LIBXSMM_INLINE
double residual_s ( float *A, unsigned int lda, unsigned int m, unsigned int n,
float *B, unsigned int ldb, unsigned int *nerrs,
unsigned int *ncorr )
{
unsigned int i, j, address, i4, j4, k4, i8, j8, k8;
double atmp, btmp, dtmp, ref, derror;
static int ntimes = 0;
*nerrs = 0;
*ncorr = 0;
derror = 0.0;
for ( j = 1 ; j<= n; j++ )
{
for ( i = 1; i <= m; i++ )
{
atmp = (double) A[ (j-1)*lda + (i-1)];
btmp = (double) B[ (j-1)*ldb + (i-1)];
ref = LIBXSMM_MAX(atmp,-atmp);
if ( atmp >= btmp ) {
dtmp = atmp - btmp;
} else {
dtmp = btmp - atmp;
}
if ( isnan(dtmp) || isinf(dtmp) )
{
if ( ++ntimes < 15 )
{
printf("Denormal bug: A(%u,%u) is %g B(%u,%u) is %g\n",i,j,atmp,i,j,btmp);
}
}
if ( (dtmp / ref > 1.0e-4) && (dtmp > 1.0e-7) )
{
*nerrs = *nerrs + 1;
if ( ++ntimes < 15 )
{
address = (j-1)*lda + (i-1);
j4 = (int)(address/(lda*4)) + 1;
i4 = (int)((address-(j4-1)*lda*4) / 4) + 1;
k4 = (address-(j4-1)*lda*4 - (i4-1)*4) + 1;
j8 = (int)(address/(lda*8)) + 1;
i8 = (int)((address-(j8-1)*lda*8) / 8) + 1;
k8 = (address-(j8-1)*lda*8 - (i8-1)*8) + 1;
printf("Bug #%i: A[%u]=A(%u,%u)=A4(%u,%u,%u)=A8(%u,%u,%u) expected=%g instead=%g err=%g\n",ntimes,address,i,j,i4,j4,k4,i8,j8,k8,atmp,btmp,dtmp);
}
} else {
if ( (*nerrs > 0) && (ntimes < 10) && (*ncorr < 40) )
{
printf("Cor #%u: A(%u,%u) expected=%g\n",*ncorr+1,i,j,atmp);
}
*ncorr = *ncorr + 1;
}
derror += dtmp;
}
}
return ( derror );
}
LIBXSMM_EXTERN_C LIBXSMM_RETARGETABLE double libxsmm_timer_duration(unsigned long long tick0, unsigned long long tick1)
{
const double d = (double)(LIBXSMM_MAX(tick1, tick0) - tick0);
#if defined(_WIN32)
LARGE_INTEGER frequency;
QueryPerformanceFrequency(&frequency);
return d / (double)frequency.QuadPart;
#elif defined(CLOCK_MONOTONIC)
return d * 1E-9;
#else
return d * 1E-6;
#endif
}
LIBXSMM_INLINE
double residual_s ( float *A, unsigned int lda, unsigned int m, unsigned int n,
float *B, unsigned int ldb, unsigned int *nerrs,
unsigned int *ncorr )
{
unsigned int i, j;
double atmp, btmp, dtmp, ref, derror;
static int ntimes = 0;
*nerrs = 0;
*ncorr = 0;
derror = 0.0;
for ( j = 1 ; j<= n; j++ )
{
for ( i = 1; i <= m; i++ )
{
atmp = (double) A[ (j-1)*lda + (i-1)];
btmp = (double) B[ (j-1)*ldb + (i-1)];
ref = LIBXSMM_MAX(atmp,-atmp);
if ( atmp >= btmp ) {
dtmp = atmp - btmp;
} else {
dtmp = btmp - atmp;
}
if ( isnan(dtmp) || isinf(dtmp) )
{
if ( ++ntimes < 15 )
{
printf("Denormal bug: A(%u,%u) is %g B(%u,%u) is %g\n",i,j,atmp,i,j,btmp);
}
}
if ( dtmp / ref > 1.0e-4 )
{
*nerrs = *nerrs + 1;
if ( ++ntimes < 15 )
{
printf("Bug #%d: A(%u,%u) expected=%g instead=%g err=%g\n",ntimes,i,j,atmp,btmp,dtmp);
}
} else {
if ( (*nerrs > 0) && (ntimes < 10) && (*ncorr < 40) )
{
printf("Cor #%u: A(%u,%u) expected=%g\n",*ncorr+1,i,j,atmp);
}
*ncorr = *ncorr + 1;
}
derror += dtmp;
}
}
return ( derror );
}
LIBXSMM_API void libxsmm_trace(FILE* stream, unsigned int depth, const int* filter_threadid, const int* filter_mindepth, const int* filter_maxnsyms)
{
#if defined(LIBXSMM_TRACE)
unsigned int depth1 = depth + 1, threadid;
const char *const name = libxsmm_trace_info(&depth1, &threadid,
filter_threadid, filter_mindepth, filter_maxnsyms);
if (name && *name) { /* implies actual other results to be valid */
const int depth0 = LIBXSMM_MAX(0 != filter_mindepth ? *filter_mindepth : internal_trace_mindepth, 0);
assert(0 != stream/*otherwise fprintf handle the error*/);
if ((0 == filter_threadid && 0 > internal_trace_threadid) || (0 != filter_threadid && 0 > *filter_threadid)) {
fprintf(stream, "%*s%s@%u\n", (int)(depth1 - depth0), "", name, threadid);
}
else {
fprintf(stream, "%*s%s\n", (int)(depth1 - depth0), "", name);
}
}
#else /* suppress warning */
LIBXSMM_UNUSED(stream); LIBXSMM_UNUSED(depth);
LIBXSMM_UNUSED(filter_threadid);
LIBXSMM_UNUSED(filter_mindepth);
LIBXSMM_UNUSED(filter_maxnsyms);
#endif
}
LIBXSMM_API int libxsmm_cpuid_x86(void)
{
int target_arch = LIBXSMM_STATIC_TARGET_ARCH;
unsigned int eax = 0, ebx = 0, ecx = 0, edx = 0;
LIBXSMM_CPUID_X86(0, eax, ebx, ecx, edx);
if (1 <= eax) { /* CPUID */
LIBXSMM_CPUID_X86(1, eax, ebx, ecx, edx);
/* XSAVE/XGETBV(0x04000000), OSXSAVE(0x08000000) */
if (0x0C000000 == (0x0C000000 & ecx)) {
/* Check for CRC32 (this is not a proper test for SSE 4.2 as a whole!) */
if (0x00100000 == (0x00100000 & ecx)) {
target_arch = LIBXSMM_X86_SSE4;
}
LIBXSMM_XGETBV(0, eax, edx);
if (0x00000006 == (0x00000006 & eax)) { /* OS XSAVE 256-bit */
if (0x000000E0 == (0x000000E0 & eax)) { /* OS XSAVE 512-bit */
LIBXSMM_CPUID_X86(7, eax, ebx, ecx, edx);
/* AVX512F(0x00010000), AVX512CD(0x10000000) */
if (0x10010000 == (0x10010000 & ebx)) { /* Common */
/* AVX512DQ(0x00020000), AVX512BW(0x40000000), AVX512VL(0x80000000) */
if (0xC0020000 == (0xC0020000 & ebx)) { /* SKX (Core) */
if (0x00000800 == (0x00000800 & ecx)) { /* ICL (CORE) */
target_arch = LIBXSMM_X86_AVX512_ICL;
}
else { /* SKX (CORE) */
target_arch = LIBXSMM_X86_AVX512_CORE;
}
}
/* AVX512PF(0x04000000), AVX512ER(0x08000000) */
else if (0x0C000000 == (0x0C000000 & ebx)) {
if (0x0000000C == (0x0000000C & edx)) { /* KNM (MIC) */
target_arch = LIBXSMM_X86_AVX512_KNM;
}
else { /* KNL (MIC) */
target_arch = LIBXSMM_X86_AVX512_MIC;
}
}
else { /* Common */
target_arch = LIBXSMM_X86_AVX512;
}
}
}
else if (0x10000000 == (0x10000000 & ecx)) { /* AVX(0x10000000) */
if (0x00001000 == (0x00001000 & ecx)) { /* FMA(0x00001000) */
target_arch = LIBXSMM_X86_AVX2;
}
else {
target_arch = LIBXSMM_X86_AVX;
}
}
}
}
}
/* check if procedure obviously failed to detect the highest available instruction set extension */
assert(LIBXSMM_STATIC_TARGET_ARCH <= target_arch);
return LIBXSMM_MAX(target_arch, LIBXSMM_STATIC_TARGET_ARCH);
}
LIBXSMM_INLINE LIBXSMM_RETARGETABLE internal_code_type* internal_init(void)
{
/*const*/internal_code_type* result;
int i;
#if !defined(LIBXSMM_OPENMP)
# if !defined(LIBXSMM_NOSYNC)
static int internal_reglock_check = 1; /* setup the locks in a thread-safe fashion */
assert(sizeof(internal_reglock) == (INTERNAL_REGLOCK_COUNT * sizeof(*internal_reglock)));
if (1 == LIBXSMM_ATOMIC_LOAD(&internal_reglock_check, LIBXSMM_ATOMIC_SEQ_CST)) {
LIBXSMM_ATOMIC_ADD_FETCH(&internal_reglock_check, 1, LIBXSMM_ATOMIC_SEQ_CST);
if (2 == internal_reglock_check) {
for (i = 0; i < INTERNAL_REGLOCK_COUNT; ++i) LIBXSMM_LOCK_INIT(internal_reglock + i);
LIBXSMM_ATOMIC_STORE_ZERO(&internal_reglock_check, LIBXSMM_ATOMIC_SEQ_CST);
}
}
while (0 != internal_reglock_check); /* wait until locks are initialized */
for (i = 0; i < INTERNAL_REGLOCK_COUNT; ++i) LIBXSMM_LOCK_ACQUIRE(internal_reglock + i);
# endif
#else
# pragma omp critical(internal_reglock)
#endif
{
result = LIBXSMM_ATOMIC_LOAD(&internal_registry, LIBXSMM_ATOMIC_SEQ_CST);
if (0 == result) {
int init_code;
/* set internal_target_archid */
libxsmm_set_target_arch(getenv("LIBXSMM_TARGET"));
{ /* select prefetch strategy for JIT */
const char *const env_prefetch = getenv("LIBXSMM_PREFETCH");
if (0 == env_prefetch || 0 == *env_prefetch) {
#if (0 > LIBXSMM_PREFETCH) /* permitted by LIBXSMM_PREFETCH_AUTO */
internal_prefetch = (LIBXSMM_X86_AVX512_MIC != internal_target_archid
? LIBXSMM_PREFETCH_NONE : LIBXSMM_PREFETCH_AL2BL2_VIA_C);
#else
internal_prefetch = LIBXSMM_MAX(INTERNAL_PREFETCH, 0);
#endif
}
else { /* user input considered even if LIBXSMM_PREFETCH_AUTO is disabled */
switch (atoi(env_prefetch)) {
case 2: internal_prefetch = LIBXSMM_PREFETCH_SIGONLY; break;
