static void
test_simple (int max, int (*adder) (OrcProgram *, int, const char *))
{
OrcProgram *p;
int v;
OrcCompileResult result;
p = orc_program_new ();
/* dummy program so compile doesn't barf */
orc_program_add_destination (p, 2, "d1");
orc_program_add_source (p, 2, "s1");
orc_program_append_str (p, "addw", "d1", "d1", "s1");
/* we've alreay added one of those */
if (adder == orc_program_add_destination || adder == orc_program_add_source)
max--;
/* Check we can add up to the claimed max */
for (v = 0; v < max; v++)
(*adder) (p, 2, names + v);
result = orc_program_compile (p);
if (ORC_COMPILE_RESULT_IS_FATAL (result))
error = TRUE;
orc_program_reset (p);
/* Check we can not add one more */
(*adder) (p, 2, names + v);
result = orc_program_compile (p);
if (ORC_COMPILE_RESULT_IS_SUCCESSFUL (result))
error = TRUE;
orc_program_free (p);
}
void
orc_memcpy_u32 (guint32 * d1, const guint32 * s1, int n)
{
OrcExecutor _ex, *ex = &_ex;
static int p_inited = 0;
static OrcProgram *p = 0;
void (*func) (OrcExecutor *);
if (!p_inited) {
orc_once_mutex_lock ();
if (!p_inited) {
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_name (p, "orc_memcpy_u32");
orc_program_set_backup_function (p, _backup_orc_memcpy_u32);
orc_program_add_destination (p, 4, "d1");
orc_program_add_source (p, 4, "s1");
orc_program_append (p, "copyl", ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1);
result = orc_program_compile (p);
}
p_inited = TRUE;
orc_once_mutex_unlock ();
}
ex->program = p;
ex->n = n;
ex->arrays[ORC_VAR_D1] = d1;
ex->arrays[ORC_VAR_S1] = (void *) s1;
func = p->code_exec;
func (ex);
}
OrcProgram *
orc_program_new_from_static_bytecode (const orc_uint8 *bytecode)
{
OrcProgram *p;
p = orc_program_new ();
orc_bytecode_parse_function (p, bytecode);
return p;
}
OrcProgram *
orc_test_get_program_for_opcode (OrcStaticOpcode *opcode)
{
OrcProgram *p;
char s[40];
int flags = 0;
int args[4] = { -1, -1, -1, -1 };
int n_args = 0;
p = orc_program_new ();
if (opcode->flags & ORC_STATIC_OPCODE_ACCUMULATOR) {
args[n_args++] =
orc_program_add_accumulator (p, opcode->dest_size[0], "d1");
} else {
args[n_args++] =
orc_program_add_destination (p, opcode->dest_size[0], "d1");
}
if (opcode->dest_size[1] != 0) {
args[n_args++] =
orc_program_add_destination (p, opcode->dest_size[1], "d2");
}
if (opcode->flags & ORC_STATIC_OPCODE_SCALAR) {
if (opcode->src_size[1] == 0) {
args[n_args++] =
orc_program_add_constant (p, opcode->src_size[0], 1, "c1");
} else {
args[n_args++] =
orc_program_add_source (p, opcode->src_size[0], "s1");
args[n_args++] =
orc_program_add_constant (p, opcode->src_size[1], 1, "c1");
if (opcode->src_size[2] != 0) {
args[n_args++] =
orc_program_add_constant (p, opcode->src_size[1], 1, "c1");
}
}
} else {
args[n_args++] =
orc_program_add_source (p, opcode->src_size[0], "s1");
args[n_args++] =
orc_program_add_source (p, opcode->src_size[1], "s2");
}
if ((opcode->flags & ORC_STATIC_OPCODE_FLOAT_SRC) ||
(opcode->flags & ORC_STATIC_OPCODE_FLOAT_DEST)) {
flags = ORC_TEST_FLAGS_FLOAT;
}
sprintf(s, "test_%s", opcode->name);
orc_program_set_name (p, s);
orc_program_append_2 (p, opcode->name, 0, args[0], args[1],
args[2], args[3]);
return p;
}
void
test_opcode_const (OrcStaticOpcode *opcode)
{
OrcProgram *p;
char s[40];
int ret;
int flags = 0;
int args[4] = { -1, -1, -1, -1 };
int n_args = 0;
if (opcode->src_size[1] == 0) {
return;
}
p = orc_program_new ();
if (opcode->flags & ORC_STATIC_OPCODE_ACCUMULATOR) {
args[n_args++] =
orc_program_add_accumulator (p, opcode->dest_size[0], "d1");
} else {
args[n_args++] =
orc_program_add_destination (p, opcode->dest_size[0], "d1");
}
if (opcode->dest_size[1] != 0) {
args[n_args++] =
orc_program_add_destination (p, opcode->dest_size[1], "d2");
}
args[n_args++] =
orc_program_add_source (p, opcode->src_size[0], "s1");
args[n_args++] =
orc_program_add_constant (p, opcode->src_size[1], 1, "c1");
if (opcode->src_size[2]) {
args[n_args++] =
orc_program_add_constant (p, opcode->src_size[2], 1, "c2");
}
if ((opcode->flags & ORC_STATIC_OPCODE_FLOAT_SRC) ||
(opcode->flags & ORC_STATIC_OPCODE_FLOAT_DEST)) {
flags = ORC_TEST_FLAGS_FLOAT;
}
sprintf(s, "test_const_%s", opcode->name);
orc_program_set_name (p, s);
orc_program_append_2 (p, opcode->name, 0, args[0], args[1],
args[2], args[3]);
ret = orc_test_compare_output_full (p, flags);
if (!ret) {
printf("FAIL: %s const\n", opcode->name);
error = TRUE;
}
orc_program_free (p);
}
/**
* orc_program_new_ds:
* @size1: size of destination array members
* @size2: size of source array members
*
* Create a new OrcProgram, with a destination named "d1" and
* one source named "s1".
