static boolean
sp_vbuf_allocate_vertices(struct vbuf_render *vbr,
ushort vertex_size, ushort nr_vertices)
{
struct softpipe_vbuf_render *cvbr = softpipe_vbuf_render(vbr);
unsigned size = vertex_size * nr_vertices;
if (cvbr->vertex_buffer_size < size) {
align_free(cvbr->vertex_buffer);
cvbr->vertex_buffer = align_malloc(size, 16);
cvbr->vertex_buffer_size = size;
}
cvbr->vertex_size = vertex_size;
cvbr->nr_vertices = nr_vertices;
return cvbr->vertex_buffer != NULL;
}
static boolean
lp_setup_allocate_vertices(struct vbuf_render *vbr,
ushort vertex_size, ushort nr_vertices)
{
struct lp_setup_context *setup = lp_setup_context(vbr);
unsigned size = vertex_size * nr_vertices;
if (setup->vertex_buffer_size < size) {
align_free(setup->vertex_buffer);
setup->vertex_buffer = align_malloc(size, 16);
setup->vertex_buffer_size = size;
}
setup->vertex_size = vertex_size;
setup->nr_vertices = nr_vertices;
return setup->vertex_buffer != NULL;
}
static void
softpipe_resource_destroy(struct pipe_screen *pscreen,
struct pipe_resource *pt)
{
struct softpipe_screen *screen = softpipe_screen(pscreen);
struct softpipe_resource *spr = softpipe_resource(pt);
if (spr->dt) {
/* display target */
struct sw_winsys *winsys = screen->winsys;
winsys->displaytarget_destroy(winsys, spr->dt);
}
else if (!spr->userBuffer) {
/* regular texture */
align_free(spr->data);
}
FREE(spr);
}
/**
* Copy the contents of the malloc buffer to a hardware buffer.
*/
static enum pipe_error
svga_buffer_update_hw(struct svga_screen *ss, struct svga_buffer *sbuf)
{
assert(!sbuf->user);
if (!sbuf->hwbuf) {
enum pipe_error ret;
void *map;
assert(sbuf->swbuf);
if (!sbuf->swbuf)
return PIPE_ERROR;
ret = svga_buffer_create_hw_storage(ss, sbuf);
if (ret != PIPE_OK)
return ret;
pipe_mutex_lock(ss->swc_mutex);
map = ss->sws->buffer_map(ss->sws, sbuf->hwbuf, PIPE_TRANSFER_WRITE);
assert(map);
if (!map) {
pipe_mutex_unlock(ss->swc_mutex);
svga_buffer_destroy_hw_storage(ss, sbuf);
return PIPE_ERROR;
}
memcpy(map, sbuf->swbuf, sbuf->b.b.width0);
ss->sws->buffer_unmap(ss->sws, sbuf->hwbuf);
/* This user/malloc buffer is now indistinguishable from a gpu buffer */
assert(!sbuf->map.count);
if (!sbuf->map.count) {
if (sbuf->user)
sbuf->user = FALSE;
else
align_free(sbuf->swbuf);
sbuf->swbuf = NULL;
}
pipe_mutex_unlock(ss->swc_mutex);
}
return PIPE_OK;
}
/* Shutdown:
*/
void lp_rast_destroy( struct lp_rasterizer *rast )
{
unsigned i;
/* Set exit_flag and signal each thread's work_ready semaphore.
* Each thread will be woken up, notice that the exit_flag is set and
* break out of its main loop. The thread will then exit.
*/
rast->exit_flag = TRUE;
for (i = 0; i < rast->num_threads; i++) {
pipe_semaphore_signal(&rast->tasks[i].work_ready);
}
/* Wait for threads to terminate before cleaning up per-thread data.
* We don't actually call pipe_thread_wait to avoid dead lock on Windows
* per https://bugs.freedesktop.org/show_bug.cgi?id=76252 */
for (i = 0; i < rast->num_threads; i++) {
#ifdef _WIN32
pipe_semaphore_wait(&rast->tasks[i].work_done);
#else
thrd_join(rast->threads[i], NULL);
#endif
}
/* Clean up per-thread data */
for (i = 0; i < rast->num_threads; i++) {
pipe_semaphore_destroy(&rast->tasks[i].work_ready);
pipe_semaphore_destroy(&rast->tasks[i].work_done);
}
for (i = 0; i < MAX2(1, rast->num_threads); i++) {
align_free(rast->tasks[i].thread_data.cache);
}
/* for synchronizing rasterization threads */
if (rast->num_threads > 0) {
pipe_barrier_destroy( &rast->barrier );
}
lp_scene_queue_destroy(rast->full_scenes);
FREE(rast);
}
static void
nouveau_buffer_destroy(struct pipe_screen *pscreen,
struct pipe_resource *presource)
{
struct nv04_resource *res = nv04_resource(presource);
nouveau_buffer_release_gpu_storage(res);
if (res->data && !(res->status & NOUVEAU_BUFFER_STATUS_USER_MEMORY))
align_free(res->data);
nouveau_fence_ref(NULL, &res->fence);
nouveau_fence_ref(NULL, &res->fence_wr);
util_range_destroy(&res->valid_buffer_range);
FREE(res);
NOUVEAU_DRV_STAT(nouveau_screen(pscreen), buf_obj_current_count, -1);
}
static void
NineTexture9_dtor( struct NineTexture9 *This )
{
unsigned l;
DBG("This=%p\n", This);
if (This->surfaces) {
/* The surfaces should have 0 references and be unbound now. */
for (l = 0; l <= This->base.base.info.last_level; ++l)
if (This->surfaces[l])
NineUnknown_Destroy(&This->surfaces[l]->base.base);
FREE(This->surfaces);
}
if (This->managed_buffer)
