[METAL] Fix codegen for inf and erf (apache#8054)

* [METAL] Fix codegen for inf and erf Fixed Metal codegen with using `inf` constant. Constant `INFINITY` is used now instead of `inf`. Also, Metal doesn't have `erf` built-in function. So, we are using `fast_erf` from tir. User will see warning message when we will generate `fast_erf` instead of `erf`. * Apply comments * Fix clang-format * Fix lint
neo-ai · Jun 17, 2021 · d02f9a0 · d02f9a0
1 parent a39009e
commit d02f9a0
Show file tree

Hide file tree

Showing 5 changed files with 168 additions and 42 deletions.
diff --git a/include/tvm/topi/elemwise.h b/include/tvm/topi/elemwise.h
@@ -456,54 +456,60 @@ inline Tensor fast_exp(const Tensor& x, std::string name = "T_fast_exp",
 }
 
 /*!
- * \brief Fast_tanh_float implementation from Eigen
+ * \brief Fast_erf_float expression from Eigen
  * https://github.com/eigenteam/eigen-git-mirror/blob/master/unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsImpl.h#L290
+ * \param arg The input expression.
+ * \param bits The number of bits in the type.
  */
-inline Tensor fast_erf_float32(const Tensor& data, std::string name, std::string tag) {
-  auto plus_4 = make_const(DataType::Float(32), 4.f);
-  auto minus_4 = make_const(DataType::Float(32), -4.f);
+inline PrimExpr fast_erf_float_expr(PrimExpr arg, int bits) {
+  auto plus_4 = make_const(DataType::Float(bits), 4.f);
+  auto minus_4 = make_const(DataType::Float(bits), -4.f);
 
   // The monomial coefficients of the numerator polynomial (odd).
-  auto alpha_1 = make_const(DataType::Float(32), -1.60960333262415e-02f);
-  auto alpha_3 = make_const(DataType::Float(32), -2.95459980854025e-03f);
-  auto alpha_5 = make_const(DataType::Float(32), -7.34990630326855e-04f);
-  auto alpha_7 = make_const(DataType::Float(32), -5.69250639462346e-05f);
-  auto alpha_9 = make_const(DataType::Float(32), -2.10102402082508e-06f);
-  auto alpha_11 = make_const(DataType::Float(32), 2.77068142495902e-08f);
-  auto alpha_13 = make_const(DataType::Float(32), -2.72614225801306e-10f);
+  auto alpha_1 = make_const(DataType::Float(bits), -1.60960333262415e-02f);
+  auto alpha_3 = make_const(DataType::Float(bits), -2.95459980854025e-03f);
+  auto alpha_5 = make_const(DataType::Float(bits), -7.34990630326855e-04f);
+  auto alpha_7 = make_const(DataType::Float(bits), -5.69250639462346e-05f);
+  auto alpha_9 = make_const(DataType::Float(bits), -2.10102402082508e-06f);
+  auto alpha_11 = make_const(DataType::Float(bits), 2.77068142495902e-08f);
+  auto alpha_13 = make_const(DataType::Float(bits), -2.72614225801306e-10f);
 
   // The monomial coefficients of the denominator polynomial (even).
-  auto beta_0 = make_const(DataType::Float(32), -1.42647390514189e-02f);
-  auto beta_2 = make_const(DataType::Float(32), -7.37332916720468e-03f);
-  auto beta_4 = make_const(DataType::Float(32), -1.68282697438203e-03f);
-  auto beta_6 = make_const(DataType::Float(32), -2.13374055278905e-04f);
-  auto beta_8 = make_const(DataType::Float(32), -1.45660718464996e-05f);
+  auto beta_0 = make_const(DataType::Float(bits), -1.42647390514189e-02f);
+  auto beta_2 = make_const(DataType::Float(bits), -7.37332916720468e-03f);
+  auto beta_4 = make_const(DataType::Float(bits), -1.68282697438203e-03f);
+  auto beta_6 = make_const(DataType::Float(bits), -2.13374055278905e-04f);
+  auto beta_8 = make_const(DataType::Float(bits), -1.45660718464996e-05f);
+
+  // clamp x
+  auto x = tvm::max(tvm::min(arg, plus_4), minus_4);
+  auto x2 = x * x;
+
+  // Evaluate the numerator polynomial p.
+  auto p = x2 * alpha_13 + alpha_11;
+  p = x2 * p + alpha_9;
+  p = x2 * p + alpha_7;
+  p = x2 * p + alpha_5;
+  p = x2 * p + alpha_3;
+  p = x2 * p + alpha_1;
+  p = x * p;
+
+  // Evaluate the denominator polynomial p.
+  auto q = x2 * beta_8 + beta_6;
+  q = x2 * q + beta_4;
+  q = x2 * q + beta_2;
+  q = x2 * q + beta_0;
+
+  return p / q;
+}
 