*
* Returns: a pointer to an OrcProgram structure
*/
OrcProgram *
orc_program_new_ds (int size1, int size2)
{
OrcProgram *p;
p = orc_program_new ();
orc_program_add_destination (p, size1, "d1");
orc_program_add_source (p, size2, "s1");
return p;
}
/**
* orc_program_new_as:
* @size1: size of destination array members
* @size2: size of source array members
*
* Create a new OrcProgram, with an accumulator named "a1" and
* one source named "s1".
*
* Returns: a pointer to an OrcProgram structure
*/
OrcProgram *
orc_program_new_as (int size1, int size2)
{
OrcProgram *p;
p = orc_program_new ();
orc_program_add_accumulator (p, size1, "a1");
orc_program_add_source (p, size2, "s1");
return p;
}
void
orc_blend_u8 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, int n, int m)
{
OrcExecutor _ex, *ex = &_ex;
static int p_inited = 0;
static OrcProgram *p = 0;
void (*func) (OrcExecutor *);
if (!p_inited) {
orc_once_mutex_lock ();
if (!p_inited) {
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_2d (p);
orc_program_set_name (p, "orc_blend_u8");
orc_program_set_backup_function (p, _backup_orc_blend_u8);
orc_program_add_destination (p, 1, "d1");
orc_program_add_source (p, 1, "s1");
orc_program_add_constant (p, 1, 8, "c1");
orc_program_add_parameter (p, 2, "p1");
orc_program_add_temporary (p, 2, "t1");
orc_program_add_temporary (p, 2, "t2");
orc_program_append (p, "convubw", ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_D1);
orc_program_append (p, "convubw", ORC_VAR_T2, ORC_VAR_S1, ORC_VAR_D1);
orc_program_append (p, "subw", ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T1);
orc_program_append (p, "mullw", ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P1);
orc_program_append (p, "shlw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1);
orc_program_append (p, "addw", ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_T2);
orc_program_append (p, "shruw", ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C1);
orc_program_append (p, "convsuswb", ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_D1);
result = orc_program_compile (p);
}
p_inited = TRUE;
orc_once_mutex_unlock ();
}
ex->program = p;
ex->n = n;
ORC_EXECUTOR_M (ex) = m;
ex->arrays[ORC_VAR_D1] = d1;
ex->params[ORC_VAR_D1] = d1_stride;
ex->arrays[ORC_VAR_S1] = (void *) s1;
ex->params[ORC_VAR_S1] = s1_stride;
ex->params[ORC_VAR_P1] = p1;
func = p->code_exec;
func (ex);
}
void
test_opcode_src_const_n_2d (OrcStaticOpcode *opcode)
{
OrcProgram *p;
char s[40];
int ret;
int flags = 0;
if (opcode->flags & ORC_STATIC_OPCODE_SCALAR) {
return;
}
p = orc_program_new ();
if (opcode->flags & ORC_STATIC_OPCODE_ACCUMULATOR) {
orc_program_add_accumulator (p, opcode->dest_size[0], "d1");
} else {
orc_program_add_destination (p, opcode->dest_size[0], "d1");
}
if (opcode->dest_size[1] != 0) {
orc_program_add_destination (p, opcode->dest_size[1], "d2");
}
orc_program_add_source (p, opcode->src_size[0], "s1");
if (opcode->src_size[1] != 0) {
orc_program_add_source (p, opcode->src_size[1], "s2");
}
if ((opcode->flags & ORC_STATIC_OPCODE_FLOAT_SRC) ||
(opcode->flags & ORC_STATIC_OPCODE_FLOAT_DEST)) {
flags = ORC_TEST_FLAGS_FLOAT;
}
sprintf(s, "test_s_%s", opcode->name);
orc_program_set_name (p, s);
orc_program_set_2d (p);
orc_program_set_constant_n (p, 8);
if (opcode->dest_size[1] != 0) {
orc_program_append_dds_str (p, opcode->name, "d1", "d2", "s1");
} else {
orc_program_append_str (p, opcode->name, "d1", "s1", "s2");
}
ret = orc_test_compare_output_full (p, flags);
if (!ret) {
printf("FAIL: %s src_const_n_2d\n", opcode->name);
error = TRUE;
}
orc_program_free (p);
}
VipsVector *
vips_vector_new( const char *name, int dsize )
{
static int vector_number = 0;
VipsVector *vector;
int i;
if( !(vector = VIPS_NEW( NULL, VipsVector )) )
return( NULL );
vector->name = name;
vector->unique_name = g_strdup_printf( "p[%d]", vector_number++ );
vector->n_temp = 0;
