void OffloadDescriptor::gather_copyout_data()
{
OFFLOAD_TIMER_STOP(c_offload_target_compute);
OFFLOAD_TIMER_START(c_offload_target_gather_outputs);
for (int i = 0; i < m_vars_total; i++) {
bool src_is_for_mic = (m_vars[i].direction.out ||
m_vars[i].into == NULL);
switch (m_vars[i].type.src) {
case c_data_ptr_array:
break;
case c_data:
case c_void_ptr:
case c_cean_var:
if (m_vars[i].direction.out &&
!m_vars[i].flags.is_static) {
m_out.send_data(
static_cast<char*>(m_vars[i].ptr) + m_vars[i].disp,
m_vars[i].size);
}
break;
case c_dv:
break;
case c_string_ptr:
case c_data_ptr:
case c_cean_var_ptr:
case c_dv_ptr:
if (m_vars[i].free_if &&
src_is_for_mic &&
!m_vars[i].flags.is_static) {
void *buf = *static_cast<char**>(m_vars[i].ptr) -
m_vars[i].mic_offset -
(m_vars[i].flags.is_stack_buf?
0 : m_vars[i].offset);
if (buf == NULL) {
break;
}
// decrement buffer reference count
OFFLOAD_TIMER_START(c_offload_target_release_buffer_refs);
BufReleaseRef(buf);
OFFLOAD_TIMER_STOP(c_offload_target_release_buffer_refs);
}
break;
case c_func_ptr:
if (m_vars[i].direction.out) {
m_out.send_func_ptr(*((void**) m_vars[i].ptr));
}
break;
case c_dv_data:
case c_dv_ptr_data:
case c_dv_data_slice:
case c_dv_ptr_data_slice:
if (src_is_for_mic &&
m_vars[i].free_if &&
!m_vars[i].flags.is_static) {
ArrDesc *dvp = (m_vars[i].type.src == c_dv_data ||
m_vars[i].type.src == c_dv_data_slice) ?
static_cast<ArrDesc*>(m_vars[i].ptr) :
*static_cast<ArrDesc**>(m_vars[i].ptr);
void *buf = reinterpret_cast<char*>(dvp->Base) -
m_vars[i].mic_offset -
m_vars[i].offset;
if (buf == NULL) {
break;
}
// decrement buffer reference count
OFFLOAD_TIMER_START(c_offload_target_release_buffer_refs);
BufReleaseRef(buf);
OFFLOAD_TIMER_STOP(c_offload_target_release_buffer_refs);
}
break;
default:
LIBOFFLOAD_ERROR(c_unknown_var_type, m_vars[i].type.dst);
abort();
}
if (m_vars[i].into) {
switch (m_vars[i].type.dst) {
case c_data_ptr_array:
break;
case c_data:
case c_void_ptr:
case c_cean_var:
case c_dv:
break;
case c_string_ptr:
case c_data_ptr:
case c_cean_var_ptr:
case c_dv_ptr:
if (m_vars[i].direction.in &&
m_vars[i].free_if &&
!m_vars[i].flags.is_static_dstn) {
void *buf = *static_cast<char**>(m_vars[i].into) -
m_vars[i].mic_offset -
(m_vars[i].flags.is_stack_buf?
0 : m_vars[i].offset);
if (buf == NULL) {
break;
}
// decrement buffer reference count
OFFLOAD_TIMER_START(
c_offload_target_release_buffer_refs);
BufReleaseRef(buf);
OFFLOAD_TIMER_STOP(
c_offload_target_release_buffer_refs);
}
break;
case c_func_ptr:
break;
case c_dv_data:
case c_dv_ptr_data:
case c_dv_data_slice:
case c_dv_ptr_data_slice:
if (m_vars[i].free_if &&
m_vars[i].direction.in &&
!m_vars[i].flags.is_static_dstn) {
ArrDesc *dvp =
(m_vars[i].type.dst == c_dv_data_slice ||
m_vars[i].type.dst == c_dv_data) ?