align_free(This->managed_buffer);
NineBaseTexture9_dtor(&This->base);
}
static void
hook_winsys_displaytarget_destroy(struct sw_winsys* winsys,
struct sw_displaytarget* displayTarget)
{
CALLED();
struct haiku_displaytarget* haikuDisplayTarget
= cast_haiku_displaytarget(displayTarget);
if (!haikuDisplayTarget) {
ERROR("%s: Attempt to destroy non-existant display target!\n",
__func__);
return;
}
if (haikuDisplayTarget->data != NULL)
align_free(haikuDisplayTarget->data);
FREE(haikuDisplayTarget);
}
void
NineBuffer9_dtor( struct NineBuffer9 *This )
{
if (This->maps) {
while (This->nmaps) {
NineBuffer9_Unlock(This);
}
FREE(This->maps);
}
if (This->base.pool == D3DPOOL_MANAGED) {
if (This->managed.data)
align_free(This->managed.data);
if (This->managed.list.prev != NULL && This->managed.list.next != NULL)
list_del(&This->managed.list);
if (This->managed.list2.prev != NULL && This->managed.list2.next != NULL)
list_del(&This->managed.list2);
}
NineResource9_dtor(&This->base);
}
static void llvmpipe_destroy( struct pipe_context *pipe )
{
struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe );
uint i, j;
lp_print_counters();
/* This will also destroy llvmpipe->setup:
*/
if (llvmpipe->draw)
draw_destroy( llvmpipe->draw );
for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) {
pipe_surface_reference(&llvmpipe->framebuffer.cbufs[i], NULL);
}
pipe_surface_reference(&llvmpipe->framebuffer.zsbuf, NULL);
for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
pipe_sampler_view_reference(&llvmpipe->fragment_sampler_views[i], NULL);
}
for (i = 0; i < PIPE_MAX_VERTEX_SAMPLERS; i++) {
pipe_sampler_view_reference(&llvmpipe->vertex_sampler_views[i], NULL);
}
for (i = 0; i < Elements(llvmpipe->constants); i++) {
for (j = 0; j < Elements(llvmpipe->constants[i]); j++) {
pipe_resource_reference(&llvmpipe->constants[i][j], NULL);
}
}
for (i = 0; i < llvmpipe->num_vertex_buffers; i++) {
pipe_resource_reference(&llvmpipe->vertex_buffer[i].buffer, NULL);
}
lp_delete_setup_variants(llvmpipe);
align_free( llvmpipe );
}
void
lp_destroy_tile_cache(struct llvmpipe_tile_cache *tc)
{
struct pipe_screen *screen;
unsigned x, y;
for (y = 0; y < MAX_HEIGHT; y += TILE_SIZE) {
for (x = 0; x < MAX_WIDTH; x += TILE_SIZE) {
struct llvmpipe_cached_tile *tile = &tc->entries[y/TILE_SIZE][x/TILE_SIZE];
if(tile->color)
align_free(tile->color);
}
}
if (tc->transfer) {
screen = tc->transfer->texture->screen;
screen->tex_transfer_destroy(tc->transfer);
}
FREE( tc );
}
static void
svga_buffer_destroy( struct pipe_buffer *buf )
{
struct svga_screen *ss = svga_screen(buf->screen);
struct svga_buffer *sbuf = svga_buffer( buf );
assert(!p_atomic_read(&buf->reference.count));
assert(!sbuf->dma.pending);
if(sbuf->handle)
svga_buffer_destroy_host_surface(ss, sbuf);
if(sbuf->uploaded.buffer)
pipe_buffer_reference(&sbuf->uploaded.buffer, NULL);
if(sbuf->hwbuf)
svga_buffer_destroy_hw_storage(ss, sbuf);
if(sbuf->swbuf && !sbuf->user)
align_free(sbuf->swbuf);
FREE(sbuf);
}
int
main(int32 argc, char *argv[])
{
char sent[16384];
cmd_ln_t *config;
print_appl_info(argv[0]);
cmd_ln_appl_enter(argc, argv, "default.arg", defn);
unlimit();
config = cmd_ln_get();
ctloffset = cmd_ln_int32_r(config, "-ctloffset");
sentfile = cmd_ln_str_r(config, "-insent");
if ((sentfp = fopen(sentfile, "r")) == NULL)
E_FATAL_SYSTEM("Failed to open file %s for reading", sentfile);
/* Note various output directories */
if (cmd_ln_str_r(config, "-s2stsegdir") != NULL)
s2stsegdir = (char *) ckd_salloc(cmd_ln_str_r(config, "-s2stsegdir"));
if (cmd_ln_str_r(config, "-stsegdir") != NULL)
stsegdir = (char *) ckd_salloc(cmd_ln_str_r(config, "-stsegdir"));
if (cmd_ln_str_r(config, "-phsegdir") != NULL)
phsegdir = (char *) ckd_salloc(cmd_ln_str_r(config, "-phsegdir"));
if (cmd_ln_str_r(config, "-phlabdir") != NULL)
phlabdir = (char *) ckd_salloc(cmd_ln_str_r(config, "-phlabdir"));
if (cmd_ln_str_r(config, "-wdsegdir") != NULL)
wdsegdir = (char *) ckd_salloc(cmd_ln_str_r(config, "-wdsegdir"));
/* HACK! Pre-read insent without checking whether ctl could also
be read. In general, this is caused by the fact that we used
multiple files to specify resource in sphinx III. This is easy
to solve but currently I just to remove process_ctl because it
duplicates badly with ctl_process.
The call back function will take care of matching the uttfile
names. We don't need to worry too much about inconsistency.