vector->n_scanline = 0;
vector->n_source = 0;
vector->n_destination = 0;
vector->n_constant = 0;
vector->n_parameter = 0;
vector->n_instruction = 0;
for( i = 0; i < VIPS_VECTOR_SOURCE_MAX; i++ ) {
vector->s[i] = -1;
vector->sl[i] = -1;
}
vector->d1 = -1;
vector->compiled = FALSE;
#ifdef HAVE_ORC
vector->program = orc_program_new();
#ifdef DEBUG_TRACE
printf( "%s = orc_program_new();\n", vector->unique_name );
#endif /*DEBUG_TRACE*/
#endif /*HAVE_ORC*/
/* We always make d1, our callers make either a single point source, or
* for area ops, a set of scanlines.
*
* Don't check error return. orc uses 0 to mean error, but the first
* var you create will have id 0 :-( The first var is unlikely to fail
* anyway.
*/
vector->d1 = vips_vector_destination( vector, "d1", dsize );
return( vector );
}
VipsVector *
vips_vector_new( const char *name, int dsize )
{
VipsVector *vector;
int i;
if( !(vector = VIPS_NEW( NULL, VipsVector )) )
return( NULL );
vector->name = name;
vector->n_temp = 0;
vector->n_scanline = 0;
vector->n_source = 0;
vector->n_destination = 0;
vector->n_constant = 0;
vector->n_parameter = 0;
vector->n_instruction = 0;
for( i = 0; i < VIPS_VECTOR_SOURCE_MAX; i++ ) {
vector->s[i] = -1;
vector->sl[i] = -1;
}
vector->d1 = -1;
vector->compiled = FALSE;
#ifdef HAVE_ORC
vector->program = orc_program_new();
/* We always make d1, our callers make either a single point source, or
* for area ops, a set of scanlines.
*/
vector->d1 = orc_program_add_destination( vector->program,
dsize, "d1" );
vector->n_destination += 1;
#endif /*HAVE_ORC*/
return( vector );
}
int
orc_parse_full (const char *code, OrcProgram ***programs, char **log)
{
OrcParser _parser;
OrcParser *parser = &_parser;
char *init_function = NULL;
memset (parser, 0, sizeof(*parser));
parser->code = code;
parser->code_length = strlen (code);
parser->line_number = 0;
parser->p = code;
parser->opcode_set = orc_opcode_set_get ("sys");
parser->log = malloc(100);
parser->log_alloc = 100;
parser->log_size = 0;
while (parser->p[0] != 0) {
char *p;
char *end;
char *token[10];
int n_tokens;
orc_parse_get_line (parser);
p = parser->line;
end = p + strlen (p);
//printf("%d: %s\n", parser->line_number, parser->line);
while (p[0] == ' ' || p[0] == '\t') p++;
if (p[0] == 0) {
continue;
}
if (p[0] == '#') {
//printf("comment: %s\n", p+1);
continue;
}
n_tokens = 0;
while (p < end) {
if (p[0] == ' ' || p[0] == '\t' || p[0] == ',') p++;
if (p[0] == 0 || p[0] == '#') break;
token[n_tokens] = p;
while (p[0] != 0 && p[0] != ' ' && p[0] != '\t' && p[0] != ',') p++;
n_tokens++;
p[0] = 0;
p++;
}
if (n_tokens == 0) {
continue;
}
{
int i;
for(i=0;i<n_tokens;i++){
//printf("'%s' ", token[i]);
}
//printf("\n");
}
if (token[0][0] == '.') {
if (strcmp (token[0], ".function") == 0) {
if (parser->program) {
orc_parse_sanity_check (parser, parser->program);
}
parser->program = orc_program_new ();
orc_program_set_name (parser->program, token[1]);
if (parser->n_programs == parser->n_programs_alloc) {
parser->n_programs_alloc += 32;
parser->programs = realloc (parser->programs,
sizeof(OrcProgram *)*parser->n_programs_alloc);
}
parser->programs[parser->n_programs] = parser->program;
parser->n_programs++;
parser->creg_index = 1;
} else if (strcmp (token[0], ".init") == 0) {
free (init_function);
init_function = NULL;
if (n_tokens < 2) {
orc_parse_log (parser, "error: line %d: .init without function name\n",
parser->line_number);
} else {
init_function = strdup (token[1]);
}
} else if (strcmp (token[0], ".flags") == 0) {
int i;
for(i=1;i<n_tokens;i++){
if (!strcmp (token[i], "2d")) {
orc_program_set_2d (parser->program);
}
}
} else if (strcmp (token[0], ".n") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_set_constant_n (parser->program, size);
} else if (strcmp (token[0], ".m") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_set_constant_m (parser->program, size);
