int main(void)
{
vbx_test_init();
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
const int VBX_SCRATCHPAD_SIZE = this_mxp->scratchpad_size;
const int required_vectors = 4;
int N = VBX_SCRATCHPAD_SIZE / sizeof(vbx_mm_t) / required_vectors;
int PRINT_LENGTH = min( N, MAX_PRINT_LENGTH );
double scalar_time, vector_time;
int errors=0;
vbx_mxp_print_params();
printf( "\nAdd test...\n" );
printf( "Vector length: %d\n", N );
vbx_mm_t *scalar_in1 = malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *scalar_in2 = malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *scalar_out = malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *vector_in1 = vbx_shared_malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *vector_in2 = vbx_shared_malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *vector_out = vbx_shared_malloc( N*sizeof(vbx_mm_t) );
// vbx_mm_t *vector_out = vector_in2 - 5;
vbx_sp_t *v_in1 = vbx_sp_malloc( N*sizeof(vbx_sp_t) );
vbx_sp_t *v_in2 = vbx_sp_malloc( N*sizeof(vbx_sp_t) );
vbx_sp_t *v_out = vbx_sp_malloc( N*sizeof(vbx_sp_t) );
// vbx_sp_t *v_out = v_in2-5;
VBX_T(test_zero_array)( scalar_out, N );
VBX_T(test_zero_array)( vector_out, N );
VBX_T(test_init_array)( scalar_in1, N, 1 );
VBX_T(test_copy_array)( vector_in1, scalar_in1, N );
VBX_T(test_init_array)( scalar_in2, N, 1 );
VBX_T(test_copy_array)( vector_in2, scalar_in2, N );
VBX_T(test_print_array)( scalar_in1, PRINT_LENGTH );
VBX_T(test_print_array)( scalar_in2, PRINT_LENGTH );
scalar_time = test_scalar( scalar_out, scalar_in1, scalar_in2, N );
VBX_T(test_print_array)( scalar_out, PRINT_LENGTH);
vbx_dma_to_vector( v_in1, (void *)vector_in1, N*sizeof(vbx_sp_t) );
vbx_dma_to_vector( v_in2, (void *)vector_in1, N*sizeof(vbx_sp_t) );
vector_time = test_vector( v_out, v_in1, v_in2, N, scalar_time );
vbx_dma_to_host( (void *)vector_out, v_out, N*sizeof(vbx_sp_t) );
vbx_sync();
VBX_T(test_print_array)( vector_out, PRINT_LENGTH );
errors += VBX_T(test_verify_array)( scalar_out, vector_out, N );
VBX_TEST_END(errors);
return 0;
}
int main(void)
{
double scalar_time, vector_time;
int errors=0;
vbx_test_init();
vbx_mxp_print_params();
printf("\nVector FIR test...\n");
vbx_mm_t *scalar_sample = malloc( (SAMP_SIZE+NTAPS)*sizeof(vbx_mm_t) );
vbx_mm_t *scalar_coeffs = malloc( NTAPS*sizeof(vbx_mm_t) );
vbx_mm_t *scalar_out = malloc( SAMP_SIZE*sizeof(vbx_mm_t) );
vbx_mm_t *sample = vbx_shared_malloc( (SAMP_SIZE+NTAPS)*sizeof(vbx_mm_t) );
vbx_mm_t *coeffs = vbx_shared_malloc( NTAPS*sizeof(vbx_mm_t) );
vbx_mm_t *vector_out = vbx_shared_malloc( SAMP_SIZE*sizeof(vbx_mm_t) );
VBX_T(test_zero_array)( scalar_out, SAMP_SIZE );
VBX_T(test_zero_array)( vector_out, SAMP_SIZE );
VBX_T(test_init_array)( scalar_sample, SAMP_SIZE, 0xff );
VBX_T(test_copy_array)( sample, scalar_sample, SAMP_SIZE );
VBX_T(test_init_array)( scalar_coeffs, NTAPS, 1 );
VBX_T(test_copy_array)( coeffs, scalar_coeffs, NTAPS );
VBX_T(test_zero_array)( scalar_sample+SAMP_SIZE, NTAPS );
VBX_T(test_zero_array)( sample+SAMP_SIZE, NTAPS );
printf("\nSamples:\n");
VBX_T(test_print_array)( scalar_sample, min(SAMP_SIZE,MAX_PRINT_LENGTH) );
printf("\nCoefficients:\n");
VBX_T(test_print_array)( scalar_coeffs, min(NTAPS,MAX_PRINT_LENGTH) );
scalar_time = test_scalar( scalar_out, scalar_sample, scalar_coeffs);
VBX_T(test_print_array)( scalar_out, min(SAMP_SIZE,MAX_PRINT_LENGTH) );
#ifdef USE_TRANSPOSE
vector_time = test_vector_transpose( vector_out, sample, coeffs, scalar_time );
VBX_T(test_print_array)( vector_out, min(SAMP_SIZE,MAX_PRINT_LENGTH) );
errors += VBX_T(test_verify_array)( scalar_out, vector_out, SAMP_SIZE-NTAPS );
#endif //USE_TRANSPOSE
#ifdef USE_1D
vector_time = test_vector_1d( vector_out, sample, coeffs, scalar_time );
VBX_T(test_print_array)( vector_out, min(SAMP_SIZE,MAX_PRINT_LENGTH) );
errors += VBX_T(test_verify_array)( scalar_out, vector_out, SAMP_SIZE-NTAPS );
#endif //USE_1D
#ifdef USE_2D
vector_time = test_vector_2d( vector_out, sample, coeffs, scalar_time );
VBX_T(test_print_array)( vector_out, min(SAMP_SIZE,MAX_PRINT_LENGTH) );
errors += VBX_T(test_verify_array)( scalar_out, vector_out, SAMP_SIZE-NTAPS );
#endif //USE_2D
VBX_TEST_END(errors);
return 0;
}
int compare_vbx_lut_to_vbx_lut_ci(int sz, int max_print_errors)
{
int f, n, errors;
vbx_byte_t* v_pass = (vbx_byte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_pattern = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_lutc = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_group = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_sel = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_lut = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_word_t));
vbx_ubyte_t* v_idx = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_word_t));
unsigned char* lut = (unsigned char*)vbx_shared_malloc(sz*sizeof(unsigned char));
unsigned char* lut_c = (unsigned char*)vbx_shared_malloc(sz*sizeof(unsigned char));
for (n = 0; n < sz; n++) {
v_pattern[n] = n & 0xff;
}
int s, stage = 11;
for (f = 0; f < face_lbp[stage].count; f++) {
lbp_feat_t feat = face_lbp[stage].feats[f];
vbx_set_vl(sz);
int total = f;
s = 0;
while(s < stage){
total += face_lbp[s].count;
s++;
}
vbx(SVBU, VCUSTOM0, v_lutc, total, v_pattern);
vbx(SVB, VMOV, v_pass, feat.fail, 0);
/* check if pattern is in lut */
vbx(SVBU, VSHR, v_group, 5, v_pattern);
for (n = 0; n < 8; n++) {
vbx(SVB, VADD, v_sel, -n, v_group);
vbx(SVBW, VCMV_Z, v_lut, feat.lut[n], v_sel);
}
vbx(SVBWU, VAND, v_idx, 0x1f, v_pattern);
vbx(VVWB, VSHR, v_lut, v_idx, v_lut);
vbx(SVB, VAND, v_lut, 1, v_lut);
vbx(SVB, VCMV_LEZ, v_pass, feat.pass, v_lut);
vbx_dma_to_host(lut_c, v_lutc, sz*sizeof(unsigned char));
vbx_dma_to_host(lut, v_pass, sz*sizeof(unsigned char));
vbx_sync();
errors = match_array_byte(lut_c, lut, "custom_lut", sz, 1, max_print_errors, 0, 0);
}
vbx_sp_free();
vbx_shared_free(lut);
vbx_shared_free(lut_c);
return errors;
}
int main(void)
{
vbx_test_init();
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
const int VBX_SCRATCHPAD_SIZE = this_mxp->scratchpad_size;
int N = VBX_SCRATCHPAD_SIZE/sizeof(vbx_word_t)/12;
N=1024;
int PRINT_LENGTH = min(N, MAX_PRINT_LENGTH);
double scalar_time, vector_time;
int errors=0;
vbx_mxp_print_params();
printf("\nVector power test...\n");
printf("Vector length: %d\n", N);
vbx_word_t *scalar_in1 = malloc( N*sizeof(vbx_word_t) );
vbx_word_t *scalar_in2 = malloc( N*sizeof(vbx_word_t) );
vbx_word_t *scalar_out = malloc( N*sizeof(vbx_word_t) );
vbx_word_t *vector_in1 = vbx_shared_malloc( N*sizeof(vbx_word_t) );
vbx_word_t *vector_in2 = vbx_shared_malloc( N*sizeof(vbx_word_t) );
vbx_word_t *vector_out = vbx_shared_malloc( N*sizeof(vbx_word_t) );
if(vector_out==NULL){
printf("malloc_failed\n");
return 1;
}
test_zero_array_word( scalar_out, N );
test_zero_array_word( vector_out, N );
test_init_array_word( scalar_in1, N, 5 );
test_copy_array_word( vector_in1, scalar_in1, N );
test_init_array_word( scalar_in2, N, 112 );
test_copy_array_word( vector_in2, scalar_in2, N );
test_print_array_word( scalar_in1, PRINT_LENGTH );
test_print_array_word( scalar_in2, PRINT_LENGTH );
scalar_time = test_scalar_power( scalar_out, scalar_in1, scalar_in2, N);
test_print_array_word( scalar_out, PRINT_LENGTH );
vector_time = test_vector_power( vector_out, vector_in1, vector_in2, N, scalar_time );
test_print_array_word( vector_out, PRINT_LENGTH );
errors += test_verify_array_word( scalar_out, vector_out, N );
VBX_TEST_END(errors);
return 0;
}
int compare_vbx_lbp_ci_to_scalar_patterns(unsigned short* img, int width, int height, int max_print_errors)
{
int j, errors = 0;
unsigned char** scalar_patterns = test_scalar_patterns(img, 0, width, height);
vbx_ubyte_t* v_in = (vbx_ubyte_t*)vbx_sp_malloc(3*width*sizeof(vbx_word_t));
vbx_ubyte_t* v_top = (vbx_byte_t*)vbx_sp_malloc(width*sizeof(vbx_byte_t));
vbx_ubyte_t* v_bot = (vbx_byte_t*)vbx_sp_malloc(width*sizeof(vbx_byte_t));
vbx_ubyte_t* v_lbp = v_bot;
unsigned char* lbp = (unsigned char*)vbx_shared_malloc(width*sizeof(unsigned char));
vbx_set_vl(width);
for(j=0; j < height - 2; j++){
vbx_dma_to_vector(v_in, img+j*width, 3*width*sizeof(unsigned char));
vbx(VVHU, VCUSTOM1, v_top, v_in, v_in+width);
vbx(VVHU, VCUSTOM1, v_bot, v_in+width, v_in+2*width);
vbx(SVHBU, VAND, v_top, 0xf0, v_top);
vbx(SVHBU, VAND, v_bot, 0x0f, v_bot);
vbx(VVBU, VADD, v_lbp, v_bot, v_top);
vbx_dma_to_host(lbp, v_lbp, width*sizeof(unsigned char));
vbx_sync();
errors = match_array_byte(lbp, scalar_patterns[0]+j*width, "custom_lbp", width-2, 1, max_print_errors, 1, j);
}
vbx_sp_free();
vbx_shared_free(lbp);
return errors;
}
int dma_bandwidth_test()
{
const int num_iter = 64;
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
int scratchpad_size = this_mxp->scratchpad_size;
uint8_t *buf = vbx_shared_malloc(scratchpad_size);
vbx_ubyte_t *v_buf = vbx_sp_malloc(scratchpad_size);
vbx_timestamp_t time_start, time_stop;
int i;
int len;
int to_host;
int errors = 0;
vbx_mxp_print_params();
// dma_alignment_bytes gives DMA master data bus width in bytes.
