fix(dnn/fallback): delete ConvBias* opr param of conv_bias algo

GitOrigin-RevId: ee5a6874fb3698b0c79698c1bb4f90997be715a1

dnn/src/aarch64/conv_bias/fp16/algos.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/aarch64/conv_bias/fp16/algos.h"
@@ -22,7 +23,7 @@ using namespace aarch64;
 MIDOUT_DECL(megdnn_aarch64_conv_bias_stride2_conv2357_fp16)
 
 bool ConvBiasImpl::AlgoF16DirectStride2::usable(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_aarch64_conv_bias_stride2_conv2357_fp16, 0, 0) {
         auto&& fm = param.filter_meta;
@@ -47,7 +48,7 @@ bool ConvBiasImpl::AlgoF16DirectStride2::usable(
 }
 
 size_t ConvBiasImpl::AlgoF16DirectStride2::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_aarch64_conv_bias_stride2_conv2357_fp16, 0, 1) {
         auto wbundle = arm_common::MultithreadDirectConvCommon<
                 dt_float16, __fp16>::get_bundle_stride(param, m_large_group);
@@ -59,7 +60,7 @@ size_t ConvBiasImpl::AlgoF16DirectStride2::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF16DirectStride2::dispatch_kerns(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_aarch64_conv_bias_stride2_conv2357_fp32, 0, 2) {
         return get_kimpls(param);
     }

dnn/src/aarch64/conv_bias/fp16/algos.h

@@ -19,6 +19,7 @@ namespace aarch64 {
 class ConvBiasImpl::AlgoF16DirectStride2 final : public AlgoBase {
     SmallVector<NCBKern> get_kimpls(const NCBKernSizeParam& param) const;
     bool m_large_group;
+
 public:
     AlgoF16DirectStride2(bool large_group) : m_large_group(large_group) {}
     bool is_reproducible() const override { return true; }
@@ -26,15 +27,12 @@ public:
         return m_large_group ? "ARMV8F16STRD2_LARGE_GROUP"
                              : "ARMV8F16STRD2_SMALL_GROUP";
     }
-
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
 
-    SmallVector<NCBKern> dispatch_kerns(FallbackConvBiasImpl*,
-                                        const NCBKernSizeParam&) const override;
+    SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam&) const override;
 };
 }  // namespace aarch64
 }  // namespace megdnn

dnn/src/aarch64/conv_bias/fp32/algos.cpp

@@ -22,7 +22,7 @@ using namespace aarch64;
 
 MIDOUT_DECL(megdnn_aarch64_conv_bias_stride2_conv2357_fp32)
 bool ConvBiasImpl::AlgoF32DirectStride2::usable(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_aarch64_conv_bias_stride2_conv2357_fp32, 0, 0) {
         auto&& fm = param.filter_meta;
@@ -47,7 +47,7 @@ bool ConvBiasImpl::AlgoF32DirectStride2::usable(
 }
 
 size_t ConvBiasImpl::AlgoF32DirectStride2::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_aarch64_conv_bias_stride2_conv2357_fp32, 0, 1) {
         auto wbundle = arm_common::MultithreadDirectConvCommon<
                 float, float>::get_bundle_stride(param, m_large_group);
@@ -58,7 +58,7 @@ size_t ConvBiasImpl::AlgoF32DirectStride2::get_workspace(
 }
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF32DirectStride2::dispatch_kerns(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_aarch64_conv_bias_stride2_conv2357_fp32, 0, 2) {
         return get_kimpls(param);
     }

dnn/src/aarch64/conv_bias/fp32/algos.h

@@ -23,6 +23,7 @@ using FallbackConvBiasImpl = fallback::ConvBiasImpl;
 class ConvBiasImpl::AlgoF32DirectStride2 final : public AlgoBase {
     SmallVector<NCBKern> get_kimpls(const NCBKernSizeParam& param) const;
     bool m_large_group;
+
 public:
     AlgoF32DirectStride2(bool large_group) : m_large_group(large_group) {}
     bool is_reproducible() const override { return true; }
@@ -31,14 +32,12 @@ public:
                              : "ARMV8F32STRD2_SMALL_GROUP";
     }
 
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
 
-    SmallVector<NCBKern> dispatch_kerns(FallbackConvBiasImpl*,
-                                        const NCBKernSizeParam&) const override;
+    SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam&) const override;
 };
 
 }  // namespace aarch64

dnn/src/aarch64/conv_bias/int8/algos.cpp

@@ -30,9 +30,8 @@ using megdnn::arm_common::TypeCvtOp;
 /* ===================== matrix mul algo ===================== */
 
 bool ConvBiasImpl::AlgoS8MatrixMul::usable(
-        FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
-    MEGDNN_MARK_USED_VAR(opr);
     auto&& fm = param.filter_meta;
     return param.src_type.enumv() == DTypeEnum::QuantizedS8 &&
            param.dst_type.enumv() == DTypeEnum::QuantizedS8 &&

dnn/src/aarch64/conv_bias/int8/algos.h

@@ -13,6 +13,7 @@
 
 #include "src/aarch64/conv_bias/opr_impl.h"
 #include "src/fallback/conv_bias/opr_impl.h"
+#include "src/common/opr_delegate.h"
 
 namespace megdnn {
 namespace aarch64 {
@@ -27,21 +28,21 @@ public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "S8MATMUL"; }
 
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override {
+    size_t get_workspace(const NCBKernSizeParam& param) const override {
         return get_bundle(param).total_size_in_bytes();
     }
     SmallVector<NCBKern> dispatch_kerns(
-            FallbackConvBiasImpl*, const NCBKernSizeParam& param) const override {
+            const NCBKernSizeParam& param) const override {
         size_t group = param.filter_meta.group;
         return {{kimpl, {group, 1_z, 1_z}}};
     }
     //! select matmul to the highest preference
-    bool is_preferred(FallbackConvBiasImpl* opr,
-                      const NCBKernSizeParam& param) const override {
-        return static_cast<arm_common::ConvBiasImpl*>(opr)
+    bool is_preferred(const NCBKernSizeParam& param) const override {
+        static CpuOprDelegationStorage<1> storage;
+        auto conv_bias_opr = storage.get<ConvBias, 0>();
+        return static_cast<ConvBiasImpl*>(conv_bias_opr)
                 ->is_matmul_quantized_prefer(param);
     }
 };

dnn/src/aarch64/conv_bias/quint8/algos.cpp

@@ -32,9 +32,8 @@ using megdnn::arm_common::TypeCvtOp;
 /* ===================== matrix mul algo ===================== */
 
 bool ConvBiasImpl::AlgoQU8MatrixMul::usable(
-        FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
-    MEGDNN_MARK_USED_VAR(opr);
     auto&& fm = param.filter_meta;
     return param.src_type.enumv() == DTypeEnum::Quantized8Asymm &&
            param.dst_type.enumv() == DTypeEnum::Quantized8Asymm &&

dnn/src/aarch64/conv_bias/quint8/algos.h

@@ -13,6 +13,7 @@
 
 #include "src/aarch64/conv_bias/opr_impl.h"
 #include "src/fallback/conv_bias/opr_impl.h"
+#include "src/common/opr_delegate.h"
 
 namespace megdnn {
 namespace aarch64 {
@@ -27,22 +28,21 @@ public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "QU8MATMUL"; }
 
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override {
+    size_t get_workspace(const NCBKernSizeParam& param) const override {
         return get_bundle(param).total_size_in_bytes();
     }
     SmallVector<NCBKern> dispatch_kerns(
-            FallbackConvBiasImpl*,
             const NCBKernSizeParam& param) const override {
         size_t group = param.filter_meta.group;
         return {{kimpl, {group, 1_z, 1_z}}};
     }
     //! select matmul to the highest preference
-    bool is_preferred(FallbackConvBiasImpl* opr,
-                      const NCBKernSizeParam& param) const override {
-        return static_cast<arm_common::ConvBiasImpl*>(opr)
+    bool is_preferred(const NCBKernSizeParam& param) const override {
+        static CpuOprDelegationStorage<1> storage;
+        auto conv_bias_opr = storage.get<ConvBias, 0>();
+        return static_cast<ConvBiasImpl*>(conv_bias_opr)
                 ->is_matmul_quantized_prefer(param);
     }
 };

dnn/src/arm_common/conv_bias/f16/algos.cpp

@@ -27,10 +27,9 @@ using namespace arm_common;
 /* ======================= AlgoFP16WinogradF23 ======================== */
 
 bool ConvBiasImpl::AlgoFP16WinogradF23::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp16, 0, 0) {
         using Strategy = winograd::winograd_2x3_4x4_f16;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -38,13 +37,13 @@ bool ConvBiasImpl::AlgoFP16WinogradF23::usable(
                                       strategy, m_tile_size, param)
                                       .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -69,10 +68,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP16WinogradF23,
 /* ======================= AlgoFP16WinogradF45 ======================== */
 
 bool ConvBiasImpl::AlgoFP16WinogradF45::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp16, 1, 0) {
         using Strategy = winograd::winograd_4x5_1x1_f16;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -80,13 +78,13 @@ bool ConvBiasImpl::AlgoFP16WinogradF45::usable(
                                       strategy, m_tile_size, param)
                                       .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 4 &&
+                 param.output_block_size == 4 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 5) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -109,10 +107,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP16WinogradF45,
 /* ======================= AlgoFP16WinogradF63 ======================== */
 
 bool ConvBiasImpl::AlgoFP16WinogradF63::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp16, 2, 0) {
         using Strategy = winograd::winograd_6x3_1x1_f16;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -120,13 +117,13 @@ bool ConvBiasImpl::AlgoFP16WinogradF63::usable(
                                       strategy, m_tile_size, param)
                                       .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 6 &&
+                 param.output_block_size == 6 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -149,10 +146,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP16WinogradF63,
 /* ======================= AlgoFP16WinogradF23_8x8 ======================== */
 
 bool ConvBiasImpl::AlgoFP16WinogradF23_8x8::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp16, 3, 0) {
         if (param.filter_meta.icpg % 8 != 0 || param.filter_meta.ocpg % 8 != 0)
             return false;
@@ -166,13 +162,13 @@ bool ConvBiasImpl::AlgoFP16WinogradF23_8x8::usable(
                         .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
                m_matmul_algo->packmode() == PackMode::NO_PACK &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK8)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -197,7 +193,7 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP16WinogradF23_8x8,
 MIDOUT_DECL(megdnn_arm_common_conv_bias_fp16_kimpl)
 
 bool ConvBiasImpl::AlgoF16Direct::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_fp16_kimpl, 0, 0) {
         auto&& fm = param.filter_meta;
@@ -227,7 +223,7 @@ bool ConvBiasImpl::AlgoF16Direct::usable(
 }
 
 size_t ConvBiasImpl::AlgoF16Direct::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_fp16_kimpl, 0, 1) {
         auto wbundle =
                 MultithreadDirectConvCommon<dt_float16, __fp16>::get_bundle(
@@ -310,7 +306,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoF16Direct::get_kimpls(
 }
 
 SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoF16Direct::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_fp16_kimpl, 0, 1) {
         return get_kimpls(param);
     }
@@ -321,7 +317,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoF16Direct::dispatch_kerns(
 /* ===================== stride-1 algo ===================== */
 
 bool ConvBiasImpl::AlgoF16DirectStride1::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_fp16_kimpl, 1, 0) {
         auto&& fm = param.filter_meta;
@@ -425,7 +421,7 @@ ConvBiasImpl::AlgoF16DirectStride1::get_kimpls(
 }
 
 size_t ConvBiasImpl::AlgoF16DirectStride1::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_fp16_kimpl, 1, 1) {
         auto bundle = MultithreadDirectConvCommon<
                 dt_float16, __fp16>::get_bundle_stride(param, m_large_group);
@@ -437,7 +433,7 @@ size_t ConvBiasImpl::AlgoF16DirectStride1::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF16DirectStride1::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_fp16_kimpl, 1, 2) {
         return get_kimpls(param);
     }

dnn/src/arm_common/conv_bias/f16/algos.h

@@ -88,14 +88,12 @@ public:
         return m_large_group ? "F16DIRECT_LARGE_GROUP"
                              : "F16DIRECT_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
 
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -109,12 +107,10 @@ public:
     const char* name() const override {
         return m_large_group ? "F16STRD1_LARGE_GROUP" : "F16STRD1_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 

dnn/src/arm_common/conv_bias/fp32/algos.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/arm_common/conv_bias/fp32/algos.h"
@@ -30,9 +31,8 @@ using namespace arm_common;
 /* ======================= AlgoFP32WinogradF23_4x4 ======================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF23_4x4::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
-    MEGDNN_MARK_USED_VAR(opr);
     MEGDNN_MARK_USED_VAR(param);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp32, 0, 0) {
         if (param.filter_meta.icpg % 4 != 0 || param.filter_meta.ocpg % 4 != 0)
@@ -47,13 +47,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF23_4x4::usable(
                         .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
                m_matmul_algo->packmode() == PackMode::NO_PACK &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK4)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -76,10 +76,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF23_4x4,
 /* ======================= AlgoFP32WinogradF63 ======================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF63::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp32, 1, 0) {
         using Strategy = winograd::winograd_6x3_1x1_f;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -87,13 +86,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF63::usable(
                                       strategy, m_tile_size, param)
                                       .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 6 &&
+                 param.output_block_size == 6 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -116,10 +115,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF63,
 /* ======================= AlgoFP32WinogradF54 ======================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF54::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp32, 2, 0) {
         using Strategy = winograd::winograd_5x4_1x1_f;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -127,13 +125,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF54::usable(
                                       strategy, m_tile_size, param)
                                       .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 5 &&
+                 param.output_block_size == 5 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 4) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -156,10 +154,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF54,
 /* ======================= AlgoFP32WinogradF45 ======================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF45::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp32, 3, 0) {
         using Strategy = winograd::winograd_4x5_1x1_f;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -167,13 +164,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF45::usable(
                                       strategy, m_tile_size, param)
                                       .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 4 &&
+                 param.output_block_size == 4 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 5) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -196,10 +193,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF45,
 /* ======================= AlgoFP32WinogradF63_4x4 ======================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF63_4x4::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp32, 4, 0) {
         if (param.filter_meta.icpg % 4 != 0 || param.filter_meta.ocpg % 4 != 0)
             return false;
@@ -213,13 +209,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF63_4x4::usable(
                         .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
                m_matmul_algo->packmode() == PackMode::NO_PACK &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 6 &&
+                 param.output_block_size == 6 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK4)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -244,9 +240,8 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF63_4x4,
 /* =================== AlgoFP32WinogradF23_4x4_NCHW44 =================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF23_4x4_NCHW44::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
-    MEGDNN_MARK_USED_VAR(opr);
     MEGDNN_MARK_USED_VAR(param);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp32,
                  midout_iv("AlgoFP32WinogradF23_4x4_NCHW44"_hash)) {
@@ -262,13 +257,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF23_4x4_NCHW44::usable(
         return m_matmul_algo->usable(matmul_param) &&
                m_matmul_algo->packmode() ==
                        fallback::MatrixMulImpl::AlgoBase::PackMode::NO_PACK &&
-               (opr->param().format == param::ConvBias::Format::NCHW44 ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW44 ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW44_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK4)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -291,10 +286,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF23_4x4_NCHW44,
 /* =================== AlgoFP32WinogradF63_4x4_NCHW44 ===================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF63_4x4_NCHW44::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_arm_common_winograd_fp32,
                  midout_iv("AlgoFP32WinogradF63_4x4_NCHW44"_hash)) {
         if (param.filter_meta.icpg % 4 != 0 || param.filter_meta.ocpg % 4 != 0)
@@ -309,13 +303,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF63_4x4_NCHW44::usable(
         return m_matmul_algo->usable(matmul_param) &&
                m_matmul_algo->packmode() ==
                        fallback::MatrixMulImpl::AlgoBase::PackMode::NO_PACK &&
-               (opr->param().format == param::ConvBias::Format::NCHW44 ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW44 ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW44_WINOGRAD &&
-                 opr->param().output_block_size == 6 &&
+                 param.output_block_size == 6 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK4)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -341,7 +335,7 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF63_4x4_NCHW44,
 MIDOUT_DECL(megdnn_arm_common_conv_bias_f32_kimpl);
 
 bool ConvBiasImpl::AlgoF32Direct::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 0, 0) {
         auto&& fm = param.filter_meta;
@@ -370,7 +364,7 @@ bool ConvBiasImpl::AlgoF32Direct::usable(
     return false;
 }
 size_t ConvBiasImpl::AlgoF32Direct::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 0, 1) {
         auto wbundle = MultithreadDirectConvCommon<float, float>::get_bundle(
                 param, m_large_group);
@@ -409,7 +403,8 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoF32Direct::get_kimpls(
             }
             for (size_t ic = 0; ic < IC; ic++) {
                 MultithreadDirectConvCommon<float, float>::copy_padding_kern(
-                        bundle, kern_param, ncb_index, {ncb_index.thread_id, 0, ic});
+                        bundle, kern_param, ncb_index,
+                        {ncb_index.thread_id, 0, ic});
             }
             for (size_t oc = 0; oc < OC; oc++) {
                 MultithreadDirectConvCommon<float, float>::do_conv_kern(
@@ -449,7 +444,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoF32Direct::get_kimpls(
 }
 
 SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoF32Direct::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 0, 1) {
         return get_kimpls(param);
     }
@@ -458,7 +453,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoF32Direct::dispatch_kerns(
 }
 /* ===================== stride-1 algo ===================== */
 bool ConvBiasImpl::AlgoF32DirectStride1::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 1, 1) {
         auto&& fm = param.filter_meta;
@@ -484,7 +479,7 @@ bool ConvBiasImpl::AlgoF32DirectStride1::usable(
 }
 
 size_t ConvBiasImpl::AlgoF32DirectStride1::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 1, 1) {
         auto bundle =
                 MultithreadDirectConvCommon<float, float>::get_bundle_stride(
@@ -575,7 +570,7 @@ ConvBiasImpl::AlgoF32DirectStride1::get_kimpls(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF32DirectStride1::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 1, 2) {
         return get_kimpls(param);
     }
@@ -586,7 +581,7 @@ ConvBiasImpl::AlgoF32DirectStride1::dispatch_kerns(
 /* ===================== stride-2 algo ===================== */
 
 bool ConvBiasImpl::AlgoF32DirectStride2::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 2, 0) {
         auto&& fm = param.filter_meta;
@@ -611,7 +606,7 @@ bool ConvBiasImpl::AlgoF32DirectStride2::usable(
     return false;
 }
 size_t ConvBiasImpl::AlgoF32DirectStride2::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 2, 1) {
         auto bundle =
                 MultithreadDirectConvCommon<float, float>::get_bundle_stride(
@@ -701,7 +696,7 @@ ConvBiasImpl::AlgoF32DirectStride2::get_kimpls(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF32DirectStride2::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_f32_kimpl, 2, 2) {
         return get_kimpls(param);
     }

dnn/src/arm_common/conv_bias/fp32/algos.h

@@ -137,13 +137,11 @@ public:
         return m_large_group ? "F32DIRECT_LARGE_GROUP"
                              : "F32DIRECT_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -157,13 +155,11 @@ public:
     const char* name() const override {
         return m_large_group ? "F32STRD1_LARGE_GROUP" : "F32STRD1_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -177,13 +173,11 @@ public:
     const char* name() const override {
         return m_large_group ? "F32STRD2_LARGE_GROUP" : "F32STRD2_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -194,13 +188,11 @@ public:
     AlgoF32DirectNCHW44() {}
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "F32_CONV_NCHW44_DIRECT"; }
-    bool usable(fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -211,13 +203,11 @@ public:
     AlgoF32DirectNCHWNCHW44() {}
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "F32_CONV_NCHW_NCHW44"; }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -227,13 +217,11 @@ class ConvBiasImpl::AlgoF32ChannelWiseNCHW44 final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "F32_CHANNEL_WISE_NCHW44"; }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 

dnn/src/arm_common/conv_bias/fp32/channel_wise_nchw44_algo.cpp

@@ -10,8 +10,8 @@
  * implied.
  */
 
-#include "src/arm_common/conv_bias/fp32/channel_wise_nchw44_kern.h"
 #include "src/arm_common/conv_bias/fp32/algos.h"
+#include "src/arm_common/conv_bias/fp32/channel_wise_nchw44_kern.h"
 #include "src/arm_common/elemwise_op.h"
 
 #include "midout.h"
@@ -26,8 +26,7 @@ using conv_fun = std::function<void(
 MIDOUT_DECL(conv_bias_fp32_channel_wise_nchw44)
 
 bool ConvBiasImpl::AlgoF32ChannelWiseNCHW44::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
-        AlgoSelectionStrategy) const {
+        const NCBKernSizeParam& param, AlgoSelectionStrategy) const {
     auto&& fm = param.filter_meta;
     auto FH = fm.spatial[0];
     size_t OC = fm.ocpg;
@@ -49,13 +48,13 @@ bool ConvBiasImpl::AlgoF32ChannelWiseNCHW44::usable(
 }
 
 size_t ConvBiasImpl::AlgoF32ChannelWiseNCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam&) const {
+        const NCBKernSizeParam&) const {
     return 0;
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF32ChannelWiseNCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     const constexpr size_t pack_group_size = 4_z;
     auto fm = param.filter_meta;
     const int batch = param.n;

dnn/src/arm_common/conv_bias/fp32/f32_direct_nchw44_algo.cpp

@@ -159,8 +159,7 @@ static void do_conv_kern(const WorkspaceBundle& bundle,
 }  // namespace
 
 /* ===================== stride1 algo ===================== */
-bool ConvBiasImpl::AlgoF32DirectNCHW44::usable(fallback::ConvBiasImpl*,
-                                               const NCBKernSizeParam& param,
+bool ConvBiasImpl::AlgoF32DirectNCHW44::usable(const NCBKernSizeParam& param,
                                                AlgoSelectionStrategy) const {
     auto&& fm = param.filter_meta;
     auto fh = fm.spatial[0];
@@ -182,13 +181,13 @@ bool ConvBiasImpl::AlgoF32DirectNCHW44::usable(fallback::ConvBiasImpl*,
 }
 
 size_t ConvBiasImpl::AlgoF32DirectNCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF32DirectNCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto fm = param.filter_meta;
     const int batch = param.n;
     const int group = fm.group;

dnn/src/arm_common/conv_bias/fp32/f32_direct_nchw_nchw44_algo.cpp

@@ -188,8 +188,7 @@ static void do_conv_kern(const WorkspaceBundle& bundle,
 }  // namespace
 
 bool ConvBiasImpl::AlgoF32DirectNCHWNCHW44::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
-        AlgoSelectionStrategy) const {
+        const NCBKernSizeParam& param, AlgoSelectionStrategy) const {
     auto&& fm = param.filter_meta;
     auto fh = fm.spatial[0];
     int oc = fm.ocpg;
@@ -209,13 +208,13 @@ bool ConvBiasImpl::AlgoF32DirectNCHWNCHW44::usable(
 }
 
 size_t ConvBiasImpl::AlgoF32DirectNCHWNCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoF32DirectNCHWNCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto fm = param.filter_meta;
     const int batch = param.n;
     const int group = fm.group;

dnn/src/arm_common/conv_bias/int8/algos.cpp

@@ -28,7 +28,7 @@ using namespace arm_common;
 MIDOUT_DECL(megdnn_arm_common_conv_bias_int8)
 /* ===================== stride1 algo ===================== */
 bool ConvBiasImpl::AlgoS8DirectStride1::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible = direct_int8_stride1::can_conv_direct_stride1_int8(param);
     auto fm = param.filter_meta;
@@ -40,7 +40,7 @@ bool ConvBiasImpl::AlgoS8DirectStride1::usable(
     return avaible;
 }
 bool ConvBiasImpl::AlgoS8DirectStride1::is_preferred(
-        megdnn::fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     auto&& fm = param.filter_meta;
     auto FH = fm.spatial[0];
     auto OC = fm.ocpg;
@@ -53,14 +53,14 @@ bool ConvBiasImpl::AlgoS8DirectStride1::is_preferred(
 }
 
 size_t ConvBiasImpl::AlgoS8DirectStride1::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     auto bundle = direct_int8_stride1::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoS8DirectStride1::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8, 1, 0) {
         return direct_int8_stride1::get_kimpls(param, m_large_group);
     }
@@ -70,20 +70,20 @@ ConvBiasImpl::AlgoS8DirectStride1::dispatch_kerns(
 
 /* ===================== stride1 algo ===================== */
 bool ConvBiasImpl::AlgoS8ChanWiseStride1NCHW44::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy) const {
     return channel_wise_nchw44::stride1::is_available(param);
 }
 
 size_t ConvBiasImpl::AlgoS8ChanWiseStride1NCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     auto bundle = channel_wise_nchw44::stride1::get_bundle(param);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoS8ChanWiseStride1NCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8,
                  midout_iv("AlgoS8ChanWiseStride1NCHW44"_hash)) {
         return channel_wise_nchw44::stride1::get_kimpls(param);
@@ -94,20 +94,20 @@ ConvBiasImpl::AlgoS8ChanWiseStride1NCHW44::dispatch_kerns(
 
 /* ===================== stride2 algo ===================== */
 bool ConvBiasImpl::AlgoS8ChanWiseStride2NCHW44::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy) const {
     return channel_wise_nchw44::stride2::is_available(param);
 }
 
 size_t ConvBiasImpl::AlgoS8ChanWiseStride2NCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     auto bundle = channel_wise_nchw44::stride2::get_bundle(param);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoS8ChanWiseStride2NCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8,
                  midout_iv("AlgoS8ChanWiseStride2NCHW44"_hash)) {
         return channel_wise_nchw44::stride2::get_kimpls(param);
@@ -118,7 +118,7 @@ ConvBiasImpl::AlgoS8ChanWiseStride2NCHW44::dispatch_kerns(
 
 /* ===================== stride2 algo ===================== */
 bool ConvBiasImpl::AlgoS8DirectStride2::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible = direct_int8_stride2::can_conv_direct_stride2_int8(param);
     if (algo_selection_strategy ==
@@ -130,14 +130,14 @@ bool ConvBiasImpl::AlgoS8DirectStride2::usable(
 }
 
 size_t ConvBiasImpl::AlgoS8DirectStride2::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto bundle = direct_int8_stride2::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoS8DirectStride2::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8, 1, 1) {
         return direct_int8_stride2::get_kimpls(param, m_large_group);
     }
@@ -148,7 +148,7 @@ ConvBiasImpl::AlgoS8DirectStride2::dispatch_kerns(
 #if __ARM_FEATURE_DOTPROD
 /* ===================== dot stride1 algo ======================== */
 bool ConvBiasImpl::AlgoDotS8DirectStride1::usable(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible =
             direct_dotprod_int8_stride1::can_conv_direct_stride1_int8(param);
@@ -163,14 +163,14 @@ bool ConvBiasImpl::AlgoDotS8DirectStride1::usable(
 }
 
 size_t ConvBiasImpl::AlgoDotS8DirectStride1::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto bundle = direct_dotprod_int8_stride1::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoDotS8DirectStride1::dispatch_kerns(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8, 2, 1) {
         return direct_dotprod_int8_stride1::get_kimpls(param, m_large_group);
     }
@@ -180,7 +180,7 @@ ConvBiasImpl::AlgoDotS8DirectStride1::dispatch_kerns(
 
 /* ===================== dot stride2 algo ======================== */
 bool ConvBiasImpl::AlgoDotS8DirectStride2::usable(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible =
             direct_dotprod_int8_stride2::can_conv_direct_stride2_int8(param);
@@ -193,14 +193,14 @@ bool ConvBiasImpl::AlgoDotS8DirectStride2::usable(
 }
 
 size_t ConvBiasImpl::AlgoDotS8DirectStride2::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     auto bundle = direct_dotprod_int8_stride2::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoDotS8DirectStride2::dispatch_kerns(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8, 2, 2) {
         return direct_dotprod_int8_stride2::get_kimpls(param, m_large_group);
     }
@@ -212,7 +212,7 @@ ConvBiasImpl::AlgoDotS8DirectStride2::dispatch_kerns(
 /* ======================= AlgoS8WinogradF23_8x8 ======================== */
 
 bool ConvBiasImpl::AlgoS8WinogradF23_8x8::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     if (param.filter_meta.icpg % 8 != 0 || param.filter_meta.ocpg % 8 != 0)
         return false;
@@ -225,13 +225,14 @@ bool ConvBiasImpl::AlgoS8WinogradF23_8x8::usable(
                     .get_matmul_kern_param(param);
     return m_matmul_algo->usable(matmul_param) &&
            m_matmul_algo->packmode() == PackMode::NO_PACK &&
-           ((opr->param().format == param::ConvBias::Format::NCHW &&
+           ((param.filter_meta.format == param::ConvBias::Format::NCHW &&
              param.filter_type.enumv() == DTypeEnum::QuantizedS8) ||
-            (opr->param().format == param::ConvBias::Format::NCHW_WINOGRAD &&
-             opr->param().output_block_size == 2 &&
+            (param.filter_meta.format ==
+                     param::ConvBias::Format::NCHW_WINOGRAD &&
+             param.output_block_size == 2 &&
              param.winograd_matmul_format == param::MatrixMul::Format::MK8 &&
              param.filter_type.enumv() == DTypeEnum::QuantizedS16)) &&
-           opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+           !param.filter_meta.should_flip &&
            (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
             param.filter_meta.spatial[0] == 3) &&
            (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -251,7 +252,7 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoS8WinogradF23_8x8,
 
 //=========================== input int8 compute float32 =========
 bool ConvBiasImpl::AlgoS8CF32WinogradF23_4x4_NCHW44::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8,
@@ -270,14 +271,14 @@ bool ConvBiasImpl::AlgoS8CF32WinogradF23_4x4_NCHW44::usable(
                         .get_matmul_kern_param(param));
         return is_matmul_usable &&
                m_matmul_algo->packmode() == PackMode::NO_PACK &&
-               ((opr->param().format == param::ConvBias::Format::NCHW44 &&
+               ((param.filter_meta.format == param::ConvBias::Format::NCHW44 &&
                  param.filter_type.enumv() == DTypeEnum::QuantizedS8) ||
-                ((opr->param().format ==
+                ((param.filter_meta.format ==
                   param::ConvBias::Format::NCHW44_WINOGRAD) &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK4)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -302,35 +303,37 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoS8CF32WinogradF23_4x4_NCHW44,
 
 /* ======================= AlgoS8WinogradF23_8x8_NCHW44 ======================== */
 bool ConvBiasImpl::AlgoS8WinogradF23_8x8_NCHW44::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
-
     MIDOUT_BEGIN(
             megdnn_arm_common_conv_bias_int8,
-            midout_iv(
-                    "arm_common_AlgoS8WinogradF23_8x8_NCHW44::usable"_hash)) {
+            midout_iv("arm_common_AlgoS8WinogradF23_8x8_NCHW44::usable"_hash)) {
         if (param.filter_meta.icpg % 8 != 0 || param.filter_meta.ocpg % 8 != 0)
             return false;
         using Strategy = winograd::winograd_2x3_8x8_s8_nchw44;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
         auto&& matmul_param =
-            megdnn::winograd::ConvBias<Strategy, param::MatrixMul::Format::MK8>(
+                megdnn::winograd::ConvBias<Strategy,
+                                           param::MatrixMul::Format::MK8>(
                         strategy, m_tile_size, param)
                         .get_matmul_kern_param(param);
         bool is_matmul_usable = m_matmul_algo->usable(matmul_param);
         return is_matmul_usable &&
-           ((opr->param().format == param::ConvBias::Format::NCHW44 &&
+               ((param.filter_meta.format == param::ConvBias::Format::NCHW44 &&
                  param.filter_type.enumv() == DTypeEnum::QuantizedS8) ||
-            (opr->param().format == param::ConvBias::Format::NCHW44_WINOGRAD &&
-             opr->param().output_block_size == 2 &&
-             param.winograd_matmul_format == param::MatrixMul::Format::MK8 &&
+                (param.filter_meta.format ==
+                         param::ConvBias::Format::NCHW44_WINOGRAD &&
+                 param.output_block_size == 2 &&
+                 param.winograd_matmul_format ==
+                         param::MatrixMul::Format::MK8 &&
                  param.filter_type.enumv() == DTypeEnum::QuantizedS16)) &&
-           opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
                 param.filter_meta.stride[0] == 1) &&
-           (param.filter_meta.dilation[0] == param.filter_meta.dilation[1] &&
+               (param.filter_meta.dilation[0] ==
+                        param.filter_meta.dilation[1] &&
                 param.filter_meta.dilation[0] == 1) &&
                param.compute_mode == param::ConvBias::ComputeMode::DEFAULT &&
                param.src_type.enumv() == DTypeEnum::QuantizedS8 &&

dnn/src/arm_common/conv_bias/int8/algos.h

@@ -26,16 +26,13 @@ public:
     const char* name() const override {
         return m_large_group ? "S8STRD1_LARGE_GROUP" : "S8STRD1_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 
-    bool is_preferred(megdnn::fallback::ConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 
 class ConvBiasImpl::AlgoS8DirectStride2 final : public AlgoBase {
@@ -47,13 +44,11 @@ public:
     const char* name() const override {
         return m_large_group ? "S8STRD2_LARGE_GROUP" : "S8STRD2_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -62,15 +57,12 @@ public:
     AlgoS8DirectNCHW44() {}
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "S8_NCHW44_DIRECT"; }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
-            const NCBKernSizeParam& param) const override;
-    bool is_preferred(megdnn::fallback::ConvBiasImpl*,
             const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 
 class ConvBiasImpl::AlgoS8DirectNCHWNCHW44 final : public AlgoBase {
@@ -78,27 +70,22 @@ public:
     AlgoS8DirectNCHWNCHW44() {}
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "S8_CONV_NCHW_NCHW44"; }
-    bool usable(fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
-            const NCBKernSizeParam& param) const override;
-    bool is_preferred(megdnn::fallback::ConvBiasImpl*,
             const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 
 class ConvBiasImpl::AlgoS8ChanWiseStride1NCHW44 final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "S8_CHAN_WISE_STRD1_NCHW44"; }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -106,12 +93,10 @@ class ConvBiasImpl::AlgoS8ChanWiseStride2NCHW44 final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "S8_CHAN_WISE_STRD2_NCHW44"; }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -121,13 +106,11 @@ class ConvBiasImpl::AlgoDotS8DirectNCHWNCHW44 final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "ARMDOTS8_NCHW_NCHW44"; }
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam&,
+    bool usable(const NCBKernSizeParam&,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam&) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -142,13 +125,11 @@ public:
         return m_large_group ? "ARMDOTS8STRD1_LARGE_GROUP"
                              : "ARMDOTS8STRD1_SMALL_GROUP";
     }
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam&,
+    bool usable(const NCBKernSizeParam&,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam&) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -163,13 +144,11 @@ public:
                              : "ARMDOTS8STRD2_SMALL_GROUP";
     }
 
