8.1.2. Data Types - 8.1.2. Data Types - 5.3 English - 57404

AOCL User Guide (57404)
Document ID
57404
Release Date
2026-05-13
Version
5.3 English

AOCL-DLP supports multiple precision formats to balance accuracy and performance:

Table 8.1 Data Type Terminology#

Terminology

C Type

Bits

Typical Use

f32

float

32

Training, high-accuracy inference

f16

float16

16

Memory-efficient inference (native on Zen5+)

bf16

bfloat16

16

Training and inference (good range, lower precision)

s8 / u8

int8_t / uint8_t

8

Quantized weights and activations

s4 / u4

packed in int8_t

4

Extreme weight quantization

s32 / u32

int32_t / uint32_t

32

Accumulation and intermediate results

Supported GEMM Data Type Combinations:

Input A

Input B

Output C

Accumulator

Function Suffix

u8/s8

s8

s32/s8/u8/f32/bf16

s32

<u8|s8>s8s32o<s32|s8|u8|f32|bf16>

bf16/f32

s8

s32/s8/u8/f32/bf16

s32

<bf16|f32>s8s32o<s32|s8|u8|f32|bf16>

bf16

s4/u4

f32/bf16

f32

bf16<s4|u4>f32o<f32|bf16>

bf16

bf16

f32/bf16

f32

bf16bf16f32o<f32|bf16>

f32

f32

f32

f32

f32f32f32of32

f16

f16

f16

f16

f16f16f16of16

u8

s4

s32

s32

u8s4s32os32

Note

  • u8s4s32os32 only has reorder and get_reorder_buf_size APIs (no GEMM).

  • s8s8s32o<f32|bf16> also has _sym_quant variants for symmetric quantization.

  • Mixed-precision reorder: f32obf16 (converts f32 input to bf16 reordered output).

For detailed information about data types, see the Types API Reference and the Library Overview Wiki.