[TOPI][CUDA] Enable vectorization on fp16 type (apache#4867)

- This allows to better utilize the memory bandwidth

- Note that not all cases are vectorized for fp16 datatype. For
  instance, when the size is not a multiple of 1024, the inner loop
  may be an expression that cannot be vectorized. In this case, a
  small inner loop is still benefical for latency hidding.

Signed-off-by: Wei Pan <[email protected]>

topi/python/topi/cuda/injective.py

@@ -40,13 +40,20 @@ def schedule_injective_from_existing(sch, out):
     num_thread = tvm.target.Target.current(allow_none=False).max_num_threads
     max_block = 256
 
+    # vectorize on fp16 data type. This allows to better utilize the memory
+    # bandwidth.
+    vector_width = 4 if out.dtype == "float16" else 1
+
     try:
         const_size = util.get_const_int(util.prod(out.shape))
-        max_block = 256
-        need_block_split = const_size > max_block * num_thread
+        need_block_split = const_size > max_block * num_thread * vector_width
     except ValueError:
         need_block_split = False
 
+    if vector_width > 1:
+        fused, v = sch[out].split(fused, vector_width)
+        sch[out].vectorize(v)
+
     if need_block_split:
         xo, xi = sch[out].split(fused, factor=num_thread * max_block)
         bx, tx = sch[out].split(xi, factor=num_thread)

topi/tests/python/test_topi_relu.py

@@ -20,11 +20,20 @@
 import tvm
 import topi
 from topi.util import get_const_tuple
-
+from tvm.contrib.nvcc import parse_compute_version
 from common import get_all_backend
 
-def verify_relu(m, n):
-    A = tvm.placeholder((m, n), name='A')
+def skip_test(dtype, device):
+    if dtype == "float16" and device == "cuda":
+        major, minor = parse_compute_version(tvm.gpu(0).compute_version)
+        # fp16 starts from 5.3
+        if major < 6 or (major == 5 and minor < 3):
+            print("skip because gpu does not support fp16")
+            return True
+    return False
+
+def verify_relu(m, n, dtype="float32"):
+    A = tvm.placeholder((m, n), name='A', dtype=dtype)
     B = topi.nn.relu(A)
 
     a_np = np.random.uniform(low=-1.0, high=1.0, size=get_const_tuple(A.shape)).astype(A.dtype)
@@ -35,6 +44,8 @@ def check_device(device):
         if not ctx.exist:
             print("Skip because %s is not enabled" % device)
             return
+        if skip_test(dtype, device):
+            return
         print("Running on target: %s" % device)
         with tvm.target.create(device):
             s = topi.generic.schedule_elemwise(B)
@@ -87,12 +98,12 @@ def _prelu_numpy(x, W):
     tvm.testing.assert_allclose(b.asnumpy(), out_np, rtol=1e-5)
 
 def test_relu():
-    verify_relu(10, 128)
+    verify_relu(10, 128, "float32")
+    verify_relu(128, 64, "float16")
 
 def test_schedule_big_array():
     verify_relu(1024 * 100 , 512)
 
-
 def test_leaky_relu():
     verify_leaky_relu(100, 0.1)
 

topi/tests/python/test_topi_tensor.py

@@ -19,6 +19,16 @@
 import tvm
 import topi
 from tvm.contrib.pickle_memoize import memoize
+from tvm.contrib.nvcc import parse_compute_version
+
+def skip_test(dtype, device):
+    if dtype == "float16" and device == "cuda":
+        major, minor = parse_compute_version(tvm.gpu(0).compute_version)
+        # fp16 starts from 5.3
+        if major < 6 or (major == 5 and minor < 3):
+            print("skip because gpu does not support fp16")
+            return True
+    return False
 
 def verify_elemwise_sum(num_args, dtype):
     shape = (3,5,4)
@@ -84,18 +94,43 @@ def check_device(device):
     for device in ["llvm"]:
         check_device(device)
 
+def verify_vectorization(n, m, dtype):
+    def check_device(device):
+        if not tvm.runtime.enabled(device):
+            print("Skip because %s is not enabled" % device)
+            return
+        if skip_test(dtype, device):
+            return
+        with tvm.target.create(device):
+            ctx = tvm.context(device, 0)
+            A = tvm.placeholder((n, m), name='A', dtype=dtype)
+            B = tvm.compute((n, m), lambda i, j:
+                             A[i, j] + tvm.const(1, A.dtype), name='B')
+            S = topi.generic.schedule_elemwise(B)
+
+            fun = tvm.build(S, [A, B], device)
+            np_A = tvm.nd.empty((n, m), A.dtype, ctx).copyfrom(
+                                np.random.uniform(size=(n, m)))
+            np_B = tvm.nd.empty((n, m), B.dtype, ctx)
+            fun(np_A, np_B)
+            tvm.testing.assert_allclose(np_B.asnumpy(), np_A.asnumpy() + 1, rtol=1e-5)
+
+    for device in ["cuda"]:
+        check_device(device)
+
+def test_vectorization():
+    verify_vectorization(128, 64, "float16")
 
 def test_elemwise_sum():
     verify_elemwise_sum(1, "float32")
     verify_elemwise_sum(5, "float32")
     verify_elemwise_sum(4, "int32")
 
-
 def test_full():
     verify_full((3,4,5), "float32", 3.14)
     verify_full((10,), "int32", 7)
 
-
 if __name__ == "__main__":
     test_elemwise_sum()
     test_full()
+    test_vectorization()