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam&,
+    bool usable(const NCBKernSizeParam&,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam&) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -178,21 +157,16 @@ public:
     AlgoDotS8Direct_NCHW44() {}
 
     bool is_reproducible() const override { return true; }
-    const char* name() const override {
-        return "ARMDOTS8DIRECT_NCHW44";
-    }
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam&,
+    const char* name() const override { return "ARMDOTS8DIRECT_NCHW44"; }
+    bool usable(const NCBKernSizeParam&,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam&) const override;
 
     SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 
-    bool is_preferred(megdnn::fallback::ConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 #endif
 

dnn/src/arm_common/conv_bias/int8/direct_dotprod_nchw44_algo.cpp

@@ -161,7 +161,7 @@ static void conv_kern(const WorkspaceBundle& bundle,
 }  // namespace
 
 bool ConvBiasImpl::AlgoDotS8Direct_NCHW44::usable(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MEGDNN_MARK_USED_VAR(algo_selection_strategy);
     auto&& fm = param.filter_meta;
@@ -199,19 +199,19 @@ bool ConvBiasImpl::AlgoDotS8Direct_NCHW44::usable(
 }
 
 bool ConvBiasImpl::AlgoDotS8Direct_NCHW44::is_preferred(
-        megdnn::fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MEGDNN_MARK_USED_VAR(param);
     return true;
 }
 
 size_t ConvBiasImpl::AlgoDotS8Direct_NCHW44::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoDotS8Direct_NCHW44::dispatch_kerns(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8,
                  midout_iv("ALGODOTS8DIRECT_NCHW44"_hash)) {
         auto fm = param.filter_meta;

dnn/src/arm_common/conv_bias/int8/direct_nchw44_algo.cpp

@@ -189,7 +189,7 @@ static void do_conv_kern(const WorkspaceBundle& bundle,
 }
 
 bool ConvBiasImpl::AlgoS8DirectNCHW44::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MEGDNN_MARK_USED_VAR(algo_selection_strategy);
     auto&& fm = param.filter_meta;
@@ -213,22 +213,20 @@ bool ConvBiasImpl::AlgoS8DirectNCHW44::usable(
 }
 
 bool ConvBiasImpl::AlgoS8DirectNCHW44::is_preferred(
-        megdnn::fallback::ConvBiasImpl* conv_bias_impl_ptr,
                 const NCBKernSizeParam& param) const {
     // TODO: benchmark and fix
-    MEGDNN_MARK_USED_VAR(conv_bias_impl_ptr);
     MEGDNN_MARK_USED_VAR(param);
     return false;
 }
 
 size_t ConvBiasImpl::AlgoS8DirectNCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoS8DirectNCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto fm = param.filter_meta;
     size_t N = param.n;
     size_t IC = fm.icpg;

dnn/src/arm_common/conv_bias/int8/direct_nchw_nchw44_algo.cpp

@@ -214,7 +214,7 @@ static void do_conv_kern(const WorkspaceBundle& bundle,
 }
 
 bool ConvBiasImpl::AlgoS8DirectNCHWNCHW44::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MEGDNN_MARK_USED_VAR(algo_selection_strategy);
     auto&& fm = param.filter_meta;
@@ -236,22 +236,20 @@ bool ConvBiasImpl::AlgoS8DirectNCHWNCHW44::usable(
 }
 
 bool ConvBiasImpl::AlgoS8DirectNCHWNCHW44::is_preferred(
-        megdnn::fallback::ConvBiasImpl* conv_bias_impl_ptr,
         const NCBKernSizeParam& param) const {
     // TODO: benchmark and fix
-    MEGDNN_MARK_USED_VAR(conv_bias_impl_ptr);
     MEGDNN_MARK_USED_VAR(param);
     return false;
 }
 
 size_t ConvBiasImpl::AlgoS8DirectNCHWNCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoS8DirectNCHWNCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto fm = param.filter_meta;
     size_t N = param.n;
     size_t OC = fm.ocpg;

dnn/src/arm_common/conv_bias/int8/dot_direct_nchw_nchw44_algo.cpp

@@ -172,8 +172,7 @@ static void do_conv_kern(const WorkspaceBundle& bundle,
 }  // namespace
 
 bool ConvBiasImpl::AlgoDotS8DirectNCHWNCHW44::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
-        AlgoSelectionStrategy) const {
+        const NCBKernSizeParam& param, AlgoSelectionStrategy) const {
     auto&& fm = param.filter_meta;
     auto fh = fm.spatial[0];
     int oc = fm.ocpg;
@@ -194,13 +193,13 @@ bool ConvBiasImpl::AlgoDotS8DirectNCHWNCHW44::usable(
 }
 
 size_t ConvBiasImpl::AlgoDotS8DirectNCHWNCHW44::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoDotS8DirectNCHWNCHW44::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto fm = param.filter_meta;
     const int batch = param.n;
     const int group = fm.group;

dnn/src/arm_common/conv_bias/int8x8x16/algos.cpp

@@ -83,7 +83,7 @@ void get_rectified_size_str2(size_t IH, size_t IW, size_t OH, size_t OW,
 
 /* ===================== direct algo ===================== */
 bool ConvBiasImpl::AlgoI8x8x16Direct::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 1, 0) {
         auto&& fm = param.filter_meta;
@@ -129,7 +129,7 @@ WorkspaceBundle ConvBiasImpl::AlgoI8x8x16Direct::get_bundle(
     return {nullptr, {part0, part1}};
 }
 size_t ConvBiasImpl::AlgoI8x8x16Direct::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 1, 1) {
         auto bundle = get_bundle(param);
         return bundle.total_size_in_bytes();
@@ -293,7 +293,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoI8x8x16Direct::get_kimpls(
 }
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoI8x8x16Direct::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 1, 2) {
         return get_kimpls(param);
     }
@@ -303,7 +303,7 @@ ConvBiasImpl::AlgoI8x8x16Direct::dispatch_kerns(
 
 /* ===================== stride-2 algo ===================== */
 bool ConvBiasImpl::AlgoI8x8x16Stride2::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 2, 0) {
         auto&& fm = param.filter_meta;
@@ -350,7 +350,7 @@ WorkspaceBundle ConvBiasImpl::AlgoI8x8x16Stride2::get_bundle(
     return {nullptr, {part0, part1}};
 }
 size_t ConvBiasImpl::AlgoI8x8x16Stride2::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 2, 1) {
         auto bundle = get_bundle(param);
         return bundle.total_size_in_bytes();
@@ -513,7 +513,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoI8x8x16Stride2::get_kimpls(
 }
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoI8x8x16Stride2::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 2, 2) {
         return get_kimpls(param);
     }
@@ -521,7 +521,7 @@ ConvBiasImpl::AlgoI8x8x16Stride2::dispatch_kerns(
     return {};
 }
 bool ConvBiasImpl::AlgoI8x8x16Stride2Filter2::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 3, 0) {
         return param.bias_mode == BiasMode::NO_BIAS &&
@@ -534,7 +534,7 @@ bool ConvBiasImpl::AlgoI8x8x16Stride2Filter2::usable(
 }
 
 size_t ConvBiasImpl::AlgoI8x8x16Stride2Filter2::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 3, 1) {
         return conv_bias::get_workspace_in_bytes_conv_int8x8x16_stride2_flt2(
                 param);
@@ -545,7 +545,7 @@ size_t ConvBiasImpl::AlgoI8x8x16Stride2Filter2::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoI8x8x16Stride2Filter2::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     // return {conv_bias::conv_int8x8x16_stride2_flt2,true};
     auto kern = [](const NCBKernParam& param, const NCBKernIndex& ncb_index) {
         MIDOUT_BEGIN(megdnn_arm_common_conv_bias_int8816_kimpl, 3, 2) {

dnn/src/arm_common/conv_bias/int8x8x16/algos.h

@@ -35,12 +35,10 @@ public:
         return m_large_group ? "I8816DIRECT_LARGE_GROUP"
                              : "I8816DIRECT_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -64,13 +62,11 @@ public:
         return m_large_group ? "I8816STRD2_LARGE_GROUP"
                              : "I8816STRD2_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -79,13 +75,11 @@ public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "I8816STRD2F2"; }
 
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 

dnn/src/arm_common/conv_bias/opr_impl.cpp

@@ -232,7 +232,7 @@ void* const ConvBiasImpl::sm_arm_common_algo_type =
         &arm_common_algo_type_storage;
 
 bool ConvBiasImpl::is_matmul_quantized_prefer(
-        const ConvBiasImpl::NCBKernSizeParam& param) {
+        const ConvBiasImpl::NCBKernSizeParam& param) const {
     // fallback::ConvBiasImpl::NCBKernParam conv_ncb_param;
     fallback::ConvBiasImpl::NCBKernSizeParam conv_ncb_param(
             param, 0, param::MatrixMul::Format::DEFAULT, {}, 0,

dnn/src/arm_common/conv_bias/opr_impl.h

@@ -27,7 +27,7 @@ public:
     SmallVector<AlgoBase*> algo_pack() override;
 
     bool is_matmul_quantized_prefer(
-            const ConvBiasImpl::NCBKernSizeParam& ncb_param) override;
+            const ConvBiasImpl::NCBKernSizeParam& ncb_param) const override;
     class AlgoPack;
 
 protected:

dnn/src/arm_common/conv_bias/quint8/algos.cpp

@@ -6,17 +6,18 @@
  *
  * 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/arm_common/conv_bias/quint8/algos.h"
+#include "midout.h"
 #include "src/arm_common/conv_bias/quint8/stride1.h"
-#include "src/arm_common/conv_bias/quint8/stride2.h"
 #include "src/arm_common/conv_bias/quint8/stride1_dotprod.h"
+#include "src/arm_common/conv_bias/quint8/stride2.h"
 #include "src/arm_common/conv_bias/quint8/stride2_dotprod.h"
 #include "src/arm_common/elemwise_op.h"
 #include "src/fallback/conv_bias/common.h"
-#include "midout.h"
 
 MIDOUT_DECL(megdnn_arm_common_conv_bias_quint8)
 
@@ -25,7 +26,7 @@ using namespace arm_common;
 
 /* ===================== stride1 algo ===================== */
 bool ConvBiasImpl::AlgoQU8DirectStride1::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible = direct_quint8_stride1::can_conv_direct_stride1_quint8(param);
     if (algo_selection_strategy ==
@@ -37,14 +38,14 @@ bool ConvBiasImpl::AlgoQU8DirectStride1::usable(
 }
 
 size_t ConvBiasImpl::AlgoQU8DirectStride1::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto bundle = direct_quint8_stride1::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoQU8DirectStride1::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_quint8, 0, 0) {
         return direct_quint8_stride1::get_kimpls(param, m_large_group);
     }
@@ -54,7 +55,7 @@ ConvBiasImpl::AlgoQU8DirectStride1::dispatch_kerns(
 
 /* ===================== stride2 algo ===================== */
 bool ConvBiasImpl::AlgoQU8DirectStride2::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible = direct_quint8_stride2::can_conv_direct_stride2_quint8(param);
     if (algo_selection_strategy ==
@@ -66,14 +67,14 @@ bool ConvBiasImpl::AlgoQU8DirectStride2::usable(
 }
 
 size_t ConvBiasImpl::AlgoQU8DirectStride2::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto bundle = direct_quint8_stride2::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
 }
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoQU8DirectStride2::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_quint8, 0, 1) {
         return direct_quint8_stride2::get_kimpls(param, m_large_group);
     }
@@ -83,7 +84,7 @@ ConvBiasImpl::AlgoQU8DirectStride2::dispatch_kerns(
 #if __ARM_FEATURE_DOTPROD
 /* ===================== stride1 algo ===================== */
 bool ConvBiasImpl::AlgoDotU8DirectStride1::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible =
             direct_dotprod_quint8_stride1::can_conv_direct_stride1_quint8(
@@ -97,7 +98,7 @@ bool ConvBiasImpl::AlgoDotU8DirectStride1::usable(
 }
 
 size_t ConvBiasImpl::AlgoDotU8DirectStride1::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto bundle =
             direct_dotprod_quint8_stride1::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
@@ -105,7 +106,7 @@ size_t ConvBiasImpl::AlgoDotU8DirectStride1::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoDotU8DirectStride1::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_quint8, 1, 0) {
         return direct_dotprod_quint8_stride1::get_kimpls(param, m_large_group);
     }
@@ -115,7 +116,7 @@ ConvBiasImpl::AlgoDotU8DirectStride1::dispatch_kerns(
 
 /* ===================== stride2 algo ===================== */
 bool ConvBiasImpl::AlgoDotU8DirectStride2::usable(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     bool avaible =
             direct_dotprod_quint8_stride2::can_conv_direct_stride2_quint8(
@@ -129,7 +130,7 @@ bool ConvBiasImpl::AlgoDotU8DirectStride2::usable(
 }
 
 size_t ConvBiasImpl::AlgoDotU8DirectStride2::get_workspace(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto bundle =
             direct_dotprod_quint8_stride2::get_bundle(param, m_large_group);
     return bundle.total_size_in_bytes();
@@ -137,7 +138,7 @@ size_t ConvBiasImpl::AlgoDotU8DirectStride2::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoDotU8DirectStride2::dispatch_kerns(
-        fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_arm_common_conv_bias_quint8, 1, 1) {
         return direct_dotprod_quint8_stride2::get_kimpls(param, m_large_group);
     }

dnn/src/arm_common/conv_bias/quint8/algos.h

@@ -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.
  */
 
 #pragma once
@@ -26,13 +27,11 @@ public:
         return m_large_group ? "QU8STRD1_LARGE_GROUP" : "QU8STRD1_SMALL_GROUP";
     }
 
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -45,13 +44,11 @@ public:
     const char* name() const override {
         return m_large_group ? "QU8STRD2_LARGE_GROUP" : "QU8STRD2_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 #if __ARM_FEATURE_DOTPROD
@@ -66,13 +63,11 @@ public:
                              : "ARMDOTU8STRD1_SMALL_GROUP";
     }
 
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 
@@ -86,13 +81,11 @@ public:
         return m_large_group ? "ARMDOTU8STRD2_LARGE_GROUP"
                              : "ARMDOTU8STRD2_SMALL_GROUP";
     }
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(fallback::ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* opr,
             const NCBKernSizeParam& param) const override;
 };
 #endif

dnn/src/armv7/conv_bias/int8/algos.cpp

@@ -26,9 +26,8 @@ using namespace armv7;
 /* ===================== matrix mul algo ===================== */
 
 bool ConvBiasImpl::AlgoS8MatrixMul::usable(
-        FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
-    MEGDNN_MARK_USED_VAR(opr);
     auto&& fm = param.filter_meta;
     return param.src_type.enumv() == DTypeEnum::QuantizedS8 &&
            param.dst_type.enumv() == DTypeEnum::QuantizedS8 &&

dnn/src/armv7/conv_bias/int8/algos.h

@@ -27,14 +27,12 @@ public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "S8MATMUL"; }
 
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override {
+    size_t get_workspace(const NCBKernSizeParam& param) const override {
         return get_bundle(param).total_size_in_bytes();
     }
     SmallVector<NCBKern> dispatch_kerns(
-            FallbackConvBiasImpl*,
             const NCBKernSizeParam& param) const override {
         size_t group = param.filter_meta.group;
         return {{kimpl, {group, 1_z, 1_z}}};

dnn/src/armv7/conv_bias/quint8/algos.cpp

@@ -26,9 +26,8 @@ using namespace armv7;
 /* ===================== matrix mul algo ===================== */
 
 bool ConvBiasImpl::AlgoQU8MatrixMul::usable(
-        FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
-    MEGDNN_MARK_USED_VAR(opr);
     auto&& fm = param.filter_meta;
     return param.src_type.enumv() == DTypeEnum::Quantized8Asymm &&
            param.dst_type.enumv() == DTypeEnum::Quantized8Asymm &&

dnn/src/armv7/conv_bias/quint8/algos.h

@@ -27,15 +27,13 @@ public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "QU8MATMUL"; }
 