*/
while (ctloffset > 0) {
if (fgets(sent, sizeof(sent), sentfp) == NULL) {
E_ERROR("EOF(%s)\n", sentfile);
break;
}
--ctloffset;
}
if ((outsentfile = cmd_ln_str_r(config, "-outsent")) != NULL) {
if ((outsentfp = fopen(outsentfile, "w")) == NULL)
E_FATAL_SYSTEM("Failed to open file %s for writing", outsentfile);
}
if ((outctlfile = cmd_ln_str_r(config, "-outctl")) != NULL) {
if ((outctlfp = fopen(outctlfile, "w")) == NULL)
E_FATAL_SYSTEM("Failed top open file %s for writing", outctlfile);
}
if ((cmd_ln_str_r(config, "-s2stsegdir") == NULL) &&
(cmd_ln_str_r(config, "-stsegdir") == NULL) &&
(cmd_ln_str_r(config, "-phlabdir") == NULL) &&
(cmd_ln_str_r(config, "-phsegdir") == NULL) &&
(cmd_ln_str_r(config, "-wdsegdir") == NULL) &&
(cmd_ln_str_r(config, "-outsent") == NULL))
E_FATAL("Missing output file/directory argument(s)\n");
/* Read in input databases */
models_init(config);
if (!feat)
feat = feat_array_alloc(kbcore_fcb(kbc), S3_MAX_FRAMES);
timers[tmr_utt].name = "U";
timers[tmr_gauden].name = "G";
timers[tmr_senone].name = "S";
timers[tmr_align].name = "A";
/* Initialize align module */
align_init(kbc->mdef, kbc->tmat, dict, config, kbc->logmath);
printf("\n");
if (cmd_ln_str_r(config, "-mllr") != NULL) {
if (kbc->mgau)
adapt_set_mllr(adapt_am, kbc->mgau, cmd_ln_str_r(config, "-mllr"), NULL,
kbc->mdef, config);
else if (kbc->ms_mgau)
model_set_mllr(kbc->ms_mgau, cmd_ln_str_r(config, "-mllr"), NULL, kbcore_fcb(kbc),
kbc->mdef, config);
else
E_WARN("Can't use MLLR matrices with .s2semi. yet\n");
}
tot_nfr = 0;
/* process_ctlfile (); */
if (cmd_ln_str_r(config, "-ctl")) {
/* When -ctlfile is speicified, corpus.c will look at -ctl_mllr to get
the corresponding MLLR for the utterance */
ctl_process(cmd_ln_str_r(config, "-ctl"),
NULL,
cmd_ln_str_r(config, "-ctl_mllr"),
cmd_ln_int32_r(config, "-ctloffset"),
cmd_ln_int32_r(config, "-ctlcount"),
utt_align, config);
}
else {
E_FATAL(" -ctl are not specified.\n");
}
if (tot_nfr > 0) {
printf("\n");
printf("TOTAL FRAMES: %8d\n", tot_nfr);
printf("TOTAL CPU TIME: %11.2f sec, %7.2f xRT\n",
tm_utt.t_tot_cpu, tm_utt.t_tot_cpu / (tot_nfr * 0.01));
printf("TOTAL ELAPSED TIME: %11.2f sec, %7.2f xRT\n",
tm_utt.t_tot_elapsed,
tm_utt.t_tot_elapsed / (tot_nfr * 0.01));
}
if (outsentfp)
fclose(outsentfp);
if (outctlfp)
fclose(outctlfp);
if (sentfp)
fclose(sentfp);
ckd_free(s2stsegdir);
ckd_free(stsegdir);
ckd_free(phsegdir);
ckd_free(wdsegdir);
feat_array_free(feat);
align_free();
models_free();
#if (! WIN32)
system("ps aguxwww | grep s3align");
#endif
cmd_ln_free_r(config);
return 0;
}
int main(int argc, char *argv[])
{
int nchann, nsample, niter, filtorder;
int k, opt;
hfilter filt = NULL;
void *buffin, *buffout;
int filein = -1, fileout = -1;
float fc = FC_DEF;
size_t buffinsize, buffoutsize;
uint32_t nchan32, ntotsample32;
int32_t datintype, datouttype;
int retval = 1;
int keepfiles = 0;
// Process command-line options
nchann = NCHANN;
nsample = NSAMPLE;
niter = NITER;
filtorder = FILTORDER;
datouttype = datintype = TYPE_DEF;
while ((opt = getopt(argc, argv, "hc:s:i:o:f:d:p:k:")) != -1) {
switch (opt) {
case 'c':
nchann = atoi(optarg);
break;
case 's':
nsample = atoi(optarg);
break;
case 'i':
niter = atoi(optarg);
break;
case 'o':
filtorder = atoi(optarg);
break;
case 'f':
fc = atof(optarg);
break;
case 'd':
datintype = atoi(optarg);
break;
case 'p':
ptype = atoi(optarg);
break;
case 'k':
keepfiles = atoi(optarg);
break;
case 'h':
default: /* '?' */
fprintf(stderr, "Usage: %s [-c numchannel] [-s numsample] [-i numiteration] "
"[-o filterorder] [-f cutoff_freq] [-d (0 for float/1 for double)]\n",
argv[0]);
exit(EXIT_FAILURE);
}
}
printf("\tfilter order: %i \tnumber of channels: %i \t\tlength of batch: %i\n",filtorder, nchann, nsample);
set_param(ptype);
datouttype = datintype;
if (ptype & RTF_COMPLEX_MASK)
datouttype |= RTF_COMPLEX_MASK;
buffinsize = sizeof_data(datintype) * nchann * nsample;
buffoutsize = sizeof_data(datouttype) * nchann * nsample;
nchan32 = nchann;
ntotsample32 = nsample*niter;
// Allocate buffers
buffin = align_alloc(16, buffinsize);
buffout = align_alloc(16, buffoutsize);
if (!buffin || !buffout) {
fprintf(stderr, "buffer allocation failed\n");
goto out;
}
// Open files for writing
filein = open(infilename, O_WRONLY|O_CREAT, S_IRWXU);
fileout = open(outfilename, O_WRONLY|O_CREAT, S_IRWXU);
