int main(int argc, char **argv)
{
uint32_t rows, cols;
uint32_t *a1, *a3, *e3;
int const radius = STENCIL_RADIUS;
// check command line args
if (argc == 1) {
rows = 100; // default rows
cols = 100; // default cols
} else if (argc == 3) {
rows = atoi(argv[1]);
if (rows <= 0 || rows > 1023) {
usage(argv[1]);
return 0;
}
cols = atoi(argv[2]);
if (cols <= 0 || cols > 1023) {
usage(argv[2]);
return 0;
}
} else {
usage(argv[0]);
return 0;
}
printf("Running with rows = %u, cols = %u\n", rows, cols);
printf("Initializing arrays\n");
fflush(stdout);
a1 = (uint32_t *)malloc((rows + radius * 2) * (cols + radius * 2) * sizeof(uint32_t));
a3 = (uint32_t *)malloc((rows + radius * 2) * (cols + radius * 2) * sizeof(uint32_t));
memset(a1, 0, (rows + radius * 2) * (cols + radius * 2) * sizeof(uint32_t));
memset(a3, 0, (rows + radius * 2) * (cols + radius * 2) * sizeof(uint32_t));
for (uint32_t r = 0; r < rows; r++)
for (uint32_t c = 0; c < cols; c++) {
a1[(r+radius)*(cols + radius * 2)+(c+radius)] = ((r+radius) << 8) | (c+radius);
}
CHtHif *pHtHif = new CHtHif();
int unitCnt = pHtHif->GetUnitCnt();
printf("#AUs = %d\n", unitCnt);
CHtAuUnit ** pAuUnits = new CHtAuUnit * [unitCnt];
for (int unit = 0; unit < unitCnt; unit++)
pAuUnits[unit] = new CHtAuUnit(pHtHif);
// Coprocessor memory arrays
uint32_t *cp_a1 = (uint32_t*)pHtHif->MemAlloc((rows+radius*2) * (cols+radius*2) * sizeof(uint32_t));
uint32_t *cp_a3 = (uint32_t*)pHtHif->MemAlloc((rows+radius*2) * (cols+radius*2) * sizeof(uint32_t));
if (!cp_a1 || !cp_a3) {
fprintf(stderr, "ht_cp_malloc() failed.\n");
exit(-1);
}
pHtHif->MemCpy(cp_a1, a1, (rows+radius*2) * (cols+radius*2) * sizeof(uint32_t));
pHtHif->MemSet(cp_a3, 0, (rows+radius*2) * (cols+radius*2) * sizeof(uint32_t));
// avoid bank aliasing for performance
if (unitCnt > 16 && !(unitCnt & 1)) unitCnt -= 1;
printf("stride = %d\n", unitCnt);
fflush(stdout);
// Send calls to units
uint32_t rowsPerUnit = (cols + unitCnt - 1) / unitCnt;
for (int unit = 0; unit < unitCnt; unit++) {
uint32_t * pSrcAddr = cp_a1 + unit * (cols+radius*2) * rowsPerUnit;
uint32_t * pDstAddr = cp_a3 + unit * (cols+radius*2) * rowsPerUnit;
uint32_t unitRows = (unit+1)*rowsPerUnit > rows ? (rows - unit*rowsPerUnit) : rowsPerUnit;
pAuUnits[unit]->SendCall_htmain((uint64_t)pSrcAddr, (uint64_t)pDstAddr, unitRows, cols);
}
// generate expected results while waiting for returns
e3 = (uint32_t *)malloc((rows+radius*2) * (cols+radius*2) * sizeof(uint32_t));
memset(e3, 0, (rows + radius * 2) * (cols + radius * 2) * sizeof(uint32_t));
uint32_t coef[5] = { STENCIL_COEF2, STENCIL_COEF1, STENCIL_COEF0/2, STENCIL_COEF1, STENCIL_COEF2 };
for (uint32_t row = radius; row < rows+radius; row++) {
for (uint32_t col = radius; col < cols+radius; col++) {
uint32_t rslt = 0;
for (uint32_t c = col - STENCIL_RADIUS; c <= col + STENCIL_RADIUS; c++)