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override {
+    size_t get_workspace(const NCBKernSizeParam& param) const override {
         return get_bundle(param).total_size_in_bytes();
     }
 
     SmallVector<fallback::ConvBiasImpl::NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl* /*opr*/,
             const NCBKernSizeParam& param) const override {
         size_t group = param.filter_meta.group;
         return {{kimpl, {group, 1_z, 1_z}}};

dnn/src/fallback/conv_bias/algos.cpp

@@ -10,6 +10,7 @@
  */
 
 #include "src/fallback/conv_bias/algos.h"
+#include "megdnn/opr_param_defs.h"
 #include "src/common/opr_delegate.h"
 #include "src/fallback/conv_bias/winograd/strategy.h"
 #include "src/naive/convolution/helper.h"
@@ -21,18 +22,28 @@ using namespace fallback;
 
 namespace {
 
-param::Convolution get_param_convolution(const param::ConvBias param) {
-    param::Convolution ret{param.mode,     param.pad_h,
-                           param.pad_w,    param.stride_h,
-                           param.stride_w, param.dilate_h,
-                           param.dilate_w, param::Convolution::Sparse::DENSE,
-                           param.format};
-    return ret;
+param::Convolution get_param_convolution(
+        const ConvBiasImpl::NCBKernSizeParam& param) {
+    param::Convolution::Mode mode;
+    param::Convolution::Sparse sparse;
+    if (param.filter_meta.should_flip) {
+        mode = param::Convolution::Mode::CONVOLUTION;
+    } else {
+        mode = param::Convolution::Mode::CROSS_CORRELATION;
+    }
+    return param::Convolution{mode,
+                              param.filter_meta.padding[0],
+                              param.filter_meta.padding[1],
+                              param.filter_meta.stride[0],
+                              param.filter_meta.stride[1],
+                              param.filter_meta.dilation[1],
+                              param.filter_meta.dilation[0],
+                              sparse = param::Convolution::Sparse::DENSE,
+                              param.filter_meta.format};
 }
 
-TensorLayoutArray get_layouts(const param::ConvBias& param,
-                              const ConvBiasImpl::NCBKernSizeParam& p) {
-    megdnn_assert(param.format == param::ConvBias::Format::NCHW);
+TensorLayoutArray get_layouts(const ConvBiasImpl::NCBKernSizeParam& p) {
+    megdnn_assert(p.filter_meta.format == param::ConvBias::Format::NCHW);
     UNPACK_CONV_NCB_KERN_SIZES(p);
     MEGDNN_MARK_USED_VAR(SH);
     MEGDNN_MARK_USED_VAR(SW);
@@ -53,14 +64,14 @@ TensorLayoutArray get_layouts(const param::ConvBias& param,
     return {src_layout, filter_layout, bias_layout, dst_layout};
 }
 
-void kern_default(param::ConvBias param, const ConvBiasImpl::NCBKernParam& p) {
+void kern_default(const ConvBiasImpl::NCBKernParam& p) {
     dt_byte* workspace_ptr = static_cast<dt_byte*>(p.workspace_ptr);
 
     auto filter_meta_ptr =
             reinterpret_cast<const ConvBiasForward::CanonizedFilterMeta*>(
                     &p.filter_meta);
     auto filter_meta = *filter_meta_ptr;
-    auto layouts = get_layouts(param, p);
+    auto layouts = get_layouts(p);
 
     TensorND src{reinterpret_cast<dt_byte*>(const_cast<void*>(p.src_ptr)),
                  layouts[0]};
@@ -83,7 +94,7 @@ void kern_default(param::ConvBias param, const ConvBiasImpl::NCBKernParam& p) {
               bias.layout.dtype.enumv() ==                                     \
                       DTypeTrait<dtype::bias_dt>::enumv) &&                    \
              sfb.layout.dtype.enumv() == DTypeTrait<dtype::out_dt>::enumv &&   \
-             param.compute_mode == param::ConvBias::ComputeMode::cmode) {      \
+             p.compute_mode == param::ConvBias::ComputeMode::cmode) {          \
         func(src, filter, bias, sfb, workspace_ptr, filter_meta);              \
     }
 #define DISPATCH(in_dt, out_dt)                             \
@@ -118,7 +129,7 @@ void kern_default(param::ConvBias param, const ConvBiasImpl::NCBKernParam& p) {
 
     auto res = sfb;
     using NonlineMode = param::ConvBias::NonlineMode;
-    switch (param.nonlineMode) {
+    switch (p.nonlineMode) {
 #define cb(_mode)                                                             \
     case NonlineMode::_mode: {                                                \
         if (res.layout.dtype.category() != DTypeCategory::QUANTIZED) {        \
@@ -168,24 +179,23 @@ MIDOUT_DECL(megdnn_fallback_naive)
 /* ======================= AlgoNaive ======================== */
 
 bool ConvBiasImpl::AlgoNaive::usable(
-        ConvBiasImpl* opr, const NCBKernSizeParam&,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MIDOUT_BEGIN(megdnn_fallback_naive, 0) {
-        return opr->param().format == param::ConvBias::Format::NCHW;
+        return param.filter_meta.format == param::ConvBias::Format::NCHW;
     }
     MIDOUT_END();
     return false;
 }
 
-size_t ConvBiasImpl::AlgoNaive::get_workspace(ConvBiasImpl* opr,
-                                              const NCBKernSizeParam& p) const {
+size_t ConvBiasImpl::AlgoNaive::get_workspace(const NCBKernSizeParam& p) const {
     MIDOUT_BEGIN(megdnn_fallback_naive, 1) {
-        auto layouts = get_layouts(opr->param(), p);
+        auto layouts = get_layouts(p);
         //! When group>1 or n>1, this algo will parallel by group and n
         size_t nr_threads = p.nr_threads;
         auto conv_opr =
                 inplace_cpu_handle()->create_operator<ConvolutionForward>();
-        conv_opr->param() = get_param_convolution(opr->param());
+        conv_opr->param() = get_param_convolution(p);
         if (p.dst_type.enumv() == DTypeEnum::QuantizedS8 ||
             p.dst_type.enumv() == DTypeEnum::Quantized8Asymm) {
             TensorLayout conv_dst_layout;
@@ -201,15 +211,14 @@ size_t ConvBiasImpl::AlgoNaive::get_workspace(ConvBiasImpl* opr,
 }
 
 SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoNaive::dispatch_kerns(
-        ConvBiasImpl* opr, const NCBKernSizeParam& p) const {
-    param::ConvBias opr_param = opr->param();
-    size_t workspace_size = get_workspace(opr, p);
+        const NCBKernSizeParam& p) const {
+    size_t workspace_size = get_workspace(p);
     //! When group>1 or n>1, this algo will parallel by group and n
     size_t nr_threads = p.nr_threads;
     size_t GROUP = p.filter_meta.group;
     size_t N = p.n;
     size_t workspace_per_thread = workspace_size / nr_threads;
-    auto kern = [opr_param, workspace_per_thread](
+    auto kern = [workspace_per_thread](
                         const NCBKernParam& param,
                         const NCBKernIndex& ncb_index) {
         MIDOUT_BEGIN(megdnn_fallback_naive, 2) {
@@ -224,7 +233,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoNaive::dispatch_kerns(
             thread_param.dst_ptr = param.dst<void>(batch_id, group_id);
             thread_param.src_ptr = param.src<void>(batch_id, group_id);
             thread_param.bias_ptr = param.bias<void>(batch_id, group_id);
-            kern_default(opr_param, thread_param);
+            kern_default(thread_param);
         }
         MIDOUT_END();
     };
@@ -235,10 +244,9 @@ MIDOUT_DECL(megdnn_fallback_winograd)
 /* ======================= AlgoWinogradF32 ======================== */
 
 bool ConvBiasImpl::AlgoWinogradF32::usable(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 1, 0) {
         using Strategy = fallback::winograd::winograd_2x3_1x1_f;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -246,13 +254,13 @@ bool ConvBiasImpl::AlgoWinogradF32::usable(
                                       strategy, UNIT_TILE_SIZE, param)
                                       .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -268,7 +276,7 @@ bool ConvBiasImpl::AlgoWinogradF32::usable(
 }
 
 size_t ConvBiasImpl::AlgoWinogradF32::get_workspace(
-        ConvBiasImpl*, const NCBKernSizeParam& p) const {
+        const NCBKernSizeParam& p) const {
     MEGDNN_MARK_USED_VAR(p);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 1, 1) {
         fallback::winograd::winograd_2x3_1x1_f strategy(
@@ -284,7 +292,7 @@ size_t ConvBiasImpl::AlgoWinogradF32::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoWinogradF32::dispatch_kerns(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MEGDNN_MARK_USED_VAR(param);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 1, 2) {
         fallback::winograd::winograd_2x3_1x1_f strategy(
@@ -302,10 +310,9 @@ ConvBiasImpl::AlgoWinogradF32::dispatch_kerns(
 /* ======================= AlgoWinogradF32 4x4 ======================== */
 
 bool ConvBiasImpl::AlgoWinogradF32_4x4::usable(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 2, 0) {
         if (param.filter_meta.icpg % 4 != 0 || param.filter_meta.ocpg % 4 != 0)
             return false;
@@ -317,13 +324,13 @@ bool ConvBiasImpl::AlgoWinogradF32_4x4::usable(
                         strategy, UNIT_TILE_SIZE, param)
                         .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK4)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -339,7 +346,7 @@ bool ConvBiasImpl::AlgoWinogradF32_4x4::usable(
 }
 
 size_t ConvBiasImpl::AlgoWinogradF32_4x4::get_workspace(
-        ConvBiasImpl*, const NCBKernSizeParam& p) const {
+        const NCBKernSizeParam& p) const {
     MEGDNN_MARK_USED_VAR(p);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 2, 1) {
         fallback::winograd::winograd_2x3_4x4_f strategy(
@@ -356,7 +363,7 @@ size_t ConvBiasImpl::AlgoWinogradF32_4x4::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoWinogradF32_4x4::dispatch_kerns(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MEGDNN_MARK_USED_VAR(param);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 2, 2) {
         fallback::winograd::winograd_2x3_4x4_f strategy(
@@ -374,10 +381,9 @@ ConvBiasImpl::AlgoWinogradF32_4x4::dispatch_kerns(
 /* ======================= AlgoWinogradQS8 ======================== */
 
 bool ConvBiasImpl::AlgoWinogradQS8::usable(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 3, 0) {
         using Strategy = fallback::winograd::winograd_2x3_1x1_qs8;
         Strategy strategy(param.src_type, param.filter_type, param.dst_type);
@@ -386,13 +392,13 @@ bool ConvBiasImpl::AlgoWinogradQS8::usable(
                                       .get_matmul_kern_param(param);
 
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::DEFAULT)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -408,7 +414,7 @@ bool ConvBiasImpl::AlgoWinogradQS8::usable(
 }
 
 size_t ConvBiasImpl::AlgoWinogradQS8::get_workspace(
-        ConvBiasImpl*, const NCBKernSizeParam& p) const {
+        const NCBKernSizeParam& p) const {
     MEGDNN_MARK_USED_VAR(p);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 3, 1) {
         fallback::winograd::winograd_2x3_1x1_qs8 strategy(
@@ -424,7 +430,7 @@ size_t ConvBiasImpl::AlgoWinogradQS8::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoWinogradQS8::dispatch_kerns(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MEGDNN_MARK_USED_VAR(param);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 3, 2) {
         fallback::winograd::winograd_2x3_1x1_qs8 strategy(
@@ -442,10 +448,9 @@ ConvBiasImpl::AlgoWinogradQS8::dispatch_kerns(
 /* ======================= AlgoWinogradQS8 8x8 ======================== */
 
 bool ConvBiasImpl::AlgoWinogradQS8_8x8::usable(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 4, 0) {
         if (param.filter_meta.icpg % 8 != 0 || param.filter_meta.ocpg % 8 != 0)
             return false;
@@ -457,13 +462,13 @@ bool ConvBiasImpl::AlgoWinogradQS8_8x8::usable(
                         strategy, UNIT_TILE_SIZE, param)
                         .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK8)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -479,7 +484,7 @@ bool ConvBiasImpl::AlgoWinogradQS8_8x8::usable(
 }
 
 size_t ConvBiasImpl::AlgoWinogradQS8_8x8::get_workspace(
-        ConvBiasImpl*, const NCBKernSizeParam& p) const {
+        const NCBKernSizeParam& p) const {
     MEGDNN_MARK_USED_VAR(p);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 4, 1) {
         fallback::winograd::winograd_2x3_8x8_qs8 strategy(
@@ -496,7 +501,7 @@ size_t ConvBiasImpl::AlgoWinogradQS8_8x8::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoWinogradQS8_8x8::dispatch_kerns(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MEGDNN_MARK_USED_VAR(param);
     MIDOUT_BEGIN(megdnn_fallback_winograd, 4, 2) {
         fallback::winograd::winograd_2x3_8x8_qs8 strategy(

dnn/src/fallback/conv_bias/algos.h

@@ -22,12 +22,10 @@ class ConvBiasImpl::AlgoNaive final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "FALLBACK_NAIVE"; }
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
-    SmallVector<NCBKern> dispatch_kerns(ConvBiasImpl*,
-                                        const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
+    SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam&) const override;
 };
 
 class ConvBiasImpl::AlgoWinogradF32 final : public AlgoBase {
@@ -43,12 +41,10 @@ public:
         }
         return m_name.c_str();
     }
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
-    SmallVector<NCBKern> dispatch_kerns(ConvBiasImpl*,
-                                        const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
+    SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam&) const override;
 
 private:
     MatrixMulImpl::AlgoBase* m_matmul_algo;
@@ -69,12 +65,10 @@ public:
         }
         return m_name.c_str();
     }
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
-    SmallVector<NCBKern> dispatch_kerns(ConvBiasImpl*,
-                                        const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
+    SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam&) const override;
 
 private:
     MatrixMulImpl::AlgoBase* m_matmul_algo;
@@ -95,12 +89,10 @@ public:
         }
         return m_name.c_str();
     }
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
-    SmallVector<NCBKern> dispatch_kerns(ConvBiasImpl*,
-                                        const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
+    SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam&) const override;
 
 private:
     MatrixMulImpl::AlgoBase* m_matmul_algo;
@@ -121,12 +113,10 @@ public:
         }
         return m_name.c_str();
     }
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
-    SmallVector<NCBKern> dispatch_kerns(ConvBiasImpl*,
-                                        const NCBKernSizeParam&) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
+    SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam&) const override;
 
 private:
     MatrixMulImpl::AlgoBase* m_matmul_algo;

dnn/src/fallback/conv_bias/common.h

@@ -140,22 +140,17 @@ using BiasMode = ConvBiasForward::BiasMode;
 
 #define MEGDNN_WINOGRAD_ALGO_FUN_DECLARE()                                     \
     bool is_reproducible() const override { return true; }                     \
-    bool usable(fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,    \
+    bool usable(const NCBKernSizeParam& param,                                 \
                 AlgoSelectionStrategy algo_selection_strategy) const override; \
-    size_t get_workspace(fallback::ConvBiasImpl*,                              \
-                         const NCBKernSizeParam& param) const override;        \
-    virtual SmallVector<NCBKern> dispatch_kerns(fallback::ConvBiasImpl* opr,   \
-                                                const NCBKernSizeParam& param) \
+    size_t get_workspace(const NCBKernSizeParam& param) const override;        \
+    virtual SmallVector<NCBKern> dispatch_kerns(const NCBKernSizeParam& param) \
             const override;                                                    \
     SmallVector<TensorLayout> deduce_preprocessed_filter_layout(               \
-            fallback::ConvBiasImpl*, const NCBKernSizeParam& param)            \
-            const override;                                                    \
-    size_t get_preprocess_workspace(fallback::ConvBiasImpl*,                   \
-                                    const NCBKernSizeParam& param)             \
+            const NCBKernSizeParam& param) const override;                     \
+    size_t get_preprocess_workspace(const NCBKernSizeParam& param)             \
             const override;                                                    \
     virtual SmallVector<NCBKern> dispatch_preprocess_kerns(                    \
-            fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param)        \
-            const override;                                                    \
+            const NCBKernSizeParam& param) const override;                     \
                                                                                \
 private:                                                                       \
     fallback::MatrixMulImpl::AlgoBase* m_matmul_algo;                          \

dnn/src/fallback/conv_bias/conv1x1/algos.cpp

@@ -48,7 +48,7 @@ size_t ConvBiasImpl::AlgoConv1x1::get_oc_tile_size_heuristic(
 }
 
 size_t ConvBiasImpl::AlgoConv1x1::get_workspace(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     size_t OH = param.osz[0];
     size_t OW = param.osz[1];
     size_t compt_oc_block_size = get_oc_tile_size_heuristic(param);
@@ -90,7 +90,7 @@ size_t ConvBiasImpl::AlgoConv1x1::get_workspace(
 }
 
 SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoConv1x1::dispatch_kerns(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     SmallVector<ConvBiasImpl::NCBKern> ret_kern;
     size_t OH = param.osz[0];
     size_t OW = param.osz[1];
@@ -142,7 +142,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoConv1x1::dispatch_kerns(
 
     Conv1x1StrategyBase* conv1x1_strategy =
             Conv1x1Factory::make_conv1x1_strategy(param, pack_mode,
-                                                  opr->param().format);
+                                                  param.filter_meta.format);
 
     auto kern_packA = [this, whole_bundle, matmul_bundle, param,
                        compt_oc_block_size, conv1x1_strategy](
@@ -180,13 +180,12 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoConv1x1::dispatch_kerns(
     return ret_kern;
 }
 
-bool ConvBiasImpl::AlgoConv1x1::usable(ConvBiasImpl* opr,
-                                       const NCBKernSizeParam& param,
+bool ConvBiasImpl::AlgoConv1x1::usable(const NCBKernSizeParam& param,
                                        AlgoSelectionStrategy) const {
     MIDOUT_BEGIN(megdnn_fallback_conv1x1, 0, 2) {
-        if (opr->param().format != param::ConvBias::Format::NCHW &&
-            opr->param().format != param::ConvBias::Format::NCHW44 &&
-            opr->param().format != param::ConvBias::Format::NCHW44_DOT)
+        if (param.filter_meta.format != param::ConvBias::Format::NCHW &&
+            param.filter_meta.format != param::ConvBias::Format::NCHW44 &&
+            param.filter_meta.format != param::ConvBias::Format::NCHW44_DOT)
             return false;
 
         size_t FH = param.filter_meta.spatial[0],
@@ -199,7 +198,7 @@ bool ConvBiasImpl::AlgoConv1x1::usable(ConvBiasImpl* opr,
         if (FH != 1 || FW != 1 || PH || PW || SH != 1 || SW != 1)
             return false;
 
-        if(param.src_type.enumv() != param.filter_type.enumv()) {
+        if (param.src_type.enumv() != param.filter_type.enumv()) {
             return false;
         }
 
@@ -225,8 +224,8 @@ bool ConvBiasImpl::AlgoConv1x1::usable(ConvBiasImpl* opr,
             }
         }
 
-        if (opr->param().format == param::ConvBias::Format::NCHW44 ||
-            opr->param().format == param::ConvBias::Format::NCHW44_DOT) {
+        if (param.filter_meta.format == param::ConvBias::Format::NCHW44 ||
+            param.filter_meta.format == param::ConvBias::Format::NCHW44_DOT) {
             if (param.filter_meta.icpg < 4_z || param.filter_meta.icpg == 1 ||
                 param.filter_meta.ocpg == 1) {
                 return false;
@@ -236,13 +235,14 @@ bool ConvBiasImpl::AlgoConv1x1::usable(ConvBiasImpl* opr,
         size_t OH = param.osz[0];
         size_t OW = param.osz[1];
 
-        MatrixMulImpl::KernSizeParam matmul_param = utils::get_matmul_kern_param(
-                param, OH * OW, get_oc_tile_size_heuristic(param));
+        MatrixMulImpl::KernSizeParam matmul_param =
+                utils::get_matmul_kern_param(param, OH * OW,
+                                             get_oc_tile_size_heuristic(param));
         bool matmul_usable = m_matmul_algo->usable(matmul_param);
 
         auto pack_mode = m_matmul_algo->packmode();
         bool strategy_usable = Conv1x1Factory::can_make_conv1x1_strategy(
-                param, pack_mode, opr->param().format);
+                param, pack_mode, param.filter_meta.format);
 
         return matmul_usable && strategy_usable &&
                (param.filter_meta.dilation[0] ==
@@ -255,7 +255,7 @@ bool ConvBiasImpl::AlgoConv1x1::usable(ConvBiasImpl* opr,
 }
 
 bool ConvBiasImpl::AlgoConv1x1::is_preferred(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     size_t OH = param.osz[0];
     size_t OW = param.osz[1];
     if (OH * OW != 1) {

dnn/src/fallback/conv_bias/conv1x1/algos.h

@@ -34,14 +34,13 @@ public:
         return m_name.c_str();
     }
 
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     SmallVector<NCBKern> dispatch_kerns(
-            ConvBiasImpl* opr, const NCBKernSizeParam& param) const override;
+            const NCBKernSizeParam& param) const override;
 
-    bool is_preferred(ConvBiasImpl*, const NCBKernSizeParam&) const override;
+    bool is_preferred(const NCBKernSizeParam&) const override;
 
 protected:
     size_t get_oc_tile_size_heuristic(const NCBKernSizeParam& param) const;

dnn/src/fallback/conv_bias/conv1x1/algos_conv1x1_gemv.cpp

@@ -249,7 +249,7 @@ size_t ConvBiasImpl::AlgoConv1x1Gemv::get_oc_tile_size_heuristic(
 }
 
 size_t ConvBiasImpl::AlgoConv1x1Gemv::get_workspace(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_fallback_conv1x1_gemv,
                  midout_iv("AlgoConv1x1Gemv::get_workspace"_hash)) {
         size_t compt_oc_block_size = get_oc_tile_size_heuristic(param);
@@ -265,7 +265,7 @@ size_t ConvBiasImpl::AlgoConv1x1Gemv::get_workspace(
 
 SmallVector<ConvBiasImpl::NCBKern>
 ConvBiasImpl::AlgoConv1x1Gemv::dispatch_kerns(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     SmallVector<ConvBiasImpl::NCBKern> ret_kern;
     size_t OC = param.filter_meta.ocpg;
     size_t compt_oc_block_size = get_oc_tile_size_heuristic(param);
@@ -311,7 +311,7 @@ ConvBiasImpl::AlgoConv1x1Gemv::dispatch_kerns(
     }                                                                      \
     MIDOUT_END()
 
-    switch (opr->param().format) {
+    switch (param.filter_meta.format) {
         case param::ConvBias::Format::NCHW:
             cb1(param::ConvBias::Format::NCHW, dt_float32, dt_float32,
                 PostprocessMode::FLOAT, "NCHW::GEMV::FLOAT"_hash);
@@ -401,18 +401,18 @@ ConvBiasImpl::AlgoConv1x1Gemv::dispatch_kerns(
     return ret_kern;
 }
 
-bool ConvBiasImpl::AlgoConv1x1Gemv::usable(ConvBiasImpl* opr,
-                                           const NCBKernSizeParam& param,
+bool ConvBiasImpl::AlgoConv1x1Gemv::usable(const NCBKernSizeParam& param,
                                            AlgoSelectionStrategy) const {
     MIDOUT_BEGIN(megdnn_fallback_conv1x1_gemv,
                  midout_iv("AlgoConv1x1Gemv::usable"_hash)) {
+        auto format = param.filter_meta.format;
 #if MEGDNN_X86
-        if (opr->param().format != param::ConvBias::Format::NCHW)
+        if (format != param::ConvBias::Format::NCHW)
             return false;
 #elif MEGDNN_AARCH64 || MEGDNN_ARMV7
-        if (opr->param().format != param::ConvBias::Format::NCHW &&
-            opr->param().format != param::ConvBias::Format::NCHW44 &&
-            opr->param().format != param::ConvBias::Format::NCHW44_DOT)
+        if (format != param::ConvBias::Format::NCHW &&
+            format != param::ConvBias::Format::NCHW44 &&
+            format != param::ConvBias::Format::NCHW44_DOT)
             return false;
 #endif
 
@@ -469,13 +469,13 @@ bool ConvBiasImpl::AlgoConv1x1Gemv::usable(ConvBiasImpl* opr,
             return false;
         }
 #if MEGDNN_AARCH64 || MEGDNN_ARMV7
-        if (opr->param().format == param::ConvBias::Format::NCHW44) {
+        if (format == param::ConvBias::Format::NCHW44) {
             if (param.src_type.enumv() != DTypeEnum::Float32 &&
                 param.src_type.enumv() != DTypeEnum::Int8 &&
                 param.src_type.enumv() != DTypeEnum::QuantizedS8) {
                 return false;
             }
-        } else if (opr->param().format == param::ConvBias::Format::NCHW44_DOT) {
+        } else if (format == param::ConvBias::Format::NCHW44_DOT) {
             if (param.src_type.enumv() != DTypeEnum::Int8 &&
                 param.src_type.enumv() != DTypeEnum::QuantizedS8) {
                 return false;
@@ -492,11 +492,11 @@ bool ConvBiasImpl::AlgoConv1x1Gemv::usable(ConvBiasImpl* opr,
 }
 
 bool ConvBiasImpl::AlgoConv1x1Gemv::is_preferred(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_fallback_conv1x1_gemv,
                  midout_iv("AlgoConv1x1Gemv::is_preferred"_hash)) {
 #if (MEGDNN_ARMV7 || MEGDNN_AARCH64)
-        if (opr->param().format == param::ConvBias::Format::NCHW &&
+        if (param.filter_meta.format == param::ConvBias::Format::NCHW &&
             param.src_type.enumv() == DTypeEnum::Quantized8Asymm) {
             return false;
         }

dnn/src/fallback/conv_bias/conv1x1/algos_conv1x1_gemv.h

@@ -24,18 +24,15 @@ public:
 
     bool is_reproducible() const override { return true; }
 
-    const char* name() const override {
-        return "CONV1x1_GEMV";
-    }
+    const char* name() const override { return "CONV1x1_GEMV"; }
 
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     SmallVector<NCBKern> dispatch_kerns(
-            ConvBiasImpl* opr, const NCBKernSizeParam& param) const override;
+            const NCBKernSizeParam& param) const override;
 
-    bool is_preferred(ConvBiasImpl*, const NCBKernSizeParam&) const override;
+    bool is_preferred(const NCBKernSizeParam&) const override;
 
 protected:
     size_t get_oc_tile_size_heuristic(const NCBKernSizeParam& param) const;

dnn/src/fallback/conv_bias/im2col/algos.cpp

@@ -478,7 +478,7 @@ WorkspaceBundle ConvBiasImpl::AlgoIm2col::get_bundle(
 }
 
 size_t ConvBiasImpl::AlgoIm2col::get_workspace(
-        ConvBiasImpl*, const NCBKernSizeParam& p) const {
+        const NCBKernSizeParam& p) const {
     MIDOUT_BEGIN(megdnn_fallback_im2col, 0, 0) {
         return get_bundle(p).total_size_in_bytes();
     }
@@ -487,7 +487,7 @@ size_t ConvBiasImpl::AlgoIm2col::get_workspace(
 }
 
 SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoIm2col::dispatch_kerns(
-        ConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     MIDOUT_BEGIN(megdnn_fallback_im2col, 0, 1) {
         UNPACK_CONV_F32_NCB_KERN_SIZES(param);
         MEGDNN_MARK_USED_VAR(SH);
@@ -660,12 +660,13 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoIm2col::dispatch_kerns(
 }
 
 bool ConvBiasImpl::AlgoIm2col::usable(
-        ConvBiasImpl* opr, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MIDOUT_BEGIN(megdnn_fallback_im2col, 0, 2) {
-        if (opr->param().format != param::ConvBias::Format::NCHW &&
-            opr->param().format != param::ConvBias::Format::NCHW44_DOT &&
-            opr->param().format != param::ConvBias::Format::NCHW44) {
+        auto format = param.filter_meta.format;
+        if (format != param::ConvBias::Format::NCHW &&
+            format != param::ConvBias::Format::NCHW44_DOT &&
+            format != param::ConvBias::Format::NCHW44) {
             return false;
         }
 
@@ -695,8 +696,8 @@ bool ConvBiasImpl::AlgoIm2col::usable(
         }
         fallback::MatrixMulImpl::AlgoBase::MatmulDescription mdesc =
                 m_matmul_algo->matmul_description();
-        if (opr->param().format == param::ConvBias::Format::NCHW44 ||
-            opr->param().format == param::ConvBias::Format::NCHW44_DOT) {
+        if (format == param::ConvBias::Format::NCHW44 ||
+            format == param::ConvBias::Format::NCHW44_DOT) {
             //! current NCHW44 im2col only support DEFAULT mode matmul
             if (mdesc.packmode != Pack_Mode::DEFAULT) {
                 return false;

dnn/src/fallback/conv_bias/im2col/algos.h

@@ -15,6 +15,8 @@
 #include "src/common/utils.h"
 #include "src/fallback/conv_bias/opr_impl.h"
 #include "src/fallback/matrix_mul/opr_impl.h"
+#include "src/common/opr_delegate.h"
+
 
 namespace megdnn {
 namespace fallback {
@@ -54,16 +56,18 @@ public:
         }
         return m_name.c_str();
     }
-    bool usable(ConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(ConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     SmallVector<NCBKern> dispatch_kerns(
-            ConvBiasImpl* opr, const NCBKernSizeParam& param) const override;
-    bool is_preferred(fallback::ConvBiasImpl* opr,
+            const NCBKernSizeParam& param) const override;
+    bool is_preferred(
                       const NCBKernSizeParam& param) const override {
         if (param.src_type.category() == DTypeCategory::QUANTIZED) {
-            return opr->is_matmul_quantized_prefer(param);
+            static CpuOprDelegationStorage<1> storage;
+            auto conv_bias_opr = storage.get<ConvBias, 0>();
+            return static_cast<ConvBiasImpl*>(conv_bias_opr)
+                    ->is_matmul_quantized_prefer(param);
         }
         auto&& fm = param.filter_meta;
         auto OC = fm.ocpg, IC = fm.icpg;

dnn/src/fallback/conv_bias/opr_impl.cpp

@@ -54,7 +54,6 @@ class ConvBiasImpl::AlgoPack : NonCopyableObj {
 
 public:
     AlgoPack() {
-        
         refhold.emplace_back(new AlgoConv1x1Gemv());
         all_algos.emplace_back(refhold.back().get());
 
@@ -121,7 +120,7 @@ bool ConvBiasImpl::is_naive_algo(ConvBiasImpl::Algorithm* algo) {
 }
 
 #define NCB_ALGO_FUNC(name, algo, param) \
-    static_cast<AlgoBase*>(algo)->name(this, param)
+    static_cast<AlgoBase*>(algo)->name(param)
 
 void ConvBiasImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                         _megdnn_tensor_in bias, _megdnn_tensor_in z,
@@ -243,11 +242,10 @@ ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_heuristic_with_ncb(
         const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
         bool reproducible) {
     for (auto i : get_all_algorithms_with_ncb(param)) {
-        size_t need_workspace = NCB_ALGO_FUNC(get_workspace, i, param);
         if (static_cast<AlgoBase*>(i)->usable_reproducible(
-                    this, param, AlgoSelectionStrategy::HEURISTIC,
-                    reproducible) &&
-            need_workspace <= workspace_limit_in_bytes) {
+                    param, AlgoSelectionStrategy::HEURISTIC, reproducible) &&
+            NCB_ALGO_FUNC(get_workspace, i, param) <=
+                    workspace_limit_in_bytes) {
             return i;
         }
     }
@@ -392,8 +390,8 @@ std::vector<ConvBiasImpl::Algorithm*> ConvBiasImpl::get_all_algorithms_with_ncb(
     std::vector<Algorithm*> algos;
     std::vector<Algorithm*> prefer_algos;
     for (auto&& algo : algo_pack()) {
-        if (algo->usable(this, param, AlgoSelectionStrategy::FULL_RUN)) {
-            if (algo->is_preferred(this, param)) {
+        if (algo->usable(param, AlgoSelectionStrategy::FULL_RUN)) {
+            if (algo->is_preferred(param)) {
                 prefer_algos.push_back(algo);
             } else {
                 algos.push_back(algo);

dnn/src/fallback/conv_bias/opr_impl.h

@@ -193,7 +193,7 @@ public:
 
     //! move arm_common to fallback
     virtual bool is_matmul_quantized_prefer(
-            const ConvBiasImpl::NCBKernSizeParam& ncb_param) {
+            const ConvBiasImpl::NCBKernSizeParam& ncb_param) const {
         MEGDNN_MARK_USED_VAR(ncb_param);
         return true;
     };
@@ -209,43 +209,39 @@ public:
     public:
         virtual ~AlgoBase() = default;
         virtual bool usable(
-                ConvBiasImpl* opr, const NCBKernSizeParam& param,
+                const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const = 0;
-        virtual size_t get_workspace(ConvBiasImpl* opr,
-                                     const NCBKernSizeParam& param) const = 0;
+        virtual size_t get_workspace(const NCBKernSizeParam& param) const = 0;
 
         virtual SmallVector<NCBKern> dispatch_kerns(
-                ConvBiasImpl* opr, const NCBKernSizeParam& param) const = 0;
+                const NCBKernSizeParam& param) const = 0;
 
         virtual SmallVector<NCBKern> dispatch_preprocess_kerns(
-                ConvBiasImpl*, const NCBKernSizeParam&) const {
+                const NCBKernSizeParam&) const {
             return {};
         };
 