static_cast<ArrDesc*>(m_vars[i].into) :
*static_cast<ArrDesc**>(m_vars[i].into);
void *buf = reinterpret_cast<char*>(dvp->Base) -
m_vars[i].mic_offset -
m_vars[i].offset;
if (buf == NULL) {
break;
}
// decrement buffer reference count
OFFLOAD_TIMER_START(
c_offload_target_release_buffer_refs);
BufReleaseRef(buf);
OFFLOAD_TIMER_STOP(
c_offload_target_release_buffer_refs);
}
break;
default:
LIBOFFLOAD_ERROR(c_unknown_var_type, m_vars[i].type.dst);
abort();
}
}
}
OFFLOAD_DEBUG_TRACE(2, "OUT buffer @ p %p size %lld\n",
m_out.get_buffer_start(),
m_out.get_buffer_size());
OFFLOAD_DEBUG_DUMP_BYTES(2,
m_out.get_buffer_start(),
m_out.get_buffer_size());
OFFLOAD_DEBUG_TRACE_1(1, get_offload_number(), c_offload_copyout_data,
"Total copyout data sent to host: [%lld] bytes\n",
m_out.get_tfr_size());
OFFLOAD_TIMER_STOP(c_offload_target_gather_outputs);
}
void OffloadDescriptor::scatter_copyin_data()
{
OFFLOAD_TIMER_START(c_offload_target_scatter_inputs);
OFFLOAD_DEBUG_TRACE(2, "IN buffer @ %p size %lld\n",
m_in.get_buffer_start(),
m_in.get_buffer_size());
OFFLOAD_DEBUG_DUMP_BYTES(2, m_in.get_buffer_start(),
m_in.get_buffer_size());
// receive data
for (int i = 0; i < m_vars_total; i++) {
bool src_is_for_mic = (m_vars[i].direction.out ||
m_vars[i].into == NULL);
void** ptr_addr = src_is_for_mic ?
static_cast<void**>(m_vars[i].ptr) :
static_cast<void**>(m_vars[i].into);
int type = src_is_for_mic ? m_vars[i].type.src :
m_vars[i].type.dst;
bool is_static = src_is_for_mic ?
m_vars[i].flags.is_static :
m_vars[i].flags.is_static_dstn;
void *ptr = NULL;
if (m_vars[i].flags.alloc_disp) {
int64_t offset = 0;
m_in.receive_data(&offset, sizeof(offset));
m_vars[i].offset = -offset;
}
if (VAR_TYPE_IS_DV_DATA_SLICE(type) ||
VAR_TYPE_IS_DV_DATA(type)) {
ArrDesc *dvp = (type == c_dv_data_slice || type == c_dv_data)?
reinterpret_cast<ArrDesc*>(ptr_addr) :
*reinterpret_cast<ArrDesc**>(ptr_addr);
ptr_addr = reinterpret_cast<void**>(&dvp->Base);
}
// Set pointer values
switch (type) {
case c_data_ptr_array:
{
int j = m_vars[i].ptr_arr_offset;
int max_el = j + m_vars[i].count;
char *dst_arr_ptr = (src_is_for_mic)?
*(reinterpret_cast<char**>(m_vars[i].ptr)) :
reinterpret_cast<char*>(m_vars[i].into);
for (; j < max_el; j++) {
if (src_is_for_mic) {
m_vars[j].ptr =
dst_arr_ptr + m_vars[j].ptr_arr_offset;
}
else {
m_vars[j].into =
dst_arr_ptr + m_vars[j].ptr_arr_offset;
}
}
}
break;
case c_data:
case c_void_ptr:
case c_cean_var:
case c_dv:
break;
case c_string_ptr:
case c_data_ptr:
case c_cean_var_ptr:
case c_dv_ptr:
if (m_vars[i].alloc_if) {
void *buf;
if (m_vars[i].flags.sink_addr) {
m_in.receive_data(&buf, sizeof(buf));
}
else {
buf = m_buffers.front();
m_buffers.pop_front();
}
if (buf) {
if (!is_static) {
if (!m_vars[i].flags.sink_addr) {
// increment buffer reference
OFFLOAD_TIMER_START(c_offload_target_add_buffer_refs);
BufferAddRef(buf);
OFFLOAD_TIMER_STOP(c_offload_target_add_buffer_refs);
}
add_ref_count(buf, 0 == m_vars[i].flags.sink_addr);
}
ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset +
(m_vars[i].flags.is_stack_buf ?
0 : m_vars[i].offset);
}
*ptr_addr = ptr;
}
else if (m_vars[i].flags.sink_addr) {
void *buf;
m_in.receive_data(&buf, sizeof(buf));
void *ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset +
(m_vars[i].flags.is_stack_buf ?