double bytes_per_sec = \
(((double) this_mxp->core_freq) * this_mxp->dma_alignment_bytes);
double max_megabytes_per_sec = bytes_per_sec/(1024*1024);
printf("\nMax available bandwidth = %s Megabytes/s\n",
vbx_eng(max_megabytes_per_sec, 4));
printf("\n");
for (to_host = 0; to_host < 2; to_host++) {
for (len = 32; len <= scratchpad_size ; len *= 2) {
printf("DMA %s, %d bytes\n", to_host ? "write" : "read", len);
vbx_timestamp_start();
if (to_host) {
time_start = vbx_timestamp();
for (i = 0; i < num_iter; i++) {
vbx_dma_to_host(buf, v_buf, len);
}
vbx_sync();
time_stop = vbx_timestamp();
} else {
time_start = vbx_timestamp();
for (i = 0; i < num_iter; i++) {
vbx_dma_to_vector(v_buf, buf, len);
}
vbx_sync();
time_stop = vbx_timestamp();
}
print_dma_bandwidth(time_start, time_stop, len, num_iter,
max_megabytes_per_sec);
printf("\n");
}
printf("\n");
}
vbx_shared_free(buf);
vbx_sp_free();
return errors;
}
int deep_vector_copy_ext_test()
{
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
int retval;
int num_test;
int total_errors = 0;
const int NUM_TESTS = TEST_DEEP_MM_NUM_TESTS;
int NB = this_mxp->scratchpad_size * 10;
int NT = NB / sizeof(vbx_mm_t);
vbx_mm_t *v = vbx_shared_malloc( NB );
srand( 0x1a84c92a );
int i;
for( num_test=0; num_test < NUM_TESTS ; num_test++ ) {
// initialize the whole working space
for( i=0; i<NT; i++ ) {
v[i] = i & MSK;
}
// choose random src/dest/length:
// -- randomly pick the dest
// -- set a window size of 2*K around the dest
// -- randomly pick the src within the window
// -- randomly pick the length, subject to end-of-scratchpad
// -- this 'window' rule increases probability of overlaps
// -- rough distribution: 30% short (pipeline) overlaps, 20% long overlaps, 50% no overlap
int K, N1, N2, NN;
N1 = rand() % NT;
K = 1 + rand() % ((N1 > 0)? min(min(N1, NT-N1), 1024): min(NT, 1024));
N2 = N1 - K + rand() % (2*K);
NN = rand() % (NT - max(N1,N2));
vbx_mm_t *dst = v + N1;
vbx_mm_t *src = v + N2;
printf("test:%d src:0x%08x dst:0x%08x len:%08d", num_test, N1, N2, NN );
// do the copy
retval = VBX_T(vbw_vec_copy_ext)( dst, src, NN );
vbx_sync();
printf(" retval:0x%04x\n",retval);
// ensure the copy was done properly
int errors = verify_copy(v, 0, N1, 0, "head")
+ verify_copy(v, N1, NN+N1, (N2-N1), "copy")
+ verify_copy(v, NN+N1, NT, 0, "tail");
total_errors += errors;
if( errors ) {
//break;
}
}
return total_errors;
}
int compare_ScalarLBPRestrictedPatterns_to_SATBinaryPattern(unsigned short *vbx_img, int log, int width, int height, int max_print_errors)
{
int l, i, j, cell, errors = 0;
/* generate patterns */
unsigned short **sums = ScalarLBPRestrictedSums(vbx_img, width, height, log);
unsigned char **patterns = ScalarLBPRestrictedPatterns(sums, width, height, log);
unsigned char **sat_patterns = (unsigned char**)vbx_shared_malloc((log+1)*sizeof(unsigned char*));
for (l=0; l<log+1; l++) {
sat_patterns[l] = (unsigned char*)vbx_shared_malloc(width*height*sizeof(unsigned char));
}
unsigned int *iImg, *iiImg;
iImg = (unsigned int *)vbx_shared_malloc(width*height*sizeof(unsigned int));
iiImg = (unsigned int *)vbx_shared_malloc(width*height*sizeof(unsigned int));
gen_integrals(vbx_img, iImg, iiImg, width, height);
image_t lbp_img = {iImg, {width, height}};
for (l=0; l<log+1; l++) {
cell = 1 << l;
lbp_feat_t lbp_feat = {{{0, 0}, {cell, cell}}, 0, 0, {0, 0, 0, 0, 0, 0, 0, 0}};
for (j = 0; j < height - (3*cell-1); j++) {
for (i = 0; i < width - (3*cell-1); i++) {
pair_t lbp_p = {i, j};
sat_patterns[l][j*width+i] = SATBinaryPattern(lbp_img, &lbp_feat, lbp_p);
}
}
}
/* test patterns vs sat binary patterns */
for (l=0; l<log+1; l++) {
cell = 1 << l;
for (j = 0; j < height - (3*cell-1); j++) {
errors += match_array_byte(patterns[l] + j*width, sat_patterns[l] + j*width, "patterns", width - (3*cell-1), 1, 0, max_print_errors, 1, j);
if (errors > max_print_errors){
max_print_errors = 0;
}
}
}
return errors;
}
vbx_mtx_fdct_t *
vbx_mtx_fdct_init( dt *coeff_v, dt *image )
{
const int BIG_TILE_SIZE = NUM_TILE_X * NUM_TILE_Y * DCT_SIZE;
const int num_bytes = BIG_TILE_SIZE * sizeof(dt);
const int co_bytes = NUM_TILE_X* DCT_SIZE *sizeof(dt);
//compute coeffs matrix in double and truncated to dt
int i, j;
double s;
for (i = 0; i < BLOCK_SIZE; i++) {
s = (i == 0) ? sqrt(0.125) : 0.5;
for (j = 0; j < BLOCK_SIZE; j++) {
c2[i][j] = s * cos((double) ((PI / 8.0) * i * j + 0.5));
cs[i][j] = (dt) (c2[i][j] * SHIFT_DOUBLE + 0.499999);
}
}
vbx_sp_push();
vbx_mtx_fdct_t *v = vbx_shared_malloc( sizeof(vbx_mtx_fct_t) );
v->vcoeff = (vbx_half_t *)vbx_sp_malloc( co_bytes );
v->vprods = (vbx_half_t *)vbx_sp_malloc( num_bytes );
#if USE_ACCUM_FLAGS
v->vaccum = (vbx_half_t *)vbx_sp_malloc( num_bytes );
v->vflags = (vbx_half_t *)vbx_sp_malloc( num_bytes );
#endif
// interleave ordering to ensure no false hazards
v->vblock[2] = (vbx_half_t *)vbx_sp_malloc( num_bytes );
v->vimage[0] = (vbx_half_t *)vbx_sp_malloc( num_bytes );
v->vblock[0] = (vbx_half_t *)vbx_sp_malloc( num_bytes );
v->vimage[1] = (vbx_half_t *)vbx_sp_malloc( num_bytes );
v->vblock[1] = (vbx_half_t *)vbx_sp_malloc( num_bytes );
if( !v->vblock[1] ) {
VBX_PRINTF( "ERROR: out of memory.\n" );
VBX_EXIT(-1);
}
vbx_dma_to_vector( v->vcoeff, coeff_v, co_bytes );
int row;
for( row=0; row < BLOCK_SIZE; row++ ) {
getBigTileImageY(v->vimage[v->db],image,row);
}
#if USE_ACCUM_FLAGS
// create a flag vector first element 0, next 'BLOCK_SIZE-1' element non-zero, etc
vbx_set_vl( NUM_TILE_X * BLOCK_SIZE * NUM_TILE_Y * BLOCK_SIZE - (BLOCK_SIZE-1) );
vbx( SEH, VAND, v->vflags, BLOCK_SIZE-1, 0 );
#endif
return v;
}
int compare_vbx_lbp_ci_to_scalar_patterns(unsigned short* img, int log, int width, int height, int max_print_errors)
{
int j, l, cell, max_cell, errors = 0;
unsigned char** scalar_patterns = test_scalar_patterns(img, log, width, height);
max_cell = 1<<log;
vbx_uhalf_t* v_in = (vbx_uhalf_t*)vbx_sp_malloc((1+2*max_cell)*width*sizeof(vbx_half_t));
vbx_uhalf_t* v_top = (vbx_half_t*)vbx_sp_malloc(width*sizeof(vbx_half_t));
vbx_uhalf_t* v_bot = (vbx_half_t*)vbx_sp_malloc(width*sizeof(vbx_half_t));
vbx_ubyte_t* v_lbp = (vbx_ubyte_t*)v_bot;
unsigned char* lbp = (unsigned char*)vbx_shared_malloc(width*sizeof(unsigned char));
vbx_set_vl(width);
for(l = 0; l < 1; l++){
cell = 1<<l;
for(j=0; j < height - 2*cell; j++){
vbx_dma_to_vector(v_in, img+j*width, (1+2*cell)*width*sizeof(unsigned short));
vbx(VVHU, VCUSTOM1, v_top, v_in, v_in+(1*cell)*width);
vbx(VVHU, VCUSTOM1, v_bot, v_in+(1*cell)*width, v_in+(2*cell)*width);
vbx(SVHBU, VAND, (vbx_ubyte_t*)v_top, 0xf0, v_top);
vbx(SVHBU, VAND, (vbx_ubyte_t*)v_bot, 0x0f, v_bot);
vbx(VVBU, VADD, v_lbp, v_bot, v_top);
vbx_dma_to_host(lbp, v_lbp, width*sizeof(unsigned char));
vbx_sync();
errors += match_array_byte(lbp, scalar_patterns[l]+j*width, "custom_lbp", width-2*cell, 1, 0, max_print_errors, 1, j);
if (errors > max_print_errors){
max_print_errors = 0;
}
}
}
vbx_sp_free();
vbx_shared_free(lbp);
return errors;
}
int compare_LBPRestrictedCI_to_test_scalar_patterns(unsigned short* vbx_img, unsigned char* vbx_img8, int log, int width, int height, int max_print_errors)
{
int i, j, l, cell, errors = 0, cell_errors, row_errors;
/* generate patterns */
unsigned char **patterns = (unsigned char**)malloc((log+1)*sizeof(unsigned char*));
for (l=0; l<log+1; l++) {
patterns[l] = (unsigned char*)vbx_shared_malloc(height*width*sizeof(unsigned char));
}
/* LBPRestrictedCI28(patterns, vbx_img8, width, height, log); */
LBPRestricted_CI_column_8(patterns, vbx_img8, width, height, log);
/* LBPRestricted_CI_column_8_scratch(patterns, vbx_img8, width, height, log); */
