Skip to content

libxs_memory.md

Latest commit

 

History

History
106 lines (78 loc) · 3.87 KB

libxs_memory.md

File metadata and controls

106 lines (78 loc) · 3.87 KB

Memory Comparison Benchmark

Benchmarks libxs_diff, libxs_memcmp, and the C standard library's memcmp for both large contiguous comparisons and many small fixed-size comparisons. A Fortran variant compares libxs_diff against Fortran array intrinsics.

Building

cd samples/memory
make GNU=1

Produces memcmp.x (C) and, if a Fortran compiler is found, memcmpf.x and matcpyf.x.

Usage (C)

./memcmp.x [elsize [stride [nelems [niters]]]]
Argument Default Description
elsize INSIZE/MAXSIZE (compile, 64) Element comparison size in bytes
stride max(MAXSIZE, elsize) Distance between successive elements
nelems 2 GB / stride Number of elements
niters 5 Repetitions (arithmetic average)

Environment Variables

Variable Default Description
STRIDED 0 0: one-call when stride==elsize, else loop; 1: seq loop; >=2: random
CHECK 0 Non-zero: validation pass (inject diffs, cross-check implementations)
MISMATCH 0 0: equal buffers; 1: first byte; 2: middle; 3: last; >=4: random
OFFSET 0 Shift buffer b off its natural alignment by this many bytes

Compile-Time Knobs

Macro Default Description
MAXSIZE 64 Stride floor (elements spaced >= this far)
INSIZE MAXSIZE Default element size when argument is omitted

Example

# 32-byte elements, sequential strided access, 10 repetitions
./memcmp.x 32 0 0 10

# 8-byte elements, random traversal, 3 repetitions
STRIDED=2 ./memcmp.x 8 0 0 3

# contiguous (single-call) comparison, ~2 GB
./memcmp.x 64 64 0 5

# mismatch at last byte, buffer b misaligned by 3 bytes
MISMATCH=3 OFFSET=3 ./memcmp.x 32 0 0 5

Environment variables and compile-time macros apply only to memcmp.x.

Usage (Fortran -- memcmpf)

./memcmpf.x [nelements [nrepeat]]

Element type is hard-coded as INTEGER(4). Compares ~2 GB by default. Reports per-iteration times (ms) and average throughput (MB/s).

Usage (Fortran -- matcpyf)

./matcpyf.x [m [n [ldi [ldo [nrepeat [nmb]]]]]]
Argument Default Description
m 4096 First matrix dimension
n m Second matrix dimension
ldi m (enforced m) Leading dimension of input
ldo ldi Leading dimension of output
nrepeat 2 Number of repetitions
nmb 2048 Memory budget in MB

Benchmarks libxs_matcopy and libxs_matcopy_task for matrix copy and zeroing. When the matrix is square and ldi==ldo, also benchmarks libxs_otrans and libxs_otrans_task for out-of-place transpose.

What Is Measured

Three comparison implementations are benchmarked:

  • libxs_diff -- SIMD-dispatched short-buffer comparison (SSE/AVX2/AVX-512). Limited to elsize < 256 bytes, per-element.
  • libxs_memcmp -- SIMD-dispatched memcmp replacement. Single call for contiguous buffers, or per-element for strided access.
  • stdlib memcmp -- C library implementation under the same pattern.

Each kernel gets freshly initialized data before its timed section. An untimed warm-up pass runs before the first measured iteration. The summary reports min / median / max across iterations plus peak MB/s. Reported MB/s counts both buffers (2 * nbytes / time).