Add support for float16 to the python pairwise distance api (#547)

Authors:
  - Ben Frederickson (https://github.com/benfred)

Approvers:
  - Corey J. Nolet (https://github.com/cjnolet)

URL: #547

cpp/src/distance/pairwise_distance_c.cpp

@@ -29,7 +29,7 @@
 
 namespace {
 
-template <typename T>
+template <typename T, typename DistT>
 void _pairwise_distance(cuvsResources_t res,
                         DLManagedTensor* x_tensor,
                         DLManagedTensor* y_tensor,
@@ -40,7 +40,7 @@ void _pairwise_distance(cuvsResources_t res,
   auto res_ptr = reinterpret_cast<raft::resources*>(res);
 
   using mdspan_type           = raft::device_matrix_view<T const, int64_t, raft::row_major>;
-  using distances_mdspan_type = raft::device_matrix_view<T, int64_t, raft::row_major>;
+  using distances_mdspan_type = raft::device_matrix_view<DistT, int64_t, raft::row_major>;
 
   auto x_mds         = cuvs::core::from_dlpack<mdspan_type>(x_tensor);
   auto y_mds         = cuvs::core::from_dlpack<mdspan_type>(y_tensor);
@@ -71,9 +71,14 @@ extern "C" cuvsError_t cuvsPairwiseDistance(cuvsResources_t res,
     }
 
     if (x_dt.bits == 32) {
-      _pairwise_distance<float>(res, x_tensor, y_tensor, distances_tensor, metric, metric_arg);
+      _pairwise_distance<float, float>(
+        res, x_tensor, y_tensor, distances_tensor, metric, metric_arg);
+    } else if (x_dt.bits == 16) {
+      _pairwise_distance<half, float>(
+        res, x_tensor, y_tensor, distances_tensor, metric, metric_arg);
     } else if (x_dt.bits == 64) {
-      _pairwise_distance<double>(res, x_tensor, y_tensor, distances_tensor, metric, metric_arg);
+      _pairwise_distance<double, double>(
+        res, x_tensor, y_tensor, distances_tensor, metric, metric_arg);
     } else {
       RAFT_FAIL("Unsupported DLtensor dtype: %d and bits: %d", x_dt.code, x_dt.bits);
     }

python/cuvs/cuvs/distance/distance.pyx

@@ -100,7 +100,10 @@ def pairwise_distance(X, Y, out=None, metric="euclidean", metric_arg=2.0,
     n = y_cai.shape[0]
 
     if out is None:
-        out = device_ndarray.empty((m, n), dtype=y_cai.dtype)
+        output_dtype = y_cai.dtype
+        if np.issubdtype(y_cai.dtype, np.float16):
+            output_dtype = np.float32
+        out = device_ndarray.empty((m, n), dtype=output_dtype)
     out_cai = wrap_array(out)
 
     x_k = x_cai.shape[1]
@@ -119,7 +122,7 @@ def pairwise_distance(X, Y, out=None, metric="euclidean", metric_arg=2.0,
     y_dt = y_cai.dtype
     d_dt = out_cai.dtype
 
-    if x_dt != y_dt or x_dt != d_dt:
+    if x_dt != y_dt:
         raise ValueError("Inputs must have the same dtypes")
 
     cdef cydlpack.DLManagedTensor* x_dlpack = \

python/cuvs/cuvs/test/test_distance.py

@@ -40,7 +40,7 @@
     ],
 )
 @pytest.mark.parametrize("inplace", [True, False])
-@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+@pytest.mark.parametrize("dtype", [np.float32, np.float64, np.float16])
 def test_distance(n_rows, n_cols, inplace, metric, dtype):
     input1 = np.random.random_sample((n_rows, n_cols))
     input1 = np.asarray(input1).astype(dtype)
@@ -55,7 +55,10 @@ def test_distance(n_rows, n_cols, inplace, metric, dtype):
         norm = np.sum(input1, axis=1)
         input1 = (input1.T / norm).T
 
-    output = np.zeros((n_rows, n_rows), dtype=dtype)
+    output_dtype = dtype
+    if np.issubdtype(dtype, np.float16):
+        output_dtype = np.float32
+    output = np.zeros((n_rows, n_rows), dtype=output_dtype)
 
     if metric == "inner_product":
         expected = np.matmul(input1, input1.T)
@@ -76,4 +79,8 @@ def test_distance(n_rows, n_cols, inplace, metric, dtype):
 
     actual = output_device.copy_to_host()
 
-    assert np.allclose(expected, actual, atol=1e-3, rtol=1e-3)
+    tol = 1e-3
+    if np.issubdtype(dtype, np.float16):
+        tol = 1e-1
+
+    assert np.allclose(expected, actual, atol=tol, rtol=tol)