[Paddle inference]conv2d_add_act_pass support cutlass

paddle/fluid/inference/api/analysis_predictor.cc

@@ -945,11 +945,7 @@ bool AnalysisPredictor::PrepareExecutor() {
         // gpu
         if (!config_.custom_pass_only_) {
           for (const auto &gpu_pass : kPirGpuPasses) {
-            auto pass = pir::PassRegistry::Instance().Get(gpu_pass);
-            if (pass->name() == "matmul_add_act_fuse_pass") {
-              pass->Set("use_cutlass", new bool(config_.use_cutlass_));
-            }
-            pass_pm.AddPass(std::move(pass));
+            pass_pm.AddPass(pir::PassRegistry::Instance().Get(gpu_pass));
           }
         }
 
@@ -990,6 +986,14 @@ bool AnalysisPredictor::PrepareExecutor() {
             std::make_unique<pir::PassManager::IRPrinterOption>(
                 ir_printing_conditions, ir_printing_conditions));
       }
+      // set attr
+      for (const auto &pass : pass_pm.passes()) {
+        if (pass->name() == "matmul_add_act_fuse_pass" ||
+            pass->name() == "conv2d_add_act_fuse_pass" ||
+            pass->name() == "conv2d_add_fuse_pass") {
+          pass->Set("use_cutlass", new bool(config_.use_cutlass_));
+        }
+      }
       pass_pm.Run(pir_program_.get());
 
       // Apply some basic passes required by the framework

paddle/fluid/pir/dialect/operator/ir/op_attribute.h

@@ -23,6 +23,9 @@
 
 namespace paddle {
 namespace dialect {
+// __force_backend__ in ["gpu","gpudnn","cpu",""]
+inline const char kForceBackendAttr[] = "__force_backend__";
+
 class IntArrayAttribute : public pir::Attribute {
  public:
   using Attribute::Attribute;

paddle/fluid/pir/transforms/general/remove_redundant_transpose_pass.cc

@@ -55,6 +55,32 @@ class RemoveRedundantTransposePattern : public paddle::drr::DrrPatternBase {
   }
 };
 
+class RemoveInvalidTransposePattern : public paddle::drr::DrrPatternBase {
+ public:
+  std::string name() const override { return "RemoveInvalidTransposePattern"; }
+  uint32_t benefit() const override { return 1; }
+
+  void operator()(paddle::drr::DrrPatternContext *ctx) const override {
+    paddle::drr::SourcePattern pat = ctx->SourcePattern();
+    const auto &transpose =
+        pat.Op("pd_op.transpose", {{"perm", pat.Attr("perm")}});
+    pat.Tensor("ret") = transpose(pat.Tensor("arg_transpose"));
+    pat.AddConstraint([this](const paddle::drr::MatchContext &match_ctx) {
+      const auto &perm = match_ctx.Attr<std::vector<int>>("perm");
+      std::vector<int> dst_vector(perm.size());
+      std::iota(dst_vector.begin(), dst_vector.end(), 0);
+      for (size_t i = 0; i < perm.size(); i++) {
+        if (perm[i] != dst_vector[i]) {
+          return false;
+        }
+      }
+      return true;
+    });
+    paddle::drr::ResultPattern res = pat.ResultPattern();
+    res.Tensor("ret").Assign(res.Tensor("arg_transpose"));
+  }
+};
+
 class RemoveRedundantTransposePass : public pir::PatternRewritePass {
  public:
   RemoveRedundantTransposePass()
@@ -63,6 +89,7 @@ class RemoveRedundantTransposePass : public pir::PatternRewritePass {
   pir::RewritePatternSet InitializePatterns(pir::IrContext *context) override {
     pir::RewritePatternSet ps(context);
     ps.Add(paddle::drr::Create<RemoveRedundantTransposePattern>(context));
+    ps.Add(paddle::drr::Create<RemoveInvalidTransposePattern>(context));
     return ps;
   }
 };