static void
_backup_orc_blend_u8 (OrcExecutor * ex)
{
int i;
int j;
int n = ex->n;
int m = ex->params[ORC_VAR_A1];
orc_int8 var0;
orc_int8 *ptr0;
orc_int8 var4;
const orc_int8 *ptr4;
const orc_int8 var16 = 8;
const int var24 = ex->params[24];
orc_int16 var32;
orc_int16 var33;
orc_int16 var34;
orc_int16 var35;
orc_int16 var36;
orc_int16 var37;
orc_int16 var38;
for (j = 0; j < m; j++) {
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
for (i = 0; i < n; i++) {
var0 = *ptr0;
var4 = *ptr4;
ptr4++;
/* 0: convubw */
var32 = (orc_uint8) var0;
/* 1: convubw */
var33 = (orc_uint8) var4;
/* 2: subw */
var34 = var33 - var32;
/* 3: mullw */
var35 = (var34 * var24) & 0xffff;
/* 4: shlw */
var36 = var32 << var16;
/* 5: addw */
var37 = var36 + var35;
/* 6: shruw */
var38 = ((orc_uint16) var37) >> var16;
/* 7: convsuswb */
var0 = ORC_CLAMP_UB (var38);
*ptr0 = var0;
ptr0++;
}
}
}
void
orc_blend_u8 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, int n, int m)
{
int i;
int j;
orc_int8 var0;
orc_int8 *ptr0;
orc_int8 var4;
const orc_int8 *ptr4;
const orc_int8 var16 = 8;
const int var24 = p1;
orc_int16 var32;
orc_int16 var33;
orc_int16 var34;
orc_int16 var35;
orc_int16 var36;
orc_int16 var37;
orc_int16 var38;
for (j = 0; j < m; j++) {
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
for (i = 0; i < n; i++) {
var0 = *ptr0;
var4 = *ptr4;
ptr4++;
/* 0: convubw */
var32 = (orc_uint8) var0;
/* 1: convubw */
var33 = (orc_uint8) var4;
/* 2: subw */
var34 = var33 - var32;
/* 3: mullw */
var35 = (var34 * var24) & 0xffff;
/* 4: shlw */
var36 = var32 << var16;
/* 5: addw */
var37 = var36 + var35;
/* 6: shruw */
var38 = ((orc_uint16) var37) >> var16;
/* 7: convsuswb */
var0 = ORC_CLAMP_UB (var38);
*ptr0 = var0;
ptr0++;
}
}
}
int
orc_array_compare (OrcArray *array1, OrcArray *array2, int flags)
{
if ((flags & ORC_TEST_FLAGS_FLOAT)) {
if (array1->element_size == 4) {
int j;
for(j=0;j<array1->m;j++){
float *a, *b;
int i;
a = ORC_PTR_OFFSET (array1->data, j*array1->stride);
b = ORC_PTR_OFFSET (array2->data, j*array2->stride);
for (i=0;i<array1->n;i++){
if (isnan(a[i]) && isnan(b[i])) continue;
if (a[i] == b[i]) continue;
if (fabs(a[i] - b[i]) < MIN_NONDENORMAL) continue;
return FALSE;
}
}
return TRUE;
} else if (array1->element_size == 8) {
int j;
for(j=0;j<array1->m;j++){
double *a, *b;
int i;
a = ORC_PTR_OFFSET (array1->data, j*array1->stride);
b = ORC_PTR_OFFSET (array2->data, j*array2->stride);
for (i=0;i<array1->n;i++){
if (isnan(a[i]) && isnan(b[i])) continue;
if (a[i] == b[i]) continue;
if (abs(a[i] - b[i]) < MIN_NONDENORMAL_D) continue;
return FALSE;
}
}
return TRUE;
}
} else {
if (memcmp (array1->alloc_data, array2->alloc_data,
array1->alloc_len) == 0) {
return TRUE;
}
}
return FALSE;
}
OrcArray *
orc_array_new (int n, int m, int element_size, int misalignment)
{
OrcArray *ar;
void *data;
#ifdef HAVE_POSIX_MEMALIGN
int ret;
#endif
ar = malloc (sizeof(OrcArray));
memset (ar, 0, sizeof(OrcArray));
ar->n = n;
ar->m = m;
ar->element_size = element_size;
ar->stride = (n*element_size + EXTEND_STRIDE);
ar->stride = (ar->stride + (ALIGNMENT-1)) & (~(ALIGNMENT-1));