if ((filein < 0) || (fileout < 0)) {
fprintf(stderr, "File opening failed\n");
goto out;
}
fprintf(stdout, "datin=%i datout=%i ptype=%i\n", datintype, datouttype, ptype );
// write filter params on fileout
if (write(fileout, &ptype, sizeof(ptype)) == -1 ||
write(fileout, &numlen, sizeof(numlen)) == -1 ||
write(fileout, num, numlen*sizeof_data(ptype)) == -1 ||
write(fileout, &denumlen, sizeof(denumlen)) == -1 ||
write(fileout, denum, denumlen*sizeof_data(ptype)) == -1 ||
write(filein, &datintype, sizeof(datintype)) == -1 ||
write(filein, &nchan32, sizeof(nchan32)) == -1 ||
write(filein, &ntotsample32, sizeof(ntotsample32)) == -1 ||
write(fileout, &datouttype, sizeof(datouttype)) == -1 ||
write(fileout, &nchan32, sizeof(nchan32)) == -1 ||
write(fileout, &ntotsample32, sizeof(ntotsample32)) == -1)
goto out;
// set signals (ramps)
set_signals(nchann, nsample, datintype, buffin);
// create filters
filt = rtf_create_filter(nchann, datintype, numlen, num, denumlen, denum, ptype);
if (!filt) {
fprintf(stderr,"Creation of filter failed\n");
goto out;
}
// Test the type
if (datintype != rtf_get_type(filt, 1)
|| datouttype != rtf_get_type(filt, 0)) {
fprintf(stderr, "Unexpected data type\n"
"expected: in=%i out=%i\n"
"returned: in=%i out=%i\n",
datintype, datouttype,
rtf_get_type(filt, 1),
rtf_get_type(filt, 0));
goto out;
}
// Filter chunks of data and write input and output on files
for (k=0; k<niter; k++) {
rtf_filter(filt, buffin, buffout, nsample);
if ( write(filein, buffin, buffinsize) == -1
|| write(fileout, buffout, buffoutsize) == -1 ) {
fprintf(stderr,"Error while writing file\n");
break;
}
}
retval = 0;
out:
rtf_destroy_filter(filt);
align_free(buffin);
align_free(buffout);
if (filein != -1)
close(filein);
if (fileout != -1)
close(fileout);
#if HAVE_MATLAB
if (retval == 0)
retval = compare_results( (datintype & RTF_PRECISION_MASK) ? 1e-12 : 1e-4);
#endif
if (!keepfiles) {
unlink(infilename);
unlink(outfilename);
}
return retval;
}
PIPE_ALIGN_STACK
static boolean
test_one(unsigned verbose,
FILE *fp,
const struct pipe_blend_state *blend,
struct lp_type type)
{
struct gallivm_state *gallivm;
LLVMValueRef func = NULL;
blend_test_ptr_t blend_test_ptr;
boolean success;
const unsigned n = LP_TEST_NUM_SAMPLES;
int64_t cycles[LP_TEST_NUM_SAMPLES];
double cycles_avg = 0.0;
unsigned i, j;
const unsigned stride = lp_type_width(type)/8;
if(verbose >= 1)
dump_blend_type(stdout, blend, type);
gallivm = gallivm_create();
func = add_blend_test(gallivm, blend, type);
gallivm_compile_module(gallivm);
blend_test_ptr = (blend_test_ptr_t)gallivm_jit_function(gallivm, func);
success = TRUE;
{
uint8_t *src, *dst, *con, *res, *ref;
src = align_malloc(stride, stride);
dst = align_malloc(stride, stride);
con = align_malloc(stride, stride);
res = align_malloc(stride, stride);
ref = align_malloc(stride, stride);
for(i = 0; i < n && success; ++i) {
int64_t start_counter = 0;
int64_t end_counter = 0;
random_vec(type, src);
random_vec(type, dst);
random_vec(type, con);
{
double fsrc[LP_MAX_VECTOR_LENGTH];
double fdst[LP_MAX_VECTOR_LENGTH];
double fcon[LP_MAX_VECTOR_LENGTH];
double fref[LP_MAX_VECTOR_LENGTH];
read_vec(type, src, fsrc);
read_vec(type, dst, fdst);
read_vec(type, con, fcon);
for(j = 0; j < type.length; j += 4)
compute_blend_ref(blend, fsrc + j, fdst + j, fcon + j, fref + j);
write_vec(type, ref, fref);
}
start_counter = rdtsc();
blend_test_ptr(src, dst, con, res);
end_counter = rdtsc();
cycles[i] = end_counter - start_counter;
if(!compare_vec(type, res, ref)) {
success = FALSE;
if(verbose < 1)
dump_blend_type(stderr, blend, type);
fprintf(stderr, "MISMATCH\n");
fprintf(stderr, " Src: ");
dump_vec(stderr, type, src);
fprintf(stderr, "\n");
fprintf(stderr, " Dst: ");
dump_vec(stderr, type, dst);
fprintf(stderr, "\n");
fprintf(stderr, " Con: ");
dump_vec(stderr, type, con);
fprintf(stderr, "\n");
fprintf(stderr, " Res: ");
dump_vec(stderr, type, res);
fprintf(stderr, "\n");
fprintf(stderr, " Ref: ");
dump_vec(stderr, type, ref);
fprintf(stderr, "\n");
}
}
align_free(src);
align_free(dst);
align_free(con);
align_free(res);
align_free(ref);
}
/*
* Unfortunately the output of cycle counter is not very reliable as it comes
* -- sometimes we get outliers (due IRQs perhaps?) which are
* better removed to avoid random or biased data.