rslt += a1[row*(cols+radius*2) + c] * coef[c - col + STENCIL_RADIUS];
for (uint32_t r = row - STENCIL_RADIUS; r <= row + STENCIL_RADIUS; r++)
rslt += a1[r*(cols+radius*2) + col] * coef[r - row + STENCIL_RADIUS];
e3[row*(cols+radius*2) + col] = rslt >> 8;
}
}
// Wait for returns
for (int unit = 0; unit < unitCnt; unit++) {
while (!pAuUnits[unit]->RecvReturn_htmain())
usleep(1000);
}
pHtHif->MemCpy(a3, cp_a3, (rows+radius*2) * (cols+radius*2) * sizeof(uint32_t));
// check results
int err_cnt = 0;
for (uint32_t col = 0; col < (cols+radius*2); col++) {
for (uint32_t row = 0; row < (rows+radius*2); row++) {
if (a3[row*(cols+radius*2) + col] != e3[row*(cols+radius*2) + col]) {
printf("a3[row=%u, col=%u] is %u, should be %u\n",
row, col, a3[row*(cols+radius*2) + col], e3[row*(cols+radius*2) + col]);
err_cnt++;
}
}
}
if (err_cnt)
printf("FAILED: detected %d issues!\n", err_cnt);
else
printf("PASSED\n");
// free memory
free(a1);
free(a3);
pHtHif->MemFree(cp_a1);
pHtHif->MemFree(cp_a3);
delete pHtHif;
return err_cnt;
}
int main(int argc, char **argv) {
int debug = 0;
int mat_int_len, temp;
#ifdef HT_MODEL
default_size = 15;
#endif
#ifdef HT_SYSC
default_size = 15;
#endif
#ifdef HT_VSIM
default_size = 3;
#endif
uint64_t i, j, k;
uint64_t aRow, aCol, bRow, bCol;
uint64_t *a1, *a2, *a3;
// check command line args
if (argc <= 2) {
// Defaults
aRow = default_size;
aCol = default_size;
bRow = default_size;
bCol = default_size;
if (argc==2) {
if (atoi(argv[1]) == 1 || atoi(argv[1]) == 0) {
debug = atoi(argv[1]);
} else {
usage(argv[0]);
}
debug = atoi(argv[1])==1 ? 1 : 0;
}
} else if (argc == 5 || argc == 6) {
// Grab Command Line Values
aRow = atoi(argv[1]);
aCol = atoi(argv[2]);
bRow = atoi(argv[3]);
bCol = atoi(argv[4]);
if (argc==6)
debug = atoi(argv[5])==1 ? 1 : 0;
if (aRow <= 0 || aCol <= 0 || bRow <= 0 || bCol <= 0) {
usage(argv[0]);
return 0;
}
} else {
usage(argv[0]);
return 0;
}
if (aCol != bRow) {
printf("ERROR: Number of columns in Matrix A (%lld) does not equal the number of rows in Matrix B (%lld)\n\n",
(long long)aCol, (long long)bRow);
exit(1);
}
printf("Running with: Matrix A: (%lldx%lld) and Matrix B (%lldx%lld)\n",
(long long)aRow, (long long)aCol, (long long)bRow, (long long)bCol);
printf("Initializing arrays\n");
fflush(stdout);
a1 = (uint64_t *)malloc(aRow * aCol * 8);
a2 = (uint64_t *)malloc(bRow * bCol * 8);
a3 = (uint64_t *)malloc(aRow * bCol * 8);
memset(a3, 0, aRow * bCol * 8);
/*
* Matrix A and B are stored differently in memory (to make reads more efficient?)
* The following are 3x3 examples: (A counts from 8 -> 0, B counts from 0 -> 8)
*
* Actual Numbers: Memory Locations (by Index)
*
* A: 8 7 6 A: 0 3 6
* 5 4 5 1 4 7
* 2 1 0 2 5 8
*
* B: 0 1 2 B: 0 1 2
* 3 4 5 3 4 5
* 6 7 8 6 7 8
*
* I did this to try to make A Rows available on strides, as well as B Columns on strides...
