Skip to content

Commit

Permalink
Back out "Support Half/BFloat16 in max_pool2d (#7829)"
Browse files Browse the repository at this point in the history 
Differential Revision: D68647398

Pull Request resolved: #7955
  • Loading branch information
@pncosta22
pncosta22 authored Jan 25, 2025
1 parent 1ecab49 commit 25395b9
Show file tree
Hide file tree
Showing 2 changed files with 65 additions and 73 deletions.
2 changes: 1 addition & 1 deletion kernels/portable/cpu/op_max_pool2d_with_indices.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -70,7 +70,7 @@ std::tuple<Tensor&, Tensor&> max_pool2d_with_indices_out(
ret_val);

ScalarType in_type = in.scalar_type();
ET_SWITCH_REALHBF16_TYPES(
ET_SWITCH_REAL_TYPES(
in_type, ctx, "max_pool2d_with_indices.out", CTYPE, [&]() {
apply_kernel_2d_reduce_then_map_fn<CTYPE>(
[](const CTYPE in_val,
Expand Down
136 changes: 64 additions & 72 deletions kernels/test/op_max_pool2d_with_indices_test.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -40,81 +40,73 @@ class OpMaxPool2DWithIndicesOutTest : public OperatorTest {
out,
indices);
}

template <exec_aten::ScalarType DTYPE>
void test_4d_dtype() {
torch::executor::testing::TensorFactory<DTYPE> tf;
torch::executor::testing::TensorFactory<exec_aten::ScalarType::Long> tfLong;

exec_aten::Tensor self = tf.make(
{2, 3, 5, 5},
{28.75, -38.875, -7.0, -13.5, 70.75, 53.75, 69.625, 97.375,
25.375, 99.5, -72.125, -87.25, 79.25, 42.0, -24.75, -15.5,
12.5, -86.0, 85.5, -0.25, 67.125, 77.0, 53.375, -61.125,
50.0, 3.875, 42.25, -37.375, 51.0, -60.875, 87.0, 32.25,
73.5, 68.875, -84.375, -98.75, -30.125, 94.25, 1.625, -86.25,
-56.5, -68.0, 74.25, -51.25, 8.125, 71.375, -53.125, 4.875,
77.5, -89.875, 4.5, -46.5, -46.375, -92.625, -85.5, -23.0,
-8.875, -12.0, -46.625, -88.625, 66.75, 87.75, 90.25, -45.0,
-78.125, 63.25, 28.75, 28.125, -30.375, 17.75, -16.0, 5.0,
11.125, 88.625, -47.625, 72.25, 32.0, -7.625, 61.625, -63.125,
-22.75, 83.125, -40.375, -78.25, 49.5, -39.125, -89.625, 47.875,
-61.375, 7.75, 16.875, -96.375, -22.5, 8.5, 74.25, 12.75,
90.125, 73.875, -71.75, -10.0, 41.25, 1.125, 10.375, -34.625,
29.75, -27.5, 26.625, 81.0, -8.875, 17.625, 84.375, -23.625,
-53.875, -26.0, -67.375, -90.75, 16.375, 45.625, 99.5, 56.25,
-87.625, -65.5, -79.75, 31.875, 79.75, 6.375, 44.625, -55.25,