*/
{
double sum = 0.0, sum2 = 0.0;
double avg, std;
unsigned m;
for(i = 0; i < n; ++i) {
sum += cycles[i];
sum2 += cycles[i]*cycles[i];
}
avg = sum/n;
std = sqrtf((sum2 - n*avg*avg)/n);
m = 0;
sum = 0.0;
for(i = 0; i < n; ++i) {
if(fabs(cycles[i] - avg) <= 4.0*std) {
sum += cycles[i];
++m;
}
}
cycles_avg = sum/m;
}
if(fp)
write_tsv_row(fp, blend, type, cycles_avg, success);
gallivm_free_function(gallivm, func, blend_test_ptr);
gallivm_destroy(gallivm);
return success;
}
/*
* Test one LLVM unary arithmetic builder function.
*/
static boolean
test_unary(unsigned verbose, FILE *fp, const struct unary_test_t *test)
{
struct gallivm_state *gallivm;
LLVMValueRef test_func;
unary_func_t test_func_jit;
boolean success = TRUE;
int i, j;
int length = lp_native_vector_width / 32;
float *in, *out;
in = align_malloc(length * 4, length * 4);
out = align_malloc(length * 4, length * 4);
/* random NaNs or 0s could wreak havoc */
for (i = 0; i < length; i++) {
in[i] = 1.0;
}
gallivm = gallivm_create();
test_func = build_unary_test_func(gallivm, test);
gallivm_compile_module(gallivm);
test_func_jit = (unary_func_t) gallivm_jit_function(gallivm, test_func);
for (j = 0; j < (test->num_values + length - 1) / length; j++) {
int num_vals = ((j + 1) * length <= test->num_values) ? length :
test->num_values % length;
for (i = 0; i < num_vals; ++i) {
in[i] = test->values[i+j*length];
}
test_func_jit(out, in);
for (i = 0; i < num_vals; ++i) {
float ref = test->ref(in[i]);
double error, precision;
bool pass;
if (util_inf_sign(ref) && util_inf_sign(out[i]) == util_inf_sign(ref)) {
error = 0;
} else {
error = fabs(out[i] - ref);
}
precision = error ? -log2(error/fabs(ref)) : FLT_MANT_DIG;
pass = precision >= test->precision;
if (isnan(ref)) {
continue;
}
if (!pass || verbose) {
printf("%s(%.9g): ref = %.9g, out = %.9g, precision = %f bits, %s\n",
test->name, in[i], ref, out[i], precision,
pass ? "PASS" : "FAIL");
}
if (!pass) {
success = FALSE;
}
}
}
gallivm_free_function(gallivm, test_func, test_func_jit);
gallivm_destroy(gallivm);
align_free(in);
align_free(out);
return success;
}
int main(int argc, char *argv[])
{
int nchann, nsample, niter;
int k, opt, ns;
hfilter filt = NULL;
void *buffin, *buffout;
int filein = -1, fileout = -1;
size_t samsize;
uint32_t nchan32;
uint32_t type;
uint32_t r = 4;
int retval = 1;
int keepfiles = 0;
// Process command-line options
nchann = NCHANN;
nsample = NSAMPLE;
niter = NITER;
type = TYPE_DEF;
while ((opt = getopt(argc, argv, "hc:s:i:t:r:k:")) != -1) {
switch (opt) {
case 'c':
nchann = atoi(optarg);
break;
case 's':
nsample = atoi(optarg);
break;
case 'i':
niter = atoi(optarg);
break;
case 'r':
r = atoi(optarg);
break;
case 't':
type = atoi(optarg);
break;
case 'k':
keepfiles = atoi(optarg);
break;
case 'h':
default: /* '?' */
fprintf(stderr, "Usage: %s [-c numchannel] [-s numsample] [-i numiteration] "
"[-t (0 for float/1 for double)]\n",
argv[0]);
exit(EXIT_FAILURE);
}
}
printf("\tnumber of channels: %i \t\tlength of batch: %i\n", nchann, nsample);
samsize = sizeof_data(type) * nchann;
nchan32 = nchann;
// Allocate buffers
buffin = align_alloc(16, nsample*samsize);
buffout = align_alloc(16, nsample*samsize);
if (!buffin || !buffout) {
fprintf(stderr, "buffer allocation failed\n");
goto out;
}
// Open files for writing
filein = open(infilename, O_WRONLY|O_CREAT, S_IRWXU);
fileout = open(outfilename, O_WRONLY|O_CREAT, S_IRWXU);
if ((filein < 0) || (fileout < 0)) {
fprintf(stderr, "File opening failed\n");
goto out;
}
// write filter params on fileout
if (write(filein, &type, sizeof(type)) == -1 ||
write(filein, &nchan32, sizeof(nchan32)) == -1 ||
write(fileout, &type, sizeof(type)) == -1 ||
write(fileout, &nchan32, sizeof(nchan32)) == -1 ||
write(fileout, &r, sizeof(r)) == -1)
goto out;
// create filters
filt = rtf_create_downsampler(nchann, type, r);
if (!filt) {
fprintf(stderr,"Creation of filter failed\n");
goto out;
}
// Filter chunks of data and write input and output on files
for (k=0; k<niter; k++) {
set_signals(nchann, nsample, type, buffin);
ns = rtf_filter(filt, buffin, buffout, nsample);
if ( write(filein, buffin, nsample*samsize) == -1
|| write(fileout, buffout, ns*samsize) == -1 ) {
fprintf(stderr,"Error while writing file\n");
break;
}
}
retval = 0;
out:
rtf_destroy_filter(filt);
align_free(buffin);
align_free(buffout);
if (filein != -1)
close(filein);
if (fileout != -1)
close(fileout);
#if HAVE_MATLAB
if (retval == 0)
retval = compare_results(0.02);
#endif
if (!keepfiles) {
unlink(infilename);
unlink(outfilename);
}
return retval;
}
static void
malloc_buffer_destroy(struct pb_buffer *buf)
{
align_free(malloc_buffer(buf)->data);
FREE(buf);
}
/* Returns a pointer to a memory area representing a window into the
* resource's data.