*
*/
// Fill Matrix A
k = 0;
for (i = 0; i < aCol; i++) {
for (j = aRow; j > 0; j--) {
a1[k] = (aRow*aCol)-aCol*(aRow-j)-i-1;
k++;
}
}
// Fill Matrix B
for (i = 0; i < bRow*bCol; i++) {
a2[i] = i;
}
//Print Matrices
mat_int_len = 1;
temp = 0;
for(i = 0; i < aRow*bCol; i++) {
temp = num_length(a1[i]);
mat_int_len = (temp > mat_int_len) ? temp : mat_int_len;
}
printf("Matrix A:\n");
for (i = 0; i < aRow; i++) {
for (j = 0; j < aCol; j++) {
printf("%*lld ", mat_int_len, (long long)a1[i+aRow*j]);
}
printf("\n");
}
printf("\n\n");
mat_int_len = 1;
temp = 0;
for(i = 0; i < aRow*bCol; i++) {
temp = num_length(a1[i]);
mat_int_len = (temp > mat_int_len) ? temp : mat_int_len;
}
printf("Matrix B:\n");
for(i = 0; i < bRow*bCol; i++) {
if (i > 0) {
if (i%bCol == 0) {
printf("\n");
}
}
printf("%*lld ", mat_int_len, (long long)a2[i]);
}
printf("\n\n");
// Debug - Print Matrix Values at memory locations
if (debug) {
printf("A - MEM\n");
for (i = 0; i < aRow*aCol; i++) {
printf("%lld - %lld\n", (long long)i, (long long)a1[i]);
}
printf("B - MEM\n");
for (i = 0; i < bRow*bCol; i++) {
printf("%lld - %lld\n", (long long)i, (long long)a2[i]);
}
}
CHtHif *pHtHif = new CHtHif();
// Coprocessor memory arrays
uint64_t *cp_a1 = (uint64_t *)pHtHif->MemAllocAlign(4 * 1024, aRow * aCol * sizeof(uint64_t));
uint64_t *cp_a2 = (uint64_t *)pHtHif->MemAllocAlign(4 * 1024, bRow * bCol * sizeof(uint64_t));
uint64_t *cp_a3 = (uint64_t *)pHtHif->MemAllocAlign(4 * 1024, aRow * bCol * sizeof(uint64_t));
pHtHif->MemCpy(cp_a1, a1, aRow * aCol * sizeof(uint64_t));
pHtHif->MemCpy(cp_a2, a2, bRow * bCol * sizeof(uint64_t));
pHtHif->MemSet(cp_a3, 0, aRow * bCol * sizeof(uint64_t));
int unitCnt = pHtHif->GetUnitCnt();
CHtAuUnit ** pAuUnits = new CHtAuUnit * [unitCnt];
for (int unitId = 0; unitId < unitCnt; unitId += 1)
pAuUnits[unitId] = new CHtAuUnit(pHtHif);
printf("#AUs = %d\n", unitCnt);
// avoid bank aliasing for performance
if (unitCnt > 16 && !(unitCnt & 1)) unitCnt -= 1;
printf("stride = %d\n", unitCnt);
fflush(stdout);
pHtHif->SendAllHostMsg(MA_BASE, (uint64_t)cp_a1);
pHtHif->SendAllHostMsg(MB_BASE, (uint64_t)cp_a2);
pHtHif->SendAllHostMsg(MC_BASE, (uint64_t)cp_a3);
pHtHif->SendAllHostMsg(MC_ROW, (uint32_t)aRow);
pHtHif->SendAllHostMsg(MC_COL, (uint32_t)bCol);
pHtHif->SendAllHostMsg(COMMON, (uint32_t)aCol);
for (int unit = 0; unit < unitCnt; unit++)
pAuUnits[unit]->SendCall_htmain(unit /*rowOffset*/, unitCnt /*stride*/);
for (int unit = 0; unit < unitCnt; unit++) {
while (!pAuUnits[unit]->RecvReturn_htmain())
usleep(1000);
fflush(stdout);
}
pHtHif->MemCpy(a3, cp_a3, aRow * bCol * sizeof(uint64_t));
// Print Resulting Matrix
mat_int_len = 1;
temp = 0;
for(i = 0; i < aRow*bCol; i++) {
temp = num_length(a3[i]);
mat_int_len = (temp > mat_int_len) ? temp : mat_int_len;
}
printf("\nMatrix C:\n");
for(i = 0; i < aRow*bCol; i++) {
if (i > 0) {
if (i%bCol == 0) {
printf("\n");
}
}
printf("%*lld ", mat_int_len, (long long)a3[i]);
}
printf("\n\n");
if (debug) {
printf("C - MEM\n");
for (i = 0; i < aRow*bCol; i++) {
printf("%lld - %lld\n", (long long)i, (long long)a3[i]);
}
}
// Do error checking
int err_cnt = 0;
uint64_t rowNum = 0, colNum = 0, calcNum = 0, eleNum = 0;
uint64_t *val;
val = (uint64_t *)malloc(aRow * bCol * 8);
memset(val, 0, aRow * bCol * 8);
// Calculate the resulting matrix to check against coprocessor results
for (rowNum = 0; rowNum < aRow; rowNum++) {
for (colNum = 0; colNum < bCol; colNum++) {
for (calcNum = 0; calcNum < aCol; calcNum++) {
val[eleNum] += a1[rowNum+(calcNum*aRow)] * a2[colNum+(calcNum*bCol)];
