[cherry-pick][bug fix] cp pr64879 and pr64859 (#64897)

* Fix match constraint of matmul_add_act_fuse_pass (#64879)

* cp pr64859

---------

Co-authored-by: Yuanle Liu <yuanlehome@163.com>

Author: bukejiyu

.gitignore

@@ -106,3 +106,7 @@ paddle/phi/kernels/fusion/cutlass/cutlass_kernels/fpA_intB_gemm/autogen_tmp/*
 paddle/fluid/pybind/static_op_function.*
 paddle/fluid/pybind/ops_api.cc
 python/paddle/tensor/tensor.pyi
+paddle/phi/kernels/fusion/cutlass/conv2d/build
+paddle/phi/kernels/fusion/cutlass/conv2d/cutlass
+paddle/phi/kernels/fusion/cutlass/gemm_epilogue/build
+paddle/phi/kernels/fusion/cutlass/gemm_epilogue/cutlass

paddle/fluid/pir/transforms/gpu/matmul_add_act_fuse_pass.cc

@@ -60,11 +60,17 @@ class MatmulAddPattern : public paddle::drr::DrrPatternBase {
 
     pat.AddConstraint([&](const paddle::drr::MatchContext &match_ctx) {
       auto w_dtype = pir::GetDataTypeFromValue(match_ctx.Tensor("w"));
-      if (!w_dtype.isa<pir::Float16Type>() &&
-          !w_dtype.isa<pir::BFloat16Type>() &&
-          !w_dtype.isa<pir::Float32Type>() &&
-          !w_dtype.isa<pir::Float64Type>()) {
-        return false;
+      if (fused_op_name_ == paddle::dialect::GemmEpilogueOp::name()) {
+        if (!w_dtype.isa<pir::Float16Type>() &&
+            !w_dtype.isa<pir::BFloat16Type>()) {
+          return false;
+        }
+      } else {
+        if (!w_dtype.isa<pir::Float16Type>() &&
+            !w_dtype.isa<pir::Float32Type>() &&
+            !w_dtype.isa<pir::Float64Type>()) {
+          return false;
+        }
       }
       auto w_dims = pir::GetShapeFromValue(match_ctx.Tensor("w"));
       auto x_dims = pir::GetShapeFromValue(match_ctx.Tensor("x"));

paddle/phi/kernels/fusion/cutlass/gemm_epilogue_kernel.cu

@@ -204,6 +204,5 @@ PD_REGISTER_KERNEL(gemm_epilogue,
                    GPU,
                    ALL_LAYOUT,
                    phi::fusion::cutlass_internal::GemmEpilogueKernel,
-                   float,
                    phi::dtype::bfloat16,
                    phi::dtype::float16) {}