*
* This may or may not be the _actual_ memory area of the resource. However
* when calling nouveau_buffer_transfer_unmap, if it wasn't the actual memory
* area, the contents of the returned map are copied over to the resource.
*
* The usage indicates what the caller plans to do with the map:
*
* WRITE means that the user plans to write to it
*
* READ means that the user plans on reading from it
*
* DISCARD_WHOLE_RESOURCE means that the whole resource is going to be
* potentially overwritten, and even if it isn't, the bits that aren't don't
* need to be maintained.
*
* DISCARD_RANGE means that all the data in the specified range is going to
* be overwritten.
*
* The strategy for determining what kind of memory area to return is complex,
* see comments inside of the function.
*/
static void *
nouveau_buffer_transfer_map(struct pipe_context *pipe,
struct pipe_resource *resource,
unsigned level, unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **ptransfer)
{
struct nouveau_context *nv = nouveau_context(pipe);
struct nv04_resource *buf = nv04_resource(resource);
struct nouveau_transfer *tx = MALLOC_STRUCT(nouveau_transfer);
uint8_t *map;
int ret;
if (!tx)
return NULL;
nouveau_buffer_transfer_init(tx, resource, box, usage);
*ptransfer = &tx->base;
if (usage & PIPE_TRANSFER_READ)
NOUVEAU_DRV_STAT(nv->screen, buf_transfers_rd, 1);
if (usage & PIPE_TRANSFER_WRITE)
NOUVEAU_DRV_STAT(nv->screen, buf_transfers_wr, 1);
/* If we are trying to write to an uninitialized range, the user shouldn't
* care what was there before. So we can treat the write as if the target
* range were being discarded. Furthermore, since we know that even if this
* buffer is busy due to GPU activity, because the contents were
* uninitialized, the GPU can't care what was there, and so we can treat
* the write as being unsynchronized.
*/
if ((usage & PIPE_TRANSFER_WRITE) &&
!util_ranges_intersect(&buf->valid_buffer_range, box->x, box->x + box->width))
usage |= PIPE_TRANSFER_DISCARD_RANGE | PIPE_TRANSFER_UNSYNCHRONIZED;
if (usage & PIPE_TRANSFER_PERSISTENT)
usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
if (buf->domain == NOUVEAU_BO_VRAM) {
if (usage & NOUVEAU_TRANSFER_DISCARD) {
/* Set up a staging area for the user to write to. It will be copied
* back into VRAM on unmap. */
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
buf->status &= NOUVEAU_BUFFER_STATUS_REALLOC_MASK;
nouveau_transfer_staging(nv, tx, TRUE);
} else {
if (buf->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
/* The GPU is currently writing to this buffer. Copy its current
* contents to a staging area in the GART. This is necessary since
* not the whole area being mapped is being discarded.
*/
if (buf->data) {
align_free(buf->data);
buf->data = NULL;
}
nouveau_transfer_staging(nv, tx, FALSE);
nouveau_transfer_read(nv, tx);
} else {
/* The buffer is currently idle. Create a staging area for writes,
* and make sure that the cached data is up-to-date. */
if (usage & PIPE_TRANSFER_WRITE)
nouveau_transfer_staging(nv, tx, TRUE);
if (!buf->data)
nouveau_buffer_cache(nv, buf);
}
}
return buf->data ? (buf->data + box->x) : tx->map;
} else
if (unlikely(buf->domain == 0)) {
return buf->data + box->x;
}
/* At this point, buf->domain == GART */
if (nouveau_buffer_should_discard(buf, usage)) {
int ref = buf->base.reference.count - 1;
nouveau_buffer_reallocate(nv->screen, buf, buf->domain);
if (ref > 0) /* any references inside context possible ? */
nv->invalidate_resource_storage(nv, &buf->base, ref);
}
/* Note that nouveau_bo_map ends up doing a nouveau_bo_wait with the
* relevant flags. If buf->mm is set, that means this resource is part of a
* larger slab bo that holds multiple resources. So in that case, don't
* wait on the whole slab and instead use the logic below to return a
* reasonable buffer for that case.
*/
ret = nouveau_bo_map(buf->bo,
buf->mm ? 0 : nouveau_screen_transfer_flags(usage),
nv->client);
if (ret) {
FREE(tx);
return NULL;
}
map = (uint8_t *)buf->bo->map + buf->offset + box->x;
/* using kernel fences only if !buf->mm */
if ((usage & PIPE_TRANSFER_UNSYNCHRONIZED) || !buf->mm)
return map;
/* If the GPU is currently reading/writing this buffer, we shouldn't
* interfere with its progress. So instead we either wait for the GPU to
* complete its operation, or set up a staging area to perform our work in.
