({ev.opr, OprPropertyFlag::NONE});
+        auto&& new_attr = ev.opr->node_prop().attribute();
+        auto&& ins = m_oprs_inserted.insert({ev.opr, OprPropertyFlag::NONE});
         mgb_assert(ins.second);
 
         if (need_src_opr && !new_attr.src_opr) {
@@ -296,20 +290,22 @@ OptState::OptState(
 
         auto csp = m_cur_iter_opr_stream_prop_type;
         if (csp.prop_type != cg::SeqCompNodeOptimizer::StreamPropType::NONE) {
-            for (auto i: ev.opr->output())
+            for (auto i : ev.opr->output())
                 m_comp_node_opt.register_stream_var(i, csp);
         }
     };
-    m_on_opr_insert_handler = graph.comp_graph()->event().register_receiver<
-        cg::event::OprInserted>(on_opr_insert);
+    m_on_opr_insert_handler =
+            graph.comp_graph()
+                    ->event()
+                    .register_receiver<cg::event::OprInserted>(on_opr_insert);
 }
 
-void OptState::on_var_replaced(VarNode *src, VarNode *dst, const char *msg) {
+void OptState::on_var_replaced(VarNode* src, VarNode* dst, const char* msg) {
     if (src->contain_flag(VarNode::Flag::VOLATILE_CONTENT)) {
         // this can only happen in auto_replace_outputs()
         mgb_assert(dst->contain_flag(VarNode::Flag::VOLATILE_CONTENT) &&
-                src->owner_opr()->dyn_typeinfo() ==
-                dst->owner_opr()->dyn_typeinfo());
+                   src->owner_opr()->dyn_typeinfo() ==
+                           dst->owner_opr()->dyn_typeinfo());
         mgb_assert(!msg);
         return;
     }
@@ -362,7 +358,7 @@ void OptState::on_var_replaced(VarNode *src, VarNode *dst, const char *msg) {
                 return f & (InferType::RT_STATIC | InferType::CONST);
             };
             if (!(norm(it0.shape) == norm(it1.shape) &&
-                norm(it0.value) <= norm(it1.value))) {
+                  norm(it0.value) <= norm(it1.value))) {
                 suc = false;
                 fail_chks.push_back("infer-type");
             }
@@ -407,22 +403,21 @@ void OptState::on_var_replaced(VarNode *src, VarNode *dst, const char *msg) {
 
 #if MGB_ENABLE_LOGGING
     if (msg && m_owner_optimizer->verbosity()) {
-        m_log_msg.
-            append("\n ").
-            append(std::to_string(m_log_nr_item)).
-            append(": ").
-            append(src->owner_opr()->cname()).
-            append(" => ").
-            append(dst->owner_opr()->cname()).
-            append(" (").
-            append(msg).
-            append(")");
-    }
-    ++ m_log_nr_item;
+        m_log_msg.append("\n ")
+                .append(std::to_string(m_log_nr_item))
+                .append(": ")
+                .append(src->owner_opr()->cname())
+                .append(" => ")
+                .append(dst->owner_opr()->cname())
+                .append(" (")
+                .append(msg)
+                .append(")");
+    }
+    ++m_log_nr_item;
 #endif
 }
 
-size_t OptState::flush_log(const char *title) {
+size_t OptState::flush_log(const char* title) {
     if (m_owner_optimizer->verbosity() >= 2) {
         if (m_log_msg.empty()) {
             m_log_msg = mgb_cstr_log(" no var replacement logged");
@@ -435,42 +430,40 @@ size_t OptState::flush_log(const char *title) {
     return ret;
 }
 
-void OptState::call_with_opr(OperatorNodeBase *opr, thin_function<void(void)> func,
+void OptState::call_with_opr(OperatorNodeBase* opr,
+                             thin_function<void(void)> func,
                              OprPropertyFlag opr_property_flag) {
     auto src_opr = cg::get_opr_root_source_opr(opr);
     auto opr_priority = opr->node_prop().attribute().priority;
     auto stream_prop_type = m_comp_node_opt.stream_prop_type(opr->output(0));
     ThinHashMap<OperatorNodeBase*, OprPropertyFlag> oprs_inserted;
 
