[eyoon1131]

Summary

SimX currently builds and runs with any VX_CFG_LMEM_LOG_SIZE while VX_mem_unit.sv statically rejects two LMEM sizes: one that the LMEM base address is not aligned with, and, with WGMMA, one above 64 KB, the most the 16-bit smem descriptor offset can address. VX_MEM_LMEM_BASE_ADDR (0xFFFF0000 / 0x1FFFF0000) is a multiple of 64 KB but not of anything larger, so LMEM_LOG_SIZE > 16 is an invalid configuration. rtlsim fails under this config at startup, but SimX builds and still runs, leading to silent errors.

This PR mirrors those asserts in SimX, so an invalid configuration fails to build in both models.

Symptom

On xlen32 SimX with LMEM_LOG_SIZE=17, the CTA dispatcher grants slots up to 128 KB but slot bases above 64 KB (0xFFFF0000 + 64 KB = 0x1_0000_0000) truncate to 0x0000xxxx, which is the IO window. Those CTAs' shared-memory traffic overwrites the console rings.

Repro from a configured xlen32 build (rerun SimX with LMEM_LOG_SIZE=16 after):

CFG="-DVX_CFG_NUM_THREADS=16 -DVX_CFG_NUM_WARPS=16 -DVX_CFG_LMEM_LOG_SIZE=17"
make -C tests/regression/sgemm2 clean
CONFIGS="$CFG" make -C tests/regression/sgemm2
CONFIGS="$CFG" ./ci/blackbox.sh --driver=rtlsim   --app=sgemm2 --perf=7 --args="-n256 -t4 -c256"
CONFIGS="$CFG" ./ci/blackbox.sh --driver=simx   --app=sgemm2 --perf=7 --args="-n256 -t4 -c256"

-c256 gives 8 KB of LMEM per CTA. With 16 threads each CTA is one warp, so 8 CTAs are resident at 64 KB LMEM and 16 at 128 KB.

driver	LMEM_LOG_SIZE	result	cycles	LMEM reqs
simx	16	PASSED	33.5M	41.9M
simx	17	PASSED, plus 64 garbage [#N: lost <fp32 bits> bytes] console lines	49.9M	21.0M
rtlsim	17	VX_mem_unit.sv:69: Assertion failed ... invalid parameter	-	-

Note how SimX still reports a PASS at 17, because slots 8-15 wrap to the IO window which behaves like ordinary global memory. This explains the significant increase in cycles and the halving in LMEM reqs, as well as the console output.

Change

sim/simx/core.cpp: static_asserts mirroring VX_mem_unit.sv:59-69:

• with TCU_WGMMA_ENABLE: LMEM is enabled, and LMEM_LOG_SIZE <= 16
  (descriptor offset field);
• with LMEM_ENABLE: the LMEM size is a multiple of MEM_BLOCK_SIZE, and
  VX_MEM_LMEM_BASE_ADDR is aligned to the LMEM size.

They are in the Core constructor beside LocalMem::Create. SimX has no mem-unit object, so the core is where the RTL's VX_mem_unit sub-blocks (LMEM, LMEM switch, LSU adapters, coalescers) are instantiated.

Testing

xlen32, build/, SimX:

• -DVX_CFG_LMEM_LOG_SIZE=17: fails to build with core.cpp:133: static assertion failed: LMEM base address must be aligned to the LMEM size.
• -DVX_CFG_EXT_TCU_ENABLE -DVX_CFG_TCU_WGMMA_ENABLE -DVX_CFG_LMEM_LOG_SIZE=17: fails with core.cpp:129: static assertion failed: LMEM exceeds the 16-bit smem descriptor offset field.
• Default and -DVX_CFG_LMEM_LOG_SIZE=16: build. sgemm2 passes as before.
• The asserts generate no code, so every configuration that builds produces the same result as before. The catalog's widest configuration is NUM_THREADS=32, which gives 64 KB, so no catalog configuration trips an assert.
• There's no regression test for the assert itself since a test would be an expected build failure.