support bfloat16 in python (facebookresearch#4037)

Summary:

Add support in python

Differential Revision: D66074156

Author: mdouze

faiss/gpu/test/torch_test_contrib_gpu.py

@@ -345,13 +345,16 @@ def test_knn_gpu(self, use_cuvs=False):
     def test_knn_gpu_cuvs(self):
         self.test_knn_gpu(use_cuvs=True)
 
-    def test_knn_gpu_datatypes(self, use_cuvs=False):
+    def test_knn_gpu_datatypes(self, use_cuvs=False, use_bf16=False):
         torch.manual_seed(10)
         d = 10
         nb = 1024
-        nq = 5
+        nq = 50
         k = 10
         res = faiss.StandardGpuResources()
+        if use_bf16 and not res.supportsBFloat16CurrentDevice():
+            print("WARNING bfloat16 not supported -- test not executed")
+            return
 
         # make GT on torch cpu and test using IndexFlatL2
         xb = torch.rand(nb, d, dtype=torch.float32)
@@ -361,29 +364,46 @@ def test_knn_gpu_datatypes(self, use_cuvs=False):
         index.add(xb)
         gt_D, gt_I = index.search(xq, k)
 
-        xb_c = xb.cuda().half()
-        xq_c = xq.cuda().half()
+        # convert to float16
+        if use_bf16:
+            xb_c = xb.cuda().bfloat16()
+            xq_c = xq.cuda().bfloat16()
+        else:
+            xb_c = xb.cuda().half()
+            xq_c = xq.cuda().half()
 
         # use i32 output indices
         D = torch.zeros(nq, k, device=xb_c.device, dtype=torch.float32)
         I = torch.zeros(nq, k, device=xb_c.device, dtype=torch.int32)
 
         faiss.knn_gpu(res, xq_c, xb_c, k, D, I, use_cuvs=use_cuvs)
 
-        self.assertTrue(torch.equal(I.long().cpu(), gt_I))
-        self.assertLess((D.float().cpu() - gt_D).abs().max(), 1.5e-3)
+        ndiff = (I.cpu() != gt_I).sum().item()
+        MSE = ((D.float().cpu() - gt_D) ** 2).sum().item()
+        if use_bf16:
+            # 57 -- bf16 is not as accurate as fp16
+            self.assertLess(ndiff, 80)
+            # 0.00515
+            self.assertLess(MSE, 8e-3)
+        else:
+            # 5
+            self.assertLess(ndiff, 10)
+            # 8.565e-5
+            self.assertLess(MSE, 1e-4)
 
         # Test using numpy
-        D = np.zeros((nq, k), dtype=np.float32)
-        I = np.zeros((nq, k), dtype=np.int32)
+        if not use_bf16:   # bf16 not supported by numpy
+            # use i32 output indices
+            D = np.zeros((nq, k), dtype=np.float32)
+            I = np.zeros((nq, k), dtype=np.int32)
 
-        xb_c = xb.half().numpy()
-        xq_c = xq.half().numpy()
+            xb_c = xb.half().numpy()
+            xq_c = xq.half().numpy()
 
         faiss.knn_gpu(res, xq_c, xb_c, k, D, I, use_cuvs=use_cuvs)
 
-        self.assertTrue(torch.equal(torch.from_numpy(I).long(), gt_I))
-        self.assertLess((torch.from_numpy(D) - gt_D).abs().max(), 1.5e-3)
+    def test_knn_gpu_bf16(self):
+        self.test_knn_gpu_datatypes(use_bf16=True)
 
 
 class TestTorchUtilsPairwiseDistanceGpu(unittest.TestCase):