-    auto swap_properties = [&,
-        need_src_opr = opr_property_flag & OprPropertyFlag::SOURCE_OPR,
-        need_priority = opr_property_flag & OprPropertyFlag::PRIORITY] {
-        if (need_src_opr) {
-            std::swap(m_cur_iter_src_opr, src_opr);
-        }
-        if (need_priority) {
-            std::swap(m_cur_iter_opr_priority, opr_priority);
-        }
-        std::swap(m_cur_iter_opr_stream_prop_type, stream_prop_type);
-        std::swap(m_opr_property_flag, opr_property_flag);
-        std::swap(m_oprs_inserted, oprs_inserted);
-    };
+    auto swap_properties =
+            [&, need_src_opr = opr_property_flag & OprPropertyFlag::SOURCE_OPR,
+             need_priority = opr_property_flag & OprPropertyFlag::PRIORITY] {
+                if (need_src_opr) {
+                    std::swap(m_cur_iter_src_opr, src_opr);
+                }
+                if (need_priority) {
+                    std::swap(m_cur_iter_opr_priority, opr_priority);
+                }
+                std::swap(m_cur_iter_opr_stream_prop_type, stream_prop_type);
+                std::swap(m_opr_property_flag, opr_property_flag);
+                std::swap(m_oprs_inserted, oprs_inserted);
+            };
     MGB_TRY {
         swap_properties();
         func();
-    } MGB_FINALLY({
-        swap_properties();
-    });
+    }
+    MGB_FINALLY({ swap_properties(); });
 }
 
 /* ================ RecursiveSubGraphRewriteHelper ================ */
-RecursiveSubGraphRewriteHelper::
-~RecursiveSubGraphRewriteHelper() noexcept = default;
+RecursiveSubGraphRewriteHelper::~RecursiveSubGraphRewriteHelper() noexcept =
+        default;
 
-RecursiveSubGraphRewriteHelper::RecursiveSubGraphRewriteHelper(OptState &state):
-    m_opt_state{state}, m_rewriter{state.graph().make_rewriter()}
-{
-}
+RecursiveSubGraphRewriteHelper::RecursiveSubGraphRewriteHelper(OptState& state)
+        : m_opt_state{state}, m_rewriter{state.graph().make_rewriter()} {}
 
 void RecursiveSubGraphRewriteHelper::apply() {
     using namespace std::placeholders;
@@ -479,8 +472,8 @@ void RecursiveSubGraphRewriteHelper::apply() {
     m_rewriter.apply_inplace();
 }
 
-void RecursiveSubGraphRewriteHelper::on_opr(OperatorNodeBase *opr) {
-    auto on_new_opr = [this](OperatorNodeBase *opr) {
+void RecursiveSubGraphRewriteHelper::on_opr(OperatorNodeBase* opr) {
+    auto on_new_opr = [this](OperatorNodeBase* opr) {
         auto repl_opr = m_rewriter.auto_replace_outputs(opr);
         return on_new_opr_check_should_process(opr, repl_opr);
     };
@@ -493,8 +486,8 @@ void RecursiveSubGraphRewriteHelper::on_opr(OperatorNodeBase *opr) {
         return;
 
     mgb_assert(m_opr_stack.empty());
-    m_opr_stack.push_back({
-            orig_out, m_rewriter.get_var(orig_out)->owner_opr()});
+    m_opr_stack.push_back(
+            {orig_out, m_rewriter.get_var(orig_out)->owner_opr()});
 
     bool first = true;
     while (!m_opr_stack.empty()) {
@@ -515,9 +508,9 @@ void RecursiveSubGraphRewriteHelper::on_opr(OperatorNodeBase *opr) {
         if (should_process) {
             auto trans = process_opr(cur_out);
             if (trans.valid()) {
-                m_opr_stack.push_back({
-                        cur_frame.orig_var, trans->result->owner_opr()});
-                for (auto i: reverse_adaptor(trans->internal)) {
+                m_opr_stack.push_back(
+                        {cur_frame.orig_var, trans->result->owner_opr()});
+                for (auto i : reverse_adaptor(trans->internal)) {
                     if (i)
                         m_opr_stack.push_back({i, i->owner_opr()});
                 }
@@ -532,7 +525,7 @@ void RecursiveSubGraphRewriteHelper::on_opr(OperatorNodeBase *opr) {
 
         auto src = cur_frame.orig_var;
         if (m_rewriter.get_var(src) != cur_out) {
-            const char *msg = nullptr;
+            const char* msg = nullptr;
             if (m_opr_stack.empty()) {
                 msg = m_log_msg.c_str();
             }
@@ -550,11 +543,12 @@ void RecursiveSubGraphRewriteHelper::on_opr(OperatorNodeBase *opr) {
 
 GraphOptimizer::~GraphOptimizer() noexcept = default;
 
-class GraphOptimizer::VarReplaceMapStorage :public UserDataContainer::UserData {
+class GraphOptimizer::VarReplaceMapStorage
+        : public UserDataContainer::UserData {
     MGB_TYPEINFO_OBJ_DECL;
 
