[Prim][PIR] gelu forward sink (#58981)

kevincheng2 · web-flow · commit 3aab90132c96 · 2023-11-20T10:28:16.000+08:00
* prim gelu op sink

* prim gelu op sink

* update code

* pir gelu sink c++

* pir gelu sink c++

* process accuracy

* adapter windows

* adapter windows

* adapter windows
diff --git a/paddle/fluid/pir/dialect/op_generator/decomp_interface_gen_op_list.py b/paddle/fluid/pir/dialect/op_generator/decomp_interface_gen_op_list.py
@@ -25,6 +25,7 @@
     "relu",
     "softmax",
     "layer_norm",
+    "gelu",
 ]
 
 # come into effect in generated file op_decomp.cc
@@ -36,6 +37,7 @@
     "relu",
     "softmax",
     "layer_norm",
+    "gelu",
 ]
 
 
diff --git a/paddle/fluid/primitive/composite/composite.h b/paddle/fluid/primitive/composite/composite.h
@@ -188,6 +188,35 @@ std::tuple<Tensor, Tensor, Tensor> layer_norm_decomp(
   return std::make_tuple(out, mean_, variance);
 }
 
+template <typename T>
+Tensor gelu_decomp(const Tensor& x, bool approximate) {
+  const double PM_2_SQRTPI = 1.12837916709551257390; /* 2/sqrt(pi) */
+  const double PM_SQRT1_2 = 0.70710678118654752440;  /* 1/sqrt(2) */
+
+  auto org_dtype = x.dtype();
+  auto half = full<T>(phi::vectorize(x.dims()), 0.5, org_dtype);
+  auto one = full<T>(phi::vectorize(x.dims()), 1.0, org_dtype);
+  if (approximate) {
+    // gelu(x) = 0.5 * x * (1 + tanh(sqrt(2 / \pi) * (x + 0.044715 * x^{3})))
+    auto kAlpha =
+        full<T>(phi::vectorize(x.dims()), PM_2_SQRTPI * PM_SQRT1_2, org_dtype);
+    auto GELU_CONSTANT = full<T>(phi::vectorize(x.dims()), 0.044715, org_dtype);
+    auto x_pow3 =
+        elementwise_pow<T>(x, full<T>(phi::vectorize(x.dims()), 3, org_dtype));
+    auto tanh_out = tanh<T>(kAlpha * (x + x_pow3 * GELU_CONSTANT));
+
+    auto res = x * half * (one + tanh_out);
+    return res;
+  } else {
+    // gelu(x) = 0.5 * x *  (1 + erf(x / sqrt(2)))
+    auto M_SQRT1_2T = full<T>(phi::vectorize(x.dims()), PM_SQRT1_2, org_dtype);
+    auto erf_out = one + erf<T>(x * M_SQRT1_2T);
+
+    auto res = x * half * erf_out;
+    return res;
+  }
+}
+
 }  // namespace details
 
 }  // namespace primitive
diff --git a/python/paddle/decomposition/rules.py b/python/paddle/decomposition/rules.py
@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from paddle import _pir_ops
 
 from .primitives import *  # noqa: F403
 from .register import register_decomp
@@ -37,31 +36,6 @@ def mean(x, axis, keepdim):
     return res
 
