[LANG] Fixed semantics of NaN in float comparisons (#281)

lib/ir/dispatch.cc

@@ -275,7 +275,7 @@ ir::value *dispatch::greater_than(ir::value *input, ir::value *other, ir::builde
   ir::type *scalar_ty = input->get_type()->get_scalar_ty();
   // float > float
   if (scalar_ty->is_floating_point_ty())
-    return builder->create_fcmpUGT(input, other);
+    return builder->create_fcmpOGT(input, other);
   // int > int
   else if (scalar_ty->is_integer_ty())
     return builder->create_icmpSGT(input, other);
@@ -287,7 +287,7 @@ ir::value *dispatch::greater_equal(ir::value *input, ir::value *other, ir::build
   ir::type *scalar_ty = input->get_type()->get_scalar_ty();
   // float >= float
   if (scalar_ty->is_floating_point_ty())
-    return builder->create_fcmpUGE(input, other);
+    return builder->create_fcmpOGE(input, other);
   // int >= int
   else if (scalar_ty->is_integer_ty())
     return builder->create_icmpSGE(input, other);
@@ -299,7 +299,7 @@ ir::value *dispatch::less_than(ir::value *input, ir::value *other, ir::builder *
   ir::type *scalar_ty = input->get_type()->get_scalar_ty();
   // float < float
   if (scalar_ty->is_floating_point_ty())
-    return builder->create_fcmpULT(input, other);
+    return builder->create_fcmpOLT(input, other);
   // int < int
   else if (scalar_ty->is_integer_ty())
     return builder->create_icmpSLT(input, other);
@@ -311,7 +311,7 @@ ir::value *dispatch::less_equal(ir::value *input, ir::value *other, ir::builder
   ir::type *scalar_ty = input->get_type()->get_scalar_ty();
   // float < float
   if (scalar_ty->is_floating_point_ty())
-    return builder->create_fcmpULE(input, other);
+    return builder->create_fcmpOLE(input, other);
   // int < int
   else if (scalar_ty->is_integer_ty())
     return builder->create_icmpSLE(input, other);
@@ -323,7 +323,7 @@ ir::value *dispatch::equal(ir::value *input, ir::value *other, ir::builder *buil
   ir::type *scalar_ty = input->get_type()->get_scalar_ty();
   // float == float
   if (scalar_ty->is_floating_point_ty())
-    return builder->create_fcmpUEQ(input, other);
+    return builder->create_fcmpOEQ(input, other);
   // int == int
   else if (scalar_ty->is_integer_ty())
     return builder->create_icmpEQ(input, other);

python/test/unit/language/test_core.py

@@ -69,7 +69,7 @@ def _test_unary(dtype_x, expr, torch_expr=None, device='cuda'):
     triton.testing.assert_almost_equal(z_ref, z_tri)
 
 
-def _test_binary(dtype_x, dtype_y, expr, device='cuda'):
+def _test_binary(dtype_x, dtype_y, expr, mode_x='real', mode_y='real', device='cuda'):
     SIZE = 128
     # define the kernel / launch-grid
     @triton.jit
@@ -84,6 +84,8 @@ def _test_binary(dtype_x, dtype_y, expr, device='cuda'):
     # inputs
     x = triton.testing.random(SIZE, dtype=cvt[dtype_x], device=device)
     y = triton.testing.random(SIZE, dtype=cvt[dtype_y], device=device)
+    if mode_x == 'nan': x[:] = float('nan')
+    if mode_y == 'nan': y[:] = float('nan')
     # reference result
     z_ref = eval(expr)
     # triton result
@@ -126,14 +128,25 @@ def test_bitwise_op(dtype_x, dtype_y, expr, device='cuda'):
 # ---------------
 # test compare ops
 # ---------------
-@pytest.mark.parametrize("dtype_x, dtype_y, expr", [
-    (dtype_x, dtype_y, f' x {op} y') \
-    for op in ['==', '!=', '>', '<', '>=', '<='] \
+ops = ['==', '!=', '>', '<', '>=', '<=']
+@pytest.mark.parametrize("dtype_x, dtype_y, expr, mode_x, mode_y", \
+# real
+[
+    (dtype_x, dtype_y, f' x {op} y', 'real', 'real') \
+    for op in ops \
     for dtype_x in dtypes \
     for dtype_y in dtypes
+] + \
+# NaNs
+[('float32', 'float32', f' x {op} y', mode_x, mode_y) \
+    for op in ops
+    for mode_x, mode_y in [('nan' , 'real'), 
+                           ('real', 'nan'), 
+                           ('nan' , 'nan')]
+
 ])
-def test_compare_op(dtype_x, dtype_y, expr, device='cuda'):
-    _test_binary(dtype_x, dtype_y, expr, device=device)
+def test_compare_op(dtype_x, dtype_y, expr, mode_x, mode_y, device='cuda'):
+    _test_binary(dtype_x, dtype_y, expr, mode_x=mode_x, mode_y=mode_y, device=device)
 