-    public:
-        ThinHashMap<VarNode*, VarNode*> map;
+public:
+    ThinHashMap<VarNode*, VarNode*> map;
 };
 MGB_TYPEINFO_OBJ_IMPL(GraphOptimizer::VarReplaceMapStorage);
 
@@ -565,7 +559,7 @@ GraphOptimizer& GraphOptimizer::add_pass(std::unique_ptr<Pass> pass) {
     return *this;
 }
 
-SubGraph GraphOptimizer::apply(const SubGraph &graph) const {
+SubGraph GraphOptimizer::apply(const SubGraph& graph) const {
     RealTimer timer;
     OptState state{this, graph};
 
@@ -574,38 +568,38 @@ SubGraph GraphOptimizer::apply(const SubGraph &graph) const {
     // first update output var shapes of all oprs
     state.graph().iter(cg::update_output_var_shapes);
 
-    auto &&opt = graph.comp_graph()->options();
+    auto&& opt = graph.comp_graph()->options();
     auto orig_setting = opt.graph_opt_level;
-    Pass *cur_pass = nullptr;
+    Pass* cur_pass = nullptr;
     MGB_MARK_USED_VAR(cur_pass);
     MGB_TRY {
-        for (auto &&i: m_passes) {
+        for (auto&& i : m_passes) {
             state.set_var_replace_check_flag(VarReplaceCheckFlag::CHECK_ALL);
             cur_pass = i.get();
             opt.graph_opt_level = 1;
             i->apply(state);
             tot_nr_replace += state.flush_log(
-                    mgb_ssprintf_log(
-                        "apply optimization pass %s:", i->name()).c_str());
+                    mgb_ssprintf_log("apply optimization pass %s:", i->name())
+                            .c_str());
         }
-    } MGB_CATCH(std::exception &exc, {
+    }
+    MGB_CATCH(std::exception & exc, {
         mgb_log_error("error while applying optimization pass %s: %s",
-                cur_pass->name(), exc.what());
+                      cur_pass->name(), exc.what());
         opt.graph_opt_level = orig_setting;
         throw;
     })
-    MGB_FINALLY(
-        opt.graph_opt_level = orig_setting
-    );
+    MGB_FINALLY(opt.graph_opt_level = orig_setting);
     if (verbosity() >= 1) {
-        mgb_log_debug("graph optimization: applied %zu passes, "
+        mgb_log_debug(
+                "graph optimization: applied %zu passes, "
                 "total %zu var(s) replaced; time=%.2fms",
                 m_passes.size(), tot_nr_replace, timer.get_msecs());
     }
     return state.graph();
 }
 
-const GraphOptimizer& GraphOptimizer::apply_inplace(VarNodeArray &vars) const {
+const GraphOptimizer& GraphOptimizer::apply_inplace(VarNodeArray& vars) const {
     if (m_passes.empty()) {
         // this check is necessary, since OptState would clear
         // var_replace_map()
@@ -613,7 +607,7 @@ const GraphOptimizer& GraphOptimizer::apply_inplace(VarNodeArray &vars) const {
     }
 
     auto g = apply({{vars.begin(), vars.end()}});
-    for (size_t i = 0; i < vars.size(); ++ i) {
+    for (size_t i = 0; i < vars.size(); ++i) {
         vars[i] = g.endpoint_vars()[i].node();
     }
     return *this;
@@ -653,7 +647,7 @@ GraphOptimizer& GraphOptimizer::add_preset_passes(
 #if MGB_JIT
     bool need_jit = false;
     if (comp_graph_opt && (std::abs(comp_graph_opt->graph_opt_level) >= 3 ||
-            comp_graph_opt->graph_opt.jit)) {
+                           comp_graph_opt->graph_opt.jit)) {
         need_jit = true;
     }
     if (need_jit && after_grad) {
@@ -679,7 +673,6 @@ GraphOptimizer& GraphOptimizer::add_preset_passes(
         add_passes_for_optimize_options(*inference_opt);
     }
 
-
     if (inference_opt) {
         // merge params to reduce loading time and graph overhead
         add_pass<ParamMergePass>();
@@ -689,15 +682,16 @@ GraphOptimizer& GraphOptimizer::add_preset_passes(
 }
 
 const ThinHashMap<VarNode*, VarNode*>& GraphOptimizer::var_replace_map(
-        ComputingGraph &graph) {
-    auto storage = graph.options().user_data.get_user_data_or_create<
-        VarReplaceMapStorage>();
+        ComputingGraph& graph) {
+    auto storage =
+            graph.options()
+                    .user_data.get_user_data_or_create<VarReplaceMapStorage>();
     return storage->map;
 }
 