unsigned char **scalar_patterns = test_scalar_patterns(vbx_img, log, width, height);
/* test sums vs scalar sums */
for (l=0; l<log+1; l++) {
cell = 1 << l;
cell_errors = 0;
printf("Testing cell %d\n", cell);
for (j = 0; j < height - (3*cell-1); j++) {
row_errors = match_array_byte(scalar_patterns[l]+j*width, patterns[l]+j*width, "restricted patterns", width - (3*cell-1), 1, 0, max_print_errors, 1, j);
if (row_errors) {
printf("errors in row %d\n", j);
}
cell_errors += row_errors;
errors += row_errors;
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
printf("Total errors: %d\n\n", cell_errors);
}
return errors;
}
int test_lbp_ci(unsigned short* img, int width, int height)
{
vbx_uhalf_t* v_a1 = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_b1 = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_1h = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_a2 = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_b2 = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_2h = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_a4 = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_b4 = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_uhalf_t* v_4h = (vbx_uhalf_t*)vbx_sp_malloc(width*sizeof(vbx_uhalf_t));
vbx_ubyte_t* v_1b = (vbx_ubyte_t*)vbx_sp_malloc(width*sizeof(vbx_ubyte_t));
vbx_ubyte_t* v_2b = (vbx_ubyte_t*)vbx_sp_malloc(width*sizeof(vbx_ubyte_t));
vbx_ubyte_t* v_4b = (vbx_ubyte_t*)vbx_sp_malloc(width*sizeof(vbx_ubyte_t));
unsigned short* lbp1h = (unsigned short*)vbx_shared_malloc(width*sizeof(unsigned short));
unsigned short* lbp2h = (unsigned short*)vbx_shared_malloc(width*sizeof(unsigned short));
unsigned short* lbp4h = (unsigned short*)vbx_shared_malloc(width*sizeof(unsigned short));
unsigned char* lbp1b = (unsigned char*)vbx_shared_malloc(width*sizeof(unsigned char));
unsigned char* lbp2b = (unsigned char*)vbx_shared_malloc(width*sizeof(unsigned char));
unsigned char* lbp4b = (unsigned char*)vbx_shared_malloc(width*sizeof(unsigned char));
img = img + width;
vbx_dma_to_vector(v_a1, img, width*sizeof(unsigned short));
vbx_dma_to_vector(v_b1, img + width, width*sizeof(unsigned short));
vbx_dma_to_vector(v_a2, img, width*sizeof(unsigned short));
vbx_dma_to_vector(v_b2, img + width, width*sizeof(unsigned short));
vbx_dma_to_vector(v_a4, img, width*sizeof(unsigned short));
vbx_dma_to_vector(v_b4, img + width, width*sizeof(unsigned short));
vbx_sync();
int i;
int m = 48;
for(i=0; i<m; i++){
v_a1[i] = 0;
v_b1[i] = 0;
v_a2[i] = 0;
v_b2[i] = 0;
v_a4[i] = 0;
v_b4[i] = 0;
}
int n = 12;
int src_a1[] = {0, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
int src_b1[] = {0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
int src_a2[] = {0, 0, 0, 2, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
int src_b2[] = {0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
int src_a4[] = {0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 2, 0, 0, 0, 0};
int src_b4[] = {0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0};
for(i=0; i<16; i++){
v_a1[i] = src_a1[i];
v_b1[i] = src_b1[i];
v_a2[i] = src_a2[i];
v_b2[i] = src_b2[i];
v_a4[i] = src_a4[i];
v_b4[i] = src_b4[i];
}
vbx_set_vl(width);
vbx(VVHU, VCUSTOM1, v_1h, v_a1, v_b1);
vbx(VVHU, VCUSTOM2, v_2h, v_a2, v_b2);
vbx(VVHU, VCUSTOM3, v_4h, v_a4, v_b4);
vbx(VVHB, VADD, v_1b, v_1h, ((vbx_byte_t*)v_1h) + 1);
vbx(VVHB, VADD, v_2b, v_2h, ((vbx_byte_t*)v_2h) + 1);
vbx(VVHB, VADD, v_4b, v_4h, ((vbx_byte_t*)v_4h) + 1);
vbx_dma_to_host(lbp1h, v_1h, width*sizeof(unsigned short));
vbx_dma_to_host(lbp2h, v_2h, width*sizeof(unsigned short));
vbx_dma_to_host(lbp4h, v_4h, width*sizeof(unsigned short));
vbx_dma_to_host(lbp1b, v_1b, width*sizeof(unsigned char));
vbx_dma_to_host(lbp2b, v_2b, width*sizeof(unsigned char));
vbx_dma_to_host(lbp4b, v_4b, width*sizeof(unsigned char));
vbx_sync();
test_print_array_half(v_a1, n);
test_print_array_half(v_b1, n);
test_print_hex_array_half(lbp1h, n);
test_print_hex_array_byte(lbp1b, n);
test_print_array_half(v_a2, n);
test_print_array_half(v_b2, n);
test_print_hex_array_half(lbp2h, n);
test_print_hex_array_byte(lbp2b, n);
test_print_array_half(v_a4, n);
test_print_array_half(v_b4, n);
test_print_hex_array_half(lbp4h, n);
test_print_hex_array_byte(lbp4b, n);
vbx_sp_free();
vbx_shared_free(lbp1h);
vbx_shared_free(lbp2h);
vbx_shared_free(lbp4h);
vbx_shared_free(lbp1b);
vbx_shared_free(lbp2b);
vbx_shared_free(lbp4b);
return 0;
}
//FIXME stride for match not implemented
int compare_LBPPassStage_to_restricted(unsigned short *vbx_img, int log, lbp_stage_t lbp_stage, int window, int width, int height, int max_print_errors)
{
int l, i, j, cell, errors = 0;
unsigned char** scalar_patterns = test_scalar_patterns(vbx_img, log, width, height);
unsigned char *pass, *vbx_pass;
pass = (unsigned char*)vbx_shared_malloc(width*height*sizeof(unsigned char));
vbx_pass = (unsigned char*)vbx_shared_malloc(width*height*sizeof(unsigned char));
vbx_byte_t** v_lbp =(vbx_byte_t**)vbx_shared_malloc((log+1)*sizeof(vbx_byte_t*));
for (l=0; l<log+1; l++) {
v_lbp[l] = (vbx_byte_t*)vbx_sp_malloc((window+1)*width*sizeof(vbx_byte_t));
}
vbx_byte_t* v_lut = (vbx_byte_t*)vbx_sp_malloc(width*sizeof(vbx_byte_t));
vbx_byte_t* v_stage = (vbx_byte_t*)vbx_sp_malloc(width*sizeof(vbx_byte_t));
vbx_byte_t* v_pattern;
lbp_feat_t feat;
int dx, dy, dw, f;
for (l=0; l<log+1; l++) {
vbx_dma_to_vector(v_lbp[l]+width, scalar_patterns[l], (window)*width*sizeof(unsigned char));
}
vbx_sync();
for(j=0; j < height-(window+1); j++) {
for (l=0; l<log+1; l++) {
vbx_set_vl(width * window);
vbx(VVB, VMOV, v_lbp[l], v_lbp[l]+width, NULL);
vbx_dma_to_vector(v_lbp[l] + window*width, scalar_patterns[l]+(j+window)*width, width*sizeof(unsigned char));
}
vbx_set_vl(width-(window+1));
vbx(SVB, VMOV, v_stage, 0, NULL);
for (f = 0; f < lbp_stage.count; f++) {
feat = lbp_stage.feats[f];
dx = feat.pos.src.x;
dy = feat.pos.src.y;
dw = feat.pos.size.x;
v_pattern = v_lbp[dw>>1]+(dy*width+dx);
vbx(SVBU, VLBPLUT, v_lut, f, v_pattern);
vbx(VVB, VADD, v_stage, v_stage, v_lut);
}
vbx(SVB, VMOV, v_lut, 0, NULL);
vbx(SVB, VCMV_GEZ, v_lut, 1, v_stage);
vbx_dma_to_host(vbx_pass + j*width, v_lut, (width-(window+1))*sizeof(unsigned char));
vbx_sync();
}
unsigned int *iImg, *iiImg;
iImg = (unsigned int *)vbx_shared_malloc(width*height*sizeof(unsigned int));
iiImg = (unsigned int *)vbx_shared_malloc(width*height*sizeof(unsigned int));
gen_integrals(vbx_img, iImg, iiImg, width, height);
image_t lbp_img = {iImg, {width, height}};
for (j = 0; j < height - (window + 1); j++) {
for (i = 0; i < width - (window + 1); i++) {
pair_t lbp_p = {i, j};
pass[j*width+i] = LBPPassStage(lbp_img, lbp_stage, lbp_p);
}
}
/* test pass vs vbx pass */
for (j = 0; j < height - (window + 1); j++) {
errors += match_array_byte(vbx_pass + j*width, pass + j*width, "pass stage", width - (window + 1), 1, 0, max_print_errors, 1, j);
if (errors > max_print_errors){
max_print_errors = 0;
}
}
return errors;
}
int compare_scalar_BLIP2_to_vector_BLIP(unsigned short* img, pixel* vbx_input, int width, int height, int max_print_errors, int scale_factor)
{
int j, errors = 0;
int scaled_width, scaled_height;
/* scale facetor v/v+1, v is between 1-10 */
scaled_width = width*scale_factor/(scale_factor+1);
scaled_height= height*scale_factor/(scale_factor+1);
unsigned short *scaled_img, *vbx_img, *vbx_scaled_img;
unsigned char *vbx_img8, *vbx_scaled_img8;
unsigned int *iImg, *iiImg, *vbx_iImg, *vbx_iiImg;
scaled_img = (unsigned short*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned short));
iImg = (unsigned int*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned int));