-5.5, -68.875, -38.625, 54.125, -3.125, 5.75, 29.25, -39.5,
26.75, 68.25, -24.625, -53.0, 51.0, 90.625, 65.375, 43.875,
90.875, -41.625, 99.875, 6.375, -31.25, -94.0});
::std::vector<int64_t> kernel_size_vec = {2, 2};
exec_aten::ArrayRef<int64_t> kernel_size = exec_aten::ArrayRef<int64_t>(
kernel_size_vec.data(), kernel_size_vec.size());
::std::vector<int64_t> stride_vec = {1, 1};
exec_aten::ArrayRef<int64_t> stride =
exec_aten::ArrayRef<int64_t>(stride_vec.data(), stride_vec.size());
::std::vector<int64_t> padding_vec = {0, 0};
exec_aten::ArrayRef<int64_t> padding =
exec_aten::ArrayRef<int64_t>(padding_vec.data(), padding_vec.size());
::std::vector<int64_t> dilation_vec = {1, 1};
exec_aten::ArrayRef<int64_t> dilation =
exec_aten::ArrayRef<int64_t>(dilation_vec.data(), dilation_vec.size());
bool ceil_mode = false;
exec_aten::Tensor out = tf.zeros({2, 3, 4, 4});
exec_aten::Tensor indices = tfLong.zeros({2, 3, 4, 4});
exec_aten::Tensor out_expected = tf.make(
{2, 3, 4, 4},
{69.625, 97.375, 97.375, 99.5, 69.625, 97.375, 97.375, 99.5,
12.5, 79.25, 85.5, 85.5, 77.0, 77.0, 85.5, 85.5,
87.0, 73.5, 73.5, 68.875, 87.0, 94.25, 94.25, 68.875,
-30.125, 94.25, 94.25, 8.125, 71.375, 74.25, 77.5, 77.5,
4.5, -8.875, -12.0, -46.625, 87.75, 90.25, 90.25, -45.0,
87.75, 90.25, 90.25, 17.75, 63.25, 28.75, 88.625, 88.625,
83.125, 83.125, 61.625, 61.625, 83.125, 83.125, 47.875, 49.5,
16.875, 47.875, 47.875, 74.25, 90.125, 90.125, 73.875, 74.25,
41.25, 81.0, 81.0, 29.75, 84.375, 81.0, 81.0, 17.625,
84.375, 45.625, 99.5, 99.5, 16.375, 45.625, 99.5, 99.5,
54.125, 54.125, 5.75, 29.25, 54.125, 68.25, 68.25, 29.25,
90.625, 90.625, 68.25, 90.875, 99.875, 99.875, 65.375, 90.875});
exec_aten::Tensor indices_expected = tfLong.make(
{2, 3, 4, 4},
{6, 7, 7, 9, 6, 7, 7, 9, 16, 12, 18, 18, 21, 21, 18, 18,
5, 7, 7, 8, 5, 12, 12, 8, 11, 12, 12, 19, 20, 17, 23, 23,
0, 6, 7, 8, 11, 12, 12, 13, 11, 12, 12, 19, 15, 16, 23, 23,
6, 6, 3, 3, 6, 6, 12, 9, 15, 12, 12, 19, 21, 21, 22, 19,
0, 7, 7, 4, 10, 7, 7, 9, 10, 17, 18, 18, 16, 17, 18, 18,
6, 6, 8, 9, 6, 12, 12, 9, 16, 16, 12, 19, 21, 21, 17, 19});
op_max_pool2d_with_indices_out(
self, kernel_size, stride, padding, dilation, ceil_mode, out, indices);
EXPECT_TENSOR_CLOSE(out, out_expected);
EXPECT_TENSOR_CLOSE(indices, indices_expected);
}
};

