Mesa Offline Compute Test

Tests linking to Mesa 26.0.6 and measures its offline GPU computation power via Vulkan compute shaders.

Why Vulkan instead of GLES/EGL?

On this device (Termux on aarch64), the GLES/EGL path defaults to the llvmpipe software renderer (CPU-based, no X display). However, Vulkan gives direct access to the hardware GPU:

• Adreno 7c+ Gen 3 (freedreno Vulkan ICD) — hardware GPU compute
• llvmpipe — software fallback

The test uses the Vulkan surfaceless platform, which works without any display server — true offline computation.

How It Works

Two tests are performed:

Test 1: Vector Addition (Correctness)

Dispaches 1024 work items doing c[i] = a[i] + b[i]. Verifies all results match expected values (3× loop unrolling). Ensures the compute pipeline, SSBOs, and dispatch are working correctly.

Test 2: Heavy Compute Benchmark (Performance)

Dispaches a large grid (4M work items on GPU, 64K on CPU), each running 1024 iterations of heavy FP math (sin, cos, sqrt, multiplication, addition). Reports:

• Execution time per run
• Million items processed per second
• Estimated GFLOPS (based on ~5 FP ops per iteration)

Results (Adreno 7c+ Gen 3)

Metric	Value
Work items	4,194,304
Inner iterations	1,024
Total FP ops (est.)	~20 billion
Avg time	0.44 s
Throughput	9.50 M items/s
Estimated GFLOPS	48.66
Output	Deterministic, all finite

Build & Run

# Build
make

# Run
./mesa_offline_compute_test

# Or build directly
gcc -o mesa_offline_compute_test mesa_offline_compute_test.c \
    -I$PREFIX/include -L$PREFIX/lib -lvulkan -lm

Dependencies

• mesa (Vulkan ICDs: freedreno, swrast)
• vulkan-loader / vulkan-headers
• glslang (for SPIR-V shader compilation)
• xxd (to embed SPIR-V as C arrays)