-VarNode* GraphOptimizer::var_replace_lookup(VarNode *var) {
-    auto &&map = var_replace_map(*(var->owner_graph()));
-    for (; ; ) {
+VarNode* GraphOptimizer::var_replace_lookup(VarNode* var) {
+    auto&& map = var_replace_map(*(var->owner_graph()));
+    for (;;) {
         auto iter = map.find(var);
         if (iter == map.end())
             return var;
@@ -705,7 +699,6 @@ VarNode* GraphOptimizer::var_replace_lookup(VarNode *var) {
     }
 }
 
-
 const GraphOptimizer& GraphOptimizer::add_passes_for_optimize_options(
         const cg::GraphCommonOptimizeOptions& options) {
     return add_passes_for_optimize_options(
@@ -723,12 +716,14 @@ const GraphOptimizer& GraphOptimizer::add_passes_for_optimize_options(
             options.disable_##_option(); \
         }                                \
     }
-    
-    cb(fuse_preprocess, {add_pass(FuseNCHW4Int8Preprocess::make());});
+
+    cb(fuse_preprocess, {
+        add_pass(FuseNCHW4Int8Preprocess::make());
+        add_pass<FuseWarpPerspectiveDimshufflePass>();
+    });
     cb(f16_io_comp, { add_pass(ConvertF32ToF16Pass::make(false)); });
     cb(f16_io_f32_comp, { add_pass(ConvertF32ToF16Pass::make(true)); });
 
-
     cb(nchw4, {
         add_pass<FuseConvBiasNonlinPass>();
         add_pass<FuseConvBiasZPass>();
@@ -763,6 +758,7 @@ const GraphOptimizer& GraphOptimizer::add_passes_for_optimize_options(
         add_pass<ShuffleShuffleRemovePass>();
         add_pass<RemoveRedundantTypeCvtPass>();
         add_pass(FuseNCHW4Int8Preprocess::make());
+        add_pass<FuseWarpPerspectiveDimshufflePass>();
     });
     cb(chwn4, {
         add_pass<FuseConvBiasNonlinPass>();
@@ -790,9 +786,9 @@ const GraphOptimizer& GraphOptimizer::add_passes_for_optimize_options(
 /* ================ ConstVarPropogateBase ================ */
 
 ConstVarPropogate::AddOprResult ConstVarPropogate::add_opr(
-        OperatorNodeBase *opr) {
+        OperatorNodeBase* opr) {
     using ProfFlag = OperatorNodeBase::NodeProp::Flag;
-    auto &&info = m_oprinfo[opr];
+    auto&& info = m_oprinfo[opr];
     if (info.processed)
         return info.result;
     info.processed = true;
@@ -819,15 +815,14 @@ ConstVarPropogate::AddOprResult ConstVarPropogate::add_opr(
     if (opr->input().empty())
         return make_ret();
 
-    if (opr->node_prop().contain(
-                ProfFlag::FORCE_UPDATE_INPUT_VAR |
-                ProfFlag::IMPURE_FUNC)) {
+    if (opr->node_prop().contain(ProfFlag::FORCE_UPDATE_INPUT_VAR |
+                                 ProfFlag::IMPURE_FUNC)) {
         return make_ret();
     }
 
     size_t max_input_size = 0;
     ret.all_const_inp = true;
-    for (auto i: opr->input()) {
+    for (auto i : opr->input()) {
         auto io = i->owner_opr();
         auto iter = m_oprinfo.find(io);
         if (iter == m_oprinfo.end()) {
@@ -835,7 +830,7 @@ ConstVarPropogate::AddOprResult ConstVarPropogate::add_opr(
             iter = m_oprinfo.find(io);
             mgb_assert(iter != m_oprinfo.end());
         }
-        auto &&src = iter->second;
+        auto&& src = iter->second;
         if (src.is_const) {
             update_max(max_input_size, src.max_size);
             ret.has_const_inp = true;

src/gopt/impl/fuse_nchw4_int8_preprocess.cpp

@@ -19,6 +19,7 @@
 #include "megbrain/opr/utility.h"
 #include "megbrain/serialization/opr_shallow_copy.h"
 #include "megbrain/serialization/serializer.h"
+#include "megbrain/opr/imgproc.h"
 
