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);
}
void
vips_vector_constant( VipsVector *vector, char *name, int value, int size )
{
#ifdef HAVE_ORC
char *sname;
if( size == 1 )
sname = "b";
else if( size == 2 )
sname = "w";
else if( size == 4 )
sname = "l";
else {
printf( "vips_vector_constant: bad constant size\n" );
/* Not really correct, heh.
*/
sname = "x";
}
if( value > 0 )
vips_snprintf( name, 256, "c%d%s", value, sname );
else
vips_snprintf( name, 256, "cm%d%s", -value, sname );
if( orc_program_find_var_by_name( vector->program, name ) == -1 ) {
orc_program_add_constant( vector->program, size, value, name );
vector->n_constant += 1;
}
#endif /*HAVE_ORC*/
}
void
test_opcode_const (OrcStaticOpcode *opcode)
{
OrcProgram *p;
char s[40];
int ret;
if (opcode->src_size[1] == 0) {
return;
}
p = orc_program_new_ds (opcode->dest_size[0], opcode->src_size[0]);
orc_program_add_constant (p, opcode->src_size[1], 1, "c1");
sprintf(s, "test_c_%s", opcode->name);
orc_program_set_name (p, s);
orc_program_append_str (p, opcode->name, "d1", "s1", "c1");
ret = orc_test_compare_output (p);
if (!ret) {
error = TRUE;
}
orc_program_free (p);
}
int
orc_program_add_constant_float (OrcProgram *program, int size,
float value, const char *name)
{
orc_union32 u;
u.f = value;
return orc_program_add_constant (program, size, u.i, name);
}
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);
}
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;
}
}
static int
add_constant (OrcProgram *program, int size, const char *name)
{
return orc_program_add_constant (program, size, 0, name);
}