case 3: internal_prefetch = LIBXSMM_PREFETCH_BL2_VIA_C; break;
case 4: internal_prefetch = LIBXSMM_PREFETCH_AL2; break;
case 5: internal_prefetch = LIBXSMM_PREFETCH_AL2_AHEAD; break;
case 6: internal_prefetch = LIBXSMM_PREFETCH_AL2BL2_VIA_C; break;
case 7: internal_prefetch = LIBXSMM_PREFETCH_AL2BL2_VIA_C_AHEAD; break;
case 8: internal_prefetch = LIBXSMM_PREFETCH_AL2_JPST; break;
case 9: internal_prefetch = LIBXSMM_PREFETCH_AL2BL2_VIA_C_JPST; break;
default: internal_prefetch = LIBXSMM_PREFETCH_NONE;
}
}
}
libxsmm_hash_init(internal_target_archid);
libxsmm_gemm_diff_init(internal_target_archid);
init_code = libxsmm_gemm_init(internal_target_archid, internal_prefetch);
#if defined(__TRACE)
{
int filter_threadid = 0, filter_mindepth = 1, filter_maxnsyms = 0;
const char *const env_trace_init = getenv("LIBXSMM_TRACE");
if (EXIT_SUCCESS == init_code && 0 != env_trace_init && 0 != *env_trace_init) {
char buffer[32];
if (1 == sscanf(env_trace_init, "%32[^,],", buffer)) {
sscanf(buffer, "%i", &filter_threadid);
}
if (1 == sscanf(env_trace_init, "%*[^,],%32[^,],", buffer)) {
sscanf(buffer, "%i", &filter_mindepth);
}
if (1 == sscanf(env_trace_init, "%*[^,],%*[^,],%32s", buffer)) {
sscanf(buffer, "%i", &filter_maxnsyms);
}
else {
filter_maxnsyms = -1; /* all */
}
}
init_code = libxsmm_trace_init(filter_threadid - 1, filter_mindepth, filter_maxnsyms);
}
#endif
if (EXIT_SUCCESS == init_code) {
assert(0 == internal_registry_keys && 0 == internal_registry); /* should never happen */
result = (internal_code_type*)libxsmm_malloc(LIBXSMM_REGSIZE * sizeof(internal_code_type));
internal_registry_keys = (internal_regkey_type*)libxsmm_malloc(LIBXSMM_REGSIZE * sizeof(internal_regkey_type));
if (0 != result && 0 != internal_registry_keys) {
const char *const env_verbose = getenv("LIBXSMM_VERBOSE");
internal_statistic_mnk = (unsigned int)(pow((double)(LIBXSMM_MAX_MNK), 0.3333333333333333) + 0.5);
internal_statistic_sml = 13; internal_statistic_med = 23;
if (0 != env_verbose && 0 != *env_verbose) {
internal_verbose_mode = atoi(env_verbose);
}
#if !defined(NDEBUG)
else {
internal_verbose_mode = 1; /* quiet -> verbose */
}
#endif
for (i = 0; i < LIBXSMM_REGSIZE; ++i) result[i].pmm = 0;
/* omit registering code if JIT is enabled and if an ISA extension is found
* which is beyond the static code path used to compile the library
*/
#if defined(LIBXSMM_BUILD)
# if (0 != LIBXSMM_JIT) && !defined(__MIC__)
/* check if target arch. permits execution (arch. may be overridden) */
if (LIBXSMM_STATIC_TARGET_ARCH <= internal_target_archid &&
(LIBXSMM_X86_AVX > internal_target_archid /* jit is not available */
/* condition allows to avoid JIT (if static code is good enough) */
|| LIBXSMM_STATIC_TARGET_ARCH == internal_target_archid))
# endif
{ /* opening a scope for eventually declaring variables */
/* setup the dispatch table for the statically generated code */
# include <libxsmm_dispatch.h>
}
#endif
atexit(libxsmm_finalize);
LIBXSMM_ATOMIC_STORE(&internal_registry, result, LIBXSMM_ATOMIC_SEQ_CST);
}
else {
#if !defined(NDEBUG) && defined(__TRACE) /* library code is expected to be mute */
fprintf(stderr, "LIBXSMM: failed to allocate code registry!\n");
#endif
libxsmm_free(internal_registry_keys);
libxsmm_free(result);
}
}
#if !defined(NDEBUG) && defined(__TRACE) /* library code is expected to be mute */
else {
fprintf(stderr, "LIBXSMM: failed to initialize sub-component (error #%i)!\n", init_code);
}
#endif
}
}
#if !defined(LIBXSMM_OPENMP) && !defined(LIBXSMM_NOSYNC) /* release locks */
for (i = 0; i < INTERNAL_REGLOCK_COUNT; ++i) LIBXSMM_LOCK_RELEASE(internal_reglock + i);
#endif
assert(result);
return result;
}
int main(int argc, char* argv[])
{
const int m = (1 < argc ? atoi(argv[1]) : 16);
const int n = (2 < argc ? atoi(argv[2]) : m);
const int unsigned ldi = LIBXSMM_MAX(3 < argc ? atoi(argv[3]) : 0, m);
const int unsigned ldo = LIBXSMM_MAX(4 < argc ? atoi(argv[4]) : 0, m);
const int unroll = (5 < argc ? atoi(argv[5]) : 1);
const int prefetch = (6 < argc ? atoi(argv[6]) : 0);
const int flags = ((7 < argc && 0 != atoi(argv[7])) ? LIBXSMM_MATCOPY_FLAG_ZERO_SOURCE : 0);
const int iters = (8 < argc ? atoi(argv[8]) : 1);
/* we should modify to test all data-types */
const libxsmm_mcopy_descriptor* desc;
libxsmm_xmcopyfunction kernel;
libxsmm_descriptor_blob blob;
libxsmm_timer_tickint l_start;
libxsmm_timer_tickint l_end;
int error = 0, i, j;
ELEM_TYPE *a, *b;
double copy_time;
printf("This is a tester for JIT matcopy kernels!\n");
desc = libxsmm_mcopy_descriptor_init(&blob, sizeof(ELEM_TYPE),
m, n, ldo, ldi, flags, prefetch, &unroll);
a = (ELEM_TYPE*)malloc(n * ldi * sizeof(ELEM_TYPE));
b = (ELEM_TYPE*)malloc(n * ldo * sizeof(ELEM_TYPE));
for (i = 0; i < n; i++) {
for (j = 0; j < m; j++) {
a[j+ldi*i] = (ELEM_TYPE)rand();
if (0 != (LIBXSMM_MATCOPY_FLAG_ZERO_SOURCE & flags)) {
b[j+ldo*i] = (ELEM_TYPE)rand();
}
}
}
/* test dispatch call */
kernel = libxsmm_dispatch_mcopy(desc);
if (kernel == 0) {
printf("JIT error -> exit!!!!\n");
exit(EXIT_FAILURE);
}
/* let's call */
kernel(a, &ldi, b, &ldo, &a[128]);
l_start = libxsmm_timer_tick();
for (i = 0; i < iters; ++i) {
kernel(a, &ldi, b, &ldo, &a[128]);
}
l_end = libxsmm_timer_tick();
copy_time = libxsmm_timer_duration(l_start, l_end);
for (i = 0; i < n; ++i) {
for (j = 0; j < m; ++j) {
if (0 != (LIBXSMM_MATCOPY_FLAG_ZERO_SOURCE & flags)) {
if (LIBXSMM_NEQ(b[j+ldo*i], 0)) {
printf("ERROR!!!\n");
i = n;
error = 1;
break;
}
}
else if (LIBXSMM_NEQ(a[j+ldi*i], b[j+ldo*i])) {
printf("ERROR!!!\n");
i = n;
error = 1;
break;
}
}
}
if (error == 0) {
printf("CORRECT copy!!!!\n");
printf("Time taken is\t%.5f seconds\n", copy_time);
}
return EXIT_SUCCESS;
}
#include <stdlib.h>
#include <stdio.h>
#if defined(LIBXSMM_OFFLOAD_TARGET)
# pragma offload_attribute(pop)
#endif
#if !defined(REAL_TYPE)
# define REAL_TYPE double
#endif
LIBXSMM_RETARGETABLE void init(int seed, REAL_TYPE *LIBXSMM_RESTRICT dst, double scale, libxsmm_blasint nrows, libxsmm_blasint ncols, libxsmm_blasint ld);
LIBXSMM_RETARGETABLE void init(int seed, REAL_TYPE *LIBXSMM_RESTRICT dst, double scale, libxsmm_blasint nrows, libxsmm_blasint ncols, libxsmm_blasint ld)
{
const libxsmm_blasint minval = seed, addval = (nrows - 1) * ld + (ncols - 1);
const libxsmm_blasint maxval = LIBXSMM_MAX(LIBXSMM_ABS(minval), addval);
const double norm = 0 != maxval ? (scale / maxval) : scale;
libxsmm_blasint i, j;
#if defined(_OPENMP)
# pragma omp parallel for
#endif
for (i = 0; i < ncols; ++i) {
for (j = 0; j < nrows; ++j) {
const libxsmm_blasint k = i * ld + j;
const double value = (double)(k + minval);
dst[k] = (REAL_TYPE)(norm * (value - 0.5 * addval));
}
}
}
int main(int argc, char* argv[])
{
const char *const filename = (1 < argc ? argv[1] : "mhd_image.mhd");
/* take some block-sizes, which are used to test leading dimensions */
const int bw = LIBXSMM_MAX(2 < argc ? atoi(argv[2]) : 64, 1);
const int bh = LIBXSMM_MAX(3 < argc ? atoi(argv[3]) : 64, 1);
size_t ndims = 3, size[3], pitch[3], offset[3], ncomponents, header_size, extension_size;
libxsmm_mhd_elemtype type;
char data_filename[1024];
void* data = 0;
int result;
/* Read header information; function includes various sanity checks. */
result = libxsmm_mhd_read_header(filename, sizeof(data_filename),
data_filename, &ndims, size, &ncomponents, &type,
&header_size, &extension_size);