 using namespace mgb;
 using namespace gopt;
@@ -443,4 +444,244 @@ void FuseNCHW4Int8Preprocess::apply(OptState& state) const {
     };
     state.graph().iter(on_opr);
     rewriter.apply_inplace();
+}
+
+/* ==================== FuseWarpPerspectiveDimshufflePass ================= */
+const char* FuseWarpPerspectiveDimshufflePass::name() const {
+    return mgb_cstr_log("Fuse warp perspective dimshuffle pass");
+}
+
+void FuseWarpPerspectiveDimshufflePass::apply(OptState& opt) const {
+    auto rewriter = opt.graph().make_rewriter();
+    auto uniq_reader_check = UniqReaderCheck{opt.graph()};
+
+    auto make_new_warp = [&rewriter](opr::WarpPerspective* warp,
+                                     opr::WarpPerspective::Param new_param,
+                                     megdnn::DType dst_dtype,
+                                     SymbolVar& new_warp) {
+        OperatorNodeConfig new_config(dst_dtype);
+        if (warp->input().size() == 3) {
+            auto src = rewriter.get_var(warp->input(0)),
+                 mat = rewriter.get_var(warp->input(1)),
+                 out_shape = rewriter.get_var(warp->input(2));
+            new_warp = opr::WarpPerspective::make(src, mat, out_shape,
+                                                  new_param, new_config);
+        } else {
+            mgb_assert(warp->input().size() == 4);
+            auto src = rewriter.get_var(warp->input(0)),
+                 mat = rewriter.get_var(warp->input(1)),
+                 mat_idx = rewriter.get_var(warp->input(2)),
+                 out_shape = rewriter.get_var(warp->input(3));
+            new_warp = opr::WarpPerspective::make(src, mat, mat_idx, out_shape,
+                                                  new_param, new_config);
+        }
+    };
+
+    auto is_warp_nchw = [&uniq_reader_check](OperatorNodeBase* bottom_opr,
+                                             OperatorNodeBase*& top_opr) {
+        // check warp
+        auto warp = try_cast_as_op<opr::WarpPerspective>(bottom_opr);
+        if (warp == nullptr)
+            return false;
+        auto inp_dtype = warp->input(0)->dtype();
+        bool is_u8_or_qu8 = inp_dtype.enumv() == DTypeEnum::Quantized8Asymm ||
+                            inp_dtype.enumv() == DTypeEnum::Uint8;
+
+        bool is_nchw = warp->param().format ==
+                       megdnn::param::WarpPerspective::Format::NCHW;
+        if (!(is_u8_or_qu8 && is_nchw))
+            return false;
+        if (!uniq_reader_check(warp->input(0)))
+            return false;
+
+        top_opr = warp;
+        return true;
+    };
+
+    auto is_warp_nhwc2nchw = [&uniq_reader_check](OperatorNodeBase* bottom_opr,
+                                                  OperatorNodeBase*& top_opr) {
+        // check shuffle
+        auto shuffle = try_cast_as_op<opr::Dimshuffle>(bottom_opr);
+        if (shuffle == nullptr)
+            return false;
+        auto&& shuffle_param = shuffle->param();
+        if (shuffle_param.pattern_len != 4)
+            return false;
+        bool is_nhwc2nchw = shuffle_param.pattern[0] == 0 &&
+                            shuffle_param.pattern[1] == 3 &&
+                            shuffle_param.pattern[2] == 1 &&
+                            shuffle_param.pattern[3] == 2;
+        if (!is_nhwc2nchw)
+            return false;
+        if (!uniq_reader_check(shuffle->input(0)))
+            return false;
+
+        // check warp
+        auto warp = try_cast_as_op<opr::WarpPerspective>(
+                shuffle->input(0)->owner_opr());
+        if (warp == nullptr)
+            return false;
+        auto inp_dtype = warp->input(0)->dtype();
+        bool is_u8_or_qu8 = inp_dtype.enumv() == DTypeEnum::Quantized8Asymm ||
+                            inp_dtype.enumv() == DTypeEnum::Uint8;
+        bool is_nhwc = warp->param().format ==
+                       megdnn::param::WarpPerspective::Format::NHWC;
+        if (!(is_u8_or_qu8 && is_nhwc))
+            return false;
+
+        top_opr = warp;
+        return true;
+    };
+
+    auto try_warp_nchw_typecvt = [&rewriter, &uniq_reader_check, &is_warp_nchw,
+                                  &make_new_warp](OperatorNodeBase* opr) {
+        // check typecvt
+        auto typecvt = try_cast_as_op<opr::TypeCvt>(opr);
+        if (typecvt == nullptr)
+            return false;
+        bool is_to_f32 =
+                typecvt->output(0)->dtype().enumv() == DTypeEnum::Float32;
+        if (!is_to_f32)
+            return false;
+        if (!uniq_reader_check(typecvt->input(0)))
+            return false;
+
+        OperatorNodeBase* top_opr = nullptr;
+        if (!is_warp_nchw(typecvt->input(0)->owner_opr(), top_opr))
+            return false;
+        auto warp = try_cast_as_op<opr::WarpPerspective>(top_opr);
+        SymbolVar new_warp;
+        make_new_warp(warp, warp->param(), opr->output()[0]->dtype(), new_warp);
+
+        rewriter.replace_var(opr->output(0), new_warp.node(),
+                             mgb_cstr_log("replace warp + typecvt"
+                                          "fuse warp_dimshuffle(NCHW)"));
+
+        return true;