+/*!
+ * \brief Fast_erf_float expression from Eigen
+ */
+inline Tensor fast_erf_float32(const Tensor& data, std::string name, std::string tag) {
   return compute(
-      data->shape,
-      [&](const Array<Var>& i) {
-        // clamp x
-        auto x = tvm::max(tvm::min(data(i), plus_4), minus_4);
-        auto x2 = x * x;
-
-        // Evaluate the numerator polynomial p.
-        auto p = x2 * alpha_13 + alpha_11;
-        p = x2 * p + alpha_9;
-        p = x2 * p + alpha_7;
-        p = x2 * p + alpha_5;
-        p = x2 * p + alpha_3;
-        p = x2 * p + alpha_1;
-        p = x * p;
-
-        // Evaluate the denominator polynomial p.
-        auto q = x2 * beta_8 + beta_6;
-        q = x2 * q + beta_4;
-        q = x2 * q + beta_2;
-        q = x2 * q + beta_0;
-
-        return p / q;
-      },
-      name, tag);
+      data->shape, [&](const Array<Var>& i) { return fast_erf_float_expr(data(i), 32); }, name,
+      tag);
 }
 
 /*!

diff --git a/src/target/source/codegen_metal.cc b/src/target/source/codegen_metal.cc
@@ -302,6 +302,26 @@ void CodeGenMetal::VisitExpr_(const CallNode* op, std::ostream& os) {  // NOLINT
   }
 }
 
+void CodeGenMetal::VisitExpr_(const FloatImmNode* op, std::ostream& os) {  // NOLINT(*)
+  std::ostringstream temp;
+  if (std::isinf(op->value)) {
+    if (op->value < 0) {
+      temp << "-";
+    }
+    temp << "INFINITY";
+  } else if (std::isnan(op->value)) {
+    temp << "NAN";
+  } else {
+    temp << std::scientific << op->value;
+    if (op->dtype.bits() == 32)
+      temp << 'f';
+    else if (op->dtype.bits() == 16)
+      temp << 'h';
+  }
+  MarkConst(temp.str());
+  os << temp.str();
+}
+
 runtime::Module BuildMetal(IRModule mod, Target target) {
   using tvm::runtime::Registry;
   bool output_ssa = false;

diff --git a/src/target/source/codegen_metal.h b/src/target/source/codegen_metal.h
@@ -51,8 +51,8 @@ class CodeGenMetal final : public CodeGenC {
   void PrintVecElemStore(const std::string& vec, DataType t, int i, const std::string& value) final;
   // overload visitor
   void VisitExpr_(const BroadcastNode* op, std::ostream& os) final;  // NOLINT(*)
-  // overload visitor
-  void VisitExpr_(const CallNode* op, std::ostream& os) final;  // NOLINT(*)
+  void VisitExpr_(const CallNode* op, std::ostream& os) final;       // NOLINT(*)
+  void VisitExpr_(const FloatImmNode* op, std::ostream& os) final;
   // reuse parent's function.
   using CodeGenC::PrintType;
 

diff --git a/src/target/source/intrin_rule_metal.cc b/src/target/source/intrin_rule_metal.cc
@@ -22,6 +22,7 @@
  * \brief Metal intrinsic rules.
  */
 #include <tvm/tir/op_attr_types.h>
+#include <tvm/topi/elemwise.h>
 