/* Allocate data according to the header information. */
if (EXIT_SUCCESS == result) {
size_t typesize;
pitch[0] = (size[0] + bw - 1) / bw * bw;
pitch[1] = (size[1] + bh - 1) / bh * bh;
pitch[2] = size[2];
/* center the image inside of the (pitched) buffer */
offset[0] = (pitch[0] - size[0]) / 2;
offset[1] = (pitch[1] - size[1]) / 2;
offset[2] = 0;
if (0 != libxsmm_mhd_typename(type, &typesize, 0/*ctypename*/)) {
const size_t nelements = pitch[0] * (1 < ndims ? (pitch[1] * (2 < ndims ? pitch[2] : 1)) : 1);
data = malloc(ncomponents * nelements * typesize);
}
else {
result = EXIT_FAILURE;
}
}
/* Read the data according to the header into the allocated buffer. */
if (EXIT_SUCCESS == result) {
result = libxsmm_mhd_read(data_filename,
offset, size, pitch, ndims, ncomponents, header_size, type, 0/*type_data*/, data,
0/*handle_element*/, 0/*extension*/, 0/*extension_size*/);
}
/* Write the data into a new file; update header_size. */
if (EXIT_SUCCESS == result) {
result = libxsmm_mhd_write("mhd_test.mhd", 0/*offset*/, pitch, pitch,
ndims, ncomponents, type, data, &header_size,
0/*extension_header*/,
0/*extension*/,
0/*extension_size*/);
}
/* Check the written data against the buffer. */
if (EXIT_SUCCESS == result) {
result = libxsmm_mhd_read(data_filename,
offset, size, pitch, ndims, ncomponents, 0/*header_size*/,
type, 0/*type_data*/, data, libxsmm_mhd_element_comparison,
0/*extension*/, 0/*extension_size*/);
}
/* Deallocate the buffer. */
free(data);
return result;
}
int main(int argc, char* argv[])
{
const libxsmm_blasint m = 1 < argc ? atoi(argv[1]) : 4096;
const libxsmm_blasint n = 2 < argc ? atoi(argv[2]) : m;
const libxsmm_blasint lda = LIBXSMM_MAX(3 < argc ? atoi(argv[3]) : 0, m);
const libxsmm_blasint ldb = LIBXSMM_MAX(4 < argc ? atoi(argv[4]) : 0, n);
REAL_TYPE *const a = (REAL_TYPE*)malloc(lda * n * sizeof(REAL_TYPE));
REAL_TYPE *const b = (REAL_TYPE*)malloc(ldb * m * sizeof(REAL_TYPE));
const unsigned int size = m * n * sizeof(REAL_TYPE);
unsigned long long start;
libxsmm_blasint i, j;
double duration;
fprintf(stdout, "m=%i n=%i lda=%i ldb=%i size=%.fMB (%s)\n", m, n, lda, ldb,
1.0 * size / (1 << 20), 8 == sizeof(REAL_TYPE) ? "DP" : "SP");
for (i = 0; i < n; ++i) {
for (j = 0; j < m; ++j) {
a[i*lda+j] = initial_value(i, j, lda);
}
}
start = libxsmm_timer_tick();
libxsmm_transpose_oop(b, a, sizeof(REAL_TYPE), m, n, lda, ldb);
libxsmm_transpose_oop(a, b, sizeof(REAL_TYPE), n, m, ldb, lda);
duration = libxsmm_timer_duration(start, libxsmm_timer_tick());
for (i = 0; i < n; ++i) {
for (j = 0; j < m; ++j) {
if (0 < fabs(a[i*lda+j] - initial_value(i, j, lda))) {
i = n + 1;
break;
}
}
}
if (i <= n) {
if (0 < duration) {
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", size / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
}
else {
fprintf(stderr, "Validation failed!\n");
}
#if defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)
{
double mkl_duration;
start = libxsmm_timer_tick();
LIBXSMM_CONCATENATE(mkl_, LIBXSMM_TPREFIX(REAL_TYPE, omatcopy))('C', 'T', m, n, 1, a, lda, b, ldb);
LIBXSMM_CONCATENATE(mkl_, LIBXSMM_TPREFIX(REAL_TYPE, omatcopy))('C', 'T', n, m, 1, b, ldb, a, lda);
mkl_duration = libxsmm_timer_duration(start, libxsmm_timer_tick());
if (0 < mkl_duration) {
fprintf(stdout, "\tMKL: %.1fx\n", duration / mkl_duration);
}
}
#endif
free(a);
free(b);
return EXIT_SUCCESS;
}
int main(int argc, char* argv[])
{
const int insize = (1 < argc ? atoi(argv[1]) : 0);
const int incrmt = (2 < argc ? atoi(argv[2]) : 0);
const int nelems = (3 < argc ? atoi(argv[3]) : 0);
const int niters = (4 < argc ? atoi(argv[4]) : 1);
const int elsize = (0 >= insize ? LIBXSMM_DESCRIPTOR_SIGSIZE : insize);
const int stride = (0 >= incrmt ? LIBXSMM_MAX(LIBXSMM_DESCRIPTOR_MAXSIZE, elsize) : LIBXSMM_MAX(incrmt, elsize));
const size_t n = (0 >= nelems ? (((size_t)2 << 30/*2 GB*/) / stride) : ((size_t)nelems));
unsigned char *input, *icopy = NULL, *ilast = NULL;
int result = EXIT_SUCCESS;
size_t nbytes, size, nrpt;
if (0 < niters) {
size = n;
nrpt = niters;
}
else {
size = LIBXSMM_MAX(LIBXSMM_ABS(niters), 1);
nrpt = n;
}
nbytes = size * stride;
input = (unsigned char*)(0 != nbytes ? malloc(nbytes) : NULL);
if (NULL != input) {
unsigned char *const ref = input + (size - 1) * stride; /* last item */
libxsmm_timer_tickint start;
size_t i, j = 0;
/* initialize the input data */
for (i = 0; i < nbytes; ++i) input[i] = LIBXSMM_MOD2(i, 128);
for (i = 0; i < (size_t)elsize; ++i) ref[i] = 255;
{ /* benchmark libxsmm_diff_n */
#if defined(USE_HASH)
const unsigned int hashref = libxsmm_hash(ref, elsize, 0/*seed*/);
#endif
start = libxsmm_timer_tick();
for (i = 0; i < nrpt; ++i) {
#if !defined(USE_HASH)
j = libxsmm_diff_n(ref, input, (unsigned char)elsize, (unsigned char)stride,
(unsigned int)LIBXSMM_MIN(i, size)/*hint*/, (unsigned int)size);
#else
const unsigned char* tst = input;
for (j = 0; j < size; ++j) {
const unsigned int hashtst = libxsmm_hash(tst, elsize, 0/*seed*/);
if (hashref == hashtst && 0 == libxsmm_diff(ref, tst, (unsigned char)elsize)) {
break;
}
tst += stride;
}
#endif
}
printf("libxsmm_diff_n:\t\t%.8f s\n", libxsmm_timer_duration(start, libxsmm_timer_tick()));
}
if (size == (j + 1) && 0 == memcmp(ref, input + j * stride, elsize)) { /* benchmark libxsmm_memcmp */
icopy = (unsigned char*)(elsize == stride ? malloc(nbytes) : NULL);
if (NULL != icopy) {
ilast = icopy + (size - 1) * stride; /* last item */
memcpy(icopy, input, nbytes);
start = libxsmm_timer_tick();
for (i = 0; i < nrpt; ++i) {
j += libxsmm_memcmp(input, icopy, nbytes); /* take result of every execution */
/* memcmp may be pure and without touching input it is not repeated (nrpt) */
ilast[i%elsize] = 255;
}
printf("libxsmm_memcmp:\t\t%.8f s\n", libxsmm_timer_duration(start, libxsmm_timer_tick()));
result += (int)j * ((int)stride / ((int)stride + 1)); /* ignore result */
}
}
else {
result = EXIT_FAILURE;
}
if (NULL != icopy) { /* benchmark stdlib's memcmp */
LIBXSMM_ASSERT(NULL != ilast);
start = libxsmm_timer_tick();
for (i = 0; i < nrpt; ++i) {
j += memcmp(input, icopy, nbytes); /* take result of every execution */
/* memcmp is likely pure and without touching input it is not repeated (nrpt) */
ilast[i%elsize] = 255;
}
printf("stdlib memcmp:\t\t%.8f s\n", libxsmm_timer_duration(start, libxsmm_timer_tick()));
result += (int)j * ((int)stride / ((int)stride + 1)); /* ignore result */
free(icopy);
}
free(input);
}
else {
result = EXIT_FAILURE;
}
return result;
}
int main(int argc, char* argv[])
{
const int ncalls = 1000000;
#if defined(_OPENMP)
const int max_nthreads = omp_get_max_threads();
#else
const int max_nthreads = 1;
#endif
const int ncycles = LIBXSMM_MAX(1 < argc ? atoi(argv[1]) : 100, 1);
const int max_nallocs = LIBXSMM_CLMP(2 < argc ? atoi(argv[2]) : 4, 1, MAX_MALLOC_N);
const int nthreads = LIBXSMM_CLMP(3 < argc ? atoi(argv[3]) : 1, 1, max_nthreads);
unsigned int nallocs = 0, nerrors = 0;
int r[MAX_MALLOC_N], i;
/* generate set of random number for parallel region */
for (i = 0; i < (MAX_MALLOC_N); ++i) r[i] = rand();
/* count number of calls according to randomized scheme */
for (i = 0; i < ncycles; ++i) {
nallocs += r[i%(MAX_MALLOC_N)] % max_nallocs + 1;
}
assert(0 != nallocs);
fprintf(stdout, "Running %i cycles with max. %i malloc+free (%u calls) using %i thread%s...\n",
ncycles, max_nallocs, nallocs, 1 >= nthreads ? 1 : nthreads, 1 >= nthreads ? "" : "s");