         //! get the layouts of weight_prerocess dst
         virtual SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
-                ConvBiasImpl*, const NCBKernSizeParam&) const {
+                const NCBKernSizeParam&) const {
             return {};
         };
 
         //! get the workspace when weight_prerocess
-        virtual size_t get_preprocess_workspace(ConvBiasImpl*,
-                                                const NCBKernSizeParam&) const {
+        virtual size_t get_preprocess_workspace(const NCBKernSizeParam&) const {
             return 0_z;
         };
 
         //! Temporarily used to identify whether the matmul algorithm is
         //! is_preferred.
-        virtual bool is_preferred(ConvBiasImpl*,
-                                  const NCBKernSizeParam&) const {
+        virtual bool is_preferred(const NCBKernSizeParam&) const {
             return false;
         }
-        bool usable_reproducible(ConvBiasImpl* opr,
-                                 const NCBKernSizeParam& param,
+        bool usable_reproducible(const NCBKernSizeParam& param,
                                  AlgoSelectionStrategy algo_selection_strategy,
                                  bool reproducible = true) const {
             return (!reproducible || is_reproducible()) &&
-                   usable(opr, param, algo_selection_strategy);
+                   usable(param, algo_selection_strategy);
         }
     };
 

dnn/src/fallback/conv_bias/winograd/winograd.h

@@ -501,9 +501,10 @@ public:
         Strategy strategy = m_strategy;
         SmallVector<NCBKern> kerns;
         auto filter_process_kern =
-                [strategy, bundle, &preprocessed_dst](
+                [strategy, bundle, &preprocessed_dst, this](
                         const NCBKernParam& ncb_param,
                         const NCBKernIndex& ncb_index) mutable {
+                    MEGDNN_MARK_USED_VAR(this);
                     MIDOUT_BEGIN(megdnn_fallback_conv_bias_winograd_common,
                                  midout_iv("filter_preprocess"_hash)) {
                         bundle.set(ncb_param.workspace_ptr);
@@ -569,9 +570,10 @@ public:
              param.filter_meta.format == param::ConvBias::Format::NCHW88 ||
              param.filter_meta.format == param::ConvBias::Format::NCHW44)) {
             auto filter_process_kern =
-                    [strategy = m_strategy, bundle_top, bundle_compute](
+                    [strategy = m_strategy, bundle_top, bundle_compute, this](
                             const NCBKernParam& ncb_param,
                             const NCBKernIndex& ncb_index) mutable {
+                        MEGDNN_MARK_USED_VAR(this);
                         MIDOUT_BEGIN(megdnn_fallback_conv_bias_winograd_common,
                                      midout_iv("filter_process"_hash)) {
                             bundle_top.set(ncb_param.workspace_ptr);
@@ -594,9 +596,10 @@ public:
         }
         auto winograd_compute_kern =
                 [strategy = m_strategy, bundle_top, bundle_compute, matmul_algo,
-                 matmul_param, unit_tile_size,
-                 unit_oc_size](const NCBKernParam& ncb_param,
+                 matmul_param, unit_tile_size, unit_oc_size,
+                 this](const NCBKernParam& ncb_param,
                        const NCBKernIndex& ncb_index) mutable {
+                    MEGDNN_MARK_USED_VAR(this);
                     MIDOUT_BEGIN(megdnn_fallback_conv_bias_winograd_common,
                                  midout_iv("winograd_compute"_hash)) {
                         bundle_top.set(ncb_param.workspace_ptr);
@@ -730,15 +733,15 @@ public:
 
 #define MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(_class, _strategy, _midout_flag,  \
                                             _matmul_format)                   \
-    size_t ConvBiasImpl::_class::get_workspace(                              \
-            fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {  \
+    size_t ConvBiasImpl::_class::get_workspace(const NCBKernSizeParam& param) \
+            const {                                                           \
         MEGDNN_WINOGRADS_ALGO_FUN_DEFINE(_class, get_workspace_size,          \
                                          _strategy, _midout_flag,             \
                                          _matmul_format);                     \
         return 0;                                                             \
     }                                                                         \
     size_t ConvBiasImpl::_class::get_preprocess_workspace(                    \
-            fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {  \
+            const NCBKernSizeParam& param) const {                            \
         MEGDNN_WINOGRADS_ALGO_FUN_DEFINE(                                     \
                 _class, get_preprocess_workspace_size, _strategy,             \
                 _midout_flag, _matmul_format);                                \
@@ -746,7 +749,7 @@ public:
     }                                                                         \
     SmallVector<TensorLayout>                                                 \
     ConvBiasImpl::_class::deduce_preprocessed_filter_layout(                  \
-            fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {  \
+            const NCBKernSizeParam& param) const {                            \
         MEGDNN_WINOGRADS_ALGO_FUN_DEFINE(                                     \
                 _class, deduce_preprocessed_filter_layout, _strategy,         \
                 _midout_flag, _matmul_format);                                \
@@ -754,14 +757,14 @@ public:
     }                                                                         \
     SmallVector<ConvBiasImpl::NCBKern>                                        \
     ConvBiasImpl::_class::dispatch_preprocess_kerns(                          \
-            fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {  \
+            const NCBKernSizeParam& param) const {                            \
         MEGDNN_WINOGRADS_ALGO_FUN_DEFINE(_class, get_preprocess_kerns,        \
                                          _strategy, _midout_flag,             \
                                          _matmul_format);                     \
         return {};                                                            \
     }                                                                         \
     SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::_class::dispatch_kerns(  \
-            fallback::ConvBiasImpl*, const NCBKernSizeParam& param) const {  \
+            const NCBKernSizeParam& param) const {                            \
         MEGDNN_WINOGRADS_ALGO_FUN_DEFINE(_class, get_kerns, _strategy,        \
                                          _midout_flag, _matmul_format);       \
         return {};                                                            \

dnn/src/fallback/convolution/algos.cpp

@@ -164,7 +164,7 @@ void kern_direct(const NCBKernParam& param) {
 /* ===================== fallback algo ===================== */
 
 bool ConvolutionImpl::AlgoFallback::usable(
-        ConvolutionImpl*, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     auto&& fm = param.filter_meta;
     return fm.format == param::Convolution::Format::NCHW &&
@@ -175,7 +175,7 @@ bool ConvolutionImpl::AlgoFallback::usable(
 }
 
 size_t ConvolutionImpl::AlgoFallback::get_workspace(
-        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     auto FH = param.filter_meta.spatial[0], FW = param.filter_meta.spatial[1];
     size_t nr_threads = param.nr_threads;
     if (param.filter_meta.should_flip) {
@@ -190,11 +190,11 @@ size_t ConvolutionImpl::AlgoFallback::get_workspace(
 
 SmallVector<ConvolutionImpl::NCBKern>
 ConvolutionImpl::AlgoFallback::dispatch_kern(
-        ConvolutionImpl* opr, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     size_t group = param.filter_meta.group;
     size_t N = param.n;
     size_t nr_threads = param.nr_threads;
-    size_t workspace_per_thread = get_workspace(opr, param) / nr_threads;
+    size_t workspace_per_thread = get_workspace( param) / nr_threads;
     auto kern_fallback = [workspace_per_thread](const NCBKernParam& p,
                                                 const NCBKernIndex& ncb_index) {
         UNPACK_CONV_F32_NCB_KERN_SIZES(p);
@@ -218,7 +218,7 @@ ConvolutionImpl::AlgoFallback::dispatch_kern(
 /* ===================== naive algo ===================== */
 
 bool ConvolutionImpl::AlgoNaive::usable(
-        ConvolutionImpl*, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     bool ret = false;
 
@@ -241,7 +241,7 @@ bool ConvolutionImpl::AlgoNaive::usable(
 }
 
 SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoNaive::dispatch_kern(
-        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     size_t N = param.n;
     size_t group = param.filter_meta.group;
 #define cb(dt, cmode, compute_type)                                      \
@@ -289,75 +289,42 @@ SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoNaive::dispatch_kern(
 
 /* ===================== default algo ===================== */
 
-ConvolutionImpl::AlgoDefault::AlgoDefault(fallback::ConvBiasImpl* conv_bias_opr,
-                                          ConvBiasImpl::AlgoBase* algorithm)
-        : m_conv_bias_opr(conv_bias_opr), m_algorithm(algorithm) {
+ConvolutionImpl::AlgoDefault::AlgoDefault(ConvBiasImpl::AlgoBase* algorithm)
+        : m_algorithm(algorithm) {
     megdnn_assert_internal(algorithm);
     m_name = ssprintf("CONVOLUTION_DEFAULT_%s", m_algorithm->name());
 }
 
 ConvBiasImpl::NCBKernSizeParam
-ConvolutionImpl::AlgoDefault::AlgoDefault::init_convbias_opr_and_param(
-        ConvBiasImpl* conv_bias_opr, const NCBKernSizeParam& param) {
+ConvolutionImpl::AlgoDefault::init_conv_bias_param(
+        const NCBKernSizeParam& param) {
     DType bias_type = param.dst_type;
     if (bias_type.category() == DTypeCategory::QUANTIZED) {
         bias_type = dtype::QuantizedS32(
                 mul_scale(param.src_type, param.filter_type));
     }
-
-    ::ConvBiasImpl::NCBKernSizeParam conv_bias_size_param(
-            param, 0, param::MatrixMul::Format::DEFAULT, bias_type, 0,
-            BiasMode::NO_BIAS, param::ConvBias::NonlineMode::IDENTITY);
-    // nonline mode
-    conv_bias_opr->param().nonlineMode = conv_bias_size_param.nonlineMode;
-    // convolution mode
-    if (conv_bias_size_param.filter_meta.should_flip) {
-        conv_bias_opr->param().mode = param::ConvolutionV0::Mode::CONVOLUTION;
-    } else {
-        conv_bias_opr->param().mode =
-                param::ConvolutionV0::Mode::CROSS_CORRELATION;
-    }
-    // sparse
-    if (conv_bias_size_param.filter_meta.group > 1) {
-        conv_bias_opr->param().sparse = param::ConvolutionV0::Sparse::GROUP;
-    } else {
-        conv_bias_opr->param().sparse = param::ConvolutionV0::Sparse::DENSE;
-    }
-    // format
-    conv_bias_opr->param().format = conv_bias_size_param.filter_meta.format;
-    // pad stride dilate
-    conv_bias_opr->param().pad_h = conv_bias_size_param.filter_meta.padding[0];
-    conv_bias_opr->param().pad_w = conv_bias_size_param.filter_meta.padding[1];
-    conv_bias_opr->param().stride_h =
-            conv_bias_size_param.filter_meta.stride[0];
-    conv_bias_opr->param().stride_w =
-            conv_bias_size_param.filter_meta.stride[1];
-    conv_bias_opr->param().dilate_h =
-            conv_bias_size_param.filter_meta.dilation[0];
-    conv_bias_opr->param().dilate_w =
-            conv_bias_size_param.filter_meta.dilation[1];
-    // output_block_size
-    conv_bias_opr->param().output_block_size =
-            conv_bias_size_param.output_block_size;
-    // compute_mode
-    conv_bias_opr->param().compute_mode = conv_bias_size_param.compute_mode;
-
-    return conv_bias_size_param;
+    return {param,
+            0,
+            param::MatrixMul::Format::DEFAULT,
+            bias_type,
+            0,
+            BiasMode::NO_BIAS,
+            param::ConvBias::NonlineMode::IDENTITY};
 }
 
 bool ConvolutionImpl::AlgoDefault::is_preferred(
-        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     ::ConvBiasImpl::NCBKernSizeParam conv_bias_param =
-            init_convbias_opr_and_param(m_conv_bias_opr, param);
-    return m_algorithm->is_preferred(m_conv_bias_opr, conv_bias_param);
+            init_conv_bias_param(param);
+    return m_algorithm->is_preferred(conv_bias_param);
 }
 
 bool ConvolutionImpl::AlgoDefault::usable(
-        ConvolutionImpl*, const NCBKernSizeParam& param,
+         const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     ::ConvBiasImpl::NCBKernSizeParam conv_bias_param =
-            init_convbias_opr_and_param(m_conv_bias_opr, param);
-    return m_algorithm->usable(m_conv_bias_opr, conv_bias_param,
+            init_conv_bias_param(param);
+    return m_algorithm->usable(conv_bias_param,
                                static_cast<ConvBiasImpl::AlgoSelectionStrategy>(
                                        algo_selection_strategy));
 }
@@ -365,69 +332,62 @@ bool ConvolutionImpl::AlgoDefault::usable(
 WorkspaceBundle ConvolutionImpl::AlgoDefault::get_bundle(
         const NCBKernSizeParam& param) const {
     ::ConvBiasImpl::NCBKernSizeParam conv_bias_param =
-            init_convbias_opr_and_param(m_conv_bias_opr, param);
-    m_conv_bias_opr->execution_policy() = {m_algorithm};
+            init_conv_bias_param(param);
     return WorkspaceBundle(nullptr, {m_algorithm->get_workspace(
-                                            m_conv_bias_opr, conv_bias_param)});
+                                            conv_bias_param)});
 }
 
 size_t ConvolutionImpl::AlgoDefault::get_workspace(
-        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
 size_t ConvolutionImpl::AlgoDefault::get_preprocess_workspace(
-        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     ::ConvBiasImpl::NCBKernSizeParam conv_bias_param =
-            init_convbias_opr_and_param(m_conv_bias_opr, param);
-    m_conv_bias_opr->execution_policy() = {m_algorithm};
-    return m_algorithm->get_preprocess_workspace(m_conv_bias_opr,
-                                                 conv_bias_param);
+            init_conv_bias_param(param);
+    return m_algorithm->get_preprocess_workspace(conv_bias_param);
 }
 
 SmallVector<TensorLayout>
 ConvolutionImpl::AlgoDefault::deduce_preprocessed_filter_layout(
-        ConvolutionImpl*, const NCBKernSizeParam& param) const {
+         const NCBKernSizeParam& param) const {
     ::ConvBiasImpl::NCBKernSizeParam conv_bias_param =
-            init_convbias_opr_and_param(m_conv_bias_opr, param);
-    m_conv_bias_opr->execution_policy() = {m_algorithm};
-    return m_algorithm->deduce_preprocessed_filter_layout(m_conv_bias_opr,
-                                                          conv_bias_param);
+            init_conv_bias_param( param);
+    return m_algorithm->deduce_preprocessed_filter_layout(conv_bias_param);
 }
 
 //! Return the implement preprocess kernel
 SmallVector<ConvolutionImpl::NCBKern>
 ConvolutionImpl::AlgoDefault::get_preprocess_kimpl(
-        ::ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase* algo,
+         ConvBiasImpl::AlgoBase* algo,
         const NCBKernSizeParam& param) {
     MIDOUT_BEGIN(megdnn_fallback_conv, midout_iv("get_preprocess_kimpl"_hash)) {
         // construct the conv_bias kern param
         ::ConvBiasImpl::NCBKernParam conv_bias_param;
-        ::ConvBiasImpl::NCBKernSizeParam conv_bias_size_param =
-                init_convbias_opr_and_param(conv_bias_opr, param);
         static_cast<::ConvBiasImpl::NCBKernSizeParam&>(conv_bias_param) =
-                conv_bias_size_param;
+                init_conv_bias_param(param);
         auto conv_bias_preprocess_kerns =
-                algo->dispatch_preprocess_kerns(conv_bias_opr, conv_bias_param);
+                algo->dispatch_preprocess_kerns(conv_bias_param);
         SmallVector<ConvolutionImpl::NCBKern> convolution_preprocess_kerns;
 
         //! Set the conv_bias param using convolution param
-        auto set_copy_param_filter_workspace_ptr =
+        auto set_param_filter_workspace_ptr =
                 [](const NCBKernParam& conv_param,
-                   ::ConvBiasImpl::NCBKernParam& copied_param) {
-                    copied_param.filter_ptr = conv_param.filter_ptr;
-                    copied_param.workspace_ptr = conv_param.workspace_ptr;
-                    copied_param.workspace_size = conv_param.workspace_size;
+                   ::ConvBiasImpl::NCBKernParam& conv_bias_param) {
+                    conv_bias_param.filter_ptr = conv_param.filter_ptr;
+                    conv_bias_param.workspace_ptr = conv_param.workspace_ptr;
+                    conv_bias_param.workspace_size = conv_param.workspace_size;
                 };
         for (size_t i = 0; i < conv_bias_preprocess_kerns.size(); i++) {
             auto kernel = conv_bias_preprocess_kerns[i];
             //! If the kerenl batch parallel
-            auto run = [=](const NCBKernParam& p,
-                           const NCBKernIndex& ncb_index) {
-                auto copy_param = conv_bias_param;
-                set_copy_param_filter_workspace_ptr(p, copy_param);
-                kernel.kern(copy_param,
-                            {ncb_index.thread_id, ncb_index.ndrange_id});
+            auto run = [param = conv_bias_param, kernel,
+                        &set_param_filter_workspace_ptr](
+                               const NCBKernParam& p,
+                               const NCBKernIndex& ncb_index) mutable {
+                set_param_filter_workspace_ptr(p, param);
+                kernel.kern(param, {ncb_index.thread_id, ncb_index.ndrange_id});
             };
             convolution_preprocess_kerns.push_back({run, kernel.global_size});
         }
@@ -438,38 +398,35 @@ ConvolutionImpl::AlgoDefault::get_preprocess_kimpl(
 
