LIBXSMM_API
void libxsmm_generator_pgemm_kernel( libxsmm_generated_code* io_generated_code,
const libxsmm_pgemm_descriptor* i_packed_pgemm_desc,
int i_arch, ... ) {
const char *const cpuid = libxsmm_cpuid_name( i_arch );
/* add instruction set mismatch check to code, header */
libxsmm_generator_isa_check_header( io_generated_code, cpuid );
/* generate kernel */
if ( LIBXSMM_X86_AVX <= i_arch ) {
#if defined(GARBAGE_PARAMETERS)
unsigned int iunroll, junroll, loopi, loopj;
va_list args;
va_start(args, i_arch);
iunroll = va_arg(args, unsigned int);
junroll = va_arg(args, unsigned int);
loopi = va_arg(args, unsigned int);
loopj = va_arg(args, unsigned int);
va_end(args);
libxsmm_generator_packed_gemm_avx_avx512_kernel( io_generated_code, i_packed_pgemm_desc, cpuid, iunroll, junroll, loopi, loopj );
#else
libxsmm_generator_packed_gemm_avx_avx512_kernel( io_generated_code, i_packed_pgemm_desc, cpuid );
#endif
} else { /* TODO fix this error */
/* @TODO change int based architecture value */
LIBXSMM_API
void libxsmm_generator_transpose_kernel( libxsmm_generated_code* io_generated_code,
const libxsmm_trans_descriptor* i_trans_desc,
const char* i_arch ) {
/* add instruction set mismatch check to code, header */
libxsmm_generator_isa_check_header( io_generated_code, i_arch );
/* generate kernel */
if ( (strcmp(i_arch, "skx") == 0) ||
(strcmp(i_arch, "knm") == 0) ||
(strcmp(i_arch, "knl") == 0) ||
(strcmp(i_arch, "hsw") == 0) ||
(strcmp(i_arch, "snb") == 0) ) {
libxsmm_generator_transpose_avx_avx512_kernel( io_generated_code, i_trans_desc, i_arch );
} else {
/* TODO fix this error */
LIBXSMM_HANDLE_ERROR( io_generated_code, LIBXSMM_ERR_ARCH );
return;
}
/* add instruction set mismatch check to code, footer */
libxsmm_generator_isa_check_footer( io_generated_code, i_arch );
}