#if defined(LIBXSMM_OFFLOAD_TARGET)
# pragma offload target(LIBXSMM_OFFLOAD_TARGET)
#endif
{
const char *const longlife_env = getenv("LONGLIFE");
const int enable_longlife = ((0 == longlife_env || 0 == *longlife_env) ? 0 : atoi(longlife_env));
void *const longlife = (0 == enable_longlife ? 0 : malloc_offsite((MAX_MALLOC_MB) << 20));
unsigned long long d0, d1 = 0;
libxsmm_scratch_info info;
/* run non-inline function to measure call overhead of an "empty" function */
const unsigned long long t0 = libxsmm_timer_tick();
for (i = 0; i < ncalls; ++i) {
libxsmm_init(); /* subsequent calls are not doing any work */
}
d0 = libxsmm_timer_diff(t0, libxsmm_timer_tick());
#if defined(_OPENMP)
# pragma omp parallel for num_threads(nthreads) private(i) default(none) shared(r) reduction(+:d1,nerrors)
#endif
for (i = 0; i < ncycles; ++i) {
const int count = r[i%(MAX_MALLOC_N)] % max_nallocs + 1;
void* p[MAX_MALLOC_N];
int j;
assert(count <= MAX_MALLOC_N);
for (j = 0; j < count; ++j) {
const int k = (i * count + j) % (MAX_MALLOC_N);
const size_t nbytes = (r[k] % (MAX_MALLOC_MB) + 1) << 20;
const unsigned long long t1 = libxsmm_timer_tick();
p[j] = libxsmm_aligned_scratch(nbytes, 0/*auto*/);
d1 += libxsmm_timer_diff(t1, libxsmm_timer_tick());
if (0 != p[j]) {
memset(p[j], j, nbytes);
}
else {
++nerrors;
}
}
for (j = 0; j < count; ++j) {
libxsmm_free(p[j]);
}
}
libxsmm_free(longlife);
if (0 != d0 && 0 != d1 && 0 < nallocs) {
const double dcalls = libxsmm_timer_duration(0, d0);
const double dalloc = libxsmm_timer_duration(0, d1);
const double alloc_freq = 1E-3 * nallocs / dalloc;
const double empty_freq = 1E-3 * ncalls / dcalls;
fprintf(stdout, "\tallocation+free calls/s: %.1f kHz\n", alloc_freq);
fprintf(stdout, "\tempty calls/s: %.1f MHz\n", 1E-3 * empty_freq);
fprintf(stdout, "\toverhead: %.1fx\n", empty_freq / alloc_freq);
}
if (EXIT_SUCCESS == libxsmm_get_scratch_info(&info) && 0 < info.size) {
fprintf(stdout, "\nScratch: %.f MB (mallocs=%lu, pools=%u",
1.0 * info.size / (1 << 20), (unsigned long int)info.nmallocs, info.npools);
if (1 < nthreads) fprintf(stdout, ", threads=%i)\n", nthreads); else fprintf(stdout, ")\n");
libxsmm_release_scratch(); /* suppress LIBXSMM's termination message about scratch */
}
}
if (0 == nerrors) {
fprintf(stdout, "Finished\n");
return EXIT_SUCCESS;
}
else {
fprintf(stdout, "FAILED (%u errors)\n", nerrors);
return EXIT_FAILURE;
}
}
int main(void)
{
const libxsmm_blasint m[] = { 1, 1, 1, 1, 2, 3, 5, 5, 5, 16, 63, 16, 75, 2507 };
const libxsmm_blasint n[] = { 1, 7, 7, 7, 2, 3, 1, 1, 1, 16, 31, 500, 130, 1975 };
const libxsmm_blasint ldi[] = { 1, 1, 1, 9, 2, 3, 5, 8, 8, 16, 64, 16, 87, 3000 };
const libxsmm_blasint ldo[] = { 1, 7, 8, 8, 2, 3, 1, 1, 4, 16, 32, 512, 136, 3072 };
const int start = 0, ntests = sizeof(m) / sizeof(*m);
libxsmm_blasint max_size_a = 0, max_size_b = 0;
unsigned int nerrors = 0;
ELEM_TYPE *a = 0, *b = 0;
int test;
for (test = start; test < ntests; ++test) {
const libxsmm_blasint size_a = ldi[test] * n[test], size_b = ldo[test] * m[test];
assert(m[test] <= ldi[test] && n[test] <= ldo[test]);
max_size_a = LIBXSMM_MAX(max_size_a, size_a);
max_size_b = LIBXSMM_MAX(max_size_b, size_b);
}
a = (ELEM_TYPE*)libxsmm_malloc((size_t)(max_size_a * sizeof(ELEM_TYPE)));
b = (ELEM_TYPE*)libxsmm_malloc((size_t)(max_size_b * sizeof(ELEM_TYPE)));
assert(0 != a && 0 != b);
LIBXSMM_MATINIT(ELEM_TYPE, 42, a, max_size_a, 1, max_size_a, 1.0);
LIBXSMM_MATINIT(ELEM_TYPE, 0, b, max_size_b, 1, max_size_b, 1.0);
for (test = start; test < ntests; ++test) {
unsigned int testerrors = (EXIT_SUCCESS == libxsmm_otrans(
b, a, sizeof(ELEM_TYPE), m[test], n[test],
ldi[test], ldo[test]) ? 0 : 1);
if (0 == testerrors) {
libxsmm_blasint i, j;
for (i = 0; i < n[test]; ++i) {
for (j = 0; j < m[test]; ++j) {
const libxsmm_blasint u = i * ldi[test] + j;
const libxsmm_blasint v = j * ldo[test] + i;
testerrors += (LIBXSMM_FEQ(a[u], b[v]) ? 0u : 1u);
}
}
}
if (nerrors < testerrors) {
nerrors = testerrors;
}
}
if (0 == nerrors) { /* previous results are correct and may be used to validate other tests */
for (test = start; test < ntests; ++test) {
/* prepare expected results in b (correct according to the previous test block) */
libxsmm_otrans(b, a, sizeof(ELEM_TYPE), m[test], n[test], ldi[test], ldo[test]);
if (m[test] == n[test] && ldi[test] == ldo[test]) {
unsigned int testerrors = (EXIT_SUCCESS == libxsmm_otrans(
a, a, sizeof(ELEM_TYPE), m[test], n[test],
ldi[test], ldo[test]) ? 0 : 1);
if (0 == testerrors) {
libxsmm_blasint i, j;
for (i = 0; i < n[test]; ++i) {
for (j = 0; j < m[test]; ++j) {
/* address serves both a and b since ldi and ldo are equal */
const libxsmm_blasint uv = i * ldi[test] + j;
testerrors += (LIBXSMM_FEQ(a[uv], b[uv]) ? 0u : 1u);
}
}
}
if (nerrors < testerrors) {
nerrors = testerrors;
}
}
else { /* negative tests */
nerrors = LIBXSMM_MAX(EXIT_SUCCESS != libxsmm_otrans(
a, a, sizeof(ELEM_TYPE), m[test], n[test],
ldi[test], ldo[test]) ? 0u : 1u, nerrors);
}
}
}
libxsmm_free(a);
libxsmm_free(b);
if (0 == nerrors) {
return EXIT_SUCCESS;
}
else {
# if defined(_DEBUG)
fprintf(stderr, "errors=%u\n", nerrors);
# endif
return EXIT_FAILURE;
}
}
int main(int argc, char* argv[])
{
unsigned int m=8, n=8, lda=8, ldb=8, nerrs, num, nmat, nmats, nmatd, ntest;
unsigned int layout, asize, VLEND=4, VLENS=8, bsize;
unsigned int ncorr;
int i, j;
char side, uplo, trans, diag;
unsigned int typesize8 = 8;
unsigned int typesize4 = 4;
float *sa, *sb, *sc, *sd;
double *da, *db, *dc, *dd, *tmpbuf;
double dalpha = 1.0;
float salpha;
double dtmp;
const unsigned char *cptr;
unsigned long op_count;
const libxsmm_trsm_descriptor* desc8 = NULL;
const libxsmm_trsm_descriptor* desc4 = NULL;
libxsmm_descriptor_blob blob;
union {
libxsmm_xtrsmfunction dp;
libxsmm_xtrsmfunction sp;
const void* pv;
} mykernel = { 0 };
#ifdef USE_KERNEL_GENERATION_DIRECTLY
void (*opcode_routine)();
#endif
#ifdef USE_KERNEL_GENERATION_DIRECTLY
#include <unistd.h>
#include <signal.h>
#include <malloc.h>
#include <sys/mman.h>
#include "../../src/generator_packed_trsm_avx_avx512.h"
unsigned char *routine_output;
libxsmm_generated_code io_generated_code;
int pagesize = sysconf(_SC_PAGE_SIZE);
if (pagesize == -1) fprintf(stderr,"sysconf pagesize\n");
routine_output = (unsigned char *) mmap(NULL,
BUFSIZE2, PROT_READ|PROT_WRITE,
MAP_PRIVATE|MAP_ANONYMOUS, 0,0);
if (mprotect(routine_output, BUFSIZE2,
PROT_EXEC | PROT_READ | PROT_WRITE ) == -1)
fprintf(stderr,"mprotect\n");
printf("Routine ready\n");
io_generated_code.generated_code = &routine_output[0];
io_generated_code.buffer_size = BUFSIZE2;
io_generated_code.code_size = 0;
io_generated_code.code_type = 2;
io_generated_code.last_error = 0;
#endif
if ( argc <= 3 )
{
printf("\nUSAGE: %s m n lda ldb nmat side uplo trans diag layout ntest alpha\n",argv[0]);
printf("Compact TRSM a mxn matrix of leading dimension ldb\n");
printf("This will test the jit of 1 VLEN work of nmat at a time\n");
printf("Defaults: m=n=lda=ldb=nmat=8, alpha=1.0, side=uplo='L',trans=diag='N',layout=102,ntest=1\n");
}
if ( argc > 1 ) m = atoi(argv[1]); else m = 8;
if ( argc > 2 ) n = atoi(argv[2]); else n = 8;
if ( argc > 3 ) lda= atoi(argv[3]); else lda = 8;
if ( argc > 4 ) ldb = atoi(argv[4]); else ldb = 8;
if ( argc > 5 ) nmat = atoi(argv[5]); else nmat = 8;
if ( argc > 6 ) side = argv[6][0]; else side = 'L';