ar->alloc_len = ar->stride * (m+2*EXTEND_ROWS) + (ALIGNMENT * element_size);
#ifdef HAVE_POSIX_MEMALIGN
ret = posix_memalign (&data, ALIGNMENT, ar->alloc_len);
#else
data = malloc (ar->alloc_len);
#endif
ar->alloc_data = data;
ar->data = ORC_PTR_OFFSET (ar->alloc_data,
ar->stride * EXTEND_ROWS + element_size * misalignment);
return ar;
}
int
float_compare (OrcArray *array1, OrcArray *array2, int i, int j)
{
void *ptr1 = ORC_PTR_OFFSET (array1->data,
i*array1->element_size + j*array1->stride);
void *ptr2 = ORC_PTR_OFFSET (array2->data,
i*array2->element_size + j*array2->stride);
switch (array1->element_size) {
case 4:
if (isnan(*(float *)ptr1) && isnan(*(float *)ptr2)) return TRUE;
if (*(float *)ptr1 == *(float *)ptr2) return TRUE;
if (fabs(*(float *)ptr1 - *(float *)ptr2) < MIN_NONDENORMAL) return TRUE;
return FALSE;
case 8:
/* FIXME */
return FALSE;
}
return FALSE;
}
void
orc_code_allocate_codemem (OrcCode *code, int size)
{
OrcCodeRegion *region;
OrcCodeChunk *chunk;
int aligned_size = (size + 15) & (~15);
chunk = orc_code_region_get_free_chunk (aligned_size);
region = chunk->region;
if (chunk->size > aligned_size) {
orc_code_chunk_split (chunk, aligned_size);
}
chunk->used = TRUE;
code->chunk = chunk;
code->code = ORC_PTR_OFFSET(region->write_ptr, chunk->offset);
code->exec = ORC_PTR_OFFSET(region->exec_ptr, chunk->offset);
code->code_size = size;
/* compiler->codeptr = ORC_PTR_OFFSET(region->write_ptr, chunk->offset); */
}
int
orc_array_check_out_of_bounds (OrcArray *array)
{
int i;
int j;
unsigned char *data;
data = array->alloc_data;
for(i=0;i<array->stride * EXTEND_ROWS;i++){
if (data[i] != ORC_OOB_VALUE) {
printf("OOB check failed at start-%d\n", array->stride * EXTEND_ROWS - i);
return FALSE;
}
}
for(j=0;j<array->m;j++){
data = ORC_PTR_OFFSET(array->data, array->stride * j);
for(i=array->element_size * array->n;i<array->stride;i++){
if (data[i] != ORC_OOB_VALUE) {
printf("OOB check failed on row %d, end+%d\n", j,
i - array->element_size * array->n);
return FALSE;
}
}
}
data = ORC_PTR_OFFSET (array->data, array->stride * array->m);
for(i=0;i<array->stride * EXTEND_ROWS;i++){
if (data[i] != ORC_OOB_VALUE) {
printf("OOB check failed at end+%d\n", i);
return FALSE;
}
}
return TRUE;
}
static orc_uint64
print_array_val_float (OrcArray *array, int i, int j)
{
void *ptr = ORC_PTR_OFFSET (array->data,
i*array->element_size + j*array->stride);
switch (array->element_size) {
case 4:
if (isnan(*(float *)ptr)) {
printf(" nan %08x", *(orc_uint32 *)ptr);
/* This is to get around signaling/non-signaling nans in the output */
return (*(orc_uint32 *)ptr) & 0xffbfffff;
} else {
printf(" %12.5g", *(float *)ptr);
return *(orc_int32 *)ptr;
}
case 8:
printf(" %12.5g", *(double *)ptr);
return *(orc_int64 *)ptr;
default:
printf(" ERROR");
return -1;
}
}
static orc_uint64
print_array_val_hex (OrcArray *array, int i, int j)
{
void *ptr = ORC_PTR_OFFSET (array->data,
i*array->element_size + j*array->stride);
switch (array->element_size) {
case 1:
printf(" %02x", *(orc_uint8 *)ptr);
return *(orc_int8 *)ptr;
