pyrox.dev
1diff --git a/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h b/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h
2index 7f05c2ad166..1632b595c4c 100644
3--- a/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h
4+++ b/aten/src/ATen/cpu/vec/sve/vec_bfloat16.h
5@@ -220,8 +220,12 @@ class Vectorized<BFloat16> {
6 Vectorized<BFloat16> le(const Vectorized<BFloat16>& other) const;
7 };
8
9-inline std::tuple<Vectorized<float>, Vectorized<float>> convert_bfloat16_float(
10- const Vectorized<c10::BFloat16>& a) {
11+#if defined(__GNUC__) && __GNUC__ == 14
12+// Workaround for gcc-14.2.0 ICE during RTL pass: vregs when compiling for SVE
13+__attribute__((optimize("no-tree-vectorize")))
14+#endif
15+inline std::tuple<Vectorized<float>, Vectorized<float>>
16+convert_bfloat16_float(const Vectorized<c10::BFloat16>& a) {
17 static_assert(
18 Vectorized<c10::BFloat16>::size() == 2 * Vectorized<float>::size());
19 auto zero = svreinterpret_bf16_f32(svdup_n_f32(0.0f));
20diff --git a/aten/src/ATen/native/cpu/Activation.cpp b/aten/src/ATen/native/cpu/Activation.cpp
21index 52d5383e60f..00c9f4eb253 100644
22--- a/aten/src/ATen/native/cpu/Activation.cpp
23+++ b/aten/src/ATen/native/cpu/Activation.cpp
24@@ -26,6 +26,10 @@ namespace at::native {
25
26 namespace {
27
28+#if defined(__GNUC__) && __GNUC__ == 14 && defined(__aarch64__) && !defined(__ARM_FEATURE_SVE)
29+// Workaround for gcc-14.2.0 ICE during RTL pass: expand when compiling for NEON
30+__attribute__((optimize("no-tree-vectorize")))
31+#endif
32 static void log_sigmoid_cpu_kernel(TensorBase &output, TensorBase &buffer, const TensorBase &input) {
33 if (at::isReducedFloatingType(input.scalar_type())) {
34 AT_DISPATCH_REDUCED_FLOATING_TYPES(input.scalar_type(), "log_sigmoid_cpu", [&]() {
35diff --git a/aten/src/ATen/native/cpu/Unfold2d.cpp b/aten/src/ATen/native/cpu/Unfold2d.cpp
36index 8ef0741e77a..8c94decfff0 100644
37--- a/aten/src/ATen/native/cpu/Unfold2d.cpp
38+++ b/aten/src/ATen/native/cpu/Unfold2d.cpp
39@@ -169,6 +169,10 @@ static void unfolded2d_acc_channels_last(
40
41 /* note: due to write issues, this one cannot be parallelized as well as
42 * unfolded2d_copy */
43+#if defined(__GNUC__) && __GNUC__ == 14 && defined(__ARM_FEATURE_SVE) && !defined(__ARM_FEATURE_BF16)
44+// Workaround for gcc-14.2.0 ICE during RTL pass: vregs when compiling for SVE without BF16
45+__attribute__((optimize("no-tree-vectorize")))
46+#endif
47 void unfolded2d_acc_kernel(
48 ScalarType dtype,
49 void *finput_data,