[Prim][PIR] softmax forward sink (PaddlePaddle#58591)

* relu forward sink

* softmax sink pir

* remove relu

* pir relu sink

* remove softmax register

* remove python softmax

* update code

* merge code

---------

Co-authored-by: kevincheng2 <cheng112@gmail.com>

paddle/fluid/pir/dialect/op_generator/decomp_interface_gen_op_list.py

@@ -19,7 +19,13 @@
 
 # come into effect in generated file pd_op.h
 # manual decomp interface declare are located in manual_op.h
-decomp_interface_declare_gen_op_list = ["mean", "squeeze", "add_n", "relu"]
+decomp_interface_declare_gen_op_list = [
+    "mean",
+    "squeeze",
+    "add_n",
+    "relu",
+    "softmax",
+]
 
 # come into effect in generated file op_decomp.cc
 # manual decomp interface implementation are located in manual_op_decomp.cc
@@ -28,4 +34,5 @@
     "squeeze",
     "add_n",
     "relu",
+    "softmax",
 ]

paddle/fluid/primitive/composite/composite.h

@@ -62,6 +62,31 @@ Tensor mean_decomp(const Tensor& x, const IntArray& axis, bool keepdim) {
   }
 }
 
+template <typename T>
+Tensor softmax_decomp(const Tensor& x, const int& axis) {
+  auto org_dtype = x.dtype();
+  auto x_tmp = x;
+  auto axis_tmp = IntArray({axis});
+
+  bool need_cast =
+      org_dtype == phi::DataType::FLOAT16 || org_dtype == phi::DataType::UINT16;
+  if (need_cast) {
+    x_tmp = cast<T>(x, phi::DataType::FLOAT32);
+  }
+
+  auto max_tmp = max<T>(x_tmp, axis_tmp, true);
+  auto molecular = exp<T>(subtract<T>(x_tmp, max_tmp));
+
+  auto denominator = sum<T>(molecular, axis_tmp, molecular.dtype(), true);
+  auto res = divide<T>(molecular, denominator);
+
+  if (need_cast) {
+    return cast<T>(res, org_dtype);
+  } else {
+    return res;
+  }
+}
+
 template <typename T>
 Tensor relu_decomp(const Tensor& x) {
   return maximum<T>(x, full<T>(phi::vectorize(x.dims()), 0.0, x.dtype()));

python/paddle/decomposition/rules.py

@@ -248,33 +248,6 @@ def silu(x):
     return res if not is_amp else cast(res, dtype)
 
 
-@register_decomp('pd_op.softmax')
-def softmax(x, axis):
-    """define composite rule of op softmax"""
-    is_amp = False
-    from paddle.base.data_feeder import convert_dtype
-
-    # Softmax need fp32 compute since it has sum op in
-    dtype = convert_dtype(x.dtype)
-    if dtype in ["float16", "uint16"]:
-        is_amp = True
-        x = cast(x, "float32")
-    if not x.shape:
-        # do not return 1, to ensure gradients
-        res = exp(x - x)
-        if is_amp:
-            res = cast(res, "float16")
-        return res
-    max_temp = max(x, axis, keepdim=True)
-    max_temp.stop_gradient = True
-    molecular = exp(x - max_temp)
-    denominator = sum(molecular, axis=axis, keepdim=True)
-    res = divide(molecular, denominator)
-    if is_amp:
-        res = cast(res, dtype)
-    return res
-
-
 @register_decomp('pd_op.full_like')
 def full_like(x, fill_value, dtype, place=None):
     """define composite rule of op full_like."""