 
 # ---------------

python/triton/testing.py

@@ -1,6 +1,9 @@
 import torch
 import os
 from .code_gen import OutOfResources
+import subprocess
+import sys
+
 
 try:
     import triton._C.libtriton.cutlass as _cutlass
@@ -99,7 +102,15 @@ def random(shape, dtype, device):
     raise RuntimeError(f'Unknown dtype {dtype}')
 
 
-def do_bench(fn, warmup=25, rep=100, grad_to_none=None, percentiles=[0.2, 0.8]):
+def nvsmi(attrs):
+    attrs = ','.join(attrs)
+    cmd = ['nvidia-smi', '-i', '0', '--query-gpu=' + attrs, '--format=csv,noheader,nounits']
+    out = subprocess.check_output(cmd)
+    ret = out.decode(sys.stdout.encoding).split(',')
+    ret = [int(x) for x in ret]
+    return ret
+
+def do_bench(fn, warmup=25, rep=100, grad_to_none=None, percentiles=[0.2, 0.8], record_clocks=False):
     """
     Benchmark the runtime of the provided function. By default, return the median runtime of :code:`fn` along with
     the 20-th and 80-th performance percentile.
@@ -127,17 +138,21 @@ def do_bench(fn, warmup=25, rep=100, grad_to_none=None, percentiles=[0.2, 0.8]):
     end_event.record()
     torch.cuda.synchronize()
     estimate_ms = start_event.elapsed_time(end_event) / 5
+    # compute number of warmup and repeat
+    n_warmup = max(1, int(warmup/estimate_ms))
+    n_repeat = max(1, int(rep/estimate_ms))
     # We maintain a buffer of 256 MB that we clear
     # before each kernel call to make sure that the L2
     # doesn't contain any input data before the run
-    start_event = [torch.cuda.Event(enable_timing=True) for i in range(rep)]
-    end_event = [torch.cuda.Event(enable_timing=True) for i in range(rep)]
+    start_event = [torch.cuda.Event(enable_timing=True) for i in range(n_repeat)]
+    end_event   = [torch.cuda.Event(enable_timing=True) for i in range(n_repeat)]
+    clocks      = [None for i in range(n_repeat)]
     cache = torch.empty(int(256e6), dtype=torch.int8, device='cuda')
     # Warm-up
-    for _ in range(int(warmup / estimate_ms)):
+    for _ in range(n_warmup):
         fn()
     # Benchmark
-    for i in range(rep):
+    for i in range(n_repeat):
         # we don't want `fn` to accumulate gradient values
         # if it contains a backward pass. So we clear the
         # provided gradients
@@ -150,11 +165,12 @@ def do_bench(fn, warmup=25, rep=100, grad_to_none=None, percentiles=[0.2, 0.8]):
         start_event[i].record()
         fn()
         end_event[i].record()
+    # Record clocks
     torch.cuda.synchronize()
     times = torch.tensor([s.elapsed_time(e) for s, e in zip(start_event, end_event)])
-    percentiles = torch.quantile(times, torch.tensor(percentiles)).tolist()
     med_ms = torch.median(times).item()
     if percentiles:
+        percentiles = torch.quantile(times, torch.tensor(percentiles)).tolist()
         return tuple([med_ms] + percentiles)
     else:
         return med_ms