} else if (strcmp (token[0], ".source") == 0) {
int size = strtol (token[1], NULL, 0);
int var;
var = orc_program_add_source (parser->program, size, token[2]);
if (n_tokens > 3) {
orc_program_set_type_name (parser->program, var, token[3]);
}
} else if (strcmp (token[0], ".dest") == 0) {
int size = strtol (token[1], NULL, 0);
int var;
var = orc_program_add_destination (parser->program, size, token[2]);
if (n_tokens > 3) {
orc_program_set_type_name (parser->program, var, token[3]);
}
} else if (strcmp (token[0], ".accumulator") == 0) {
int size = strtol (token[1], NULL, 0);
int var;
var = orc_program_add_accumulator (parser->program, size, token[2]);
if (n_tokens > 3) {
orc_program_set_type_name (parser->program, var, token[3]);
}
} else if (strcmp (token[0], ".temp") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_temporary (parser->program, size, token[2]);
} else if (strcmp (token[0], ".param") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter (parser->program, size, token[2]);
} else if (strcmp (token[0], ".longparam") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter_int64 (parser->program, size, token[2]);
} else if (strcmp (token[0], ".const") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_constant_str (parser->program, size, token[3], token[2]);
} else if (strcmp (token[0], ".floatparam") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter_float (parser->program, size, token[2]);
} else if (strcmp (token[0], ".doubleparam") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter_double (parser->program, size, token[2]);
} else {
orc_parse_log (parser, "error: line %d: unknown directive: %s\n",
parser->line_number, token[0]);
}
} else {
OrcStaticOpcode *o;
unsigned int flags = 0;
int offset = 0;
if (strcmp (token[0], "x4") == 0) {
flags |= ORC_INSTRUCTION_FLAG_X4;
offset = 1;
} else if (strcmp (token[0], "x2") == 0) {
flags |= ORC_INSTRUCTION_FLAG_X2;
offset = 1;
}
o = get_opcode (parser, token[offset]);
if (o) {
int n_args = opcode_n_args (o);
int i;
if (n_tokens != 1 + offset + n_args) {
orc_parse_log (parser, "error: line %d: too %s arguments for %s (expected %d)\n",
parser->line_number, (n_tokens < 1+offset+n_args) ? "few" : "many",
token[offset], n_args);
}
for(i=offset+1;i<n_tokens;i++){
char *end;
double d;
d = strtod (token[i], &end);
if (end != token[i]) {
orc_program_add_constant_str (parser->program, 0, token[i],
token[i]);
}
}
if (n_tokens - offset == 5) {
orc_program_append_str_2 (parser->program, token[offset], flags,
token[offset+1], token[offset+2], token[offset+3], token[offset+4]);
} else if (n_tokens - offset == 4) {
orc_program_append_str_2 (parser->program, token[offset], flags,
token[offset+1], token[offset+2], token[offset+3], NULL);
} else {
orc_program_append_str_2 (parser->program, token[offset], flags,
token[offset+1], token[offset+2], NULL, NULL);
}
} else {
orc_parse_log (parser, "error: line %d: unknown opcode: %s\n",
parser->line_number,
token[offset]);
}
}
}
if (parser->program) {
orc_parse_sanity_check (parser, parser->program);
}
if (parser->line) free (parser->line);
if (log) {
*log = parser->log;
} else {
free (parser->log);
}
if (parser->programs[0]) {
parser->programs[0]->init_function = init_function;
} else {
free (init_function);
}
*programs = parser->programs;
return parser->n_programs;
}
int
main(int argc, char *argv[])
{
char *s, *d;
orc_uint8 *src, *dest;
OrcProfile prof;
OrcProfile prof_libc;
double ave, std;
double ave_libc, std_libc;
double null;
int i,j;
double cpufreq;
int unalign;
OrcProgram *p;
int level1, level2, level3;
int max;
/* const uint8_t zero = 0; */
orc_init ();
/* cpufreq = 2333e6; */
cpufreq = 1;
if (argc > 1) {
unalign = strtoul (argv[1], NULL, 0);
} else {
unalign = 0;
}
s = malloc(1024*1024*64+1024);
d = malloc(1024*1024*64+1024);
src = ORC_PTR_OFFSET(ALIGN(s,128),unalign);
dest = ALIGN(d,128);
orc_profile_init (&prof);