iiImg = (unsigned int*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned int));
vbx_scaled_img = (unsigned short*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned short));
vbx_img = (unsigned short*)vbx_shared_malloc(width*height*sizeof(unsigned short));
vbx_img8 = (unsigned char*)vbx_shared_malloc(width*height*sizeof(unsigned char));
vbx_scaled_img8 = (unsigned char*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned char));
vbx_iImg = (unsigned int*)vbx_shared_malloc(width*height*sizeof(unsigned int));
vbx_iiImg = (unsigned int*)vbx_shared_malloc(width*height*sizeof(unsigned int));
#if 0
scalar_BLIP2(img, height, width, scaled_img, scaled_height, scaled_width, scale_factor);
#else
float percent = 1.0 * (scale_factor+1) / scale_factor;
scalar_BLIP(img, height, width, scaled_img, scaled_height, scaled_width, &percent);
#endif
gen_integrals(scaled_img, iImg, iiImg, scaled_width, scaled_height);
vector_get_img(vbx_img, vbx_iImg, vbx_iiImg, vbx_input, 1, width, height, width, 1);
vector_BLIP(vbx_img, height, width, vbx_scaled_img, vbx_iImg, vbx_iiImg, scaled_height, scaled_width, scale_factor, 1);
vector_get_img8(vbx_img8, vbx_input, 1, width, height, width);
/* vector_BLIP8(vbx_img8, height, width, vbx_scaled_img8, scaled_height, scaled_width, scale_factor); */
vbx_timestamp_start();
vbx_timestamp_t time_start, time_stop;
double vbx_time;
time_start = vbx_timestamp();
#if 1
vector_BLIP8F3(vbx_img8, height, width, vbx_scaled_img8, scaled_height, scaled_width, scale_factor);
#else
vector_BLIP8F2(vbx_img8, height, width, vbx_scaled_img8, scaled_height, scaled_width, scale_factor);
#endif
time_stop = vbx_timestamp();
vbx_time = vbx_print_vector_time(time_start, time_stop, 0.0);
/* test greyscale image */
for (j = 0; j < height; j++) {
errors += match_array_half(img+j*width, vbx_img+j*width, "greyscale", width, 1, 0, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
/* test scaled image */
for (j = 0; j < scaled_height; j++) {
errors += match_array_half(scaled_img+j*scaled_width, vbx_scaled_img+j*scaled_width, "scaled greyscale", scaled_width, 1, 1, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
for (j = 0; j < scaled_height; j++) {
errors += match_array_half_byte(scaled_img+j*scaled_width, vbx_scaled_img8+j*scaled_width, "scaled greyscale8", scaled_width, 1, 1, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
#if 0
/* test scaled_integral image */
for (j = 0; j < scaled_height; j++) {
errors += match_array_word(iImg+j*scaled_width, vbx_iImg+j*scaled_width, "scaled integral", scaled_width, 1, 0, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
/* test scaled squared integral image */
for (j = 0; j < scaled_height; j++) {
errors += match_array_word(iiImg+j*scaled_width, vbx_iiImg+j*scaled_width, "scaled squared", scaled_width, 1, 0, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
#endif
/* test scaled_integral image */
return errors;
}
int main(void)
{
vbx_timestamp_t time_start, time_stop;
double scalar_time, vector_time;
input_pointer img1;
input_pointer img2;
input_pointer sc_img1;
input_pointer sc_img2;
output_pointer scalar_out;
output_pointer vector_out;
int i,j;
int total_errors = 0;
vbx_test_init();
vbx_mxp_print_params();
img1 = vbx_shared_malloc( NUM_OF_ROWS*NUM_OF_COLUMNS*sizeof(input_type) );
img2 = vbx_shared_malloc( NUM_OF_ROWS*NUM_OF_COLUMNS*sizeof(input_type) );
vector_out = vbx_shared_malloc( NUM_OF_ROWS*NUM_OF_COLUMNS*sizeof(output_type) );
sc_img1 = malloc( NUM_OF_ROWS*NUM_OF_COLUMNS*sizeof(input_type) );
sc_img2 = malloc( NUM_OF_ROWS*NUM_OF_COLUMNS*sizeof(input_type) );
scalar_out = malloc( NUM_OF_ROWS*NUM_OF_COLUMNS*sizeof(output_type) );
init_img( img1, img2 );
init_img( sc_img1, sc_img2 );
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
const int VBX_VECTOR_BYTE_LANES = this_mxp->vector_lanes * sizeof(int);
printf("\n");
printf("Num of byte lanes: %d\n", VBX_VECTOR_BYTE_LANES);
printf("Initialized data\n\n");
printf("Executing Scalar Image Blend...\n");
vbx_timestamp_start();
time_start = vbx_timestamp();
scalar_blend( scalar_out, sc_img1, sc_img2, NUM_OF_ROWS, NUM_OF_COLUMNS, CONST_BLEND );
time_stop = vbx_timestamp();
printf("Finished Scalar Image Blend\n");
scalar_time = vbx_print_scalar_time(time_start, time_stop);
printf("\nExecuting Vector Image Blend...\n");
vbx_timestamp_start();
time_start = vbx_timestamp();
vector_blend( vector_out, img1, img2, NUM_OF_ROWS, NUM_OF_COLUMNS, CONST_BLEND);
time_stop = vbx_timestamp();
printf("Finished Vector Image Blend\n");
vector_time = vbx_print_vector_time(time_start, time_stop, scalar_time);
int errors = 0;
for( j=0; j<NUM_OF_ROWS; j++ ) {
for( i = 0; i < NUM_OF_COLUMNS; i++ ) {
if( vector_out[j*NUM_OF_COLUMNS+i] != scalar_out[j*NUM_OF_COLUMNS+i] ) {
if(errors < 5)
printf( "\nFail at sample [%3d,%3d]. Scalar: %3d Vector: %3d Img1: %3d Img2: %3d",
j, i, scalar_out[j*NUM_OF_COLUMNS+i],
vector_out[j*NUM_OF_COLUMNS+i], img1[j*NUM_OF_COLUMNS+i], img2[j*NUM_OF_COLUMNS+i] );
errors++;
}
}
}
printf("\n%d errors\n", errors);
total_errors += errors;
VBX_TEST_END(total_errors);
return 0;
}
int main(void)
{
vbx_test_init();
typedef vbx_word_t vbx_mm_t;
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
const int VBX_SCRATCHPAD_SIZE = this_mxp->scratchpad_size;
int N = VBX_SCRATCHPAD_SIZE / sizeof(vbx_mm_t );
N = 20;
int M = 20;
int PRINT_LENGTH = N<MAX_PRINT_LENGTH ? N : MAX_PRINT_LENGTH ;
// int PRINT_ROWS = PRINT_LENGTH;
int PRINT_ROWS = M<MAX_PRINT_LENGTH ? N : MAX_PRINT_LENGTH;
int PRINT_COLS = PRINT_LENGTH;
double scalar_time, vector_time,vector2_time;
int errors=0;
vbx_mxp_print_params();
printf( "\nMatrix multiply test...\n" );
printf( "Matrix dimensions: %d,%d\n", N, M );
vbx_mm_t *scalar_in1 = (vbx_mm_t*)malloc( M*N*sizeof(vbx_mm_t ) );
vbx_mm_t *scalar_in2 = (vbx_mm_t*)malloc( M*N*sizeof(vbx_mm_t ) );
vbx_mm_t *scalar_out = (vbx_mm_t*)malloc( N*N*sizeof(vbx_mm_t ) );
vbx_mm_t *vector_in1 = (vbx_mm_t*)vbx_shared_malloc( M*N*sizeof(vbx_mm_t ) );
vbx_mm_t *vector_in2 = (vbx_mm_t*)vbx_shared_malloc( M*N*sizeof(vbx_mm_t ) );
vbx_mm_t *vector_out = (vbx_mm_t*)vbx_shared_malloc( N*N*sizeof(vbx_mm_t ) );
if ( scalar_in1 == NULL ||
scalar_in2 == NULL ||
scalar_out == NULL ||
vector_in1 == NULL ||
vector_in2 == NULL ||
vector_out == NULL ){
printf("Malloc failed\n");
VBX_TEST_END(1);
return 0;
}
test_zero_array_word(scalar_out, N*N );
test_zero_array_word(vector_out, N*N );
test_init_array_word( scalar_in1, M*N, 1 );
test_copy_array_word( vector_in1, scalar_in1, M*N );
test_init_array_word( scalar_in2, M*N, 999 );
//scalar_mtx_xp_MN_word( vector_in2, scalar_in2, N, N );
test_copy_array_word( vector_in2, scalar_in2, M*N );
test_print_matrix_word( scalar_in1, PRINT_COLS, PRINT_ROWS, M );
test_print_matrix_word( scalar_in2, PRINT_ROWS, PRINT_COLS, N );
//change print sizes for outputs
PRINT_ROWS=PRINT_COLS=N<PRINT_LENGTH?N:PRINT_LENGTH;
scalar_time = test_scalar( scalar_out, scalar_in1, N, M, scalar_in2, M, N);
test_print_matrix_word( scalar_out, PRINT_COLS, PRINT_ROWS, N );
vector_time = test_vector( vector_out, vector_in1, N, M, vector_in2, M, N, scalar_time );
test_print_matrix_word( vector_out, PRINT_COLS, PRINT_ROWS, N );
errors += test_verify_array_word( scalar_out, vector_out, N*N);
vector2_time = test_vector_trans( vector_out, vector_in1, N, M, vector_in2, M, N, scalar_time );
test_print_matrix_word( vector_out, PRINT_COLS, PRINT_ROWS, N );
errors += test_verify_array_word( scalar_out, vector_out, N*N);
vector2_time = test_vector_sp( vector_out, vector_in1, N, M, vector_in2, M, N, scalar_time );