case 2:
printf(" %04x", *(orc_uint16 *)ptr);
return *(orc_int16 *)ptr;
case 4:
printf(" %08x", *(orc_uint32 *)ptr);
return *(orc_int32 *)ptr;
case 8:
printf(" 0x%08x%08x", (orc_uint32)((*(orc_uint64 *)ptr)>>32),
(orc_uint32)((*(orc_uint64 *)ptr)));
return *(orc_int64 *)ptr;
default:
return -1;
}
}
void
orc_array_set_pattern_2 (OrcArray *array, OrcRandomContext *context,
int type)
{
int i,j;
switch (type) {
case ORC_PATTERN_RANDOM:
orc_random_bits (context, array->alloc_data, array->alloc_len);
break;
case ORC_PATTERN_FLOAT_SMALL:
{
if (array->element_size != 4) return;
for(j=0;j<array->m;j++){
orc_union32 *data;
int exp;
data = ORC_PTR_OFFSET(array->data, array->stride * j);
for(i=0;i<array->n;i++){
data[i].i = orc_random (context);
exp = (data[i].i & 0x7f80000) >> 23;
exp &= 0xf;
exp += 122;
data[i].i &= ~0x7f800000;
data[i].i |= (exp&0xff) << 23;
}
}
}
break;
case ORC_PATTERN_FLOAT_SPECIAL:
{
if (array->element_size != 4) return;
for(j=0;j<array->m;j++){
orc_union32 *data;
int x;
data = ORC_PTR_OFFSET(array->data, array->stride * j);
for(i=0;i<array->n;i++){
x = i&0x1f;
data[i].i = special_floats[x];
}
}
}
break;
case ORC_PATTERN_FLOAT_DENORMAL:
{
if (array->element_size != 4) return;
for(j=0;j<array->m;j++){
orc_union32 *data;
data = ORC_PTR_OFFSET(array->data, array->stride * j);
for(i=0;i<array->n;i++){
data[i].i = orc_random (context);
data[i].i &= ~0x7f800000;
}
}
}
break;
default:
break;
}
}
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;
}
void
orc_executor_emulate (OrcExecutor *ex)
{
int i;
int j;
int k;
int m, m_index;
OrcCode *code;
OrcInstruction *insn;
OrcStaticOpcode *opcode;
OrcOpcodeExecutor *opcode_ex;
void *tmpspace[ORC_N_COMPILER_VARIABLES] = { 0 };
if (ex->program) {
code = ex->program->orccode;
} else {
code = (OrcCode *)ex->arrays[ORC_VAR_A2];
}
ex->accumulators[0] = 0;
ex->accumulators[1] = 0;
ex->accumulators[2] = 0;
ex->accumulators[3] = 0;
ORC_DEBUG("emulating");
memset (&opcode_ex, 0, sizeof(opcode_ex));
if (code == NULL) {
ORC_ERROR("attempt to run program that failed to compile");
ORC_ASSERT(0);
}
if (code->is_2d) {
m = ORC_EXECUTOR_M(ex);
} else {
m = 1;
}
for(i=0;i<ORC_N_COMPILER_VARIABLES;i++){
OrcCodeVariable *var = code->vars + i;
if (var->size) {
tmpspace[i] = malloc(ORC_MAX_VAR_SIZE * CHUNK_SIZE);
}
}
opcode_ex = malloc(sizeof(OrcOpcodeExecutor)*code->n_insns);
for(j=0;j<code->n_insns;j++){
insn = code->insns + j;
opcode = insn->opcode;
opcode_ex[j].emulateN = opcode->emulateN;
opcode_ex[j].shift = 0;
if (insn->flags & ORC_INSTRUCTION_FLAG_X2) {
opcode_ex[j].shift = 1;
} else if (insn->flags & ORC_INSTRUCTION_FLAG_X4) {
opcode_ex[j].shift = 2;
}
for(k=0;k<ORC_STATIC_OPCODE_N_SRC;k++) {
OrcCodeVariable *var = code->vars + insn->src_args[k];
if (opcode->src_size[k] == 0) continue;
if (var->vartype == ORC_VAR_TYPE_CONST) {
opcode_ex[j].src_ptrs[k] = tmpspace[insn->src_args[k]];
/* FIXME hack */
load_constant (tmpspace[insn->src_args[k]], 8,
var->value.i);
} else if (var->vartype == ORC_VAR_TYPE_PARAM) {
opcode_ex[j].src_ptrs[k] = tmpspace[insn->src_args[k]];
/* FIXME hack */