for(j=0;j<10;j++){
orc_profile_start(&prof);
orc_profile_stop(&prof);
}
orc_profile_get_ave_std (&prof, &null, &std);
{
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_name (p, "orc_memcpy");
/* orc_program_set_name (p, "orc_memset"); */
orc_program_add_destination (p, 1, "d1");
orc_program_add_source (p, 1, "s1");
/* orc_program_add_parameter (p, 1, "p1"); */
orc_program_append (p, "copyb", ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1);
result = orc_program_compile (p);
if (ORC_COMPILE_RESULT_IS_FATAL (result)) {
fprintf (stderr, "Failed to compile orc_memcpy\n");
return -1;
}
}
#ifndef M_LN2
#define M_LN2 0.69314718055994530942
#endif
orc_get_data_cache_sizes (&level1, &level2, &level3);
if (level3 > 0) {
max = (log(level3)/M_LN2 - 6.0) * 10 + 20;
} else if (level2 > 0) {
max = (log(level2)/M_LN2 - 6.0) * 10 + 20;
} else {
max = 140;
}
for(i=0;i<max;i++){
double x = i*0.1 + 6.0;
int size = pow(2.0, x);
if (flush_cache) {
touch (src, (1<<18));
}
if (hot_src) {
touch (src, size);
}
if (hot_dest) {
touch (dest, size);
}
orc_profile_init (&prof);
for(j=0;j<10;j++){
OrcExecutor _ex, *ex = &_ex;
void (*func) (OrcExecutor *);
orc_profile_start(&prof);
/* orc_memcpy (dest, src, size); */
ex->program = p;
ex->n = size;
ex->arrays[ORC_VAR_D1] = dest;
ex->arrays[ORC_VAR_S1] = (void *)src;
func = p->code_exec;
func (ex);
orc_profile_stop(&prof);
if (flush_cache) {
touch (src, (1<<18));
}
if (hot_src) {
touch (src, size);
}
if (hot_dest) {
touch (dest, size);
}
}
orc_profile_init (&prof_libc);
for(j=0;j<10;j++){
orc_profile_start(&prof_libc);
memcpy (dest, src, size);
orc_profile_stop(&prof_libc);
if (flush_cache) {
touch (src, (1<<18));
}
if (hot_src) {
touch (src, size);
}
if (hot_dest) {
touch (dest, size);
}
}
orc_profile_get_ave_std (&prof, &ave, &std);
orc_profile_get_ave_std (&prof_libc, &ave_libc, &std_libc);
ave -= null;
ave_libc -= null;
/* printf("%d: %10.4g %10.4g %10.4g %10.4g (libc %10.4g)\n", i, ave, std, */
/* ave/(1<<i), cpufreq/(ave/(1<<i)), */
/* cpufreq/(ave_libc/(1<<i))); */
printf("%g %10.4g %10.4g\n", x,
cpufreq/(ave/size), cpufreq/(ave_libc/size));
/* printf("%g %10.4g %10.4g\n", x, */
/* 32*(ave/(size)), 32*(ave_libc/(size))); */
fflush (stdout);
}
orc_program_free (p);
free (s);
free (d);
return 0;
}
int
orc_parse_full (const char *code, OrcProgram ***programs, char **log)
{
OrcParser _parser;
OrcParser *parser = &_parser;
char *init_function = NULL;
memset (parser, 0, sizeof(*parser));
parser->code = code;
parser->code_length = strlen (code);
parser->line_number = 0;
parser->p = code;
parser->opcode_set = orc_opcode_set_get ("sys");
parser->log = malloc(100);
parser->log_alloc = 100;
parser->log_size = 0;
parser->log[0] = 0;
while (parser->p[0] != 0) {
char *p;
char *end;
char *token[10];
int n_tokens;
orc_parse_get_line (parser);
if (parser->program) orc_program_set_line (parser->program, parser->line_number);
p = parser->line;
end = p + strlen (p);
/* printf("%d: %s\n", parser->line_number, parser->line); */
while (p[0] == ' ' || p[0] == '\t') p++;
if (p[0] == 0) {
continue;
}
if (p[0] == '#') {
/* printf("comment: %s\n", p+1); */
continue;
}
n_tokens = 0;
while (p < end) {
while (p[0] != 0 && (p[0] == ' ' || p[0] == '\t')) p++;
if (p[0] == 0 || p[0] == '#') break;
token[n_tokens] = p;
while (p[0] != 0 && p[0] != ' ' && p[0] != '\t' && p[0] != ',') p++;
n_tokens++;
p[0] = 0;
p++;
}
if (n_tokens == 0) {
continue;
}
{
int i;
for(i=0;i<n_tokens;i++){
/* printf("'%s' ", token[i]); */
}
/* printf("\n"); */
}
if (token[0][0] == '.') {
if (strcmp (token[0], ".function") == 0) {
if (parser->program) {
orc_parse_sanity_check (parser, parser->program);
}
parser->program = orc_program_new ();
orc_program_set_name (parser->program, token[1]);
if (parser->n_programs == parser->n_programs_alloc) {
parser->n_programs_alloc += 32;
parser->programs = realloc (parser->programs,
