【Hackathon No.62】digamma, dirichlet算子FP16/BF16单测完善

paddle/phi/kernels/gpu/digamma_grad_kernel.cu

@@ -15,9 +15,16 @@
 #include "paddle/phi/kernels/digamma_grad_kernel.h"
 
 #include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/digamma_grad_kernel_impl.h"
 
-PD_REGISTER_KERNEL(
-    digamma_grad, GPU, ALL_LAYOUT, phi::DigammaGradKernel, float, double) {}
+PD_REGISTER_KERNEL(digamma_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::DigammaGradKernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/gpu/digamma_kernel.cu

@@ -15,10 +15,17 @@
 #include "paddle/phi/kernels/digamma_kernel.h"
 
 #include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/common/scalar.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/digamma_kernel_impl.h"
 
-PD_REGISTER_KERNEL(
-    digamma, GPU, ALL_LAYOUT, phi::DigammaKernel, float, double) {}
+PD_REGISTER_KERNEL(digamma,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::DigammaKernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/gpu/dirichlet_kernel.cu

@@ -112,5 +112,11 @@ struct DirichletSampler<GPUContext, T> {
 };
 }  // namespace phi
 
-PD_REGISTER_KERNEL(
-    dirichlet, GPU, ALL_LAYOUT, phi::Dirichletkernel, float, double) {}
+PD_REGISTER_KERNEL(dirichlet,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::Dirichletkernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}

paddle/phi/kernels/impl/digamma_grad_kernel_impl.h

@@ -16,6 +16,7 @@
 
 #include <unsupported/Eigen/SpecialFunctions>
 
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/kernels/funcs/for_range.h"
 
@@ -27,7 +28,11 @@ struct DigammaGradFunctor {
       : dout_(dout), x_(x), output_(output), numel_(numel) {}
 
   HOSTDEVICE void operator()(int64_t idx) const {
-    output_[idx] = dout_[idx] * Eigen::numext::polygamma(T(1), x_[idx]);
+    using MPType = typename phi::dtype::MPTypeTrait<T>::Type;
+    const MPType mp_dout = static_cast<MPType>(dout_[idx]);
+    const MPType mp_x = static_cast<MPType>(x_[idx]);
+    output_[idx] =
+        static_cast<T>(mp_dout * Eigen::numext::polygamma(MPType(1), mp_x));
   }
 
  private:

paddle/phi/kernels/impl/digamma_kernel_impl.h

@@ -16,6 +16,7 @@
 
 #include <unsupported/Eigen/SpecialFunctions>
 
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/kernels/funcs/for_range.h"
 
@@ -27,7 +28,9 @@ struct DigammaFunctor {
       : input_(input), output_(output), numel_(numel) {}
 
   HOSTDEVICE void operator()(int64_t idx) const {
-    output_[idx] = Eigen::numext::digamma(input_[idx]);
+    using MPType = typename phi::dtype::MPTypeTrait<T>::Type;
+    const MPType mp_input = static_cast<MPType>(input_[idx]);
+    output_[idx] = static_cast<T>(Eigen::numext::digamma(mp_input));
   }
 
  private:

paddle/phi/kernels/impl/dirichlet_kernel_impl.h

@@ -16,6 +16,7 @@
 
 #include <cmath>
 #include <random>
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/kernels/dirichlet_kernel.h"
 
 // ROCM hcc doesn't work well with using std:: in kernel functions
@@ -47,7 +48,10 @@ template <typename ScalarT, typename SamplerT>
 struct BaseSampler {
   SamplerT sampler_;
   HOSTDEVICE BaseSampler(const SamplerT& sampler) : sampler_(sampler) {}
-  HOSTDEVICE ScalarT sample() { return sampler_(); }
+  HOSTDEVICE ScalarT sample() {
+    // Sometimes convert float to float16/bfloat16
+    return static_cast<ScalarT>(sampler_());
+  }
 };
 
 // `sample_gamma` is d from Numpy's distributions.c, and add support for
@@ -83,33 +87,40 @@ HOSTDEVICE ScalarT
 sample_gamma(ScalarT alpha,
              BaseSampler<AccscalarT, UniformSamplerT> standard_uniform,
              BaseSampler<AccscalarT, NormalSamplerT> standard_normal) {
-  AccscalarT scale = 1.0f;
+  using MPTypeScalar = typename phi::dtype::MPTypeTrait<ScalarT>::Type;
+  using MPTypeAccscalar = typename phi::dtype::MPTypeTrait<AccscalarT>::Type;
+
+  MPTypeAccscalar mp_scale = static_cast<MPTypeAccscalar>(1.0f);
+  MPTypeScalar mp_alpha = static_cast<MPTypeScalar>(alpha);
 