 
-@register_decomp('pd_op.gelu')
-def gelu(x, approximate):
-    """define composite rule of op gelu"""
-    M_SQRT1_2 = (
-        0.70710678118654752440  # /* 1/sqrt(2) */ copy from gelu-kernel.cc
-    )
-    M_2_SQRTPI = 1.12837916709551257390  # /* 2/sqrt(pi) */
-    full_shape = x.shape if len(x.shape) == 0 else [1]
-    one = ones(full_shape, x.dtype)
-    half = full(full_shape, 0.5, x.dtype)
-    # Todo(cz): after symbol overload, add and multiply will be replaced by "+" and "*"
-    if approximate:
-        # gelu(x) = 0.5 * x * (1 + tanh(sqrt(2 / \pi) * (x + 0.044715 * x^{3})))
-        kAlpha = full(full_shape, M_2_SQRTPI * M_SQRT1_2, x.dtype)
-        GELU_CONSTANT = full(full_shape, 0.044715, x.dtype)
-        tanh_out = tanh(kAlpha * (x + GELU_CONSTANT * x * x * x))
-        out = x * half * (one + tanh_out)
-        return out
-    else:
-        # gelu(x) = 0.5 * x *  (1 + erf(x / sqrt(2)))
-        cdf = half * (one + _pir_ops.erf(x * full(x.shape, M_SQRT1_2, x.dtype)))
-        out = x * cdf
-        return out
-
-
 @register_decomp('pd_op.sqrt')
 def sqrt(x):
     """
diff --git a/test/legacy_test/test_activation_op.py b/test/legacy_test/test_activation_op.py
@@ -2691,6 +2691,8 @@ def setUp(self):
         self.public_python_api = paddle.nn.functional.gelu
         self.init_dtype()
         self.init_shape()
+        # Todo: Under float64, only this accuracy is currently supported, for further processing
+        self.fw_comp_rtol = 1e-7
         approximate = False
         np.random.seed(2048)
         x = np.random.uniform(-1, 1, self.shape).astype(self.dtype)
@@ -2713,7 +2715,7 @@ def if_enable_cinn(self):
         pass
 
     def test_check_output(self):
-        self.check_output(check_prim=True, check_pir=True, check_prim_pir=False)
+        self.check_output(check_prim=True, check_pir=True, check_prim_pir=True)
 
     def test_check_grad(self):
         if self.dtype == np.float16:
diff --git a/test/prim/pir_prim/test_sink_decomp.py b/test/prim/pir_prim/test_sink_decomp.py
@@ -149,5 +149,50 @@ def test_relu_forward(self):
             np.testing.assert_equal(ref, actual)
 
 
+class TestGeluSink(unittest.TestCase):
+    def setUp(self):
+        np.random.seed(2023)
+        self.shape_x = [8, 16, 32, 64]
+        self.x = np.random.random(self.shape_x).astype("float32")
+        self.prog = None
+
+    def base_net(self, approximate=True, flag=None):
+        if flag == "forward":
+            core._set_prim_forward_enabled(True)
+        main_program = paddle.static.Program()
+        with paddle.static.program_guard(main_program):
+            x = paddle.static.data('x', self.shape_x, dtype='float32')
+            x.stop_gradient = False
+            sum_out = F.gelu(x, approximate=approximate)
+            [new_out] = decompose(main_program, [sum_out])
+            gradients = grad(new_out, x)
+
+            exe = paddle.static.Executor()
+            [fwd, dx] = exe.run(
+                feed={'x': self.x}, fetch_list=[new_out, gradients]
+            )
+
+        whole_ops = [op.name() for op in main_program.global_block().ops]
+        self.prog = main_program
+        if flag == "forward":
+            core._set_prim_forward_enabled(False)
+            assert 'pd_op.gelu' not in whole_ops
+        else:
+            assert 'pd_op.gelu' in whole_ops
+        return fwd, dx
+
+    def test_gelu_forward_true(self):
+        res_ref = self.base_net(approximate=True)
+        res = self.base_net(approximate=True, flag="forward")
+        for ref, actual in zip(res_ref, res):
+            np.testing.assert_allclose(ref, actual, rtol=1e-6)
+
+    def test_gelu_approximate_false(self):
+        res_ref = self.base_net(approximate=False)
+        res = self.base_net(approximate=False, flag="forward")
+        for ref, actual in zip(res_ref, res):
+            np.testing.assert_allclose(ref, actual, rtol=1e-6)
+
+
 if __name__ == "__main__":
     unittest.main()

Original file line number	Diff line number	Diff line change
`@@ -25,6 +25,7 @@`
`25`	`25`	`"relu",`
`26`	`26`	`"softmax",`
`27`	`27`	`"layer_norm",`
	`28`	`+ "gelu",`
`28`	`29`	`]`
`29`	`30`
`30`	`31`	`# come into effect in generated file op_decomp.cc`
`@@ -36,6 +37,7 @@`
`36`	`37`	`"relu",`
`37`	`38`	`"softmax",`
`38`	`39`	`"layer_norm",`
	`40`	`+ "gelu",`
`39`	`41`	`]`
`40`	`42`
`41`	`43`