+    };
+
+    auto try_warp_nhwc2nchw_typecvt = [&rewriter, &uniq_reader_check,
+                                       &is_warp_nhwc2nchw,
+                                       &make_new_warp](OperatorNodeBase* opr) {
+        // check typecvt
+        auto typecvt = try_cast_as_op<opr::TypeCvt>(opr);
+        if (typecvt == nullptr)
+            return false;
+        bool is_to_f32 =
+                typecvt->output(0)->dtype().enumv() == DTypeEnum::Float32;
+        if (!is_to_f32)
+            return false;
+        if (!uniq_reader_check(typecvt->input(0)))
+            return false;
+
+        OperatorNodeBase* top_opr = nullptr;
+        if (!is_warp_nhwc2nchw(typecvt->input(0)->owner_opr(), top_opr))
+            return false;
+        auto warp = try_cast_as_op<opr::WarpPerspective>(top_opr);
+        opr::WarpPerspective::Param new_param = warp->param();
+        new_param.format = megdnn::param::WarpPerspective::Format::NHWC_NCHW;
+        SymbolVar new_warp;
+        make_new_warp(warp, new_param, opr->output()[0]->dtype(), new_warp);
+
+        rewriter.replace_var(
+                opr->output(0), new_warp.node(),
+                mgb_cstr_log("replace conv_bias + dimshuffle + "
+                             "typecvt to warp_dimshuffle(NHWC_NCHW)"));
+
+        return true;
+    };
+
+    auto try_warp_nhwc2nchw4_typecvt = [&rewriter, &uniq_reader_check,
+                                        &is_warp_nhwc2nchw,
+                                        &make_new_warp](OperatorNodeBase* opr) {
+        // check relayout
+        auto relayout = try_cast_as_op<opr::RelayoutFormat>(opr);
+        if (relayout == nullptr)
+            return false;
+        bool is_to_q8 =
+                relayout->output(0)->dtype().enumv() == DTypeEnum::QuantizedS8;
+        bool is_to_nchw2nchw4 = relayout->param().mode ==
+                                opr::RelayoutFormat::Param::Mode::NCHW_NCHW4;
+        if (!(is_to_q8 && is_to_nchw2nchw4))
+            return false;
+        if (!uniq_reader_check(relayout->input(0)))
+            return false;
+
+        OperatorNodeBase* top_opr = nullptr;
+        if (!is_warp_nhwc2nchw(relayout->input(0)->owner_opr(), top_opr))
+            return false;
+
+        auto warp = try_cast_as_op<opr::WarpPerspective>(top_opr);
+
+        bool is_small_chn = warp->input(0)->shape()[3] < 4;
+        if (!is_small_chn)
+            return false;
+
+        opr::WarpPerspective::Param new_param = warp->param();
+        new_param.format =
+                megdnn::param::WarpPerspective::Format::NHWC_NCHW4_IC_SMALL;
+
+        SymbolVar new_warp;
+        make_new_warp(warp, new_param, opr->output()[0]->dtype(), new_warp);
+
+        rewriter.replace_var(
+                opr->output(0), new_warp.node(),
+                mgb_cstr_log("replace warp + dimshuffle + relayout(NCHW_NCHW4)"
+                             "to warp_dimshuffle(NHWC_NCHW4_IC_SMALL)"));
+
+        return true;
+    };
+
+    auto try_warp_nchw2nchw4_typecvt = [&rewriter, &uniq_reader_check,
+                                        &is_warp_nchw,
+                                        &make_new_warp](OperatorNodeBase* opr) {
+        // check relayout
+        auto relayout = try_cast_as_op<opr::RelayoutFormat>(opr);
+        if (relayout == nullptr)
+            return false;
+        bool is_to_q8 =
+                relayout->output(0)->dtype().enumv() == DTypeEnum::QuantizedS8;
+        bool is_to_nchw2nchw4 = relayout->param().mode ==
+                                opr::RelayoutFormat::Param::Mode::NCHW_NCHW4;
+        if (!(is_to_q8 && is_to_nchw2nchw4))
+            return false;
+        if (!uniq_reader_check(relayout->input(0)))
+            return false;
+
+        OperatorNodeBase* top_opr = nullptr;
+        if (!is_warp_nchw(relayout->input(0)->owner_opr(), top_opr))
+            return false;
+
+        auto warp = try_cast_as_op<opr::WarpPerspective>(top_opr);
+
+        bool is_small_chn = warp->input(0)->shape()[1] < 4;
+        if (!is_small_chn)
+            return false;
+
+        opr::WarpPerspective::Param new_param = warp->param();
+        new_param.format =
+                megdnn::param::WarpPerspective::Format::NCHW_NCHW4_IC_SMALL;
+
+        SymbolVar new_warp;
+        make_new_warp(warp, new_param, opr->output()[0]->dtype(), new_warp);
+
+        rewriter.replace_var(
+                opr->output(0), new_warp.node(),
+                mgb_cstr_log("replace warp + relayout(NCHW_NCHW4)"
+                             "to warp_dimshuffle(NCHW_NCHW4_IC_SMALL)"));
+
+        return true;
+    };
+
+    auto on_opr = [&try_warp_nchw_typecvt, &try_warp_nhwc2nchw_typecvt,
+                   &try_warp_nhwc2nchw4_typecvt, &try_warp_nchw2nchw4_typecvt,
+                   &rewriter](OperatorNodeBase* opr) {
+        if (!try_warp_nchw_typecvt(opr) && !try_warp_nhwc2nchw_typecvt(opr) &&
+            !try_warp_nhwc2nchw4_typecvt(opr) &&
+            !try_warp_nchw2nchw4_typecvt(opr)) {
+            rewriter.auto_replace_outputs(opr);
+        }
+    };
+    opt.graph().iter(on_opr);
+    rewriter.apply_inplace();
 }