test_print_matrix_word( vector_out, PRINT_COLS, PRINT_ROWS, N );
errors += test_verify_array_word( scalar_out, vector_out, N*N);
vbx_shared_free(vector_out);
vbx_shared_free(vector_in2);
vbx_shared_free(vector_in1);
free(scalar_out);
free(scalar_in2);
free(scalar_in1);
//errors += orig_test();
VBX_TEST_END(errors);
return 0;
}
int VBX_T(vbw_vec_reverse_test_mm)()
{
unsigned int aN[] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 12, 15, 16, 17, 20, 25, 31, 32, 33, 35, 40, 48, 60, 61, 62, 63, 64, 64, 65,
66, 67, 68, 70, 80, 90, 99, 100, 101, 110, 128, 128, 144, 144, 160, 160, 176, 176, 192, 192, 224, 224,
256, 256, 288, 288, 320, 320, 352, 352, 384, 384, 400, 450, 512, 550, 600, 650, 700, 768, 768, 900,
900, 1023, 1024, 1200, 1400, 1600, 1800, 2048, 2048, 2100, 2200, 2300, 2400, 2500, 2600, 2700, 2800,
2900, 3000, 3100, 3200, 3300, 3400, 3500, 3600, 3700, 3800, 3900, 4000, 4096, 4096, 4100, 4200, 4300,
4400, 4500, 4600, 4700, 4800, 4900, 5000, 6000, 7000, 8000, 8192, 8192, 9000, 10000, 11000, 12000,
13000, 14000, 15000, 16000, 16384, 16384, 20000, 25000, 30000, 32767, 32768, 32768, 35000, 40000,
45000, 50000, 55000, 60000, 65000, 65535, 65536, 65536, 65537, 100000, 128000, 256000, 333333, 528374,
528374 };
int retval;
unsigned int N;
unsigned int NBYTES;
unsigned int NREPS = 100;
unsigned int NREPSFORLARGE = 10;
unsigned int i,j,k;
vbx_timestamp_t start=0,finish=0;
for( i=0; i<sizeof(aN)/4; i++ ) {
N = aN[i];
//printf( "testing with vector size %d\n", N );
if(N > 10000) NREPS = NREPSFORLARGE;
NBYTES = N*sizeof(vbx_mm_t);
vbx_mm_t *src = (vbx_mm_t *) vbx_shared_malloc( NBYTES );
vbx_mm_t *dst = (vbx_mm_t *) vbx_shared_malloc( NBYTES );
//printf("bytes alloc: %d\n", NBYTES );
if( !src ) VBX_EXIT(-1);
if( !dst ) VBX_EXIT(-1);
for ( j=0; j<N; j++ ) {
dst[j] = -1; // Fill the destination with -1
src[j] = j; // Fill the source with enumerated values
}
// VBX_T(vbw_vec_reverse_ext)( dst, src, N );
/** measure performance of function call **/
start = vbx_timestamp();
for(k=0; __builtin_expect(k<NREPS,1); k++ ) {
retval = VBX_T(vbw_vec_reverse_ext)( dst, src, N );
}
finish = vbx_timestamp();
printf( "length %d (%s):\tvbware mm f():\t%llu", N, VBX_EXPAND_AND_QUOTE(BYTEHALFWORD), (unsigned long long) vbx_mxp_cycles((finish-start)/NREPS) );
#if VERIFY_VBWARE_ALGORITHM
VBX_T(verify_vector)( src, dst, N );
#else
printf(" [VERIFY OFF]");
#endif
printf("\treturn value: %X", retval);
/** measure performance of scalar **/
vbx_mm_t *A = vbx_remap_cached( src, N*sizeof(vbx_mm_t) ); // Use cached pointers for better performance
vbx_mm_t *B = vbx_remap_cached( dst, N*sizeof(vbx_mm_t) );
start = vbx_timestamp();
for(k=0; k<NREPS; k++ ) {
unsigned int m;
for(m=0; m<N; m++) {
B[N-1-m]=A[m];
}
vbx_dcache_flush( A, N*sizeof(vbx_mm_t) ); // Make sure to read from main memory
vbx_dcache_flush( B, N*sizeof(vbx_mm_t) ); // Make sure writes are committed to memory
}
finish = vbx_timestamp();
printf( "\tscalar (cache friendly):\t%llu", (unsigned long long) vbx_mxp_cycles((finish-start)/NREPS) );
#if VERIFY_SIMPLE_ALGORITHM
VBX_T(verify_vector)( src, dst, N );
#else
printf(" [VERIFY OFF]");
#endif
printf("\tcycles\n");
vbx_shared_free(src);
vbx_shared_free(dst);
}
printf("All tests passed successfully.\n");
return 0;
}
int main(void)
{
vbx_test_init();
#if 0
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
const int VBX_SCRATCHPAD_SIZE = this_mxp->scratchpad_size;
int N = VBX_SCRATCHPAD_SIZE/sizeof(vbx_mm_t)/8;
#endif
int TEST_LENGTH = TEST_ROWS*TEST_COLS;
int NTAP_LENGTH = NTAP_ROWS*NTAP_COLS;
int PRINT_COLS = min( TEST_COLS, MAX_PRINT_LENGTH );
int PRINT_ROWS = min( TEST_ROWS, MAX_PRINT_LENGTH );
double scalar_time, vector_time;
int errors=0;
vbx_mxp_print_params();
printf( "\nMatrix FIR test...\n" );
printf( "Matrix dimensions: %d,%d\n", TEST_ROWS, TEST_COLS );
vbx_mm_t *scalar_in = malloc( TEST_LENGTH*sizeof(vbx_mm_t) );
vbx_mm_t *vector_in = vbx_shared_malloc( TEST_LENGTH*sizeof(vbx_mm_t) );
int32_t *scalar_filt = malloc( NTAP_LENGTH*sizeof(int32_t) );
int32_t *vector_filt = vbx_shared_malloc( NTAP_LENGTH*sizeof(int32_t) );
vbx_mm_t *scalar_out = malloc( TEST_LENGTH*sizeof(vbx_mm_t) );
vbx_mm_t *vector_out = vbx_shared_malloc( TEST_LENGTH*sizeof(vbx_mm_t) );
VBX_T(test_zero_array)( scalar_out, TEST_LENGTH );
VBX_T(test_zero_array)( vector_out, TEST_LENGTH );
VBX_T(test_init_array)( scalar_in, TEST_LENGTH, 1 );
VBX_T(test_copy_array)( vector_in, scalar_in, TEST_LENGTH );
test_init_array_word( scalar_filt, NTAP_LENGTH, 1 );
test_copy_array_word( vector_filt, scalar_filt, NTAP_LENGTH );
VBX_T(test_print_matrix)( scalar_in, PRINT_ROWS, PRINT_COLS, TEST_COLS );
test_print_matrix_word( scalar_filt, NTAP_ROWS, NTAP_COLS, NTAP_COLS );
scalar_time = test_scalar( scalar_out, scalar_in, scalar_filt,
TEST_ROWS, TEST_COLS, NTAP_ROWS, NTAP_COLS);
VBX_T(test_print_matrix)( scalar_out, PRINT_COLS, PRINT_ROWS, TEST_COLS );
vector_time = test_vector( vector_out, vector_in, vector_filt,
TEST_ROWS, TEST_COLS, NTAP_ROWS, NTAP_COLS, scalar_time );
VBX_T(test_print_matrix)( vector_out, PRINT_COLS, PRINT_ROWS, TEST_COLS );
int i;
for(i=0; i<TEST_ROWS-NTAP_ROWS; i++){
errors += VBX_T(test_verify_array)( scalar_out+i*TEST_COLS, vector_out+i*TEST_COLS, TEST_COLS-NTAP_COLS );
}
VBX_TEST_END(errors);
return 0;
}
int main(void)
{
vbx_test_init();
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
const int VBX_SCRATCHPAD_SIZE = this_mxp->scratchpad_size;
const int required_vectors = 4;
int N = VBX_PAD_DN(VBX_SCRATCHPAD_SIZE / sizeof(vbx_mm_t) / required_vectors, this_mxp->scratchpad_alignment_bytes);
int PRINT_LENGTH = min( N, MAX_PRINT_LENGTH );
double scalar_time, vector_time;
int errors=0;
vbx_mxp_print_params();
printf( "\nVector copy test...\n" );
printf( "Vector length: %d\n", N );
vbx_mm_t *scalar_in = malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *scalar_out = malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *vector_in = vbx_shared_malloc( N*sizeof(vbx_mm_t) );
vbx_mm_t *vector_out = vbx_shared_malloc( N*sizeof(vbx_mm_t) );
vbx_sp_t *v_out = vbx_sp_malloc( N*sizeof(vbx_sp_t) );
vbx_sp_t *v_in = vbx_sp_malloc( N*sizeof(vbx_sp_t) );
VBX_T(test_zero_array)( scalar_in, N );
VBX_T(test_zero_array)( vector_in, N );
VBX_T(test_init_array)( scalar_in, N, 1 );
VBX_T(test_copy_array)( vector_in, scalar_in, N );
scalar_time = test_scalar( scalar_out, scalar_in, N );
VBX_T(test_print_array)( scalar_out, PRINT_LENGTH );
vbx_dma_to_vector( v_in, vector_in, N*sizeof(vbx_sp_t) );
vector_time = test_vector( v_out, v_in, N, scalar_time );
vbx_dma_to_host(vector_out, v_out, N*sizeof(vbx_sp_t) );
vbx_sync();
VBX_T(test_print_array)( vector_out, PRINT_LENGTH );
errors += VBX_T(test_verify_array)( scalar_out, vector_out, N );
vbx_sp_free();
#if TEST_DEEP_SP
errors += deep_vector_copy_test();
#endif
#if DEBUG_MAKE_SP_FULL
vbx_sp_malloc(vbx_sp_getfree());
#endif
#if TEST_DEEP_MM
errors += deep_vector_copy_ext_test();
#endif
VBX_TEST_END(errors);
return 0;
}
int compare_vbx_lut_to_vbx_lut_ci(int stage, int max_print_errors)
{
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
int vci_lanes = this_mxp->vcustom0_lanes;
int sz = this_mxp->scratchpad_size/(16*sizeof(vbx_ubyte_t));
vbx_byte_t* v_pass = (vbx_byte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_pattern = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_lutc = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_group = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_sel = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_byte_t));