 //! Return the implement kernel
 SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoDefault::get_kimpl(
-        ::ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase* algo,
+        ConvBiasImpl::AlgoBase* algo,
         const NCBKernSizeParam& param) {
     MIDOUT_BEGIN(megdnn_fallback_conv, midout_iv(0)) {
         // construct the conv_bias kern param
         ::ConvBiasImpl::NCBKernParam conv_bias_param;
-        ::ConvBiasImpl::NCBKernSizeParam conv_bias_size_param =
-                init_convbias_opr_and_param(conv_bias_opr, param);
         static_cast<::ConvBiasImpl::NCBKernSizeParam&>(conv_bias_param) =
-                conv_bias_size_param;
-        auto conv_bias_kerns =
-                algo->dispatch_kerns(conv_bias_opr, conv_bias_param);
+                init_conv_bias_param(param);
+        auto&& conv_bias_kerns = algo->dispatch_kerns(conv_bias_param);
         SmallVector<ConvolutionImpl::NCBKern> convolution_kerns;
 
         //! Set the conv_bias param using convolution param
         auto set_copy_param_compute_address =
                 [](const NCBKernParam& conv_param,
-                   ::ConvBiasImpl::NCBKernParam& copied_param) {
-                    copied_param.src_ptr = conv_param.src_ptr;
-                    copied_param.filter_ptr = conv_param.filter_ptr;
-                    copied_param.dst_ptr = conv_param.dst_ptr;
-                    copied_param.workspace_ptr = conv_param.workspace_ptr;
-                    copied_param.workspace_size = conv_param.workspace_size;
+                   ::ConvBiasImpl::NCBKernParam& conv_bias_param) {
+                    conv_bias_param.src_ptr = conv_param.src_ptr;
+                    conv_bias_param.filter_ptr = conv_param.filter_ptr;
+                    conv_bias_param.dst_ptr = conv_param.dst_ptr;
+                    conv_bias_param.workspace_ptr = conv_param.workspace_ptr;
+                    conv_bias_param.workspace_size = conv_param.workspace_size;
                 };
         for (size_t i = 0; i < conv_bias_kerns.size(); i++) {
-            auto kernel = conv_bias_kerns[i];
+            auto&& kernel = conv_bias_kerns[i];
             //! If the kerenl batch parallel
-            auto run = [=](const NCBKernParam& p,
-                           const NCBKernIndex& ncb_index) {
-                auto copy_param = conv_bias_param;
-                set_copy_param_compute_address(p, copy_param);
-                kernel.kern(copy_param,
-                            {ncb_index.thread_id, ncb_index.ndrange_id});
+            auto run = [param = conv_bias_param, kernel,
+                        &set_copy_param_compute_address](
+                               const NCBKernParam& p,
+                               const NCBKernIndex& ncb_index) mutable {
+                set_copy_param_compute_address(p, param);
+                kernel.kern(param, {ncb_index.thread_id, ncb_index.ndrange_id});
             };
             convolution_kerns.push_back({run, kernel.global_size});
         }

dnn/src/fallback/convolution/algos.h

@@ -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.
  */
 
 #pragma once
@@ -75,14 +76,12 @@ class ConvolutionImpl::AlgoFallback final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "FALLBACK_ALGO"; }
-    bool usable(ConvolutionImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(ConvolutionImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
 
     SmallVector<NCBKern> dispatch_kern(
-            ConvolutionImpl* /*opr*/,
             const NCBKernSizeParam& /*param*/) const override;
 };
 
@@ -90,66 +89,55 @@ class ConvolutionImpl::AlgoNaive final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "NAIVE_ALGO"; }
-    bool usable(ConvolutionImpl* /*opr*/, const NCBKernSizeParam& /*param*/,
+    bool usable(const NCBKernSizeParam& /*param*/,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(ConvolutionImpl*,
-                         const NCBKernSizeParam&) const override {
-        return 0;
-    };
+    size_t get_workspace(const NCBKernSizeParam&) const override { return 0; };
 
     SmallVector<NCBKern> dispatch_kern(
-            ConvolutionImpl* /*opr*/,
             const NCBKernSizeParam& /*param*/) const override;
 };
 
 class ConvolutionImpl::AlgoDefault final : public AlgoBase {
-    static ConvBiasImpl::NCBKernSizeParam init_convbias_opr_and_param(
-            ConvBiasImpl* conv_bias_opr, const NCBKernSizeParam& param);
+    static ConvBiasImpl::NCBKernSizeParam init_conv_bias_param(
+            const NCBKernSizeParam& param);
     WorkspaceBundle get_bundle(const NCBKernSizeParam& param) const;
-    static SmallVector<NCBKern> get_kimpl(ConvBiasImpl* conv_bias_opr,
-                                          ConvBiasImpl::AlgoBase* algo,
+    static SmallVector<NCBKern> get_kimpl(ConvBiasImpl::AlgoBase* algo,
                                           const NCBKernSizeParam& param);
     static SmallVector<NCBKern> get_preprocess_kimpl(
-            ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase* algo,
-            const NCBKernSizeParam& param);
+            ConvBiasImpl::AlgoBase* algo, const NCBKernSizeParam& param);
 
 public:
-    AlgoDefault(fallback::ConvBiasImpl* conv_bias_opr, ConvBiasImpl::AlgoBase*);
+    AlgoDefault(ConvBiasImpl::AlgoBase*);
     bool is_reproducible() const override { return true; }
     const char* name() const override { return m_name.c_str(); }
-    bool usable(ConvolutionImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(ConvolutionImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
 
-    size_t get_preprocess_workspace(ConvolutionImpl*,
-                                    const NCBKernSizeParam&) const override;
+    size_t get_preprocess_workspace(const NCBKernSizeParam&) const override;
 
     SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
-            ConvolutionImpl*, const NCBKernSizeParam&) const override;
+            const NCBKernSizeParam&) const override;
 
     SmallVector<NCBKern> dispatch_preprocess_kern(
-            ConvolutionImpl*, const NCBKernSizeParam& param) const override {
-        return get_preprocess_kimpl(m_conv_bias_opr, m_algorithm, param);
+            const NCBKernSizeParam& param) const override {
+        return get_preprocess_kimpl(m_algorithm, param);
     }
 
     SmallVector<NCBKern> dispatch_kern(
-            ConvolutionImpl* /*opr*/,
             const NCBKernSizeParam& param) const override {
-        return get_kimpl(m_conv_bias_opr, m_algorithm, param);
+        return get_kimpl(m_algorithm, param);
     }
 
     void* type() const override { return sm_fallback_conv_algo_type; }
 
     //! select matmul to the highest preference
-    bool is_preferred(ConvolutionImpl* opr,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 
 private:
     std::string m_name;
-    fallback::ConvBiasImpl* m_conv_bias_opr;
     ConvBiasImpl::AlgoBase* m_algorithm;
 };
 

dnn/src/fallback/convolution/opr_impl.cpp

@@ -59,8 +59,7 @@ public:
                 static_cast<ConvBiasImpl*>(conv_bias_opr)->algo_pack();
         for (auto&& algorithm : conv_bias_algo) {
             // fallback algo
-            refhold.emplace_back(new AlgoDefault(
-                    static_cast<ConvBiasImpl*>(conv_bias_opr), algorithm));
+            refhold.emplace_back(new AlgoDefault(algorithm));
             all_algos.emplace_back(refhold.back().get());
         }
 
@@ -82,7 +81,7 @@ bool ConvolutionImpl::is_naive_algo(ConvolutionImpl::Algorithm* algo) {
 }
 
 #define NCB_ALGO_FUNC(name, algo, param) \
-    static_cast<AlgoBase*>(algo)->name(this, fparam)
+    static_cast<AlgoBase*>(algo)->name(param)
 
 void ConvolutionImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                            _megdnn_tensor_out dst,
@@ -131,7 +130,7 @@ size_t ConvolutionImpl::get_workspace_in_bytes(
         return naive::ConvolutionForwardImpl::get_workspace_in_bytes(
                 src, filter, dst, preprocessed_filter);
     } else {
-        return static_cast<AlgoBase*>(algo)->get_workspace(this, fparam);
+        return NCB_ALGO_FUNC(get_workspace, algo, fparam);
     }
 }
 
@@ -144,8 +143,7 @@ size_t ConvolutionImpl::get_preprocess_workspace_in_bytes(
         return naive::ConvolutionForwardImpl::get_preprocess_workspace_in_bytes(
                 src, filter, dst);
     } else {
-        return static_cast<AlgoBase*>(algo)->get_preprocess_workspace(this,
-                                                                      fparam);
+        return NCB_ALGO_FUNC(get_preprocess_workspace, algo, fparam);
     }
 }
 
@@ -158,8 +156,7 @@ SmallVector<TensorLayout> ConvolutionImpl::deduce_preprocessed_filter_layout(
         return naive::ConvolutionForwardImpl::deduce_preprocessed_filter_layout(
                 src, filter, dst);
     } else {
-        return static_cast<AlgoBase*>(algo)->deduce_preprocessed_filter_layout(
-                this, fparam);
+        return NCB_ALGO_FUNC(deduce_preprocessed_filter_layout, algo, fparam);
     }
 }
 
@@ -251,8 +248,7 @@ ConvolutionImpl::NCBKernParam ConvolutionImpl::make_ncb_kern_param(
 
 void ConvolutionImpl::exec_preprocess_with_ncb_kern(const NCBKernParam& param,
                                                     Algorithm* algo) {
-    auto kerns =
-            static_cast<AlgoBase*>(algo)->dispatch_preprocess_kern(this, param);
+    auto kerns = NCB_ALGO_FUNC(dispatch_preprocess_kern, algo, param);
     auto fallback_handle = handle();
     for (auto kernel : kerns) {
         megdnn_assert(
@@ -272,10 +268,11 @@ void ConvolutionImpl::exec_preprocess_with_ncb_kern(const NCBKernParam& param,
 
 void ConvolutionImpl::exec_with_ncb_kern(const NCBKernParam& param,
                                          Algorithm* algo) {
-    auto kerns = static_cast<AlgoBase*>(algo)->dispatch_kern(this, param);
+    auto kerns = NCB_ALGO_FUNC(dispatch_kern, algo, param);
     auto fallback_handle = handle();
     for (auto kernel : kerns) {
-        megdnn_assert(param.filter_meta.format == Param::Format::NCHW ||
+        megdnn_assert(
+                param.filter_meta.format == Param::Format::NCHW ||
                         param.filter_meta.format == Param::Format::NHWC ||
                         param.filter_meta.format == Param::Format::NCHW88 ||
                         param.filter_meta.format == Param::Format::NCHW44,
@@ -293,13 +290,11 @@ ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm_heuristic_with_ncb(
         const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
         bool reproducible) {
     for (auto i : get_all_algorithms_with_ncb(param)) {
-        size_t need_workspace =
-                static_cast<AlgoBase*>(i)->get_workspace(this, param);
         bool usable_reproducible =
                 static_cast<AlgoBase*>(i)->usable_reproducible(
-                        this, param, AlgoSelectionStrategy::HEURISTIC,
-                        reproducible);
-        if (usable_reproducible && need_workspace <= workspace_limit_in_bytes) {
+                        param, AlgoSelectionStrategy::HEURISTIC, reproducible);
+        if (usable_reproducible && NCB_ALGO_FUNC(get_workspace, i, param) <=
+                                           workspace_limit_in_bytes) {
             return i;
         }
     }
@@ -311,8 +306,8 @@ ConvolutionImpl::get_all_algorithms_with_ncb(const NCBKernSizeParam& param) {
     std::vector<Algorithm*> ret;
     std::vector<Algorithm*> prefer_algos;
     for (auto&& i : algo_pack()) {
-        if (i->usable(this, param, AlgoSelectionStrategy::FULL_RUN)) {
-            if (i->is_preferred(this, param)) {
+        if (i->usable(param, AlgoSelectionStrategy::FULL_RUN)) {
+            if (i->is_preferred(param)) {
                 prefer_algos.push_back(i);
             } else {
                 ret.push_back(i);

dnn/src/fallback/convolution/opr_impl.h

@@ -178,42 +178,38 @@ public:
     class AlgoBase : public Algorithm {
     public:
         virtual ~AlgoBase() = default;
-        virtual bool usable(ConvolutionImpl* opr, const NCBKernSizeParam& param,
+        virtual bool usable(const NCBKernSizeParam& param,
                             AlgoSelectionStrategy) const = 0;
-        virtual size_t get_workspace(ConvolutionImpl* opr,
-                                     const NCBKernSizeParam& param) const = 0;
+        virtual size_t get_workspace(const NCBKernSizeParam& param) const = 0;
         virtual SmallVector<NCBKern> dispatch_kern(
-                ConvolutionImpl* opr, const NCBKernSizeParam& param) const = 0;
+                const NCBKernSizeParam& param) const = 0;
 
         virtual SmallVector<NCBKern> dispatch_preprocess_kern(
-                ConvolutionImpl*, const NCBKernSizeParam&) const {
+                const NCBKernSizeParam&) const {
             return {};
         };
 
         //! get the layouts of weight_prerocess dst
         virtual SmallVector<TensorLayout> deduce_preprocessed_filter_layout(
-                ConvolutionImpl*, const NCBKernSizeParam&) const {
+                const NCBKernSizeParam&) const {
             return {};
         };
 
         //! get the workspace when weight_prerocess
-        virtual size_t get_preprocess_workspace(ConvolutionImpl*,
-                                                const NCBKernSizeParam&) const {
+        virtual size_t get_preprocess_workspace(const NCBKernSizeParam&) const {
             return 0_z;
         };
 
         //! Temporarily used to identify whether the matmul algorithm is
         //! is_preferred.
-        virtual bool is_preferred(ConvolutionImpl*,
-                                  const NCBKernSizeParam&) const {
+        virtual bool is_preferred(const NCBKernSizeParam&) const {
             return false;
         }
-        bool usable_reproducible(ConvolutionImpl* opr,
-                                 const NCBKernSizeParam& param,
+        bool usable_reproducible(const NCBKernSizeParam& param,
                                  AlgoSelectionStrategy algo_selection_strategy,
                                  bool reproducible = true) const {
             return (!reproducible || is_reproducible()) &&
-                   usable(opr, param, algo_selection_strategy);
+                   usable(param, algo_selection_strategy);
         }
     };
 

dnn/src/x86/conv_bias/f32/algos.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/x86/conv_bias/f32/algos.h"
@@ -104,7 +105,7 @@ void get_rectified_size(size_t IH, size_t IW, size_t OH, size_t OW, size_t FH,
 /* ===================== direct algo ===================== */
 
 bool ConvBiasImpl::AlgoDirect::usable(
-        FallbackConvBiasImpl* /*opr*/, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     auto&& fm = param.filter_meta;
     bool aviliable = fm.format == Param::Format::NCHW && fm.spatial_ndim == 2 &&
@@ -142,7 +143,7 @@ WorkspaceBundle ConvBiasImpl::AlgoDirect::get_bundle(
     return {nullptr, {part0, part1}};
 }
 size_t ConvBiasImpl::AlgoDirect::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
@@ -280,7 +281,8 @@ void ConvBiasImpl::AlgoDirect::do_conv_kern(const WorkspaceBundle& bundle,
     size_t workspace_group_id = workspace_ids[0],
            workspace_batch_id = workspace_ids[1], oc = workspace_ids[2];
     const float* sptr = kern_param.src<float>(batch_id, group_id);
-    const float* filter = kern_param.filter<float>(group_id) + oc * FH * FW * IC;
+    const float* filter =
+            kern_param.filter<float>(group_id) + oc * FH * FW * IC;
     const float* bias_ptr =
             kern_param.bias<float>(batch_id, group_id) + oc * bias_offset;
     float* dst = kern_param.dst<float>(batch_id, group_id) + oc * OH * OW;
@@ -318,7 +320,7 @@ SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoDirect::get_kimpls(
 }
 /* ===================== direct-stride2 algo ===================== */
 bool ConvBiasImpl::AlgoDirectStride2::usable(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy algo_selection_strategy) const {
     auto&& fm = param.filter_meta;
     auto FH = fm.spatial[0];
@@ -363,7 +365,7 @@ WorkspaceBundle ConvBiasImpl::AlgoDirectStride2::get_bundle(
 }
 
 size_t ConvBiasImpl::AlgoDirectStride2::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 //! Process one input channel copy padding
@@ -528,7 +530,7 @@ WorkspaceBundle ConvBiasImpl::AlgoMatrixMul::get_bundle(
 }
 
 bool ConvBiasImpl::AlgoMatrixMul::is_preferred(
-        FallbackConvBiasImpl* opr, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     auto&& fm = param.filter_meta;
     if (fm.dilation[0] != 1 || fm.dilation[1] != 1) {
         return false;
@@ -550,7 +552,7 @@ bool ConvBiasImpl::AlgoMatrixMul::is_preferred(
     int ic = find_nearest_elem<int>(fm.icpg, {4, 8, 16, 32, 64, 96, 128});
     int on = std::round(geometric_mean(param.osz[0], param.osz[1]));
     ProfileElement cur(f, oc, ic, on);
-    auto H = static_cast<HandleImpl*>(opr->handle());
+    auto H = static_cast<HandleImpl*>(inplace_cpu_handle().get());
     auto&& target = std::lower_bound(H->profile_cache().begin(),
                                      H->profile_cache().end(), cur);
     megdnn_assert_internal(target->f == cur.f);

dnn/src/x86/conv_bias/f32/algos.h

@@ -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.
  */
 
 #pragma once
@@ -37,14 +38,13 @@ public:
         return m_large_group ? "X86_CONV_BIAS_DIRECT_STRIDE1_LARGE_GROUP"
                              : "X86_CONV_BIAS_DIRECT_STRIDE1_SMALL_GROUP";
     }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
 