if ( argc > 7 ) uplo = argv[7][0]; else uplo = 'L';
if ( argc > 8 ) trans = argv[8][0]; else trans = 'N';
if ( argc > 9 ) diag = argv[9][0]; else diag = 'N';
if ( argc > 10 ) layout = atoi(argv[10]); else layout=102;
if ( argc > 11 ) ntest = atoi(argv[11]); else ntest = 1;
if ( argc > 12 ) dalpha = atof(argv[12]); else dalpha = 1.0;
salpha = (float)dalpha;
m = LIBXSMM_MAX(m,1);
n = LIBXSMM_MAX(n,1);
/* A is either mxm or nxn depending on side */
if ( (side == 'L') || (side=='l') ) asize = m; else asize = n;
lda = LIBXSMM_MAX(lda,asize);
if ( layout == 102 )
{
/* Column major: B is mxn, and stored in B format */
ldb = LIBXSMM_MAX(ldb,m);
bsize = ldb*n;
} else {
/* Row major: B is mxn, and stored in B^T format */
ldb = LIBXSMM_MAX(ldb,n);
bsize = ldb*m;
}
nmats = LIBXSMM_MAX(VLENS,nmat - (nmat%VLENS));
nmatd = LIBXSMM_MAX(VLEND,nmat - (nmat%VLEND));
nmat = LIBXSMM_MAX(nmats,nmatd);
op_count = n * m * asize;
printf("This is a real*%u tester for JIT compact TRSM kernels! (%c%c%c%c m=%u n=%u lda=%u ldb=%u layout=%u nmat=%u)\n",typesize8,side,uplo,trans,diag,m,n,lda,ldb,layout,nmat);
#ifdef USE_XSMM_GENERATED
printf("This code tests the LIBXSMM generated kernels\n");
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
printf("This code tests some predefined assembly kenrel\n");
#endif
#ifdef USE_KERNEL_GENERATION_DIRECTLY
printf("This code tests kernel generation directly\n");
#endif
#ifdef TIME_MKL
printf("This code tests MKL compact batch directly\n");
#endif
desc8 = libxsmm_trsm_descriptor_init(&blob, typesize8, m, n, lda, ldb, &dalpha, trans, diag, side, uplo, layout);
desc4 = libxsmm_trsm_descriptor_init(&blob, typesize4, m, n, lda, ldb, &salpha, trans, diag, side, uplo, layout);
#ifdef USE_XSMM_GENERATED
printf("calling libxsmm_dispatch_trsm: typesize8=%u\n",typesize8);
mykernel.dp = libxsmm_dispatch_trsm(desc8);
printf("done calling libxsmm_dispatch_trsm: typesize8=%u\n",typesize8);
if ( mykernel.dp == NULL ) printf("R8 Kernel after the create call is null\n");
mykernel.sp = libxsmm_dispatch_trsm(desc4);
if ( mykernel.sp == NULL ) printf("R4 kernel after the create call is null\n");
#endif
#ifdef USE_KERNEL_GENERATION_DIRECTLY
libxsmm_generator_trsm_kernel ( &io_generated_code, &desc8, "hsw" );
#endif
#ifndef NO_ACCURACY_CHECK
printf("mallocing matrices\n");
#endif
sa = (float *) malloc ( lda*asize*nmats*sizeof(float) );
da = (double *) malloc ( lda*asize*nmatd*sizeof(double) );
sb = (float *) malloc ( bsize*nmats*sizeof(float) );
db = (double *) malloc ( bsize*nmatd*sizeof(double) );
sc = (float *) malloc ( bsize*nmats*sizeof(float) );
dc = (double *) malloc ( bsize*nmatd*sizeof(double) );
sd = (float *) malloc ( bsize*nmats*sizeof(float) );
dd = (double *) malloc ( bsize*nmatd*sizeof(double) );
tmpbuf = (double *) malloc ( asize*VLEND*sizeof(double) );
#ifndef NO_ACCURACY_CHECK
printf("filling matrices\n");
#endif
sfill_matrix ( sa, lda, asize, asize*nmats );
#ifdef TRIANGLE_IS_IDENTITY
printf("Warning: setting triangular matrix to identity. Not good for accuracy testing\n");
dfill_identity ( da, lda, asize, asize, VLEND, nmatd/VLEND );
#else
dfill_matrix ( da, lda, asize, asize*nmatd );
#endif
sfill_matrix ( sb, bsize, bsize, nmats );
dfill_matrix ( db, bsize, bsize, nmatd );
#ifndef NO_ACCURACY_CHECK
for ( i = 0 ; i < (int)(bsize*nmats) ; i++ ) sc[i]=sb[i];
for ( i = 0 ; i < (int)(bsize*nmatd) ; i++ ) dc[i]=db[i];
for ( i = 0 ; i < (int)(bsize*nmats) ; i++ ) sd[i]=sb[i];
for ( i = 0 ; i < (int)(bsize*nmatd) ; i++ ) dd[i]=db[i];
printf("Pointing at the kernel now\n");
#endif
#ifdef USE_XSMM_GENERATED
cptr = (const unsigned char*) mykernel.pv;
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
cptr = (const unsigned char*) trsm_xct_;
#endif
#ifdef USE_KERNEL_GENERATION_DIRECTLY
cptr = (const unsigned char*) &routine_output[0];
opcode_routine = (void *) &cptr[0];
#endif
#ifndef TIME_MKL
#define DUMP_ASSEMBLY_FILE
#endif
#ifdef DUMP_ASSEMBLY_FILE
printf("Dumping assembly file\n");
FILE *fp = fopen("foo.s","w");
char buffer[80];
fputs("\t.text\n",fp);
fputs("\t.align 256\n",fp);
fputs("\t.globl trsm_xct_\n",fp);
fputs("trsm_xct_:\n",fp);
for (i = 0 ; i < 4000; i+=4 )
{
sprintf(buffer,".byte 0x%02x, 0x%02x, 0x%02x, 0x%02x\n",cptr[i],cptr[i+1],cptr[i+2],cptr[i+3]);
fputs(buffer,fp);
}
fputs("\tretq\n",fp);
fputs("\t.type trsm_xct_,@function\n",fp);
fputs("\t.size trsm_xct_,.-trsm_xct_\n",fp);
fclose(fp);
#endif
#if defined(USE_MKL_FOR_REFERENCE) || defined(TIME_MKL)
#include "mkl.h"
MKL_LAYOUT CLAYOUT = (layout == 101) ? MKL_ROW_MAJOR : MKL_COL_MAJOR;
MKL_SIDE SIDE = (side == 'R' || side == 'r') ? MKL_RIGHT : MKL_LEFT;
MKL_UPLO UPLO = (uplo == 'U' || uplo == 'u') ? MKL_UPPER : MKL_LOWER;
MKL_TRANSPOSE TRANSA = (trans == 'N' || trans == 'n') ? MKL_NOTRANS : MKL_TRANS;
MKL_DIAG DIAG = (diag == 'N' || diag == 'n') ? MKL_NONUNIT : MKL_UNIT;
MKL_COMPACT_PACK CMP_FORMAT = mkl_get_format_compact();
#if 0
MKL_COMPACT_PACK CMP_FORMAT = MKL_COMPACT_AVX;
#endif
#endif
#ifndef NO_ACCURACY_CHECK
printf("Before routine, initial B(1,1)=%g B[256]=%g\n",db[0],db[256]);
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
double one = 1.0;
#endif
double timer;
#ifdef MKL_TIMER
double tmptimer;
tmptimer = dsecnd_();
#else
unsigned long long l_start, l_end;
#endif
timer = 0.0;
for ( j = 0 ; j < (int)ntest ; j++ )
{
#ifndef TRIANGLE_IS_IDENTITY
for ( i = 0 ; i < (int)(bsize*nmatd) ; i++ ) db[i]=dd[i];
#endif
for ( i = 0 , num = 0; i < (int)nmatd ; i+= (int)VLEND, num++ )
{
double *Ap = &da[num*lda*asize*VLEND];
double *Bp = &db[num*bsize*VLEND];
#ifdef MKL_TIMER
tmptimer = dsecnd_();
#else
l_start = libxsmm_timer_tick();
#endif
#ifdef USE_XSMM_GENERATED
mykernel.dp ( Ap, Bp, tmpbuf );
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
trsm_xct_ ( Ap, Bp, &one );
#endif
#ifdef USE_KERNEL_GENERATION_DIRECTLY
(*opcode_routine)( Ap, Bp );
#endif
#ifdef TIME_MKL
mkl_dtrsm_compact ( CLAYOUT, SIDE, UPLO, TRANSA, DIAG, m, n, dalpha, da, lda, db, ldb, CMP_FORMAT, nmatd );
i+=nmatd; /* Because MKL will do everything */
#endif
#ifdef MKL_TIMER
timer += dsecnd_() - tmptimer;
#else
l_end = libxsmm_timer_tick();
timer += libxsmm_timer_duration(l_start,l_end);
#endif
}
}
timer /= ((double)ntest);
#ifndef NO_ACCURACY_CHECK
printf("Average time to get through %u matrices: %g\n",nmatd,timer);
printf("Gflops: %g\n",(double)(op_count*nmatd)/(timer*1.0e9));
printf("after routine, new B(1,1)=%g B[256]=%g\n",db[0],db[256]);
#endif
#ifdef TEST_SINGLE
printf("Before r4 routine, initial B(1,1)=%g B[256]=%g\n",sb[0],sb[256]);
for ( i = 0 , num = 0; i < nmats ; i+= VLENS, num++ )
{
float *Ap = &sa[num*lda*asize*VLENS];
float *Bp = &sb[num*bsize*VLENS];
#ifdef USE_XSMM_GENERATED
mykernel.sp ( Ap, Bp, NULL );
#endif
}
printf("after r4 routine, new B(1,1)=%g B]256]=%g\n",db[0],db[256]);
#endif
#ifndef NO_ACCURACY_CHECK
/* Call some reference code now on a copy of the B matrix (C) */
double timer2 = 0.0;
for ( j = 0 ; j < (int)ntest ; j++ )
{
#ifndef TRIANGLE_IS_IDENTITY
for ( i = 0 ; i < (int)(bsize*nmatd) ; i++ ) dc[i]=dd[i];
#endif
#ifdef MKL_TIMER
tmptimer = dsecnd_();
#else
l_start = libxsmm_timer_tick();
#endif
#ifdef USE_MKL_FOR_REFERENCE
mkl_dtrsm_compact ( CLAYOUT, SIDE, UPLO, TRANSA, DIAG, m, n, dalpha, da, lda, dc, ldb, CMP_FORMAT, nmatd );
#elif !defined(LIBXSMM_NOFORTRAN)
if ( (layout == 101) && (nmatd!=VLEND) )
{
unsigned int lay = 102, m1 = n, n1 = m;