   // Boost alpha for higher acceptance probability.
-  if (alpha < 1.0f) {
-    if (alpha == 0.f) return 0.f;
-    scale *= COMPAT_POW(1 - standard_uniform.sample(), 1.0f / alpha);
-    alpha += 1.0f;
+  if (mp_alpha < 1.0f) {
+    if (mp_alpha == 0.f) return static_cast<ScalarT>(0.f);
+    MPTypeAccscalar mp_sample =
+        static_cast<MPTypeAccscalar>(standard_uniform.sample());
+    mp_scale *= COMPAT_POW(1 - mp_sample, 1.0f / mp_alpha);
+    mp_alpha += 1.0f;
   }
 
   // This implements the acceptance-rejection method of Marsaglia and Tsang
   // (2000)
   // doi:10.1145/358407.358414
-  const AccscalarT d = alpha - 1.0f / 3.0f;
-  const AccscalarT c = 1.0f / COMPAT_SQRT(9.0f * d);
+  const MPTypeAccscalar d = mp_alpha - 1.0f / 3.0f;
+  const MPTypeAccscalar c = 1.0f / COMPAT_SQRT(9.0f * d);
   for (;;) {
-    AccscalarT x, y;
+    MPTypeAccscalar x, y;
     do {
-      x = standard_normal.sample();
+      x = static_cast<MPTypeAccscalar>(standard_normal.sample());
       y = 1.0f + c * x;
     } while (y <= 0);
-    const AccscalarT v = y * y * y;
-    const AccscalarT u = 1 - standard_uniform.sample();
-    const AccscalarT xx = x * x;
+    const MPTypeAccscalar v = y * y * y;
+    const MPTypeAccscalar u =
+        1 - static_cast<MPTypeAccscalar>(standard_uniform.sample());
+    const MPTypeAccscalar xx = x * x;
     if (u < 1.0f - 0.0331f * xx * xx)
-      return static_cast<ScalarT>(scale * d * v);
+      return static_cast<ScalarT>(mp_scale * d * v);
     if (COMPAT_LOG(u) < 0.5f * xx + d * (1.0f - v + COMPAT_LOG(v)))
-      return static_cast<ScalarT>(scale * d * v);
+      return static_cast<ScalarT>(mp_scale * d * v);
   }
 }
 

python/paddle/distribution/dirichlet.py

@@ -164,7 +164,10 @@ def _dirichlet(concentration, name=None):
     else:
         op_type = 'dirichlet'
         check_variable_and_dtype(
-            concentration, 'concentration', ['float32', 'float64'], op_type
+            concentration,
+            'concentration',
+            ['float16', 'float32', 'float64', 'uint16'],
+            op_type,
         )
         helper = LayerHelper(op_type, **locals())
         out = helper.create_variable_for_type_inference(

python/paddle/fluid/tests/unittests/distribution/test_dirichlet_op.py

@@ -20,7 +20,15 @@
 import paddle
 
 sys.path.append("../")
-from eager_op_test import OpTest
+import unittest
+
+from eager_op_test import (
+    OpTest,
+    convert_float_to_uint16,
+    convert_uint16_to_float,
+)
+
+from paddle.fluid import core
 
 paddle.enable_static()
 