sizeof(OrcProgram *)*parser->n_programs_alloc);
}
parser->programs[parser->n_programs] = parser->program;
parser->n_programs++;
parser->creg_index = 1;
} else if (strcmp (token[0], ".backup") == 0) {
if (n_tokens < 2) {
orc_parse_log (parser, "error: line %d: .backup without function name\n",
parser->line_number);
} else {
orc_program_set_backup_name (parser->program, token[1]);
}
} else if (strcmp (token[0], ".init") == 0) {
free (init_function);
init_function = NULL;
if (n_tokens < 2) {
orc_parse_log (parser, "error: line %d: .init without function name\n",
parser->line_number);
} else {
init_function = strdup (token[1]);
}
} else if (strcmp (token[0], ".flags") == 0) {
int i;
for(i=1;i<n_tokens;i++){
if (!strcmp (token[i], "2d")) {
orc_program_set_2d (parser->program);
}
}
} else if (strcmp (token[0], ".n") == 0) {
int i;
for(i=1;i<n_tokens;i++){
if (strcmp (token[i], "mult") == 0) {
if (i == n_tokens - 1) {
orc_parse_log (parser, "error: line %d: .n mult requires multiple value\n",
parser->line_number);
} else {
orc_program_set_n_multiple (parser->program,
strtol (token[1], NULL, 0));
i++;
}
} else if (strcmp (token[i], "min") == 0) {
if (i == n_tokens - 1) {
orc_parse_log (parser, "error: line %d: .n min requires multiple value\n",
parser->line_number);
} else {
orc_program_set_n_minimum (parser->program,
strtol (token[1], NULL, 0));
i++;
}
} else if (strcmp (token[i], "max") == 0) {
if (i == n_tokens - 1) {
orc_parse_log (parser, "error: line %d: .n max requires multiple value\n",
parser->line_number);
} else {
orc_program_set_n_maximum (parser->program,
strtol (token[1], NULL, 0));
i++;
}
} else if (i == n_tokens - 1) {
orc_program_set_constant_n (parser->program,
strtol (token[1], NULL, 0));
} else {
orc_parse_log (parser, "error: line %d: unknown .n token '%s'\n",
parser->line_number, token[i]);
}
}
} else if (strcmp (token[0], ".m") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_set_constant_m (parser->program, size);
} else if (strcmp (token[0], ".source") == 0) {
int size = strtol (token[1], NULL, 0);
int var;
int i;
var = orc_program_add_source (parser->program, size, token[2]);
for(i=3;i<n_tokens;i++){
if (strcmp (token[i], "align") == 0) {
if (i == n_tokens - 1) {
orc_parse_log (parser, "error: line %d: .source align requires alignment value\n",
parser->line_number);
} else {
int alignment = strtol (token[i+1], NULL, 0);
orc_program_set_var_alignment (parser->program, var, alignment);
i++;
}
} else if (i == n_tokens - 1) {
orc_program_set_type_name (parser->program, var, token[i]);
} else {
orc_parse_log (parser, "error: line %d: unknown .dest token '%s'\n",
parser->line_number, token[i]);
}
}
} else if (strcmp (token[0], ".dest") == 0) {
int size = strtol (token[1], NULL, 0);
int var;
int i;
var = orc_program_add_destination (parser->program, size, token[2]);
for(i=3;i<n_tokens;i++){
if (strcmp (token[i], "align") == 0) {
if (i == n_tokens - 1) {
orc_parse_log (parser, "error: line %d: .source align requires alignment value\n",
parser->line_number);
} else {
int alignment = strtol (token[i+1], NULL, 0);
orc_program_set_var_alignment (parser->program, var, alignment);
i++;
}
} else if (i == n_tokens - 1) {
orc_program_set_type_name (parser->program, var, token[i]);
} else {
orc_parse_log (parser, "error: line %d: unknown .source token '%s'\n",
parser->line_number, token[i]);
}
}
} else if (strcmp (token[0], ".accumulator") == 0) {
int size = strtol (token[1], NULL, 0);
int var;
var = orc_program_add_accumulator (parser->program, size, token[2]);
if (n_tokens > 3) {
orc_program_set_type_name (parser->program, var, token[3]);
}
} else if (strcmp (token[0], ".temp") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_temporary (parser->program, size, token[2]);
} else if (strcmp (token[0], ".param") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter (parser->program, size, token[2]);