 #include "../intrin_rule.h"
 
@@ -90,6 +91,24 @@ TVM_REGISTER_OP("tir.cos").set_attr<FLowerIntrinsic>("metal.FLowerIntrinsic",
 TVM_REGISTER_OP("tir.cosh")
     .set_attr<FLowerIntrinsic>("metal.FLowerIntrinsic", DispatchPureExtern<Direct>);
 
+// There is no erf function in Metal. When erf is used, we use fast_erf instead
+static PrimExpr DispatchFastErf(const PrimExpr& e) {
+  LOG(WARNING) << " Metal doesn't have built-in erf function. fast_erf will be used instead.";
+  const CallNode* call = e.as<CallNode>();
+  ICHECK(call != nullptr);
+  ICHECK_EQ(call->args.size(), 1);
+  PrimExpr arg = call->args[0];
+  int bits = arg.dtype().bits();
+  bool isFloat = arg.dtype().is_float();
+  PrimExpr res;
+  if (isFloat && (bits == 16 || bits == 32))
+    res = topi::fast_erf_float_expr(arg, bits);
+  else
+    LOG(FATAL) << "Unsupported type in Metal fast_erf";
+  return res;
+}
+TVM_REGISTER_OP("tir.erf").set_attr<FLowerIntrinsic>("metal.FLowerIntrinsic", DispatchFastErf);
+
 }  // namespace intrin
 }  // namespace codegen
 }  // namespace tvm
diff --git a/tests/python/unittest/test_target_codegen_metal.py b/tests/python/unittest/test_target_codegen_metal.py
@@ -0,0 +1,81 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+import tvm
+from tvm import te
+import numpy as np
+from tvm import topi
+import unittest
+from tvm.contrib.nvcc import have_fp16, have_int8, have_bf16
+from tvm.contrib import nvcc
+import tvm.testing
+
+tx = te.thread_axis("threadIdx.x")
+bx = te.thread_axis("blockIdx.x")
+
+
+@tvm.testing.requires_gpu
+@tvm.testing.requires_metal
+def test_metal_inf_nan():
+    target = "metal"
+
+    def check_inf_nan(dev, n, value, dtype):
+        A = te.placeholder((n,), name="A", dtype=dtype)
+        inf_value = tvm.tir.const(value, dtype=dtype)
+        C = te.compute((n,), lambda i: inf_value, name="C")
+        s = te.create_schedule(C.op)
+        s[C].bind(s[C].op.axis[0], tx)
+        fun = tvm.build(s, [A, C], target)
+        a = tvm.nd.empty((n,), A.dtype, dev)
+        c = tvm.nd.empty((n,), A.dtype, dev)
+        # Only need to test compiling here
+        fun(a, c)
+
+    dev = tvm.device(target, 0)
+
+    check_inf_nan(dev, 1, -float("inf"), "float32")
+    check_inf_nan(dev, 1, -float("inf"), "float16")
+    check_inf_nan(dev, 1, float("inf"), "float32")
+    check_inf_nan(dev, 1, float("inf"), "float16")
+    check_inf_nan(dev, 1, float("nan"), "float32")
+    check_inf_nan(dev, 1, float("nan"), "float16")
+
+
+@tvm.testing.requires_gpu
+@tvm.testing.requires_metal
+def test_metal_erf():
+    target = "metal"
+
+    def check_erf(dev, n, dtype):
+        A = te.placeholder((n,), name="A", dtype=dtype)
+        C = te.compute(A.shape, lambda *i: te.erf(A(*i)), name="C")
+        s = te.create_schedule(C.op)
+        s[C].bind(s[C].op.axis[0], tx)
+        fun = tvm.build(s, [A, C], target)
+        a = tvm.nd.empty((n,), A.dtype, dev)
+        c = tvm.nd.empty((n,), A.dtype, dev)
+        # Only need to test compiling here
+        fun(a, c)
+
+    dev = tvm.device(target, 0)
+
+    check_erf(dev, 1, "float32")
+    check_erf(dev, 1, "float16")
+
+
+if __name__ == "__main__":
+    test_metal_inf_nan()
+    test_metal_erf()