char side1='L', uplo1='L';
if ( side == 'L' || side == 'l' ) side1 = 'R';
if ( uplo == 'L' || uplo == 'l' ) uplo1 = 'U';
compact_dtrsm_ ( &lay, &side1, &uplo1, &trans, &diag, &m1, &n1, &dalpha, da, &lda, dc, &ldb, &nmatd, &VLEND );
} else {
compact_dtrsm_ ( &layout, &side, &uplo, &trans, &diag, &m, &n, &dalpha, da, &lda, dc, &ldb, &nmatd, &VLEND );
}
#endif
#ifdef MKL_TIMER
timer2 += dsecnd_() - tmptimer;
#else
l_end = libxsmm_timer_tick();
timer2 += libxsmm_timer_duration(l_start,l_end);
#endif
}
timer2 /= ((double)ntest);
printf("Reference time=%g Reference Gflops=%g\n",timer2,(op_count*nmatd)/(timer2*1.0e9));
/* Compute the residual between B and C */
dtmp = residual_d ( dc, bsize, bsize, nmatd, db, bsize, &nerrs, &ncorr );
printf("R8 %c%c%c%c m=%u n=%u lda=%u ldb=%u error: %g number of errors: %u corrects: %u",side,uplo,trans,diag,m,n,lda,ldb,dtmp,nerrs,ncorr);
if ( nerrs > 0 ) printf(" ->FAILED at %ux%u real*8 %u case",m,n,layout);
printf("\n");
#ifdef TEST_SINGLE
/* Call some reference code now on a copy of the B matrix (C) */
compact_strsm_ ( &layout, &side, &uplo, &trans, &diag, &m, &n, &salpha, sa, &lda, sc, &ldb, &nmats, &VLENS );
/* Compute the residual between B and C */
dtmp = residual_s ( sc, bsize, bsize, nmats, sb, bsize, &nerrs, &ncorr );
printf("R4 %c%c%c%c m=%u n=%u lda=%u ldb=%u error: %g number of errors: %u corrects: %u\n",side,uplo,trans,diag,m,n,lda,ldb,dtmp,nerrs,ncorr);
if ( nerrs > 0 ) printf(" ->FAILED at %ux%u real*4 case",m,n);
printf("\n");
#endif
#else
for ( j = 0, nerrs = 0 ; j < bsize*nmatd ; j++ )
{
if ( isnan(db[j]) || isinf(db[j]) )
{
if ( ++nerrs < 10 )
{
printf("WARNING: db[%d]=%g\n",j,db[j]);
}
}
}
printf("%g,real*8 %c%c%c%c m=%u n=%u lda=%u ldb=%u Denormals=%u Time=%g Gflops=%g",(op_count*nmatd)/(timer*1.0e9),side,uplo,trans,diag,m,n,lda,ldb,nerrs,timer,(op_count*nmatd)/(timer*1.0e9));
if ( nerrs > 0 ) printf(" -> FAILED at %ux%u real*8 case",m,n);
printf("\n");
#endif
free(dd);
free(sd);
free(dc);
free(sc);
free(db);
free(sb);
free(da);
free(sa);
return 0;
}
int main(int argc, char* argv[])
{
const char t = (char)(1 < argc ? *argv[1] : 'o');
const libxsmm_blasint m = (2 < argc ? atoi(argv[2]) : 4096);
#if 0 /* TODO: enable when in-place transpose is fully supported */
const libxsmm_blasint n = (3 < argc ? atoi(argv[3]) : m);
#else
const libxsmm_blasint n = (3 < argc ? (('o' == t || 'O' == t) ? atoi(argv[3]) : m) : m);
#endif
const libxsmm_blasint ldi = LIBXSMM_MAX/*sanitize ld*/(4 < argc ? atoi(argv[4]) : 0, m);
const libxsmm_blasint ldo = LIBXSMM_MAX/*sanitize ld*/(5 < argc ? atoi(argv[5]) : 0, n);
const int r = (6 < argc ? atoi(argv[6]) : 0), s = LIBXSMM_ABS(r);
const libxsmm_blasint lower = (7 < argc ? atoi(argv[7]) : 0);
libxsmm_blasint km = m, kn = n, kldi = ldi, kldo = (('o' == t || 'O' == t) ? ldo : ldi);
int result = EXIT_SUCCESS, k;
if (0 == strchr("oOiI", t)) {
fprintf(stderr, "%s [<transpose-kind:o|i>] [<m>] [<n>] [<ld-in>] [<ld-out>] [random:0|nruns] [lbound]\n", argv[0]);
exit(EXIT_FAILURE);
}
#if defined(LIBXSMM_OFFLOAD_TARGET)
# pragma offload target(LIBXSMM_OFFLOAD_TARGET)
#endif
{
const char *const env_tasks = getenv("TASKS"), *const env_check = getenv("CHECK");
const int tasks = (0 == env_tasks || 0 == *env_tasks) ? 0/*default*/ : atoi(env_tasks);
const int check = (0 == env_check || 0 == *env_check) ? 1/*default*/ : atoi(env_check);
ELEM_TYPE *const a = (ELEM_TYPE*)libxsmm_malloc((size_t)(ldi * (('o' == t || 'O' == t) ? n : ldo) * sizeof(ELEM_TYPE)));
ELEM_TYPE *const b = (ELEM_TYPE*)libxsmm_malloc((size_t)(ldo * (('o' == t || 'O' == t) ? m : ldi) * sizeof(ELEM_TYPE)));
libxsmm_timer_tickint start, duration = 0;
#if defined(USE_REFERENCE) /* benchmark against a reference */
libxsmm_timer_tickint duration2 = 0;
#endif
libxsmm_blasint i;
size_t size = 0;
#if defined(MKL_ENABLE_AVX512)
mkl_enable_instructions(MKL_ENABLE_AVX512);
#endif
fprintf(stdout, "m=%lli n=%lli ldi=%lli ldo=%lli size=%.fMB (%s, %s)\n",
(long long)m, (long long)n, (long long)ldi, (long long)ldo,
1.0 * (m * n * sizeof(ELEM_TYPE)) / (1 << 20), LIBXSMM_STRINGIFY(ELEM_TYPE),
('o' == t || 'O' == t) ? "out-of-place" : "in-place");
#if defined(_OPENMP)
# pragma omp parallel for private(i)
#endif
for (i = 0; i < n; ++i) {
libxsmm_blasint j;
for (j = 0; j < m; ++j) {
a[i*ldi+j] = initial_value(i, j, m);
}
}
if (0 != check) { /* repeatable (reference) */
srand(RAND_SEED);
}
else { /* randomized selection */
srand(libxsmm_timer_tick() % ((unsigned int)-1));
}
for (k = (0 == r ? -1 : 0); k < s && EXIT_SUCCESS == result; ++k) {
if (0 < r) {
const libxsmm_blasint rldi = 0 <= lower ? randstart(lower, ldi) : 0;
km = randstart(LIBXSMM_ABS(lower), m);
kldi = LIBXSMM_MAX(rldi, km);
if (('o' == t || 'O' == t)) {
const libxsmm_blasint rldo = 0 <= lower ? randstart(lower, ldo) : 0;
kn = randstart(LIBXSMM_ABS(lower), n);
kldo = LIBXSMM_MAX(rldo, kn);
/* trigger JIT-generated code */
OTRANS(b, a, sizeof(ELEM_TYPE), km, kn, kldi, kldo);
}
else {
#if 0 /* TODO: enable when in-place transpose is fully supported */
kn = randstart(LIBXSMM_ABS(lower), n);
#else
kn = km;
#endif
kldo = kldi;
/* trigger JIT-generated code */
ITRANS(b, sizeof(ELEM_TYPE), km, kn, kldi);
}
}
size += (size_t)(km * kn * sizeof(ELEM_TYPE));
if (('o' == t || 'O' == t)) {
if (0 == tasks) { /* library-internal parallelization */
start = libxsmm_timer_tick();
#if defined(OTRANS_THREAD)
# pragma omp parallel
OTRANS_THREAD(b, a, sizeof(ELEM_TYPE), km, kn, kldi, kldo, omp_get_thread_num(), omp_get_num_threads());
#else
result = OTRANS(b, a, sizeof(ELEM_TYPE), km, kn, kldi, kldo);
#endif
duration += libxsmm_timer_diff(start, libxsmm_timer_tick());
}
else { /* external parallelization */
start = libxsmm_timer_tick();
#if defined(_OPENMP)
# pragma omp parallel
# pragma omp single nowait
#endif
result = OTRANS(b, a, sizeof(ELEM_TYPE), km, kn, kldi, kldo);
duration += libxsmm_timer_diff(start, libxsmm_timer_tick());
}
}
else {
assert(('i' == t || 'I' == t) && kldo == kldi);
memcpy(b, a, (size_t)(kldi * kn * sizeof(ELEM_TYPE)));
if (2 > tasks) { /* library-internal parallelization */
start = libxsmm_timer_tick();
result = ITRANS(b, sizeof(ELEM_TYPE), km, kn, kldi);
duration += libxsmm_timer_diff(start, libxsmm_timer_tick());
}
else { /* external parallelization */
start = libxsmm_timer_tick();
#if defined(_OPENMP)
# pragma omp parallel
# pragma omp single
#endif
result = ITRANS(b, sizeof(ELEM_TYPE), km, kn, kldi);
duration += libxsmm_timer_diff(start, libxsmm_timer_tick());
}
}
if (0 != check) { /* check */
for (i = 0; i < km; ++i) {
libxsmm_blasint j;
for (j = 0; j < kn; ++j) {
const ELEM_TYPE u = b[i*kldo+j];
const ELEM_TYPE v = a[j*kldi+i];
if (LIBXSMM_NEQ(u, v)) {
i += km; /* leave outer loop as well */
result = EXIT_FAILURE;
break;
}
}
}
}
}
#if defined(USE_REFERENCE)
if (0 < check) { /* check shall imply reference (performance-)test */
srand(RAND_SEED); /* reproduce the same sequence as above */
for (k = (0 == r ? -1 : 0); k < s && EXIT_SUCCESS == result; ++k) {
if (0 < r) {
const libxsmm_blasint rldi = 0 <= lower ? randstart(lower, ldi) : 0;
km = randstart(LIBXSMM_ABS(lower), m);
kldi = LIBXSMM_MAX(rldi, km);
if (('o' == t || 'O' == t)) {
const libxsmm_blasint rldo = 0 <= lower ? randstart(lower, ldo) : 0;
kn = randstart(LIBXSMM_ABS(lower), n);
kldo = LIBXSMM_MAX(rldo, kn);
}
else {
#if 0 /* TODO: enable when in-place transpose is fully supported */