} else if (strcmp (token[0], ".longparam") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter_int64 (parser->program, size, token[2]);
} else if (strcmp (token[0], ".const") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_constant_str (parser->program, size, token[3], token[2]);
} else if (strcmp (token[0], ".floatparam") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter_float (parser->program, size, token[2]);
} else if (strcmp (token[0], ".doubleparam") == 0) {
int size = strtol (token[1], NULL, 0);
orc_program_add_parameter_double (parser->program, size, token[2]);
} else {
orc_parse_log (parser, "error: line %d: unknown directive: %s\n",
parser->line_number, token[0]);
}
} else {
OrcStaticOpcode *o;
unsigned int flags = 0;
int offset = 0;
if (strcmp (token[0], "x4") == 0) {
flags |= ORC_INSTRUCTION_FLAG_X4;
offset = 1;
} else if (strcmp (token[0], "x2") == 0) {
flags |= ORC_INSTRUCTION_FLAG_X2;
offset = 1;
}
o = get_opcode (parser, token[offset]);
if (o) {
int n_args = opcode_n_args (o);
int i, j;
char *args[4] = { NULL };
if (n_tokens != 1 + offset + n_args) {
orc_parse_log (parser, "error: line %d: too %s arguments for %s (expected %d)\n",
parser->line_number, (n_tokens < 1+offset+n_args) ? "few" : "many",
token[offset], n_args);
}
for(i=offset+1,j=0;i<n_tokens;i++,j++){
char *end;
double unused ORC_GNUC_UNUSED;
char varname[80];
args[j] = token[i];
unused = strtod (token[i], &end);
if (end != token[i]) {
int id;
/* make a unique name based on value and size */
snprintf (varname, sizeof (varname), "_%d.%s", opcode_arg_size(o, j), token[i]);
id = orc_program_add_constant_str (parser->program, opcode_arg_size(o, j),
token[i], varname);
/* it's possible we reused an existing variable, get its name so
* that we can refer to it in the opcode */
args[j] = parser->program->vars[id].name;
}
}
orc_program_append_str_2 (parser->program, token[offset], flags,
args[0], args[1], args[2], args[3]);
} else {
orc_parse_log (parser, "error: line %d: unknown opcode: %s\n",
parser->line_number,
token[offset]);
}
}
}
if (parser->program) {
orc_parse_sanity_check (parser, parser->program);
}
if (parser->line) free (parser->line);
if (log) {
*log = parser->log;
} else {
free (parser->log);
}
if (parser->programs && parser->programs[0]) {
parser->programs[0]->init_function = init_function;
} else {
free (init_function);
}
*programs = parser->programs;
return parser->n_programs;
}
static void
schro_motion_init_functions (SchroMotion * motion)
{
if (motion_funcs[motion->xblen >> 1].block_accumulate == NULL) {
OrcProgram *p;
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_constant_n (p, motion->xblen);
orc_program_set_2d (p);
orc_program_set_name (p, "block_acc_Xxn");
orc_program_add_destination (p, 2, "d1");
orc_program_add_source (p, 2, "s1");
orc_program_add_source (p, 1, "s2");
orc_program_add_temporary (p, 2, "t1");
orc_program_append (p, "convubw", ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_D1);
orc_program_append (p, "mullw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_S1);
orc_program_append (p, "addw", ORC_VAR_D1, ORC_VAR_D1, ORC_VAR_T1);
result = orc_program_compile (p);
if (!ORC_COMPILE_RESULT_IS_SUCCESSFUL (result)) {
SCHRO_ERROR ("compile failed");
}
motion_funcs[motion->xblen / 2].block_accumulate = p;
}
if (motion_funcs[motion->xblen >> 1].block_accumulate_scaled == NULL) {
OrcProgram *p;
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_constant_n (p, motion->xblen);
orc_program_set_2d (p);
orc_program_set_name (p, "block_acc_scaled_Xxn");
orc_program_add_destination (p, 2, "d1");
orc_program_add_source (p, 2, "s1");
orc_program_add_source (p, 1, "s2");
orc_program_add_parameter (p, 2, "p1");
orc_program_add_constant (p, 2, 32, "c1");
orc_program_add_constant (p, 2, 6, "c2");
orc_program_add_temporary (p, 2, "t1");
