diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-16 16:34:12 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-16 16:34:12 +0200 |
| commit | 78f5588d761f0df66eb11c2b79bee7c22fb0d132 (patch) | |
| tree | 302b5f567b649fc2e9d2ffad383130bc0a469f1f /src/dtype.h | |
| parent | d1ba1ebc06ef0d62ca932388fae83e6992594719 (diff) | |
| download | ffwd-78f5588d761f0df66eb11c2b79bee7c22fb0d132.tar.gz ffwd-78f5588d761f0df66eb11c2b79bee7c22fb0d132.zip | |
api: dedup scalar bf16-to-f32 into dtype.h
Diffstat (limited to 'src/dtype.h')
| -rw-r--r-- | src/dtype.h | 19 |
1 files changed, 19 insertions, 0 deletions
diff --git a/src/dtype.h b/src/dtype.h new file mode 100644 index 0000000..5c1ce8f --- /dev/null +++ b/src/dtype.h @@ -0,0 +1,19 @@ +#ifndef EMBED_DTYPE_H +#define EMBED_DTYPE_H + +#include <stdint.h> +#include <string.h> + +/* BF16 is the upper 16 bits of an IEEE-754 binary32, so widening one value is a + * left shift reinterpreted as a float. Shared by every host-side caller (the + * CPU loader, the safetensors reader); the CUDA backend keeps its own + * __nv_bfloat16 device kernels, and kernels.c keeps the SIMD buffer variant + * embed_bf16_to_f32_buf for hot, contiguous conversions. */ +static inline float embed_bf16_to_f32(uint16_t bf16) { + uint32_t bits = (uint32_t)bf16 << 16; + float f; + memcpy(&f, &bits, sizeof(f)); + return f; +} + +#endif /* EMBED_DTYPE_H */ |