vbx_ubyte_t* v_lut = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_word_t));
vbx_ubyte_t* v_idx = (vbx_ubyte_t*)vbx_sp_malloc(sz*sizeof(vbx_word_t));
if(v_idx == NULL) {
printf("failed to allocate in compare_vbx_lut_to_vbx_lut_ci\n");
}
unsigned char* lut = (unsigned char*)vbx_shared_malloc(sz*sizeof(unsigned char));
unsigned char* lut_c = (unsigned char*)vbx_shared_malloc(sz*sizeof(unsigned char));
int f, n, s, errors = 0;
for (n = 0; n < sz; n++) {
v_pattern[n] = (n & 0xff);
}
for (f = 0; f < face_lbp[stage].count; f++) {
lbp_feat_t feat = face_lbp[stage].feats[f];
vbx_set_vl(sz);
int total = f;
s = 0;
while(s < stage){
total += face_lbp[s].count;
s++;
}
if(total < 256) {
vbx(SVBU, VLBPLUT, v_lutc, total, v_pattern);
} else {
vbx(SVBS, VLBPLUT, v_lutc, total-256, v_pattern);
}
vbx(SVB, VMOV, v_pass, feat.fail, 0);
/* check if pattern is in lut */
vbx(SVBU, VSHR, v_group, 5, v_pattern);
for (n = 0; n < 8; n++) {
vbx(SVB, VADD, v_sel, -n, v_group);
vbx(SVBW, VCMV_Z, v_lut, feat.lut[n], v_sel);
}
vbx(SVBWU, VAND, v_idx, 0x1f, v_pattern);
vbx(VVWB, VSHR, v_lut, v_idx, v_lut);
vbx(SVB, VAND, v_lut, 1, v_lut);
vbx(SVB, VCMV_LEZ, v_pass, feat.pass, v_lut);
vbx_dma_to_host(lut_c, v_lutc, sz*sizeof(unsigned char));
vbx_dma_to_host(lut, v_pass, sz*sizeof(unsigned char));
vbx_sync();
errors += match_array_byte(lut, lut_c, "custom_lut", sz, 1, 0, max_print_errors, 0, 0);
}
vbx_sp_free();
vbx_shared_free(lut);
vbx_shared_free(lut_c);
return errors;
}
int main(void)
{
pixel *input;
pixel *scalar_input;
#if USE_LUMA
unsigned char *vbx_luma;
#endif
unsigned short *scalar_luma;
pixel *vbx_output;
pixel *scalar_output;
vbx_timestamp_t time_start, time_stop;
double scalar_time, vbx_time;
int x, y;
int errors = 0;
vbx_test_init();
vbx_mxp_print_params();
input = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
scalar_input = (pixel *)vbx_remap_cached(input, IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
#if USE_LUMA
vbx_luma = (unsigned char *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned char));
#endif
scalar_luma = (unsigned short *)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned short));
vbx_output = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
scalar_output = (pixel *)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
printf("\nInitializing data\n");
printf("Resolution = %dx%d\n", IMAGE_WIDTH, IMAGE_HEIGHT);
init_matrix(input, IMAGE_WIDTH, IMAGE_HEIGHT);
printf("Starting Sobel 3x3 edge-detection test\n");
#if USE_LUMA
scalar_rgb2luma(scalar_luma, scalar_input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_PITCH);
#endif
vbx_timestamp_start();
time_start = vbx_timestamp();
#if !USE_LUMA
scalar_rgb2luma(scalar_luma, scalar_input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_PITCH);
#endif
scalar_sobel_argb32_3x3(scalar_output, scalar_luma, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_PITCH, RENORM_AMOUNT);
time_stop = vbx_timestamp();
scalar_time = vbx_print_scalar_time(time_start, time_stop);
#if USE_LUMA
vbw_rgb2luma8(vbx_luma, (unsigned *)input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_PITCH);
#endif
vbx_timestamp_start();
time_start = vbx_timestamp();
#if USE_LUMA
vbw_sobel_luma8_3x3((unsigned *)vbx_output, vbx_luma, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_PITCH, RENORM_AMOUNT);
#else
vbw_sobel_argb32_3x3((unsigned *)vbx_output, (unsigned *)input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_PITCH, RENORM_AMOUNT);
#endif
time_stop = vbx_timestamp();
vbx_time = vbx_print_vector_time(time_start, time_stop, scalar_time);
for (y = 0; y < IMAGE_HEIGHT; y++) {
for (x = 0; x < IMAGE_WIDTH; x++) {
#if USE_LUMA
if (scalar_luma[y*IMAGE_WIDTH+x] != vbx_luma[y*IMAGE_WIDTH+x]) {
if (errors < MAX_PRINT_ERRORS) {
printf("Y Error at %d, %d: Expected = %02X, got = %02X\n",
y, x, scalar_luma[y*IMAGE_WIDTH+x], vbx_luma[y*IMAGE_WIDTH+x]);
}
errors++;
}
#endif
if (scalar_output[y*IMAGE_WIDTH+x].r != vbx_output[y*IMAGE_WIDTH+x].r) {
if (errors < MAX_PRINT_ERRORS) {
printf("R Error at %d, %d: Expected = %02X, got = %02X\n",
y, x, scalar_output[y*IMAGE_WIDTH+x].r, vbx_output[y*IMAGE_WIDTH+x].r);
}
errors++;
}
if (scalar_output[y*IMAGE_WIDTH+x].g != vbx_output[y*IMAGE_WIDTH+x].g) {
if (errors < MAX_PRINT_ERRORS) {
printf("G Error at %d, %d: Expected = %02X, got = %02X\n",
y, x, scalar_output[y*IMAGE_WIDTH+x].g, vbx_output[y*IMAGE_WIDTH+x].g);
}
errors++;
}
if (scalar_output[y*IMAGE_WIDTH+x].b != vbx_output[y*IMAGE_WIDTH+x].b) {
if (errors < MAX_PRINT_ERRORS) {
printf("B Error at %d, %d: Expected = %02X, got = %02X\n",
y, x, scalar_output[y*IMAGE_WIDTH+x].b, vbx_output[y*IMAGE_WIDTH+x].b);
}
errors++;
}
}
}
VBX_TEST_END(errors);
return errors;
}
int compare_scalar_BLIP2_to_vector_BLIP(unsigned short* img, pixel* vbx_input, int width, int height, int max_print_errors)
{
int j, errors = 0;
int value, scaled_width, scaled_height;
/* scale facetor v/v+1, v is between 1-10 */
value = 3; //BAD 2,5,6,8
scaled_width = width*value/(value+1);
scaled_height= height*value/(value+1);
unsigned short *scaled_img, *vbx_img, *vbx_scaled_img;
unsigned int *iImg, *iiImg, *vbx_iImg, *vbx_iiImg;
scaled_img = (unsigned short*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned short));
iImg = (unsigned int*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned int));
iiImg = (unsigned int*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned int));
vbx_scaled_img = (unsigned short*)vbx_shared_malloc(scaled_width*scaled_height*sizeof(unsigned short));
vbx_img = (unsigned short*)vbx_shared_malloc(width*height*sizeof(unsigned short));
vbx_iImg = (unsigned int*)vbx_shared_malloc(width*height*sizeof(unsigned int));
vbx_iiImg = (unsigned int*)vbx_shared_malloc(width*height*sizeof(unsigned int));
scalar_BLIP2(img, height, width, scaled_img, scaled_height, scaled_width, value);
gen_integrals(scaled_img, iImg, iiImg, scaled_width, scaled_height);
vector_get_img(vbx_img, vbx_iImg, vbx_iiImg, vbx_input, 1, width, height, width, 1);
vector_BLIP(vbx_img, height, width, vbx_scaled_img, vbx_iImg, vbx_iiImg, scaled_height, scaled_width, value, 1);
/* test greyscale image */
for (j = 0; j < height; j++) {
errors += match_array_half(img+j*width, vbx_img+j*width, "greyscale", width, 1, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
/* test scaled image */
for (j = 0; j < scaled_height; j++) {
errors += match_array_half(scaled_img+j*scaled_width, vbx_scaled_img+j*scaled_width, "scaled greyscale", scaled_width, 1, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
/* test scaled_integral image */
for (j = 0; j < scaled_height; j++) {
errors += match_array_word(iImg+j*scaled_width, vbx_iImg+j*scaled_width, "scaled integral", scaled_width, 1, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
/* test scaled squared integral image */
for (j = 0; j < scaled_height; j++) {
errors += match_array_word(iiImg+j*scaled_width, vbx_iiImg+j*scaled_width, "scaled squared", scaled_width, 1, max_print_errors, j);
if(errors > max_print_errors) {
max_print_errors = 0;
}
}
return errors;
}
int main_tile()
{
int i, j, k, l, base, block_num;
int x, y;
int time_start, time_stop;
unsigned int cycles;
double vbx_time, scalar_time;
int wrong;
int total_errors = 0;
//all of the initialization can be hard coded without any computation
vbx_mtx_fdct_t *v = vbx_mtx_fdct_init( coeff_v, image );
vbx_timestamp_start();
printf("\nGenerating initial data...\n");
dt *image = (dt *) malloc( IMAGE_WIDTH * IMAGE_HEIGHT * sizeof(dt) );
GenerateRandomImage( image, IMAGE_WIDTH, IMAGE_HEIGHT, 0/*seed*/ );
// Allocate memory to store results.