TEST_F(OpMaxPool2DWithIndicesOutTest, SanityTest4D) {
#define TEST_ENTRY(ctype, dtype) test_4d_dtype<exec_aten::ScalarType::dtype>();
ET_FORALL_FLOATHBF16_TYPES(TEST_ENTRY);
#undef TEST_ENTRY
torch::executor::testing::TensorFactory<exec_aten::ScalarType::Float> tfFloat;
torch::executor::testing::TensorFactory<exec_aten::ScalarType::Long> tfLong;

exec_aten::Tensor self = tfFloat.make(
{2, 3, 5, 5},
{28.75, -38.875, -7.0, -13.5, 70.75, 53.75, 69.625, 97.375,
25.375, 99.5, -72.125, -87.25, 79.25, 42.0, -24.75, -15.5,
12.5, -86.0, 85.5, -0.25, 67.125, 77.0, 53.375, -61.125,
50.0, 3.875, 42.25, -37.375, 51.0, -60.875, 87.0, 32.25,
73.5, 68.875, -84.375, -98.75, -30.125, 94.25, 1.625, -86.25,
-56.5, -68.0, 74.25, -51.25, 8.125, 71.375, -53.125, 4.875,
77.5, -89.875, 4.5, -46.5, -46.375, -92.625, -85.5, -23.0,
-8.875, -12.0, -46.625, -88.625, 66.75, 87.75, 90.25, -45.0,
-78.125, 63.25, 28.75, 28.125, -30.375, 17.75, -16.0, 5.0,
11.125, 88.625, -47.625, 72.25, 32.0, -7.625, 61.625, -63.125,
-22.75, 83.125, -40.375, -78.25, 49.5, -39.125, -89.625, 47.875,
-61.375, 7.75, 16.875, -96.375, -22.5, 8.5, 74.25, 12.75,
90.125, 73.875, -71.75, -10.0, 41.25, 1.125, 10.375, -34.625,
29.75, -27.5, 26.625, 81.0, -8.875, 17.625, 84.375, -23.625,
-53.875, -26.0, -67.375, -90.75, 16.375, 45.625, 99.5, 56.25,
-87.625, -65.5, -79.75, 31.875, 79.75, 6.375, 44.625, -55.25,
-5.5, -68.875, -38.625, 54.125, -3.125, 5.75, 29.25, -39.5,
26.75, 68.25, -24.625, -53.0, 51.0, 90.625, 65.375, 43.875,
90.875, -41.625, 99.875, 6.375, -31.25, -94.0});
::std::vector<int64_t> kernel_size_vec = {2, 2};
exec_aten::ArrayRef<int64_t> kernel_size = exec_aten::ArrayRef<int64_t>(
kernel_size_vec.data(), kernel_size_vec.size());
::std::vector<int64_t> stride_vec = {1, 1};
exec_aten::ArrayRef<int64_t> stride =
exec_aten::ArrayRef<int64_t>(stride_vec.data(), stride_vec.size());
::std::vector<int64_t> padding_vec = {0, 0};
exec_aten::ArrayRef<int64_t> padding =
exec_aten::ArrayRef<int64_t>(padding_vec.data(), padding_vec.size());
::std::vector<int64_t> dilation_vec = {1, 1};
exec_aten::ArrayRef<int64_t> dilation =
exec_aten::ArrayRef<int64_t>(dilation_vec.data(), dilation_vec.size());
bool ceil_mode = false;
exec_aten::Tensor out = tfFloat.zeros({2, 3, 4, 4});
exec_aten::Tensor indices = tfLong.zeros({2, 3, 4, 4});
exec_aten::Tensor out_expected = tfFloat.make(
{2, 3, 4, 4},
{69.625, 97.375, 97.375, 99.5, 69.625, 97.375, 97.375, 99.5, 12.5,
79.25, 85.5, 85.5, 77.0, 77.0, 85.5, 85.5, 87.0, 73.5,
73.5, 68.875, 87.0, 94.25, 94.25, 68.875, -30.125, 94.25, 94.25,
8.125, 71.375, 74.25, 77.5, 77.5, 4.5, -8.875, -12.0, -46.625,
87.75, 90.25, 90.25, -45.0, 87.75, 90.25, 90.25, 17.75, 63.25,
28.75, 88.625, 88.625, 83.125, 83.125, 61.625, 61.625, 83.125, 83.125,
47.875, 49.5, 16.875, 47.875, 47.875, 74.25, 90.125, 90.125, 73.875,
74.25, 41.25, 81.0, 81.0, 29.75, 84.375, 81.0, 81.0, 17.625,
84.375, 45.625, 99.5, 99.5, 16.375, 45.625, 99.5, 99.5, 54.125,
54.125, 5.75, 29.25, 54.125, 68.25, 68.25, 29.25, 90.625, 90.625,
68.25, 90.875, 99.875, 99.875, 65.375, 90.875});
exec_aten::Tensor indices_expected = tfLong.make(
{2, 3, 4, 4},
{6, 7, 7, 9, 6, 7, 7, 9, 16, 12, 18, 18, 21, 21, 18, 18,
5, 7, 7, 8, 5, 12, 12, 8, 11, 12, 12, 19, 20, 17, 23, 23,
0, 6, 7, 8, 11, 12, 12, 13, 11, 12, 12, 19, 15, 16, 23, 23,
6, 6, 3, 3, 6, 6, 12, 9, 15, 12, 12, 19, 21, 21, 22, 19,
0, 7, 7, 4, 10, 7, 7, 9, 10, 17, 18, 18, 16, 17, 18, 18,
6, 6, 8, 9, 6, 12, 12, 9, 16, 16, 12, 19, 21, 21, 17, 19});
op_max_pool2d_with_indices_out(
self, kernel_size, stride, padding, dilation, ceil_mode, out, indices);
EXPECT_TENSOR_CLOSE(out, out_expected);
EXPECT_TENSOR_CLOSE(indices, indices_expected);
}

TEST_F(OpMaxPool2DWithIndicesOutTest, SanityTest4D_2) {
Expand Down

0 comments on commit 25395b9

Please sign in to comment.