@@ -52,3 +60,89 @@ def _hypothesis_testing(self, outs):
             )[0],
             0.01,
         )
+
+
+class TestDirichletFP16Op(OpTest):
+    # Because dirichlet random sample have not gradient, we skip gradient check.
+    no_need_check_grad = True
+
+    def setUp(self):
+        self.op_type = "dirichlet"
+        self.alpha = np.array((1.0, 2.0))
+        self.sample_shape = (100000, 2)
+        self.dtype = np.float16
+
+        self.inputs = {
+            'Alpha': np.broadcast_to(self.alpha, self.sample_shape).astype(
+                self.dtype
+            )
+        }
+        self.attrs = {}
+        self.outputs = {'Out': np.zeros(self.sample_shape).astype(self.dtype)}
+
+    def test_check_output(self):
+        self.check_output_customized(self._hypothesis_testing)
+
+    def _hypothesis_testing(self, outs):
+        self.assertEqual(outs[0].shape, self.sample_shape)
+        self.assertTrue(np.all(outs[0] > 0.0))
+        self.assertLess(
+            scipy.stats.kstest(
+                outs[0][:, 0],
+                # scipy dirichlet have not cdf, use beta to replace it.
+                scipy.stats.beta(a=self.alpha[0], b=self.alpha[1]).cdf,
+            )[0],
+            0.01,
+        )
+
+
+@unittest.skipIf(
+    not core.is_compiled_with_cuda()
+    or not core.is_bfloat16_supported(core.CUDAPlace(0)),
+    "core is not complied with CUDA and not support the bfloat16",
+)
+class TestDirichletBF16Op(OpTest):
+    # Because dirichlet random sample have not gradient, we skip gradient check.
+    no_need_check_grad = True
+
+    def setUp(self):
+        self.op_type = "dirichlet"
+        self.alpha = np.array((1.0, 2.0))
+        self.sample_shape = (10000, 2)
+        self.dtype = np.uint16
+        self.np_dtype = np.float32
+
+        self.inputs = {
+            'Alpha': np.broadcast_to(self.alpha, self.sample_shape).astype(
+                self.np_dtype
+            )
+        }
+        self.attrs = {}
+        self.outputs = {
+            'Out': np.zeros(self.sample_shape).astype(self.np_dtype)
+        }
+        self.inputs['Alpha'] = convert_float_to_uint16(self.inputs['Alpha'])
+        self.outputs['Out'] = convert_float_to_uint16(self.outputs['Out'])
+        self.place = core.CUDAPlace(0)
+
+    def test_check_output(self):
+        self.check_output_with_place_customized(
+            self._hypothesis_testing, place=core.CUDAPlace(0)
+        )
+
+    def _hypothesis_testing(self, outs):
+        outs = convert_uint16_to_float(outs)
+        self.assertEqual(outs[0].shape, self.sample_shape)
+        self.assertTrue(np.all(outs[0] > 0.0))
+        self.assertLess(
+            scipy.stats.kstest(
+                outs[0][:, 0],
+                # scipy dirichlet have not cdf, use beta to replace it.
+                scipy.stats.beta(a=self.alpha[0], b=self.alpha[1]).cdf,
+            )[0],
+            0.3,  # The bfloat16 test difference is below 0.3
+        )
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_digamma_op.py

@@ -15,11 +15,12 @@
 import unittest
 
 import numpy as np
-from eager_op_test import OpTest
+from eager_op_test import OpTest, convert_float_to_uint16
 from scipy.special import psi
 
 import paddle
 from paddle import fluid, static
+from paddle.fluid import core
 
 
 class TestDigammaOp(OpTest):
@@ -55,6 +56,43 @@ def test_check_grad_normal(self):
         self.check_grad(['X'], 'Out')
 
 
+class TestDigammaFP16Op(TestDigammaOp):
+    def init_dtype_type(self):
+        self.dtype = np.float16
+
+
+@unittest.skipIf(
+    not core.is_compiled_with_cuda()
+    or not core.is_bfloat16_supported(core.CUDAPlace(0)),
+    "core is not compiled with CUDA or not support bfloat16",
+)
+class TestDigammaBF16Op(OpTest):
+    def setUp(self):
+        # switch to static
+        paddle.enable_static()
+
+        self.op_type = 'digamma'
+        self.python_api = paddle.digamma
+        self.init_dtype_type()
+        shape = (5, 32)
+        data = np.random.random(shape).astype(self.np_dtype) + 1
+        self.inputs = {'X': convert_float_to_uint16(data)}
+        result = np.ones(shape).astype(self.np_dtype)
+        result = psi(data)
+        self.outputs = {'Out': convert_float_to_uint16(result)}
+
+    def init_dtype_type(self):
+        self.dtype = np.uint16
+        self.np_dtype = np.float32
+
+    def test_check_output(self):
+        # bfloat16 needs to set the parameter place
+        self.check_output_with_place(core.CUDAPlace(0))
+
+    def test_check_grad_normal(self):
+        self.check_grad_with_place(core.CUDAPlace(0), ['X'], 'Out')
+
+
 class TestDigammaAPI(unittest.TestCase):
     def setUp(self):
         # switch to static

python/paddle/tensor/math.py

@@ -4011,7 +4011,9 @@ def digamma(x, name=None):
     if in_dygraph_mode():
         return _C_ops.digamma(x)
     else:
-        check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'digamma')
+        check_variable_and_dtype(
+            x, 'x', ['float16', 'float32', 'float64', 'uint16'], 'digamma'
+        )
         helper = LayerHelper('digamma', **locals())
         out = helper.create_variable_for_type_inference(x.dtype)
         helper.append_op(type='digamma', inputs={'X': x}, outputs={'Out': out})