orc_program_append (p, "convubw", ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_D1);
orc_program_append (p, "mullw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1);
orc_program_append (p, "addw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1);
orc_program_append (p, "shrsw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C2);
orc_program_append (p, "mullw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_S1);
orc_program_append (p, "addw", ORC_VAR_D1, ORC_VAR_D1, ORC_VAR_T1);
result = orc_program_compile (p);
if (!ORC_COMPILE_RESULT_IS_SUCCESSFUL (result)) {
SCHRO_ERROR ("compile failed");
}
motion_funcs[motion->xblen / 2].block_accumulate_scaled = p;
}
if (motion_funcs[motion->xblen >> 1].block_accumulate_dc == NULL) {
OrcProgram *p;
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_constant_n (p, motion->xblen);
orc_program_set_2d (p);
orc_program_set_name (p, "block_acc_dc_Xxn");
orc_program_add_destination (p, 2, "d1");
orc_program_add_source (p, 2, "s1");
orc_program_add_parameter (p, 2, "p1");
orc_program_add_temporary (p, 2, "t1");
orc_program_append (p, "mullw", ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_P1);
orc_program_append (p, "addw", ORC_VAR_D1, ORC_VAR_D1, ORC_VAR_T1);
result = orc_program_compile (p);
if (!ORC_COMPILE_RESULT_IS_SUCCESSFUL (result)) {
SCHRO_ERROR ("compile failed");
}
motion_funcs[motion->xblen / 2].block_accumulate_dc = p;
}
if (motion_funcs[motion->xblen >> 1].block_accumulate_avg == NULL) {
OrcProgram *p;
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_constant_n (p, motion->xblen);
orc_program_set_2d (p);
orc_program_set_name (p, "block_acc_avg_Xxn");
orc_program_add_destination (p, 2, "d1");
orc_program_add_source (p, 2, "s1");
orc_program_add_source (p, 1, "s2");
orc_program_add_source (p, 1, "s3");
orc_program_add_temporary (p, 2, "t1");
orc_program_add_temporary (p, 1, "t2");
orc_program_append (p, "avgub", ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S3);
orc_program_append (p, "convubw", ORC_VAR_T1, ORC_VAR_T2, 0);
orc_program_append (p, "mullw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_S1);
orc_program_append (p, "addw", ORC_VAR_D1, ORC_VAR_D1, ORC_VAR_T1);
result = orc_program_compile (p);
if (!ORC_COMPILE_RESULT_IS_SUCCESSFUL (result)) {
SCHRO_ERROR ("compile failed");
}
motion_funcs[motion->xblen / 2].block_accumulate_avg = p;
}
if (motion_funcs[motion->xblen >> 1].block_accumulate_biref == NULL) {
OrcProgram *p;
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_constant_n (p, motion->xblen);
orc_program_set_2d (p);
orc_program_set_name (p, "block_acc_biref_Xxn");
orc_program_add_destination (p, 2, "d1");
orc_program_add_source (p, 2, "s1");
orc_program_add_source (p, 1, "s2");
orc_program_add_source (p, 1, "s3");
orc_program_add_parameter (p, 2, "p1");
orc_program_add_parameter (p, 2, "p2");
orc_program_add_constant (p, 2, 32, "c1");
orc_program_add_constant (p, 2, 6, "c2");
orc_program_add_temporary (p, 2, "t1");
orc_program_add_temporary (p, 2, "t2");
orc_program_append (p, "convubw", ORC_VAR_T1, ORC_VAR_S2, 0);
orc_program_append (p, "mullw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1);
orc_program_append (p, "convubw", ORC_VAR_T2, ORC_VAR_S3, 0);
orc_program_append (p, "mullw", ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2);
orc_program_append (p, "addw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2);
orc_program_append (p, "addw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1);
orc_program_append (p, "shrsw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C2);
orc_program_append (p, "mullw", ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_S1);
orc_program_append (p, "addw", ORC_VAR_D1, ORC_VAR_D1, ORC_VAR_T1);
result = orc_program_compile (p);
if (!ORC_COMPILE_RESULT_IS_SUCCESSFUL (result)) {
SCHRO_ERROR ("compile failed");
}
motion_funcs[motion->xblen / 2].block_accumulate_biref = p;
}
}