kn = randstart(LIBXSMM_ABS(lower), n);
#else
kn = km;
#endif
kldo = kldi;
}
}
if (('o' == t || 'O' == t)) {
start = libxsmm_timer_tick();
OTRANS_GOLD(&km, &kn, a, &kldi, b, &kldo);
duration2 += libxsmm_timer_diff(start, libxsmm_timer_tick());
}
else {
assert(('i' == t || 'I' == t) && kldo == kldi);
memcpy(b, a, (size_t)(kldi * kn * sizeof(ELEM_TYPE)));
start = libxsmm_timer_tick();
ITRANS_GOLD(&km, &kn, b, &kldi, &kldo);
duration2 += libxsmm_timer_diff(start, libxsmm_timer_tick());
}
if (1 < check || 0 > check) { /* check */
for (i = 0; i < km; ++i) {
libxsmm_blasint j;
for (j = 0; j < kn; ++j) {
const ELEM_TYPE u = b[i*kldo+j];
const ELEM_TYPE v = a[j*kldi+i];
if (LIBXSMM_NEQ(u, v)) {
i += km; /* leave outer loop as well */
result = EXIT_FAILURE;
break;
}
}
}
}
}
}
#endif
if (EXIT_SUCCESS == result) {
const double d = libxsmm_timer_duration(0, duration);
if (0 < duration) {
/* out-of-place transpose bandwidth assumes RFO */
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", size
* ((('o' == t || 'O' == t)) ? 3 : 2) / (d * (1 << 30)));
}
if (0 == lower) {
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * (d / (0 == r ? (s + 1) : s)));
}
else {
fprintf(stdout, "\tduration: %f ms\n", 1000.0 * d);
}
#if defined(USE_REFERENCE)
if (0 < duration2) {
fprintf(stdout, "\treference: %.1fx\n", (1.0 * duration) / duration2);
}
#endif
}
else if (0 != check) { /* check */
fprintf(stderr, "Error: validation failed for m=%lli, n=%lli, ldi=%lli, and ldo=%lli!\n",
(long long)km, (long long)kn, (long long)kldi, (long long)kldo);
}
libxsmm_free(a);
libxsmm_free(b);
}
return result;
}
int main(int argc, char* argv[])
{
unsigned int m=8, n=8, k=8, lda=8, ldb=8, ldc=8, nerrs, num, nmat;
unsigned int layout, asize, bsize, ntest, ncorr;
#ifdef AVX512_TESTING
unsigned int VLEND=8, VLENS=16;
int arch=LIBXSMM_X86_AVX512_CORE;
#else
unsigned int VLEND=4, VLENS=8;
int arch=LIBXSMM_X86_AVX2;
#endif
unsigned int nmats, nmatd;
unsigned int i, j, l, iunroll, junroll, loopi, loopj;
char side='L', uplo='U', transa='N', transb='N', diag='N';
unsigned int typesize8 = 8;
unsigned int typesize4 = 4;
float *sa, *sb, *sc, *sd, *sc1;
double *da, *db, *dc, *dd, *dc1;
double dalpha = 1.0;
float salpha = (float)dalpha;
double dbeta = 1.0;
float sbeta = (float)dbeta;
double dtmp;
const unsigned char *cptr = NULL;
unsigned long op_count;
const libxsmm_pgemm_descriptor* desc8 = NULL;
const libxsmm_pgemm_descriptor* desc4 = NULL;
#ifdef USE_XSMM_GENERATED
libxsmm_descriptor_blob blob;
libxsmm_pgemm_xfunction mykernel = NULL;
#endif
#if defined(USE_KERNEL_GENERATION_DIRECTLY) && defined(__linux__)
void (*opcode_routine)();
unsigned char *routine_output;
libxsmm_generated_code io_generated_code;
int pagesize = sysconf(_SC_PAGE_SIZE);
if (pagesize == -1) fprintf(stderr,"sysconf pagesize\n");
routine_output = (unsigned char *) mmap(NULL,
BUFSIZE2, PROT_READ|PROT_WRITE,
MAP_PRIVATE|MAP_ANONYMOUS, 0,0);
if (mprotect(routine_output, BUFSIZE2,
PROT_EXEC | PROT_READ | PROT_WRITE ) == -1)
fprintf(stderr,"mprotect\n");
printf("Routine ready\n");
io_generated_code.generated_code = &routine_output[0];
io_generated_code.buffer_size = BUFSIZE2;
io_generated_code.code_size = 0;
io_generated_code.code_type = 2;
io_generated_code.last_error = 0;
#endif
printf("\nUSAGE: %s m n k lda ldb ldc nmat layout ntest transa transb iunroll junroll loopj loopi\n",argv[0]);
if ( argc <= 3 )
{
#ifdef TEST_SINGLE
printf("Compact SGEMM a C_mxn<-C_mxn+A_mxk*B_kxn matrix of leading dims lda/b/c\n");
printf("This will test the jit of 1 VLEN=%d ",VLENS);
if ( VLENS==8 ) printf("(AVX2)");
else printf("(AVX512)");
#else
printf("Compact DGEMM a C_mxn<-C_mxn+A_mxk*B_kxn matrix of leading dims lda/b/c\n");
printf("This will test the jit of 1 VLEN=%d ",VLEND);
if ( VLEND==4 ) printf("(AVX2)");
else printf("(AVX512)");
#endif
printf(" work of nmat at a time\n");
printf("Configurable: M-loop controlled by iunroll & loopi. N-loop by junroll & loopj\n");
printf("Defaults: m=n=k=lda=ldb=ldc=nmat=8, layout=102 (col major), transa=/b='N', ntest=1\n");
}
if ( argc > 1 ) m = atoi(argv[1]); else m = 8;
if ( argc > 2 ) n = atoi(argv[2]); else n = 8;
if ( argc > 3 ) k = atoi(argv[3]); else k = 8;
if ( argc > 4 ) lda= atoi(argv[4]); else lda = 8;
if ( argc > 5 ) ldb= atoi(argv[5]); else ldb = 8;
if ( argc > 6 ) ldc= atoi(argv[6]); else ldc = 8;
if ( argc > 7 ) nmat = atoi(argv[7]); else nmat = 8;
if ( argc > 8 ) layout = atoi(argv[8]); else layout=102;
if ( argc > 9 ) ntest = atoi(argv[9]); else ntest = 1;
if ( argc > 10 ) transa = argv[10][0]; else transa = 'N';
if ( argc > 11 ) transb = argv[11][0]; else transb = 'N';
if ( argc > 12 ) iunroll=atoi(argv[12]); else iunroll=0;
if ( argc > 13 ) junroll=atoi(argv[13]); else junroll=0;
if ( argc > 14 ) loopj=atoi(argv[14]); else loopj=0;
if ( argc > 15 ) loopi=atoi(argv[15]); else loopi=0;
salpha = (float)dalpha;
m = LIBXSMM_MAX(m,1);
n = LIBXSMM_MAX(n,1);
k = LIBXSMM_MAX(k,1);
ntest = LIBXSMM_MAX(ntest,1);
nmat = LIBXSMM_MAX(nmat,VLEND);
layout = LIBXSMM_MAX(LIBXSMM_MIN(layout,102),101);
if ( transa!='N' && transa!='n' && transa!='T' && transa!='t' ) transa='N';
if ( transb!='N' && transb!='n' && transb!='T' && transb!='t' ) transb='N';
lda = LIBXSMM_MAX(lda,m);
ldb = LIBXSMM_MAX(ldb,k);
ldc = LIBXSMM_MAX(ldc,m);
nmats = LIBXSMM_MAX(VLENS,nmat - (nmat%VLENS));
nmatd = LIBXSMM_MAX(VLEND,nmat - (nmat%VLEND));
#ifdef TEST_SINGLE
nmat = nmats;
#else
nmat = nmatd;
#endif
op_count = (unsigned long)(nmat * 2.0 * (double)m * (double)n * (double)k);
#ifdef TEST_SINGLE
printf("This is a real*%d tester for JIT compact SGEMM %c%c kernels! (m=%u n=%u k=%u lda=%u ldb=%u ldc=%u layout=%d nmat=%d alpha=%g beta=%g iun=%d jun=%d loopi=%d loopj=%d VLEN=%d)\n",typesize4,transa,transb,m,n,k,lda,ldb,ldc,layout,nmat,dalpha,dbeta,iunroll,junroll,loopi,loopj,VLENS);
#else
printf("This is a real*%d tester for JIT compact DGEMM %c%c kernels! (m=%u n=%u k=%u lda=%u ldb=%u ldc=%u layout=%d nmat=%d alpha=%g beta=%g iun=%d jun=%d loopi=%d loopj=%d VLEN=%d)\n",typesize8,transa,transb,m,n,k,lda,ldb,ldc,layout,nmat,dalpha,dbeta,iunroll,junroll,loopi,loopj,VLEND);
#endif
#ifdef USE_XSMM_GENERATED
printf("This code tests the LIBXSMM generated kernels\n");
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
printf("This code tests some predefined assembly kernel\n");
#endif
#if defined(USE_KERNEL_GENERATION_DIRECTLY) && defined(__linux__)
printf("This code tests kernel generation directly\n");
#endif
#ifdef TIME_MKL
printf("This code tests MKL compact batch directly\n");
#endif
#ifdef AVX512_TESTING
printf("This tests AVX512 binaries\n");
#endif
#ifdef AVX2_TESTING
printf("This tests AVX2 binaries\n");
#endif
desc8 = libxsmm_pgemm_descriptor_init(&blob, typesize8, m, n, k, lda, ldb, ldc, &dalpha, transa, transb, layout );
#ifdef TEST_SINGLE
desc4 = libxsmm_pgemm_descriptor_init(&blob, typesize4, m, n, k, lda, ldb, ldc, &dalpha, transa, transb, layout );
#endif
printf("Descriptor set\n");
#ifdef USE_XSMM_GENERATED
printf("calling libxsmm_dispatch_pgemm: typesize8=%u\n",typesize8);
mykernel = libxsmm_dispatch_pgemm(desc8);
printf("done calling libxsmm_dispatch_pgemm: typesize8=%u\n",typesize8);
if ( mykernel == NULL ) printf("R8 Kernel after the create call is null\n");
#ifdef TEST_SINGLE
mykernel = libxsmm_dispatch_pgemm(desc4);