src/gopt/include/megbrain/gopt/inference.h

@@ -173,6 +173,16 @@ namespace gopt {
     };
 
     /*!
+     * \brief fuse warp perspective and dimshuffle, quint8/uint8 to qint8/float
+     */
+    class FuseWarpPerspectiveDimshufflePass : public Pass {
+        public:
+            const char* name() const override;
+            void apply(OptState& opt) const override;
+    };
+
+
+    /*!
      * \brief fuse deconv and typecvt to a deconv opr
      */
     class FuseDeconvCvtPass : public Pass {

src/gopt/test/inference.cpp

@@ -1172,7 +1172,8 @@ TEST(TestGoptInference, ConvertFormatNHWCD4) {
     param.pad_h = param.pad_w = 1;
     auto w2 = mkcvar("w2", {4, 4, 3, 3}),
          y = opr::Convolution::make(elem, w2, param),
-         z = opr::AxisAddRemove::make(y, {opr::AxisAddRemove::AxisDesc::make_add(0)});
+         z = opr::AxisAddRemove::make(
+                 y, {opr::AxisAddRemove::AxisDesc::make_add(0)});
 
     SymbolVar y_opt, z_opt;
     auto options = gopt::OptimizeForInferenceOptions{};
@@ -3722,5 +3723,65 @@ TEST(TestGoptInference, PreProcessCase1) {
 
