Add support for bfloat16 in segment_matmul operation (#272)

This PR covers CPU implementation. --------- Co-authored-by: Matthias Fey <[email protected]>
pyg-team · Nov 2, 2023 · 44760ec · 44760ec
1 parent 2a0d558
commit 44760ec
Show file tree

Hide file tree

Showing 4 changed files with 27 additions and 12 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -5,6 +5,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ## [0.4.0] - 2023-MM-DD
 ### Added
+- Added support for `bfloat16` data type in `segment_matmul` and `grouped_matmul` (CPU only) ([#272](https://github.com/pyg-team/pyg-lib/pull/272))
 ### Changed
 - Added `--biased` parameter to run benchmarks for biased sampling ([#267](https://github.com/pyg-team/pyg-lib/pull/267))
 - Improved speed of biased sampling ([#270](https://github.com/pyg-team/pyg-lib/pull/270))

diff --git a/pyg_lib/csrc/ops/cpu/matmul_kernel.cpp b/pyg_lib/csrc/ops/cpu/matmul_kernel.cpp
@@ -297,7 +297,8 @@ std::vector<at::Tensor> grouped_matmul_kernel(const at::TensorList input,
         {input_contig[i].size(0), other_contig[i].size(-1)}));
   }
 
-  AT_DISPATCH_ALL_TYPES(
+  AT_DISPATCH_ALL_TYPES_AND2(
+      at::ScalarType::Half, at::ScalarType::BFloat16,
       input_contig.front().scalar_type(), "grouped_matmul_kernel", [&] {
         if (mkl_path_available<scalar_t>() &&
             mkl_path_possible(input_contig, other_contig)) {
@@ -413,7 +414,8 @@ at::Tensor segment_matmul_kernel(const at::Tensor& input,
   const auto other_contig = other.contiguous();
   auto out = input_contig.new_empty({input.size(0), other.size(-1)});
 
-  AT_DISPATCH_ALL_TYPES(
+  AT_DISPATCH_ALL_TYPES_AND2(
+      at::ScalarType::Half, at::ScalarType::BFloat16,
       input_contig.scalar_type(), "segment_matmul_kernel", [&] {
         const auto n = other_contig.size(-1);
         const auto k = input_contig.size(-1);

diff --git a/pyg_lib/testing.py b/pyg_lib/testing.py
@@ -37,12 +37,13 @@ def onlyTriton(func: Callable) -> Callable:
 
 
 def withCUDA(func: Callable) -> Callable:
-    def wrapper(*args, **kwargs):
-        func(*args, device=torch.device('cpu'), **kwargs)
-        if torch.cuda.is_available():
-            func(*args, device=torch.device('cuda:0'), **kwargs)
+    import pytest
 
-    return wrapper
+    devices = [torch.device('cpu')]
+    if torch.cuda.is_available():
+        devices.append(torch.device('cuda:0'))
+
+    return pytest.mark.parametrize('device', devices)(func)
 
 
 def withDataset(group: str, name: str) -> Callable:

diff --git a/test/ops/test_matmul.py b/test/ops/test_matmul.py
@@ -1,5 +1,6 @@
 import os
 
+import pytest
 import torch
 
 import pyg_lib
@@ -11,11 +12,17 @@
 
 
 @withCUDA
-def test_segment_matmul_autograd(device):
-    inputs = torch.randn((8, 16), requires_grad=True, device=device)
+@pytest.mark.parametrize('dtype', [torch.float, torch.bfloat16])
+def test_segment_matmul_autograd(dtype, device):
+    if device.type == 'cuda' and dtype == torch.bfloat16:
+        pytest.skip('CUDA does not support bfloat16')
+
+    inputs = torch.randn((8, 16), requires_grad=True, device=device,
+                         dtype=dtype)
     ptr = torch.tensor([0, 5, 8]).to(torch.device(device))
-    other = torch.randn((2, 16, 32), requires_grad=True, device=device)
-    bias = torch.randn((2, 32), requires_grad=True, device=device)
+    other = torch.randn((2, 16, 32), requires_grad=True, device=device,
+                        dtype=dtype)
+    bias = torch.randn((2, 32), requires_grad=True, device=device, dtype=dtype)
     out = pyg_lib.ops.segment_matmul(inputs, ptr, other, bias)
     assert out.size() == (8, 32)
 
@@ -31,7 +38,11 @@ def test_segment_matmul_autograd(device):
 
 
 @withCUDA
-def test_grouped_matmul_autograd(device):
+@pytest.mark.parametrize('dtype', [torch.float, torch.bfloat16])
+def test_grouped_matmul_autograd(dtype, device):
+    if device.type == 'cuda' and dtype == torch.bfloat16:
+        pytest.skip('CUDA does not support bfloat16')
+
     inputs = [
         torch.randn(5, 16, device=device, requires_grad=True),
         torch.randn(6, 9, device=device, requires_grad=True),