if ( mykernel == NULL ) printf("R4 kernel after the create call is null\n");
#endif
#endif
#if defined(USE_KERNEL_GENERATION_DIRECTLY) && defined(__linux__)
libxsmm_generator_pgemm_kernel( &io_generated_code, desc8, arch, iunroll, junroll, loopi, loopj );
#endif
#ifndef NO_ACCURACY_CHECK
printf("mallocing matrices\n");
#endif
sa = (float *) malloc ( lda*k*nmat*sizeof(float) );
da = (double *) malloc ( lda*k*nmat*sizeof(double) );
sb = (float *) malloc ( ldb*n*nmat*sizeof(float) );
db = (double *) malloc ( ldb*n*nmat*sizeof(double) );
sc1 = (float *) malloc ( ldc*n*nmat*sizeof(float) );
dc1 = (double *) malloc ( ldc*n*nmat*sizeof(double) );
sc = (float *) malloc ( ldc*n*nmat*sizeof(float) );
dc = (double *) malloc ( ldc*n*nmat*sizeof(double) );
sd = (float *) malloc ( ldc*n*nmat*sizeof(float) );
dd = (double *) malloc ( ldc*n*nmat*sizeof(double) );
#ifndef NO_ACCURACY_CHECK
printf("filling matrices\n");
#endif
sfill_matrix ( sa, lda, m, k*nmat );
sfill_matrix ( sb, ldb, k, n*nmat );
sfill_matrix ( sc, ldc, m, n*nmat );
dfill_matrix ( da, lda, m, k*nmat );
dfill_matrix ( db, ldb, k, n*nmat );
dfill_matrix ( dc, ldc, m, n*nmat );
#ifndef NO_ACCURACY_CHECK
for ( i = 0 ; i < ldc*n*nmat ; i++ ) sd[i]=sc[i];
for ( i = 0 ; i < ldc*n*nmat ; i++ ) dd[i]=dc[i];
for ( i = 0 ; i < ldc*n*nmat ; i++ ) sc1[i]=sc[i];
for ( i = 0 ; i < ldc*n*nmat ; i++ ) dc1[i]=dc[i];
printf("Pointing at the kernel now\n");
#endif
#ifdef USE_XSMM_GENERATED
cptr = (const unsigned char*) mykernel;
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
cptr = (const unsigned char*) gemm_;
#endif
#if defined(USE_KERNEL_GENERATION_DIRECTLY) && defined(__linux__)
cptr = (const unsigned char*) &routine_output[0];
opcode_routine = (void *) &cptr[0];
#endif
#ifndef TIME_MKL
# define DUMP_ASSEMBLY_FILE
#endif
#ifdef DUMP_ASSEMBLY_FILE
printf("Dumping assembly file\n");
FILE *fp = fopen("foo.s","w");
char buffer[80];
fputs("\t.text\n",fp);
fputs("\t.align 256\n",fp);
fputs("\t.globl gemm_\n",fp);
fputs("gemm_:\n",fp);
for (i = 0 ; i < 7000; i+=4 )
{
sprintf(buffer,".byte 0x%02x, 0x%02x, 0x%02x, 0x%02x\n",cptr[i],cptr[i+1],cptr[i+2],cptr[i+3]);
fputs(buffer,fp);
}
fputs("\tretq\n",fp);
fputs("\t.type gemm_,@function\n",fp);
fputs("\t.size gemm_,.-gemm_\n",fp);
fclose(fp);
#endif
#if defined(USE_MKL_FOR_REFERENCE) || defined(TIME_MKL)
# include <mkl.h>
MKL_LAYOUT CLAYOUT = (layout == 101) ? MKL_ROW_MAJOR : MKL_COL_MAJOR;
MKL_SIDE SIDE = (side == 'R' || side == 'r') ? MKL_RIGHT : MKL_LEFT;
MKL_UPLO UPLO = (uplo == 'U' || uplo == 'u') ? MKL_UPPER : MKL_LOWER;
MKL_TRANSPOSE TRANSA = (transa == 'N' || transa == 'n') ? MKL_NOTRANS : MKL_TRANS;
MKL_TRANSPOSE TRANSB = (transb == 'N' || transb == 'n') ? MKL_NOTRANS : MKL_TRANS;
MKL_DIAG DIAG = (diag == 'N' || diag == 'n') ? MKL_NONUNIT : MKL_UNIT;
MKL_COMPACT_PACK CMP_FORMAT = mkl_get_format_compact();
#if 0
MKL_COMPACT_PACK CMP_FORMAT = MKL_COMPACT_AVX;
#endif
#endif
#ifndef NO_ACCURACY_CHECK
printf("Before routine, initial A(1,1)=%g A[256]=%g\n",da[0],da[256]);
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
double one = 1.0;
#endif
double timer, firsttime = 0;
#ifdef MKL_TIMER
double tmptimer;
tmptimer = dsecnd_();
#else
unsigned long long l_start, l_end;
#endif
timer = 0.0;
for ( j = 0 ; j < (int)ntest ; j++ )
{
for ( i = 0 ; i < ldc*n*nmat ; i++ ) dc[i]=dc1[i];
for ( i = 0 , num = 0; i < (int)nmat ; i+= (int)VLEND, num++ )
{
double *Ap = &da[num*lda*k*VLEND];
double *Bp = &db[num*ldb*n*VLEND];
double *Cp = &dc[num*ldc*n*VLEND];
#ifdef MKL_TIMER
tmptimer = dsecnd_();
#else
l_start = libxsmm_timer_tick();
#endif
#if !defined(USE_XSMM_GENERATED) && !defined(USE_PREDEFINED_ASSEMBLY) && !defined(USE_KERNEL_GENERATION_DIRECTLY) && !defined(TIME_MKL) && !defined(USE_PREDEFINED_ASSEMBLY_XCT)
gen_compact_dgemm_ ( &layout, &m, &n, &k, &dalpha, Ap, &lda, Bp, &ldb, &dbeta, Cp, &ldc, &VLEND );
#endif
#ifdef USE_XSMM_GENERATED
mykernel ( Ap, Bp, Cp );
#endif
#ifdef USE_PREDEFINED_ASSEMBLY
gemm_ ( Ap, Bp, Cp );
#endif
#ifdef USE_KERNEL_GENERATION_DIRECTLY
(*opcode_routine)( Ap, Bp, Cp );
#endif
#ifdef TIME_MKL
mkl_dgemm_compact ( CLAYOUT, TRANSA, TRANSB, m, n, k, dalpha, da, lda, db, ldb, dbeta, dc, ldc, CMP_FORMAT, nmat );
i+=nmatd; /* Because MKL will do everything */
#endif
#ifdef MKL_TIMER
dtmp = dsecnd_() - tmptimer;
#else
l_end = libxsmm_timer_tick();
dtmp = libxsmm_timer_duration(l_start,l_end);
#endif
if ( j == 0 ) firsttime=dtmp;
timer += dtmp;
}
}
if ( ntest >= 100 ) {
/* Skip the first timing: super necessary if using MKL */
timer = (timer-firsttime)/((double)(ntest-1));
} else {
timer /= ((double)ntest);
}
#ifndef NO_ACCURACY_CHECK
printf("Average time to get through %u matrices: %g\n",nmat,timer);
printf("Gflops: %g\n",(double)op_count/(timer*1.0e9));
printf("after routine, new C(1,1)=%g C[256]=%g\n",dc[0],dc[256]);
#endif
#ifdef TEST_SINGLE
printf("Before r4 routine, initial C(1,1)=%g C[256]=%g\n",sc[0],sc[256]);
for ( i = 0 , num = 0; i < nmats ; i+= VLENS, num++ )
{
float *Ap = &sa[num*lda*k*VLENS];
float *Bp = &sb[num*ldb*n*VLENS];
float *Cp = &sc[num*ldc*n*VLENS];
#ifdef USE_XSMM_GENERATED
mykernel ( Ap, Bp, Cp );
#endif
}
printf("after r4 routine, new C(1,1)=%g C]256]=%g\n",dc[0],dc[256]);
#endif
#ifndef NO_ACCURACY_CHECK
/* Call some reference code now on a copy of the B matrix (C) */
double timer2 = 0.0;
for ( j = 0 ; j < (int)ntest ; j++ )
{
for ( i = 0 ; i < ldc*n*nmat ; i++ ) dd[i]=dc1[i];
#ifdef MKL_TIMER
tmptimer = dsecnd_();
#else
l_start = libxsmm_timer_tick();
#endif
#ifndef USE_MKL_FOR_REFERENCE
compact_dgemm_ ( &layout, &transa, &transb, &m, &n, &k, &dalpha, da, &lda, db, &ldb, &dbeta, dd, &ldc, &nmat, &VLEND );
#else
mkl_dgemm_compact ( CLAYOUT, TRANSA, TRANSB, m, n, k, dalpha, da, lda, db, ldb, dbeta, dd, ldc, CMP_FORMAT, nmat );
#endif
#ifdef MKL_TIMER
timer2 += dsecnd_() - tmptimer;
#else
l_end = libxsmm_timer_tick();
timer2 += libxsmm_timer_duration(l_start,l_end);
#endif
}
timer2 /= ((double)ntest);
printf("Reference time=%g Reference Gflops=%g\n",timer2,op_count/(timer2*1.0e9));
/* Compute the residual between B and C */
dtmp = residual_d ( dc, ldc, m, n*nmat, dd, ldc, &nerrs, &ncorr );
printf("R8 mnk=%u %u %u ldabc=%u %u %u error: %g number of errors: %u corrects: %u",m,n,k,lda,ldb,ldc,dtmp,nerrs,ncorr);
if ( nerrs > 0 ) printf(" ->FAILED at %ux%u real*8 %u case",m,n,layout);
printf("\n");
#ifdef TEST_SINGLE
/* Call some reference code now on a copy of the B matrix (C) */
compact_dgemm_ ( &layout, &transa, &transb, &m, &n, &k, &salpha, sa, &lda, sb, &ldb, &sbeta, sd, &ldc, &nmat, &VLENS );
/* Compute the residual between C and D */
dtmp = residual_s ( sc, ldc, m, n*nmat, sd, ldc, &nerrs, &ncorr );
printf("R4 mnk=%u %u %u ldabc=%u %u %u error: %g number of errors: %u corrects: %u",m,n,k,lda,ldb,ldc,dtmp,nerrs,ncorr);
if ( nerrs > 0 ) printf(" ->FAILED at %ux%u real*4 case",m,n);
printf("\n");
#endif
#else
for ( j = 0, nerrs = 0 ; j < lda*n*nmat; j++ )
{
if ( isnan(dc[j]) || isinf(dc[j]) )
{
if ( ++nerrs < 10 )
{
printf("WARNING: dc[%d]=%g\n",j,dc[j]);
}
}
}
printf("%g,real*8 m/n/k=%u %u %u lda-c=%u %u %u Denormals=%u Time=%g Gflops=%g",op_count/(timer*1.0e9),m,n,k,lda,ldb,ldc,nerrs,timer,op_count/(timer*1.0e9));
if ( nerrs > 0 ) printf(" -> FAILED at %ux%u real*8 case",m,n);
printf("\n");
#endif
free(dd);
free(sd);
free(dc);
free(sc);
free(dc1);
free(sc1);
free(db);
free(sb);
free(da);
free(sa);
return 0;
}