     ASSERT_TRUE(y_opt.node()->owner_opr()->same_type<opr::RelayoutFormat>());
 }
+
+TEST(TestGoptInference, WarpAndPreProcessCase) {
+    REQUIRE_GPU(1);
+    HostTensorGenerator<dtype::Uint8, RandomDistribution::UNIFORM> gen(0, 255);
+    auto cn = CompNode::load("gpu0");
+    auto graph = ComputingGraph::make();
+    graph->options().graph_opt_level = 0;
+
+    size_t n = 1;
+    size_t c = 3;
+    size_t h = 16;
+    size_t w = 16;
+    auto host_x1 = gen({n, h, w, c}, cn);
+    auto x = opr::Host2DeviceCopy::make(*graph, host_x1);
+
+    auto mat_host = std::make_shared<HostTensorND>(cn, TensorShape{n, 3, 3},
+                                                   dtype::Float32());
+    warp_perspective_mat_gen(*mat_host, n, h, w);
+    auto mat = opr::Host2DeviceCopy::make(*graph, mat_host).rename("mat");
+
+    opr::WarpPerspective::Param warp_param;
+    warp_param.format = opr::WarpPerspective::Param::Format::NHWC;
+    auto x_warp =
+            opr::WarpPerspective::make(x, mat, TensorShape{h, w}, warp_param);
+    auto x_nchw = opr::Dimshuffle::make(x_warp, {0, 3, 1, 2}, 4, cn);
+
+    auto x_u8 = opr::TypeCvt::make(x_nchw, dtype::Float32(), cn);
+    auto x_s8 = x_u8 - 128;
+    auto zero = DTypeScalar(dtype::Float32());
+    auto zero_tensor = opr::ImmutableTensor::make(*graph, zero, cn);
+    auto pad_channel_tensor =
+            opr::Broadcast::make(zero_tensor, {n, 1, h, w}, cn);
+    auto paded_x = opr::Concat::make({x_s8, pad_channel_tensor}, 1, cn)
+                           .reshape({n, 1, 4, h, w});
+
+    auto nchw4_out = opr::Dimshuffle::make(paded_x, {0, 1, 3, 4, 2}, 5, cn);
+    auto result = opr::TypeCvt::make(nchw4_out, dtype::QuantizedS8(1.f));
+
+    auto y = result;
+    SymbolVar y_opt;
+    auto options = gopt::OptimizeForInferenceOptions{};
+    options.enable_fuse_preprocess();
+    unpack_vector(gopt::optimize_for_inference({y}, options), y_opt);
+
+    ASSERT_TRUE(y_opt.node()->owner_opr()->same_type<opr::WarpPerspective>());
+
+    ASSERT_EQ(opr::WarpPerspective::Param::Format::NHWC_NCHW4_IC_SMALL,
+              find_opr<opr::WarpPerspective>(y_opt).param().format);
+
+    graph->compile({{y_opt, {}}})
+            ->to_json()
+            ->writeto_fpath(output_file(
+                    "TestGoptInference.WarpAndPreProcessCase.json"));
+
+    HostTensorND host_y_opt, host_y;
+    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-5);
+}
 #endif
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/imgproc.cpp

@@ -47,7 +47,11 @@ SymbolVar WarpPerspectiveForward::make(SymbolVar i0, SymbolVar i1, SymbolVar i2,
 }
 
 void WarpPerspectiveForward::init_output_dtype() {
-    output(0)->dtype(input(0)->dtype());
+    if (config().output_dtype().valid()) {
+        output(0)->dtype(config().output_dtype());
+    } else {
+        output(0)->dtype(input(0)->dtype());
+    }
 }
 
 void WarpPerspectiveForward::add_input_layout_constraint() {
@@ -78,23 +82,40 @@ void WarpPerspectiveForward::outshape_by_symvar_do_get_output_shape(
                    mat_idx_shp.to_string().c_str());
     }
 
-    //! The index of height, e.g.,[b, h, w, c], the height_idx = 1
-    size_t height_idx = 0;
-    if (param().format == Param::Format::NCHW ||
-        param().format == Param::Format::NCHW4) {
-        height_idx = 2;
-    } else {
-        height_idx = 1;
-    }
-
-    dest = imgshp;
-    dest[0] = matshp[0];
-    if (param().format == Param::Format::NHWCD4) {
-        dest.shape[height_idx] = oshp2d.shape[0];
-        dest.shape[height_idx + 2] = oshp2d.shape[1];
-    } else {
-        for (int i = 0; i < 2; ++i)
-            dest.shape[height_idx + i] = oshp2d.shape[i];
+    switch (param().format) {
+        case Param::Format::NCHW_NCHW4_IC_SMALL:
+        case Param::Format::NHWC_NCHW4_IC_SMALL:
+            dest.ndim = 5;
+            dest[0] = matshp[0];
+            dest.shape[1] = 1;
+            dest.shape[2] = oshp2d.shape[0];
+            dest.shape[3] = oshp2d.shape[1];
+            dest.shape[4] = 4;
+            break;
+        case Param::Format::NHWC_NCHW:
+            dest[0] = matshp[0];
+            dest.shape[1] = imgshp.shape[3];
+            dest.shape[2] = oshp2d.shape[0];
+            dest.shape[3] = oshp2d.shape[1];
+            break;
+        default:
+            size_t height_idx = 0;
+            if (param().format == Param::Format::NCHW ||
+                param().format == Param::Format::NCHW4) {
+                height_idx = 2;
+            } else {
+                height_idx = 1;
+            }
+            dest = imgshp;
+            dest[0] = matshp[0];
+            if (param().format == Param::Format::NHWCD4) {
+                dest.shape[height_idx] = oshp2d.shape[0];
+                dest.shape[height_idx + 2] = oshp2d.shape[1];
+            } else {
+                for (int i = 0; i < 2; ++i)
+                    dest.shape[height_idx + i] = oshp2d.shape[i];
+            }
+            break;
     }
 }