# cmd: bash -lc cd 'tracks/A-gpu-programming/A01.foundations-and-programming-model' && make -s run-hip GPU_ARCH=gfx950 HIP_PATH=/opt/rocm
# gpu_arch: gfx950
# captured: 2026-08-10T06:50:40Z
---
== hip/hello.exe ==
Launching 2 blocks x 4 threads = 8 GPU threads
Hello from block 1, thread 0 -> global index 4
Hello from block 1, thread 1 -> global index 5
Hello from block 1, thread 2 -> global index 6
Hello from block 1, thread 3 -> global index 7
Hello from block 0, thread 0 -> global index 0
Hello from block 0, thread 1 -> global index 1
Hello from block 0, thread 2 -> global index 2
Hello from block 0, thread 3 -> global index 3
Hello from the CPU (host) - after the GPU finished.

== hip/vector_add.exe ==
n = 16777216, block = 256, grid = 65536
kernel time      : 0.647 ms
effective BW     : 311.0 GB/s
result           : PASSED

== hip/coalescing.exe ==
stride   effective GB/s  
-------- ----------------
1        3618.1          
2        1533.3          
4        757.4           
8        382.0           
16       190.8           
32       104.4           

Expect: stride 1 (coalesced) near peak HBM; higher strides drop as
each wavefront's lanes fall into separate cache lines. Same math, worse
memory pattern -> much lower effective bandwidth.