Files differ when oblas compiled with SSE

[osresearch]

Building on x86-64 (amd ryzen 7) with gcc 9.4.0-1ubuntu1~20.04.2 the _mm256_loadu2_m128i intrinsic is missing:

cc -D_DEFAULT_SOURCE -O3 -g -std=c99 -Wall -march=native  -funroll-loops -ftree-vectorize -fno-inline  -DOBLAS_AVX -DOCTMAT_ALIGN=32  -c -o oblas.o oblas.c
In file included from oblas.c:18:
oblas_avx.c: In function ‘oaxpy’:
oblas_avx.c:56:7: warning: implicit declaration of function ‘_mm256_loadu2_m128i’; did you mean ‘_mm256_loadu_si256’? [-Wimplicit-function-declaration]
   56 |       _mm256_loadu2_m128i((__m128i *)OCT_MUL_HI[u], (__m128i *)OCT_MUL_HI[u]);
      |       ^~~~~~~~~~~~~~~~~~~
      |       _mm256_loadu_si256

Compiling oblas with SSE (after cleaning the old partial compile) completes:

 deps/oblas/liboblas.a:
-       $(MAKE) -C deps/oblas CPPFLAGS+="-DOBLAS_AVX -DOCTMAT_ALIGN=32"
+       $(MAKE) -C deps/oblas CPPFLAGS+="-DOBLAS_SSE -DOCTMAT_ALIGN=32"

But the decode process produces corrupt output in the decoded peace_and_war.txt file from offset 0x280 to 0x4ff, and from 0x780 to 0x9ff, etc.

Compiling in classic mode works fine:

 deps/oblas/liboblas.a:
-       $(MAKE) -C deps/oblas CPPFLAGS+="-DOBLAS_AVX -DOCTMAT_ALIGN=32"
+       $(MAKE) -C deps/oblas CPPFLAGS+="-DOCTMAT_ALIGN=32"

Diffing the data.rq files between the classic and SSE versions show that the corruption is likely happening in the encode process since the first block has the corrupted data present.

[sleepybishop]

Hi, thank you for the report. There are assumptions in the code about the alignment based on whether sse or avx is used so really OCTMAT_ALIGN should be derived internally. If OBLAS_SSE is used then OCTMAT_ALIGN=16 should be used in tandem.