*/
if (nouveau_buffer_busy(buf, usage & PIPE_TRANSFER_READ_WRITE)) {
if (unlikely(usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)) {
/* Discarding was not possible, must sync because
* subsequent transfers might use UNSYNCHRONIZED. */
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else
if (usage & PIPE_TRANSFER_DISCARD_RANGE) {
/* The whole range is being discarded, so it doesn't matter what was
* there before. No need to copy anything over. */
nouveau_transfer_staging(nv, tx, TRUE);
map = tx->map;
} else
if (nouveau_buffer_busy(buf, PIPE_TRANSFER_READ)) {
if (usage & PIPE_TRANSFER_DONTBLOCK)
map = NULL;
else
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else {
/* It is expected that the returned buffer be a representation of the
* data in question, so we must copy it over from the buffer. */
nouveau_transfer_staging(nv, tx, TRUE);
if (tx->map)
memcpy(tx->map, map, box->width);
map = tx->map;
}
}
if (!map)
FREE(tx);
return map;
}
void draw_vs_aos_machine_destroy( struct aos_machine *machine )
{
align_free(machine);
}
/* Migrate a linear buffer (vertex, index, constants) USER -> GART -> VRAM. */
boolean
nouveau_buffer_migrate(struct nouveau_context *nv,
struct nv04_resource *buf, const unsigned new_domain)
{
struct nouveau_screen *screen = nv->screen;
struct nouveau_bo *bo;
const unsigned old_domain = buf->domain;
unsigned size = buf->base.width0;
unsigned offset;
int ret;
assert(new_domain != old_domain);
if (new_domain == NOUVEAU_BO_GART && old_domain == 0) {
if (!nouveau_buffer_allocate(screen, buf, new_domain))
return FALSE;
ret = nouveau_bo_map(buf->bo, 0, nv->client);
if (ret)
return ret;
memcpy((uint8_t *)buf->bo->map + buf->offset, buf->data, size);
align_free(buf->data);
} else
if (old_domain != 0 && new_domain != 0) {
struct nouveau_mm_allocation *mm = buf->mm;
if (new_domain == NOUVEAU_BO_VRAM) {
/* keep a system memory copy of our data in case we hit a fallback */
if (!nouveau_buffer_data_fetch(nv, buf, buf->bo, buf->offset, size))
return FALSE;
if (nouveau_mesa_debug)
debug_printf("migrating %u KiB to VRAM\n", size / 1024);
}
offset = buf->offset;
bo = buf->bo;
buf->bo = NULL;
buf->mm = NULL;
nouveau_buffer_allocate(screen, buf, new_domain);
nv->copy_data(nv, buf->bo, buf->offset, new_domain,
bo, offset, old_domain, buf->base.width0);
nouveau_bo_ref(NULL, &bo);
if (mm)
release_allocation(&mm, screen->fence.current);
} else
if (new_domain == NOUVEAU_BO_VRAM && old_domain == 0) {
struct nouveau_transfer tx;
if (!nouveau_buffer_allocate(screen, buf, NOUVEAU_BO_VRAM))
return FALSE;
tx.base.resource = &buf->base;
tx.base.box.x = 0;
tx.base.box.width = buf->base.width0;
tx.bo = NULL;
if (!nouveau_transfer_staging(nv, &tx, FALSE))
return FALSE;
nouveau_transfer_write(nv, &tx, 0, tx.base.box.width);
nouveau_buffer_transfer_del(nv, &tx);
} else
return FALSE;
assert(buf->domain == new_domain);
return TRUE;
}
static void *
nouveau_buffer_transfer_map(struct pipe_context *pipe,
struct pipe_resource *resource,
unsigned level, unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **ptransfer)
{
struct nouveau_context *nv = nouveau_context(pipe);
struct nv04_resource *buf = nv04_resource(resource);
struct nouveau_transfer *tx = MALLOC_STRUCT(nouveau_transfer);
uint8_t *map;
int ret;
if (!tx)
return NULL;
nouveau_buffer_transfer_init(tx, resource, box, usage);
*ptransfer = &tx->base;
if (buf->domain == NOUVEAU_BO_VRAM) {
if (usage & NOUVEAU_TRANSFER_DISCARD) {
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
buf->status &= NOUVEAU_BUFFER_STATUS_REALLOC_MASK;
nouveau_transfer_staging(nv, tx, TRUE);
} else {
if (buf->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
if (buf->data) {
align_free(buf->data);
buf->data = NULL;
}
nouveau_transfer_staging(nv, tx, FALSE);
nouveau_transfer_read(nv, tx);
} else {
if (usage & PIPE_TRANSFER_WRITE)
nouveau_transfer_staging(nv, tx, TRUE);
if (!buf->data)
nouveau_buffer_cache(nv, buf);
}
}
return buf->data ? (buf->data + box->x) : tx->map;
} else
if (unlikely(buf->domain == 0)) {
return buf->data + box->x;
}
if (nouveau_buffer_should_discard(buf, usage)) {
int ref = buf->base.reference.count - 1;
nouveau_buffer_reallocate(nv->screen, buf, buf->domain);
if (ref > 0) /* any references inside context possible ? */
nv->invalidate_resource_storage(nv, &buf->base, ref);
}
ret = nouveau_bo_map(buf->bo,
buf->mm ? 0 : nouveau_screen_transfer_flags(usage),
nv->client);
if (ret) {
FREE(tx);
return NULL;
}
map = (uint8_t *)buf->bo->map + buf->offset + box->x;
/* using kernel fences only if !buf->mm */
if ((usage & PIPE_TRANSFER_UNSYNCHRONIZED) || !buf->mm)
return map;
if (nouveau_buffer_busy(buf, usage & PIPE_TRANSFER_READ_WRITE)) {
if (unlikely(usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)) {
/* Discarding was not possible, must sync because
* subsequent transfers might use UNSYNCHRONIZED. */
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else
if (usage & PIPE_TRANSFER_DISCARD_RANGE) {
nouveau_transfer_staging(nv, tx, TRUE);
map = tx->map;
} else
if (nouveau_buffer_busy(buf, PIPE_TRANSFER_READ)) {
if (usage & PIPE_TRANSFER_DONTBLOCK)
map = NULL;
else
nouveau_buffer_sync(buf, usage & PIPE_TRANSFER_READ_WRITE);
} else {
nouveau_transfer_staging(nv, tx, TRUE);
if (tx->map)
memcpy(tx->map, map, box->width);
map = tx->map;
}
}
if (!map)
FREE(tx);
return map;
}
/**
* Execute geometry shader.