// Results are computed BIGTILE_SIZE halfwords at a time.
const int BIGTILE_SIZE = NUM_TILE_X * NUM_TILE_Y * DCT_SIZE;
dt *block_s = malloc( BIGTILE_SIZE * sizeof(dt) );
dt *block_v = (dt *) vbx_shared_malloc( BIGTILE_SIZE * sizeof(dt) );
dt *coeff_v = (dt *) vbx_shared_malloc( BIGTILE_SIZE * sizeof(dt) );
//Make an uncached 1D version of the coeff matrix
for (i = 0; i < NUM_TILE_Y; i++) { // row
for (j = 0; j < BLOCK_SIZE; j++) { // row
for (k = 0; k < NUM_TILE_X; k++) { // col
for (l = 0; l < BLOCK_SIZE; l++) { // col
coeff_v[i*NUM_TILE_X*DCT_SIZE + j*DCT_SIZE + k*BLOCK_SIZE + l] = cs[j][l];
}
}
}
}
#ifdef DEBUG
printf("input matrix is:\n");
for (i = 0; i < BLOCK_SIZE; i++) {
base = i * BLOCK_SIZE;
for (j = 0; j < BLOCK_SIZE; j++) {
printf("%d ", (int) block_s[base + j]);
}
printf("\n");
}
#endif
printf("\nRunning DCT...\n");
time_start = vbx_timestamp();
for( y = 0; y < IMG_DOWN; y++ ) {
for( x = 0; x < IMG_ACROSS; x++ ) {
vbx_mtx_fdct_scalar( block_s, (dt*)cs, image, x/*start_x*/, y/*start_y*/, NUM_TILE_X, NUM_TILE_Y );
}
}
time_stop = vbx_timestamp();
cycles = time_stop - time_start;
scalar_time = (double) cycles;
scalar_time /= (double) vbx_timestamp_freq();
scalar_time *= 1000.0; //ms
vbx_timestamp_t mxp_cycles = vbx_mxp_cycles(cycles);
printf("%dx%d Block Size\n", BLOCK_SIZE, BLOCK_SIZE);
printf("Finished, scalar CPU took %0.3f ms \n", scalar_time);
printf(" CPU Cycles: %d\n", (int) mxp_cycles);
printf(" CPU Cycles per block: %f\n", mxp_cycles / ((double) (NUM_BLOCKS)));
vbx_sync(); // wait for image to be prefetched
time_start = vbx_timestamp();
for( y = 0; y < IMG_DOWN; y++ ) {
for( x = 0; x < IMG_ACROSS; x++ ) {
vbx_mtx_fdct( v, block_v, image, x/*start_x*/, y/*start_y*/, IMG_ACROSS-1,IMG_DOWN-1,NUM_TILE_X, NUM_TILE_Y );
}
}
time_stop = vbx_timestamp();
cycles = time_stop - time_start;
vbx_time = (double) cycles;
vbx_time /= (double) vbx_timestamp_freq();
vbx_time *= 1000.0; //ms
mxp_cycles = vbx_mxp_cycles(cycles);
printf("Finished, MXP took %0.3f ms \n", vbx_time);
printf(" CPU Cycles: %d\n", (int) mxp_cycles);
printf(" CPU Cycles per block: %f\n", mxp_cycles / ((double) (NUM_BLOCKS)));
printf(" Speedup: %f\n", scalar_time / vbx_time);
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
double vbx_mbps = (double) (NUM_BLOCKS) * 1000 / vbx_time; // blocks per second
printf("V%d@%dMHz: %dx%d tile, %dx%d blocks, %f blocks/s, %f megapixel/s\n",
this_mxp->vector_lanes, this_mxp->core_freq / 1000000,
NUM_TILE_Y, NUM_TILE_X,
BLOCK_SIZE, BLOCK_SIZE,
vbx_mbps, (vbx_mbps * DCT_SIZE) / 1000000);
printf("\nChecking results...\n");
wrong = 0;
for (block_num = 0; block_num < NUM_BLOCKS; block_num++) {
for (i = 0; i < BLOCK_SIZE; i++) {
base = i * BLOCK_SIZE;
for (j = 0; j < BLOCK_SIZE; j++) {
if (block_s[block_num * DCT_SIZE + base + j] != block_v[block_num * DCT_SIZE + base + j]) {
if (wrong < 5) {
printf("\nError at %d [%d,%d], result is %d, should be %d\n",
block_num, i, j, (int) block_v[block_num * DCT_SIZE + base + j],
(int) block_s[block_num * DCT_SIZE + base + j]);
}
wrong++;
}
}
}
}
printf("wrong is %d\n\n", wrong);
total_errors += wrong;
free(block_s);
vbx_shared_free(block_v);
vbx_shared_free(coeff_v);
vbx_mtx_fdct_free( v );
VBX_TEST_END(total_errors);
return (0);
}
int main(void)
{
vbx_timestamp_t time_start, time_stop;
double scalar_time, vbx_time, vbx_time_masked;
int i, j, k, l, m, n;
int errors = 0;
vbx_test_init();
vbx_mxp_print_params();
pixel *input, *scalar_input, *vbx_input, *vbx_input_masked;
uint16_t *scalar_short;
input = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
scalar_input = (pixel *)vbx_remap_cached(input, IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
scalar_short = (uint16_t *)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(uint16_t));
vbx_input = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
vbx_input_masked = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
#if UNIT
unsigned char *vbx_img8;
unsigned short *img, *vbx_img;
unsigned int *iImg, *vbx_iImg;
unsigned int *iiImg, *vbx_iiImg;
img = (unsigned short*)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned short));
vbx_img = (unsigned short*)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned short));
vbx_img8 = (unsigned char*)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned char));
iImg = (unsigned int*)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
vbx_iImg = (unsigned int*)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
iiImg = (unsigned int*)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
vbx_iiImg = (unsigned int*)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
#endif//UNIT
printf("Resolution = %dx%d\n", IMAGE_WIDTH, IMAGE_HEIGHT);
printf("Initializing data\n");
vbx_timestamp_start();
for(l = 0; l < 1; l++){
char *src;
char *sdst;
char *vdst;
char *mdst;
if(l == 0){
load_lenna(input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_lenna(vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_lenna(vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT);
printf("\nLenna\n");
src = "lenna";
sdst = "s_lenna";
vdst = "v_lenna";
mdst = "m_lenna";
}else if(l == 1){
load_ms(input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_ms(vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_ms(vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT);
printf("\nMicrosoft\n");
src = "ms";
sdst = "s_ms";
vdst = "v_ms";
mdst = "m_ms";
}else if(l == 2){
load_blank(input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_blank(vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_blank(vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT);
printf("\nblank\n");
src = "blank";
sdst = "s_blank";
vdst = "v_blank";
mdst = "m_blank";
}
#if UNIT
int window = 20;
int log=0;
while(((window/3)>>log) >= 2) log++;
errors += compare_scalar_rgb2luma_to_vbw_rgb2luma16(img, vbx_img, vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, MAX_PRINT_ERRORS);
vbw_rgb2luma8(vbx_img8, vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH);
int s;
#if LUT_CI
#if DOUBLE_LUT
printf("Testing double lut\n");
printf("Assign lbp double lut\n");
assign_lbp_lut_ci2();
int prev = errors;
printf("Cascade check\n");
/* errors += cascade_check_2w(face_lbp, face_lbp_max_stage, 256); */
/* errors += cascade_check_2h(face_lbp, face_lbp_max_stage, 256); */
errors += cascade_check_2b(face_lbp, face_lbp_max_stage, 256);
if (errors) {
printf("errors %d\n", errors-prev);
}
#else
assign_lbp_lut_ci();
printf("Testing cascade\n");
int prev = errors;
printf("lut check\n");
#if 0
#if 0
errors += lut_check(256, 0, 0, 0);
if (errors) {
printf("errors %d\n", errors-prev);
}
#elif 1
int print_errors = 0;
vbx_mxp_t *this_mxp = VBX_GET_THIS_MXP();
int vci_lanes = this_mxp->vcustom0_lanes;
int num_features = cascade_max_feature();
int input_length = 10;
int lut_length = num_features*vci_lanes;
int lut_iterations = 15;
#if 1
lut_length = input_length = 128;
lut_iterations = 13;
print_errors = 0;
errors += lut_check2(input_length, lut_length, lut_iterations, print_errors);
if (errors) {
printf("errors %d\n", errors-prev);
}
#elif 1
input_length = 64;
lut_length = input_length;
lut_iterations = 13;
print_errors = 1;
errors += lut_check2(input_length, lut_length, lut_iterations, print_errors);
if (errors) {
printf("errors %d\n", errors-prev);
}
#else
for(s = 2; s < 100; s=s+10){
errors += lut_check2(s, lut_length, lut_iterations, print_errors);
if (errors - prev > 0) {
printf("%d\terrors %d\n", s, errors-prev);
} else {
printf("%d\n", s);
}
prev = errors;
}
#endif
#else
for(s = 0; s < 2000; s=s+100){
errors += lut_check(s, 0, 0, 0);
if (errors - prev > 0) {
printf("%d\terrors %d\n", s, errors-prev);
} else {
printf("%d\n", s);
}
prev = errors;
}
#endif
#elif 1
#else
printf("check cascade\n");
prev = errors;
errors += cascade_check(face_lbp, face_lbp_max_stage, 256);
if (errors) {
printf("errors %d\n", errors-prev);
}
printf("Testing LBP LUT CI\n");
prev = errors;
for(s = 0; s < face_lbp_max_stage; s++){
errors += compare_vbx_lut_to_vbx_lut_ci(s, MAX_PRINT_ERRORS);
}
if (errors) {
printf("errors %d\n", errors-prev);
prev = errors;
}
#endif
#endif
#endif
#if 0
printf("Printing grey scale img\n");
printf("grey = [");
for (j = 0; j < IMAGE_HEIGHT; j++) {
printf("[");
for (i = 0; i < IMAGE_WIDTH; i++) {
printf("%d, ", vbx_img8[j*IMAGE_WIDTH+i]);
}
printf("],\n");
}
printf("]\n");
#endif
#if LBP_CI
printf("Testing LBP Pattern CI\n");
errors += compare_LBPRestrictedCI_to_test_scalar_patterns(vbx_img, vbx_img8, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
#endif
#if BLIP
printf("Testing BLIP\n");
for(s = 1; s < 10; s++){
errors += compare_scalar_BLIP2_to_vector_BLIP(img, vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS, s);
}
#endif
#if 0