}
eleNum++;
}
}
// Check results
for (eleNum = 0; eleNum < aRow*bCol; eleNum++) {
if (val[eleNum] != a3[eleNum]) {
err_cnt++;
printf("Found element mismatch at matrix position %lld - found value %lld, expected value %lld.\n",
(unsigned long long)eleNum, (unsigned long long)a3[eleNum], (unsigned long long)val[eleNum]);
}
}
if (err_cnt == 0) {
// Test Passed
printf("PASSED\n\n");
} else {
// Test Failed
printf("FAILED - error count %d\n\n", err_cnt);
}
// free memory
free(a1);
free(a2);
free(a3);
pHtHif->MemFreeAlign(cp_a1);
pHtHif->MemFreeAlign(cp_a2);
pHtHif->MemFreeAlign(cp_a3);
free(val);
delete pHtHif;
return err_cnt;
}
int main(int argc, char **argv)
{
uint64_t i;
uint64_t vecLen;
uint64_t *a1, *a2, *a3;
// check command line args
if (argc == 1) {
vecLen = 100; // default vecLen
} else if (argc == 2) {
vecLen = atoi(argv[1]);
if (vecLen <= 0) {
usage(argv[0]);
return 0;
}
} else {
usage(argv[0]);
return 0;
}
printf("Running with vecLen = %llu\n", (long long)vecLen);
printf("Initializing arrays\n");
fflush(stdout);
a1 = (uint64_t *)malloc(vecLen * sizeof(uint64_t));
a2 = (uint64_t *)malloc(vecLen * sizeof(uint64_t));
a3 = (uint64_t *)malloc(vecLen * sizeof(uint64_t));
memset(a3, 0, vecLen * 8);
for (i = 0; i < vecLen; i++) {
a1[i] = i;
a2[i] = 2 * i;
}
CHtHif *pHtHif = new CHtHif();
int unitCnt = pHtHif->GetUnitCnt();
printf("#AUs = %d\n", unitCnt);
// Coprocessor memory arrays
uint64_t * cp_a1 = (uint64_t*)pHtHif->MemAllocAlign(4 * 1024, vecLen * sizeof(uint64_t));
uint64_t * cp_a2 = (uint64_t*)pHtHif->MemAllocAlign(4 * 1024, vecLen * sizeof(uint64_t));
uint64_t * cp_a3 = (uint64_t*)pHtHif->MemAllocAlign(4 * 1024, vecLen * sizeof(uint64_t));
pHtHif->MemCpy(cp_a1, a1, vecLen * sizeof(uint64_t));
pHtHif->MemCpy(cp_a2, a2, vecLen * sizeof(uint64_t));
pHtHif->MemSet(cp_a3, 0, vecLen * sizeof(uint64_t));
CHtAuUnit ** pAuUnits = new CHtAuUnit * [unitCnt];
for (int unit = 0; unit < unitCnt; unit++)
pAuUnits[unit] = new CHtAuUnit(pHtHif);
// avoid bank aliasing for performance
if (unitCnt > 16 && !(unitCnt & 1)) unitCnt -= 1;
printf("stride = %d\n", unitCnt);
fflush(stdout);
pHtHif->SendAllHostMsg(OP1_ADDR, (uint64_t)cp_a1);
pHtHif->SendAllHostMsg(OP2_ADDR, (uint64_t)cp_a2);
pHtHif->SendAllHostMsg(RES_ADDR, (uint64_t)cp_a3);
pHtHif->SendAllHostMsg(VEC_LEN, (uint64_t)vecLen);
for (int unit = 0; unit < unitCnt; unit++)
pAuUnits[unit]->SendCall_htmain(unit /*offset*/, unitCnt /*stride*/);
uint64_t act_sum = 0;
uint64_t au_sum;
for (int unit = 0; unit < unitCnt; unit++) {
while (!pAuUnits[unit]->RecvReturn_htmain(au_sum))
usleep(1000);
printf("unit=%-2d: au_sum %llu \n", unit, (long long)au_sum);
fflush(stdout);
act_sum += au_sum;
}
printf("RTN: act_sum = %llu\n", (long long)act_sum);
pHtHif->MemCpy(a3, cp_a3, vecLen * sizeof(uint64_t));
// check results
int err_cnt = 0;
uint64_t exp_sum = 0;
for (i = 0; i < vecLen; i++) {
if (a3[i] != a1[i] + a2[i]) {
printf("a3[%llu] is %llu, should be %llu\n",
(long long)i, (long long)a3[i], (long long)(a1[i] + a2[i]));
err_cnt++;
}
exp_sum += a1[i] + a2[i];
//printf("i=%llu: a1=%llu + a2=%llu => a3=%llu\n",
// (long long)i, (long long)a1[i], (long long)a2[i], (long long)a3[i]);
}
if (act_sum != exp_sum) {
printf("act_sum %llu != exp_sum %llu\n", (long long)act_sum, (long long)exp_sum);
err_cnt++;
}
if (err_cnt)
printf("FAILED: detected %d issues!\n", err_cnt);
else
printf("PASSED\n");
// free memory
free(a1);
free(a2);
free(a3);
pHtHif->MemFreeAlign(cp_a1);
pHtHif->MemFreeAlign(cp_a2);
pHtHif->MemFreeAlign(cp_a3);
delete pHtHif;
return err_cnt;
}