Examples and Tutorials

Examples and Tutorials

AOCL-DLP ships with example programs in the examples/classic/ directory. Build them with:

cd aocl-dlp
mkdir build && cd build
cmake -DBUILD_EXAMPLES=ON ..
make -j$(nproc)

Compiled examples are in build/examples/classic/.

Example Catalog

Basic GEMM

Example	Description	Key concepts
simple_gemm_f32.c	Float32 matrix multiplication	Basic GEMM call, row-major layout
simple_gemm_bf16.c	BFloat16 GEMM	BF16 input type, f32 accumulation
simple_gemm_s8.c	Signed int8 GEMM	Integer quantized GEMM

Mixed Precision

Example	Description	Key concepts
simple_gemm_bf16s8.c	BF16 activations with int8 weights	Mixed-precision, on-the-fly quantization
simple_gemm_f32s8.c	F32 activations with int8 weights	Mixed-precision quantized inference

Post-Operations

Example	Description	Key concepts
simple_gemm_with_bias.c	GEMM with fused bias addition	dlp_metadata_t, BIAS post-op
simple_gemm_with_relu.c	GEMM with fused ReLU activation	ELTWISE post-op, RELU
simple_gemm_with_mish.c	F32 GEMM with fused Mish activation	aocl_gemm_f32f32f32of32, ELTWISE post-op, MISH algo_type
post_ops_combinations.c	Multiple chained post-operations	Chaining BIAS + ELTWISE, seq_vector

Quantization

Example	Description	Key concepts
quantization.c	Symmetric quantization workflow	DLP_SYMM_STAT_QUANT, sym_quant APIs
simple_gemm_s8_sym_quant.c	s8 x s8 -> f32 GEMM with symmetric static quantization	aocl_gemm_s8s8s32of32_sym_quant, post_op_grp scales, group_size
simple_gemm_per_token_quant.c	W8A8 s8 x s8 GEMM with per-token (PerM) A dequant, incl. n=1 decoder path	aocl_gemm_s8s8s32of32, SCALE post-op, DLP_PARAM_DIM_PER_TOKEN
simple_gemm_bf16s4.c	BF16 activations x s4 weights, symmetric weight-only quantization (WOQ)	aocl_gemm_bf16s4f32of32, aocl_reorder_bf16s4f32of32, pre_ops->b_scl
simple_gemm_bf16u4.c	BF16 activations x u4 weights, asymmetric WOQ with B zero-point	aocl_gemm_bf16u4f32of32, pre_ops b_scl + b_zp

Batch & Advanced

Example	Description	Key concepts
batch_gemm.c	Batch GEMM for multiple matrices	aocl_batch_gemm_*, group_count
matrix_reorder.c	Pre-reorder weights for repeated use	aocl_reorder_*, mem_format_b = 'R'
eltwise_ops.c	Standalone element-wise operations	aocl_gemm_eltwise_ops_*

Multi-Instance & Utilities

Example	Description	Key concepts
multi_instance_gemm_f32.c	Multiple GEMM instances in parallel	Thread-local settings, concurrent calls
multi_instance_gemm_u8s8.c	Multi-instance quantized GEMM	Parallel quantized inference
version.c	Query library version	dlp_version_query()

Suggested Learning Path

If you are new to AOCL-DLP, work through the examples in this order:

1. Quick Start -- Build and run your first program (inline example)
2. simple_gemm_f32.c -- Understand basic GEMM parameters
3. simple_gemm_with_bias.c -- Learn how post-ops work
4. matrix_reorder.c -- Optimize for repeated inference
5. batch_gemm.c -- Process multiple matrices efficiently
6. quantization.c -- Use integer quantization for inference

Then explore the guides for deeper understanding:

• GEMM Guide -- All data types, parameters, and reordering
• Post-Ops Guide -- Full post-operations reference
• Performance Guide -- Threading and optimization

Building Examples Against an Installed Library

If AOCL-DLP is already installed on your system, you can build examples standalone:

# Using shared library
gcc -o simple_gemm_f32 simple_gemm_f32.c -I/usr/local/include -L/usr/local/lib -laocl-dlp -lm

# Using static library
gcc -o simple_gemm_f32 simple_gemm_f32.c -I/usr/local/include -L/usr/local/lib \
    -Wl,--whole-archive -laocl-dlp_static -Wl,--no-whole-archive -lstdc++ -lm -fopenmp

See the Integration Guide for CMake-based builds and troubleshooting.