load_constant (tmpspace[insn->src_args[k]], 8,
(orc_uint64)(orc_uint32)ex->params[insn->src_args[k]] |
(((orc_uint64)(orc_uint32)ex->params[insn->src_args[k] +
(ORC_VAR_T1 - ORC_VAR_P1)])<<32));
} else if (var->vartype == ORC_VAR_TYPE_TEMP) {
opcode_ex[j].src_ptrs[k] = tmpspace[insn->src_args[k]];
} else if (var->vartype == ORC_VAR_TYPE_SRC) {
if (ORC_PTR_TO_INT(ex->arrays[insn->src_args[k]]) & (var->size - 1)) {
ORC_ERROR("Unaligned array for src%d, program %s",
(insn->src_args[k]-ORC_VAR_S1), ex->program->name);
}
opcode_ex[j].src_ptrs[k] = ex->arrays[insn->src_args[k]];
} else if (var->vartype == ORC_VAR_TYPE_DEST) {
if (ORC_PTR_TO_INT(ex->arrays[insn->src_args[k]]) & (var->size - 1)) {
ORC_ERROR("Unaligned array for dest%d, program %s",
(insn->src_args[k]-ORC_VAR_D1), ex->program->name);
}
opcode_ex[j].src_ptrs[k] = ex->arrays[insn->src_args[k]];
}
}
for(k=0;k<ORC_STATIC_OPCODE_N_DEST;k++) {
OrcCodeVariable *var = code->vars + insn->dest_args[k];
if (opcode->dest_size[k] == 0) continue;
if (var->vartype == ORC_VAR_TYPE_TEMP) {
ORC_DEBUG("dest vartype tmp %d", insn->dest_args[k]);
opcode_ex[j].dest_ptrs[k] = tmpspace[insn->dest_args[k]];
} else if (var->vartype == ORC_VAR_TYPE_ACCUMULATOR) {
opcode_ex[j].dest_ptrs[k] =
&ex->accumulators[insn->dest_args[k] - ORC_VAR_A1];
} else if (var->vartype == ORC_VAR_TYPE_DEST) {
if (ORC_PTR_TO_INT(ex->arrays[insn->dest_args[k]]) & (var->size - 1)) {
ORC_ERROR("Unaligned array for dest%d, program %s",
(insn->dest_args[k]-ORC_VAR_D1), ex->program->name);
}
opcode_ex[j].dest_ptrs[k] = ex->arrays[insn->dest_args[k]];
}
}
ORC_DEBUG("opcode %s %p %p %p", opcode->name,
opcode_ex[j].dest_ptrs[0], opcode_ex[j].src_ptrs[0],
opcode_ex[j].src_ptrs[1]);
}
ORC_DEBUG("src ptr %p stride %d", ex->arrays[ORC_VAR_S1], ex->params[ORC_VAR_S1]);
for(m_index=0;m_index<m;m_index++){
ORC_DEBUG("m_index %d m %d", m_index, m);
for(j=0;j<code->n_insns;j++){
insn = code->insns + j;
opcode = insn->opcode;
for(k=0;k<ORC_STATIC_OPCODE_N_SRC;k++) {
OrcCodeVariable *var = code->vars + insn->src_args[k];
if (opcode->src_size[k] == 0) continue;
if (var->vartype == ORC_VAR_TYPE_SRC) {
opcode_ex[j].src_ptrs[k] =
ORC_PTR_OFFSET(ex->arrays[insn->src_args[k]],
ex->params[insn->src_args[k]]*m_index);
} else if (var->vartype == ORC_VAR_TYPE_DEST) {
opcode_ex[j].src_ptrs[k] =
ORC_PTR_OFFSET(ex->arrays[insn->src_args[k]],
ex->params[insn->src_args[k]]*m_index);
}
}
for(k=0;k<ORC_STATIC_OPCODE_N_DEST;k++) {
OrcCodeVariable *var = code->vars + insn->dest_args[k];
if (opcode->dest_size[k] == 0) continue;
if (var->vartype == ORC_VAR_TYPE_DEST) {
opcode_ex[j].dest_ptrs[k] =
ORC_PTR_OFFSET(ex->arrays[insn->dest_args[k]],
ex->params[insn->dest_args[k]]*m_index);
}
}
}
for(i=0;i<ex->n;i+=CHUNK_SIZE){
for(j=0;j<code->n_insns;j++){
if (ex->n - i >= CHUNK_SIZE) {
opcode_ex[j].emulateN (opcode_ex + j, i, CHUNK_SIZE << opcode_ex[j].shift);
} else {
opcode_ex[j].emulateN (opcode_ex + j, i, (ex->n - i) << opcode_ex[j].shift);
}
}
}
}
free (opcode_ex);
for(i=0;i<ORC_N_COMPILER_VARIABLES;i++){
if (tmpspace[i]) free (tmpspace[i]);
}
}