chore(mgblar): delete winograd transform switch

GitOrigin-RevId: aa4051541ddfe7ff0cb0ba8fa50526acbc919fb1

sdk/load-and-run/src/mgblar.cpp

@@ -190,10 +190,6 @@ R"__usage__(
   --enable-jit
     Execute supported operators with JIT(now only support NVRTC). Can only be used on Nvidia GPUs.
 )__usage__"
-R"__usage__(
-  --winograd-transform
-    Execute opr replace, replace weights by winograd transform. Currently support on conv bias opr.
-)__usage__"
 R"__usage__(
   --enable-chwn4
     Execute operators with kernels implemented in MegDNN with CHWN4 tensor format. Can only be used
@@ -1391,11 +1387,6 @@ Args Args::from_argv(int argc, char **argv) {
             graph_opt.graph_opt.jit = 1;
             continue;
         }
-        if (!strcmp(argv[i], "--winograd-transform")) {
-            mgb_log_warn("enable winograd transform");
-            graph_opt.graph_opt.weight_winograd_transform = true;
-            continue;
-        }
         if (!strcmp(argv[i], "--weight-preprocess")) {
             mgb_log_warn("enable weight-preprocess optimization");
             graph_opt.graph_opt.enable_weight_preprocess();

src/core/include/megbrain/graph/cg.h

@@ -95,8 +95,6 @@ struct GraphCommonOptimizeOptions {
     //! fuse pattern like ReLU(conv_bias(x, w, b) + z) or conv_bias(x, w, b)
     //! + z -> conv_bias(x, w, b, z)
     bool fuse_conv_bias_with_z = false;
-    //! whether to enable fast-run profiled winograd opr replace
-    bool weight_winograd_transform = false;
     //! whether to enable weight preprocess, if enabled it may use more
     //! memory, default disable now, when weight preprocess is enabled, the
     //! input shape should no change
@@ -133,7 +131,6 @@ struct GraphCommonOptimizeOptions {
     SET(fuse_conv_bias_nonlinearity);
     SET(fuse_conv_bias_with_z);
     SET(fuse_preprocess);
-    SET(weight_winograd_transform);
     SET(weight_preprocess);
 #undef SET
 #define SET(_trans, _trans_capital)                                 \