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl*,
+
             const NCBKernSizeParam& param) const override {
         return get_kimpls(param);
     }
@@ -74,14 +74,13 @@ public:
         return m_large_group ? "X86_CONV_BIAS_DIRECT_STRIDE2_LARGE_GROUP"
                              : "X86_CONV_BIAS_DIRECT_STRIDE2_SMALL_GROUP";
     }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
 
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl*,
+
             const NCBKernSizeParam& param) const override {
         return get_kimpls(param);
     }
@@ -131,7 +130,7 @@ public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "X86_CONV_BIAS_MATMUL"; }
 
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy) const override {
         auto&& fm = param.filter_meta;
         return fm.format == Param::Format::NCHW && fm.spatial_ndim == 2 &&
@@ -145,15 +144,12 @@ public:
                param.nr_threads == 1_z;
     }
 
-    bool is_preferred(FallbackConvBiasImpl*,
-                      const NCBKernSizeParam&) const override;
+    bool is_preferred(const NCBKernSizeParam&) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl*,
-                         const NCBKernSizeParam& param) const override {
+    size_t get_workspace(const NCBKernSizeParam& param) const override {
         return get_bundle(param).total_size_in_bytes();
     }
     SmallVector<NCBKern> dispatch_kerns(
-            FallbackConvBiasImpl* /*opr*/,
             const NCBKernSizeParam& param) const override {
         size_t group = param.filter_meta.group;
         return {{kimpl, {group, 1_z, 1_z}}};
@@ -171,7 +167,7 @@ public:
     AlgoMkldnnConv() {}
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "MKLDNN_CONV_FP32"; }
-    bool usable(FallbackConvBiasImpl*, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy) const override {
         auto&& fm = param.filter_meta;
 
@@ -184,13 +180,9 @@ public:
         return ok;
     };
 
-    size_t get_workspace(FallbackConvBiasImpl* /*opr*/,
-                         const NCBKernSizeParam&) const override {
-        return 0;
-    }
+    size_t get_workspace(const NCBKernSizeParam&) const override { return 0; }
 
     SmallVector<NCBKern> dispatch_kerns(
-            FallbackConvBiasImpl* /*opr*/,
             const NCBKernSizeParam& /*param*/) const override {
         auto kern = [](const NCBKernParam& param,
                        const NCBKernIndex& ncb_index) {

dnn/src/x86/conv_bias/f32/winograd_algo.cpp

@@ -6,16 +6,17 @@
  *
  * 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/x86/conv_bias/f32/algos.h"
 #include "src/common/utils.h"
+#include "src/x86/conv_bias/f32/algos.h"
+#include "src/x86/conv_bias/f32/strategy.h"
 #include "src/x86/conv_bias/opr_impl.h"
 #include "src/x86/conv_bias/postprocess_helper.h"
 #include "src/x86/handle.h"
 #include "src/x86/profile.h"
-#include "src/x86/conv_bias/f32/strategy.h"
 
 #include "midout.h"
 
@@ -27,10 +28,9 @@ using namespace x86;
 /* ======================= AlgoFP32WinogradF63_8*8 ======================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF63_8x8::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_x86_winograd_fp32, 1, 0) {
         //! TODO: now nchw88 winograd only support Dense mode
         if (param.filter_meta.icpg % 8 != 0 ||
@@ -44,13 +44,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF63_8x8::usable(
                         strategy, m_tile_size, param)
                         .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW88 ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW88 ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW88_WINOGRAD &&
-                 opr->param().output_block_size == 6 &&
+                 param.output_block_size == 6 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK8)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&
@@ -74,10 +74,9 @@ MEGDNN_WINOGRAD_ALGO_FUN_DEFINE_ALL(AlgoFP32WinogradF63_8x8,
 /* ======================= AlgoFP32WinogradF23_8*8 ======================== */
 
 bool ConvBiasImpl::AlgoFP32WinogradF23_8x8::usable(
-        fallback::ConvBiasImpl* opr, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MEGDNN_MARK_USED_VAR(param);
-    MEGDNN_MARK_USED_VAR(opr);
     MIDOUT_BEGIN(megdnn_x86_winograd_fp32, 2, 0) {
         //! TODO: now nchw88 winograd only support Dense mode
         if (param.filter_meta.icpg % 8 != 0 ||
@@ -91,13 +90,13 @@ bool ConvBiasImpl::AlgoFP32WinogradF23_8x8::usable(
                         strategy, m_tile_size, param)
                         .get_matmul_kern_param(param);
         return m_matmul_algo->usable(matmul_param) &&
-               (opr->param().format == param::ConvBias::Format::NCHW88 ||
-                (opr->param().format ==
+               (param.filter_meta.format == param::ConvBias::Format::NCHW88 ||
+                (param.filter_meta.format ==
                          param::ConvBias::Format::NCHW88_WINOGRAD &&
-                 opr->param().output_block_size == 2 &&
+                 param.output_block_size == 2 &&
                  param.winograd_matmul_format ==
                          param::MatrixMul::Format::MK8)) &&
-               opr->param().mode == param::ConvBias::Mode::CROSS_CORRELATION &&
+               !param.filter_meta.should_flip &&
                (param.filter_meta.spatial[0] == param.filter_meta.spatial[1] &&
                 param.filter_meta.spatial[0] == 3) &&
                (param.filter_meta.stride[0] == param.filter_meta.stride[1] &&

dnn/src/x86/conv_bias/int8/algos.cpp

@@ -36,7 +36,7 @@ using namespace megdnn;
 using namespace x86;
 
 bool ConvBiasImpl::AlgoChanWiseAvx2Stride1Qint8::usable(
-        FallbackConvBiasImpl* /*opr*/, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     return chanwise_avx2_stride1_qint8_usable(param);
 }
@@ -66,7 +66,7 @@ WorkspaceBundle ConvBiasImpl::AlgoChanWiseAvx2Stride1Qint8::get_bundle(
 }
 
 size_t ConvBiasImpl::AlgoChanWiseAvx2Stride1Qint8::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
@@ -78,12 +78,12 @@ ConvBiasImpl::AlgoChanWiseAvx2Stride1Qint8::get_kimpls(
 }
 
 bool ConvBiasImpl::AlgoChanWiseAvx2Stride1Qint8::is_preferred(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return chanwise_avx2_stride1_qint8_preferred(param);
 }
 
 bool ConvBiasImpl::AlgoChanWiseAvx2Stride2Qint8::usable(
-        FallbackConvBiasImpl* /*opr*/, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     return chanwise_avx2_stride2_qint8_usable(param);
 }
@@ -113,7 +113,7 @@ WorkspaceBundle ConvBiasImpl::AlgoChanWiseAvx2Stride2Qint8::get_bundle(
 }
 
 size_t ConvBiasImpl::AlgoChanWiseAvx2Stride2Qint8::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
@@ -125,12 +125,12 @@ ConvBiasImpl::AlgoChanWiseAvx2Stride2Qint8::get_kimpls(
 }
 
 bool ConvBiasImpl::AlgoChanWiseAvx2Stride2Qint8::is_preferred(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return chanwise_avx2_stride2_qint8_preferred(param);
 }
 
 bool ConvBiasImpl::AlgoDirectAvx2Stride1Int8::usable(
-        FallbackConvBiasImpl* /*opr*/, const NCBKernSizeParam& param,
+        const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     return direct_avx2_stride1_int8_usable(param);
 }
@@ -170,7 +170,7 @@ WorkspaceBundle ConvBiasImpl::AlgoDirectAvx2Stride1Int8::get_bundle(
 }
 
 size_t ConvBiasImpl::AlgoDirectAvx2Stride1Int8::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
@@ -182,14 +182,13 @@ ConvBiasImpl::AlgoDirectAvx2Stride1Int8::get_kimpls(
 }
 
 bool ConvBiasImpl::AlgoDirectAvx2Stride1Int8::is_preferred(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return direct_avx2_stride1_int8_preferred(param);
 }
 
 /* ===================== avx2 int8 stride 2 ===================== */
 bool ConvBiasImpl::AlgoAVX2DirectConvStride2::usable(
-        FallbackConvBiasImpl* /*opr*/, const NCBKernSizeParam& param,
-        AlgoSelectionStrategy) const {
+        const NCBKernSizeParam& param, AlgoSelectionStrategy) const {
     return direct_avx2_stride2_int8_usable(param);
 }
 
@@ -229,7 +228,7 @@ WorkspaceBundle ConvBiasImpl::AlgoAVX2DirectConvStride2::get_bundle(
 }
 
 size_t ConvBiasImpl::AlgoAVX2DirectConvStride2::get_workspace(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return get_bundle(param).total_size_in_bytes();
 }
 
@@ -241,13 +240,12 @@ ConvBiasImpl::AlgoAVX2DirectConvStride2::get_kimpls(
 }
 
 bool ConvBiasImpl::AlgoAVX2DirectConvStride2::is_preferred(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return direct_avx2_stride2_int8_preferred(param);
 }
 
 #if MEGDNN_X86_WITH_MKL_DNN
-bool ConvBiasImpl::AlgoMkldnnQint8::usable(FallbackConvBiasImpl*,
-                                           const NCBKernSizeParam& param,
+bool ConvBiasImpl::AlgoMkldnnQint8::usable(const NCBKernSizeParam& param,
                                            AlgoSelectionStrategy) const {
     return mkldnn_qint8_usable(param);
 }
@@ -426,19 +424,18 @@ void ConvBiasImpl::AlgoMkldnnQint8::kern_mkldnn_s8x8x32(
 #undef REORDER_MEMORY
 
 bool ConvBiasImpl::AlgoMkldnnQint8::is_preferred(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return mkldnn_qint8_preferred(param);
 }
 
 /* ===================== mkldnn qint8 matmul algo ===================== */
-bool ConvBiasImpl::AlgoMkldnnMatmulQint8::usable(FallbackConvBiasImpl*,
-                                                 const NCBKernSizeParam& param,
+bool ConvBiasImpl::AlgoMkldnnMatmulQint8::usable(const NCBKernSizeParam& param,
                                                  AlgoSelectionStrategy) const {
     return mkldnn_matmul_qint8_usable(param);
 }
 
 bool ConvBiasImpl::AlgoMkldnnMatmulQint8::is_preferred(
-        FallbackConvBiasImpl*, const NCBKernSizeParam& param) const {
+        const NCBKernSizeParam& param) const {
     return mkldnn_matmul_qint8_preferred(param);
 }
 

dnn/src/x86/conv_bias/int8/algos.h

@@ -25,18 +25,15 @@ public:
     const char* name() const override {
         return "X86_CONV_BIAS_CHANWISE_AVX2_INT8_STRIDE1";
     }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl*,
             const NCBKernSizeParam& param) const override {
         return get_kimpls(param);
     }
     void* type() const override;
-    bool is_preferred(FallbackConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 
 /* ===================== avx2 stride2 chanwise algo ===================== */
@@ -49,18 +46,15 @@ public:
     const char* name() const override {
         return "X86_CONV_BIAS_CHANWISE_AVX2_INT8_STRIDE2";
     }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl*,
             const NCBKernSizeParam& param) const override {
         return get_kimpls(param);
     }
     void* type() const override;
-    bool is_preferred(FallbackConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 
 /* ===================== avx2 stride1 direct algo ===================== */
@@ -73,18 +67,15 @@ public:
     const char* name() const override {
         return "X86_CONV_BIAS_DIRECT_AVX2_INT8_STRIDE1";
     }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     virtual SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl*,
             const NCBKernSizeParam& param) const override {
         return get_kimpls(param);
     }
     void* type() const override;
-    bool is_preferred(FallbackConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 
 /* ================== avx2 int8 direct conv stride2 algo ================== */
@@ -97,18 +88,15 @@ public:
     const char* name() const override {
         return "X86_CONV_BIAS_DIRECT_AVX2_INT8_STRIDE2";
     }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy algo_selection_strategy) const override;
-    size_t get_workspace(FallbackConvBiasImpl* opr,
-                         const NCBKernSizeParam& param) const override;
+    size_t get_workspace(const NCBKernSizeParam& param) const override;
     SmallVector<NCBKern> dispatch_kerns(
-            fallback::ConvBiasImpl*,
             const NCBKernSizeParam& param) const override {
         return get_kimpls(param);
     }
     void* type() const override;
-    bool is_preferred(FallbackConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 
 #if MEGDNN_X86_WITH_MKL_DNN
@@ -122,16 +110,14 @@ public:
     AlgoMkldnnQint8() {}
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "MKLDNN_INT8"; }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl* /*opr*/,
-                         const NCBKernSizeParam& param) const override {
+    size_t get_workspace(const NCBKernSizeParam& param) const override {
         size_t nr_threads = param.nr_threads;
         return get_bundle(param).total_size_in_bytes() * nr_threads;
     }
     SmallVector<NCBKern> dispatch_kerns(
-            FallbackConvBiasImpl* /*opr*/,
             const NCBKernSizeParam& param) const override {
         size_t group = param.filter_meta.group;
         size_t n = param.n;
@@ -147,8 +133,7 @@ public:
         return {{kern, {group, n, 1_z}}};
     }
     void* type() const override;
-    bool is_preferred(FallbackConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 };
 /* ===================== mkldnn qint8 matmul algo ===================== */
 class ConvBiasImpl::AlgoMkldnnMatmulQint8 final : public AlgoBase {
@@ -160,22 +145,19 @@ class ConvBiasImpl::AlgoMkldnnMatmulQint8 final : public AlgoBase {
 public:
     bool is_reproducible() const override { return true; }
     const char* name() const override { return "MKLDNN_MATMUL_INT8"; }
-    bool usable(FallbackConvBiasImpl* opr, const NCBKernSizeParam& param,
+    bool usable(const NCBKernSizeParam& param,
                 AlgoSelectionStrategy) const override;
 
-    size_t get_workspace(FallbackConvBiasImpl* /*opr*/,
-                         const NCBKernSizeParam& param) const override {
+    size_t get_workspace(const NCBKernSizeParam& param) const override {
         return get_bundle(param).total_size_in_bytes();
     }
     SmallVector<NCBKern> dispatch_kerns(
-            FallbackConvBiasImpl* /*opr*/,
             const NCBKernSizeParam& param) const override {
         size_t group = param.filter_meta.group;
         return {{kern_mkldnn_matmul_s8x8x32, {group, 1_z, 1_z}}};
     }
     //! select matmul to the highest preference
-    bool is_preferred(FallbackConvBiasImpl*,
-                      const NCBKernSizeParam& param) const override;
+    bool is_preferred(const NCBKernSizeParam& param) const override;
 
     void* type() const override;
 };

dnn/src/x86/conv_bias/opr_impl.cpp

@@ -163,7 +163,7 @@ const char* ConvBiasImpl::get_algorithm_set_name() const {
 }
 
 bool ConvBiasImpl::is_matmul_quantized_prefer(
-        const ConvBiasImpl::NCBKernSizeParam& param) {
+        const ConvBiasImpl::NCBKernSizeParam& param) const {
     bool conv_direct_chanwise_mkldnn_usable = true;
     if (param.dst_type.enumv() == DTypeEnum::QuantizedS8 ||
         param.dst_type.enumv() == DTypeEnum::QuantizedS32) {

dnn/src/x86/conv_bias/opr_impl.h

@@ -55,7 +55,7 @@ public:
     const char* get_algorithm_set_name() const override;
 
     bool is_matmul_quantized_prefer(
-            const ConvBiasImpl::NCBKernSizeParam& ncb_param) override;
+            const ConvBiasImpl::NCBKernSizeParam& ncb_param) const override;
 };
 
 }  // namespace x86