*/
int draw_geometry_shader_run(struct draw_geometry_shader *shader,
const void *constants[PIPE_MAX_CONSTANT_BUFFERS],
const unsigned constants_size[PIPE_MAX_CONSTANT_BUFFERS],
const struct draw_vertex_info *input_verts,
const struct draw_prim_info *input_prim,
const struct tgsi_shader_info *input_info,
struct draw_vertex_info *output_verts,
struct draw_prim_info *output_prims )
{
const float (*input)[4] = (const float (*)[4])input_verts->verts->data;
unsigned input_stride = input_verts->vertex_size;
unsigned num_outputs = draw_total_gs_outputs(shader->draw);
unsigned vertex_size = sizeof(struct vertex_header) + num_outputs * 4 * sizeof(float);
unsigned num_input_verts = input_prim->linear ?
input_verts->count :
input_prim->count;
unsigned num_in_primitives =
align(
MAX2(u_decomposed_prims_for_vertices(input_prim->prim,
num_input_verts),
u_decomposed_prims_for_vertices(shader->input_primitive,
num_input_verts)),
shader->vector_length);
unsigned max_out_prims =
u_decomposed_prims_for_vertices(shader->output_primitive,
shader->max_output_vertices)
* num_in_primitives;
//Assume at least one primitive
max_out_prims = MAX2(max_out_prims, 1);
output_verts->vertex_size = vertex_size;
output_verts->stride = output_verts->vertex_size;
/* we allocate exactly one extra vertex per primitive to allow the GS to emit
* overflown vertices into some area where they won't harm anyone */
output_verts->verts =
(struct vertex_header *)MALLOC(output_verts->vertex_size *
max_out_prims *
shader->primitive_boundary);
#if 0
debug_printf("%s count = %d (in prims # = %d)\n",
__FUNCTION__, num_input_verts, num_in_primitives);
debug_printf("\tlinear = %d, prim_info->count = %d\n",
input_prim->linear, input_prim->count);
debug_printf("\tprim pipe = %s, shader in = %s, shader out = %s, max out = %d\n",
u_prim_name(input_prim->prim),
u_prim_name(shader->input_primitive),
u_prim_name(shader->output_primitive),
shader->max_output_vertices);
#endif
shader->emitted_vertices = 0;
shader->emitted_primitives = 0;
shader->vertex_size = vertex_size;
shader->tmp_output = (float (*)[4])output_verts->verts->data;
shader->fetched_prim_count = 0;
shader->input_vertex_stride = input_stride;
shader->input = input;
shader->input_info = input_info;
FREE(shader->primitive_lengths);
shader->primitive_lengths = MALLOC(max_out_prims * sizeof(unsigned));
#ifdef HAVE_LLVM
if (draw_get_option_use_llvm()) {
shader->gs_output = output_verts->verts;
if (max_out_prims > shader->max_out_prims) {
unsigned i;
if (shader->llvm_prim_lengths) {
for (i = 0; i < shader->max_out_prims; ++i) {
align_free(shader->llvm_prim_lengths[i]);
}
FREE(shader->llvm_prim_lengths);
}
shader->llvm_prim_lengths = MALLOC(max_out_prims * sizeof(unsigned*));
for (i = 0; i < max_out_prims; ++i) {
int vector_size = shader->vector_length * sizeof(unsigned);
shader->llvm_prim_lengths[i] =
align_malloc(vector_size, vector_size);
}
shader->max_out_prims = max_out_prims;
}
shader->jit_context->prim_lengths = shader->llvm_prim_lengths;
shader->jit_context->emitted_vertices = shader->llvm_emitted_vertices;
shader->jit_context->emitted_prims = shader->llvm_emitted_primitives;
}
#endif
shader->prepare(shader, constants, constants_size);
if (input_prim->linear)
gs_run(shader, input_prim, input_verts,
output_prims, output_verts);
else
gs_run_elts(shader, input_prim, input_verts,
output_prims, output_verts);
/* Flush the remaining primitives. Will happen if
* num_input_primitives % 4 != 0
*/
if (shader->fetched_prim_count > 0) {
gs_flush(shader);
}
debug_assert(shader->fetched_prim_count == 0);
/* Update prim_info:
*/
output_prims->linear = TRUE;
output_prims->elts = NULL;
output_prims->start = 0;
output_prims->count = shader->emitted_vertices;
output_prims->prim = shader->output_primitive;
output_prims->flags = 0x0;
output_prims->primitive_lengths = shader->primitive_lengths;
output_prims->primitive_count = shader->emitted_primitives;
output_verts->count = shader->emitted_vertices;
if (shader->draw->collect_statistics) {
unsigned i;
for (i = 0; i < shader->emitted_primitives; ++i) {
shader->draw->statistics.gs_primitives +=
u_decomposed_prims_for_vertices(shader->output_primitive,
shader->primitive_lengths[i]);
}
}
#if 0
debug_printf("GS finished, prims = %d, verts = %d\n",
output_prims->primitive_count,
output_verts->count);
#endif
return shader->emitted_vertices;
}
void llvmpipe_setup_destroy_context( struct setup_context *setup )
{
align_free( setup );
}
