1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
|
Workaround for libc++ issue https://github.com/llvm/llvm-project/issues/100802
"reference to __host__ function 'memcpy' in __device__ function"
--- a/c10/util/Half.h
+++ b/c10/util/Half.h
@@ -227,7 +227,7 @@ C10_HOST_DEVICE inline float fp16_ieee_to_fp32_value(uint16_t h) {
// const float exp_scale = 0x1.0p-112f;
constexpr uint32_t scale_bits = (uint32_t)15 << 23;
float exp_scale_val = 0;
- std::memcpy(&exp_scale_val, &scale_bits, sizeof(exp_scale_val));
+ memcpy(&exp_scale_val, &scale_bits, sizeof(exp_scale_val));
const float exp_scale = exp_scale_val;
const float normalized_value =
fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
@@ -298,8 +298,8 @@ inline uint16_t fp16_ieee_from_fp32_value(float f) {
constexpr uint32_t scale_to_inf_bits = (uint32_t)239 << 23;
constexpr uint32_t scale_to_zero_bits = (uint32_t)17 << 23;
float scale_to_inf_val = 0, scale_to_zero_val = 0;
- std::memcpy(&scale_to_inf_val, &scale_to_inf_bits, sizeof(scale_to_inf_val));
- std::memcpy(
+ memcpy(&scale_to_inf_val, &scale_to_inf_bits, sizeof(scale_to_inf_val));
+ memcpy(
&scale_to_zero_val, &scale_to_zero_bits, sizeof(scale_to_zero_val));
const float scale_to_inf = scale_to_inf_val;
const float scale_to_zero = scale_to_zero_val;
|