errors += compare_LBPRestrictedSums_to_test_scalar_sums_byte(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_LBPRestrictedSums2_to_test_scalar_sums_half(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_ScalarLBPRestrictedSums_to_test_scalar_sums_half(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_ScalarLBPRestrictedPatterns_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_LBPRestrictedPatterns2_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_LBPRestricted_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
/* overflow issues -- using bytes changes lbp pattern */
errors += compare_LBPRestrictedPatterns_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
/* requires SKIP_INTEGRALS 0 */
errors += compare_gen_integrals_to_vector_get_img(img, iImg, iiImg, vbx_img, vbx_iImg, vbx_iiImg, vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
/* redundant test, compare to test_scalar_patterns instead */
errors += compare_ScalarLBPRestrictedPatterns_to_SATBinaryPattern(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_SATBinaryPattern_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_LBPPassStage_to_restricted(vbx_img, log, face_lbp[0], window, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
#endif
#else // UNIT
#if PRINT
print_python_pixel(scalar_input, src, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
time_start = vbx_timestamp();
scalar_rgb2luma(scalar_short, input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH);
scalar_face_detect_luma(scalar_short, input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, sdst);
time_stop = vbx_timestamp();
scalar_time = vbx_print_scalar_time(time_start, time_stop);
#if PRINT
print_python_pixel(scalar_input, sdst, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
printf("\nVector");
time_start = vbx_timestamp();
vector_face_detect((pixel *)vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, 0, vdst);
time_stop = vbx_timestamp();
vbx_time = vbx_print_vector_time(time_start, time_stop, scalar_time);
#if PRINT
print_python_pixel(vbx_input, vdst, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
printf("\nVector Masked");
time_start = vbx_timestamp();
vector_face_detect((pixel *)vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, 1, mdst);
time_stop = vbx_timestamp();
vbx_time_masked = vbx_print_vector_time(time_start, time_stop, scalar_time);
#if PRINT
print_python_pixel(vbx_input_masked, mdst, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
/* errors += match_array_pixel(input, vbx_input, "vector", IMAGE_WIDTH, IMAGE_HEIGHT, 0, MAX_PRINT_ERRORS, 0); */
/* errors += match_array_pixel(input, vbx_input_masked, "masked", IMAGE_WIDTH, IMAGE_HEIGHT, 0, MAX_PRINT_ERRORS, 0); */
errors += match_array_pixel(vbx_input, vbx_input_masked, "masked", IMAGE_WIDTH, IMAGE_HEIGHT, 0, MAX_PRINT_ERRORS, 0);
#endif // UNIT
}
VBX_TEST_END(errors);
return errors;
}
int main(void)
{
vbx_timestamp_t time_start, time_stop;
double scalar_time, vbx_time, vbx_time_masked;
int i, j, k, l, m, n;
int errors = 0;
vbx_test_init();
vbx_mxp_print_params();
pixel *input, *scalar_input, *vbx_input, *vbx_input_masked;
uint16_t *scalar_short;
input = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
scalar_input = (pixel *)vbx_remap_cached(input, IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
scalar_short = (uint16_t *)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(uint16_t));
vbx_input = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
vbx_input_masked = (pixel *)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(pixel));
#if UNIT
unsigned short *img, *vbx_img;
unsigned int *iImg, *vbx_iImg;
unsigned int *iiImg, *vbx_iiImg;
img = (unsigned short*)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned short));
vbx_img = (unsigned short*)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned short));
iImg = (unsigned int*)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
vbx_iImg = (unsigned int*)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
iiImg = (unsigned int*)malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
vbx_iiImg = (unsigned int*)vbx_shared_malloc(IMAGE_WIDTH*IMAGE_HEIGHT*sizeof(unsigned int));
#endif//UNIT
printf("Resolution = %dx%d\n", IMAGE_WIDTH, IMAGE_HEIGHT);
printf("Initializing data\n");
vbx_timestamp_start();
for(l = 0; l < 1; l++){
char *src;
char *sdst;
char *vdst;
char *mdst;
if(l == 0){
load_lenna(input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_lenna(vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_lenna(vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT);
printf("\nLenna\n");
src = "lenna";
sdst = "s_lenna";
vdst = "v_lenna";
mdst = "m_lenna";
}else if(l == 1){
load_ms(input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_ms(vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_ms(vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT);
printf("\nMicrosoft\n");
src = "ms";
sdst = "s_ms";
vdst = "v_ms";
mdst = "m_ms";
}else if(l == 2){
load_blank(input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_blank(vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT);
load_blank(vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT);
printf("\nblank\n");
src = "blank";
sdst = "s_blank";
vdst = "v_blank";
mdst = "m_blank";
}
#if UNIT
int window = 20;
int log=0;
while(((window/3)>>log) >= 2) log++;
#if LUT_CI
/* errors += compare_vbx_lut_to_vbx_lut_ci(1024, MAX_PRINT_ERRORS); */
#endif
#if LBP_CI
errors += compare_vbx_lbp_ci_to_scalar_patterns(vbx_img, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
#endif
errors += compare_scalar_rgb2luma_to_vbw_rgb2luma16(img, vbx_img, vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, MAX_PRINT_ERRORS);
/* errors += compare_LBPRestrictedSums_to_test_scalar_sums_byte(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS); */
/* errors += compare_LBPRestrictedSums2_to_test_scalar_sums_half(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS); */
/* errors += compare_ScalarLBPRestrictedSums_to_test_scalar_sums_half(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS); */
/* errors += compare_ScalarLBPRestrictedPatterns_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS); */
/* errors += compare_LBPRestrictedPatterns2_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS); */
errors += compare_LBPRestricted_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
#if 0
/* overflow issues -- using bytes changes lbp pattern */
errors += compare_LBPRestrictedPatterns_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
/* requires SKIP_INTEGRALS 0 */
errors += compare_gen_integrals_to_vector_get_img(img, iImg, iiImg, vbx_img, vbx_iImg, vbx_iiImg, vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
/* currently last values have errors if the scaled images size is not an integer, width * f/ (f+1) */
errors += compare_scalar_BLIP2_to_vector_BLIP(img, vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
/* redundant test, compare to test_scalar_patterns instead */
errors += compare_ScalarLBPRestrictedPatterns_to_SATBinaryPattern(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
errors += compare_SATBinaryPattern_to_test_scalar_patterns(vbx_img, log, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
#endif
errors += compare_LBPPassStage_to_restricted(vbx_img, log, face_lbp[0], window, IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS);
#else // UNIT
#if PRINT
print_python_pixel(scalar_input, src, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
time_start = vbx_timestamp();
scalar_rgb2luma(scalar_short, input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH);
scalar_face_detect_luma(scalar_short, input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, sdst);
time_stop = vbx_timestamp();
scalar_time = vbx_print_scalar_time(time_start, time_stop);
#if PRINT
print_python_pixel(scalar_input, sdst, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
printf("\nVector");
time_start = vbx_timestamp();
vector_face_detect((pixel *)vbx_input, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, 0, vdst);
time_stop = vbx_timestamp();
vbx_time = vbx_print_vector_time(time_start, time_stop, scalar_time);
#if PRINT
print_python_pixel(vbx_input, vdst, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
printf("\nVector Masked");
time_start = vbx_timestamp();
vector_face_detect((pixel *)vbx_input_masked, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_WIDTH, 1, mdst);
time_stop = vbx_timestamp();
vbx_time_masked = vbx_print_vector_time(time_start, time_stop, scalar_time);
#if PRINT
print_python_pixel(vbx_input_masked, mdst, IMAGE_WIDTH, IMAGE_HEIGHT);
#endif
/* errors += match_array_pixel(input, vbx_input, "vector", IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS, 0); */
/* errors += match_array_pixel(input, vbx_input_masked, "masked", IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS, 0); */
errors += match_array_pixel(vbx_input, vbx_input_masked, "masked", IMAGE_WIDTH, IMAGE_HEIGHT, MAX_PRINT_ERRORS, 0);
#endif // UNIT
}
VBX_TEST_END(errors);
return errors;
}