0 : m_vars[i].offset);
*ptr_addr = ptr;
}
break;
case c_func_ptr:
break;
case c_dv_data:
case c_dv_ptr_data:
case c_dv_data_slice:
case c_dv_ptr_data_slice:
if (m_vars[i].alloc_if) {
void *buf;
if (m_vars[i].flags.sink_addr) {
m_in.receive_data(&buf, sizeof(buf));
}
else {
buf = m_buffers.front();
m_buffers.pop_front();
}
if (buf) {
if (!is_static) {
if (!m_vars[i].flags.sink_addr) {
// increment buffer reference
OFFLOAD_TIMER_START(c_offload_target_add_buffer_refs);
BufferAddRef(buf);
OFFLOAD_TIMER_STOP(c_offload_target_add_buffer_refs);
}
add_ref_count(buf, 0 == m_vars[i].flags.sink_addr);
}
ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset + m_vars[i].offset;
}
*ptr_addr = ptr;
}
else if (m_vars[i].flags.sink_addr) {
void *buf;
m_in.receive_data(&buf, sizeof(buf));
ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset + m_vars[i].offset;
*ptr_addr = ptr;
}
break;
default:
LIBOFFLOAD_ERROR(c_unknown_var_type, type);
abort();
}
// Release obsolete buffers for stack of persistent objects
if (type = c_data_ptr &&
m_vars[i].flags.is_stack_buf &&
!m_vars[i].direction.bits &&
m_vars[i].alloc_if &&
m_vars[i].size != 0) {
for (int j=0; j < m_vars[i].size; j++) {
void *buf;
m_in.receive_data(&buf, sizeof(buf));
BufferReleaseRef(buf);
ref_data.erase(buf);
}
}
// Do copyin
switch (m_vars[i].type.dst) {
case c_data_ptr_array:
break;
case c_data:
case c_void_ptr:
case c_cean_var:
if (m_vars[i].direction.in &&
!m_vars[i].flags.is_static_dstn) {
int64_t size;
int64_t disp;
char* ptr = m_vars[i].into ?
static_cast<char*>(m_vars[i].into) :
static_cast<char*>(m_vars[i].ptr);
if (m_vars[i].type.dst == c_cean_var) {
m_in.receive_data((&size), sizeof(int64_t));
m_in.receive_data((&disp), sizeof(int64_t));
}
else {
size = m_vars[i].size;
disp = 0;
}
m_in.receive_data(ptr + disp, size);
}
break;
case c_dv:
if (m_vars[i].direction.bits ||
m_vars[i].alloc_if ||
m_vars[i].free_if) {
char* ptr = m_vars[i].into ?
static_cast<char*>(m_vars[i].into) :
static_cast<char*>(m_vars[i].ptr);
m_in.receive_data(ptr + sizeof(uint64_t),
m_vars[i].size - sizeof(uint64_t));
}
break;
case c_string_ptr:
case c_data_ptr:
case c_cean_var_ptr:
case c_dv_ptr:
case c_dv_data:
case c_dv_ptr_data:
case c_dv_data_slice:
case c_dv_ptr_data_slice:
break;
case c_func_ptr:
if (m_vars[i].direction.in) {
m_in.receive_func_ptr((const void**) m_vars[i].ptr);
}
break;
default:
LIBOFFLOAD_ERROR(c_unknown_var_type, m_vars[i].type.dst);
abort();
}
}
OFFLOAD_TRACE(1, "Total copyin data received from host: [%lld] bytes\n",
m_in.get_tfr_size());
OFFLOAD_TIMER_STOP(c_offload_target_scatter_inputs);
OFFLOAD_TIMER_START(c_offload_target_compute);
}
void OffloadDescriptor::offload(
uint32_t buffer_count,
void** buffers,
void* misc_data,
uint16_t misc_data_len,
void* return_data,
uint16_t return_data_len
)
{
FunctionDescriptor *func = (FunctionDescriptor*) misc_data;
const char *name = func->data;
OffloadDescriptor ofld;
char *in_data = 0;
char *out_data = 0;
char *timer_data = 0;
console_enabled = func->console_enabled;
timer_enabled = func->timer_enabled;
offload_report_level = func->offload_report_level;
offload_number = func->offload_number;
ofld.set_offload_number(func->offload_number);
#ifdef SEP_SUPPORT
if (sep_monitor) {
if (__sync_fetch_and_add(&sep_counter, 1) == 0) {
OFFLOAD_DEBUG_TRACE(2, "VTResumeSampling\n");
VTResumeSampling();
}
}
#endif // SEP_SUPPORT
OFFLOAD_DEBUG_TRACE_1(2, ofld.get_offload_number(),
c_offload_start_target_func,
"Offload \"%s\" started\n", name);
// initialize timer data
OFFLOAD_TIMER_INIT();
OFFLOAD_TIMER_START(c_offload_target_total_time);
OFFLOAD_TIMER_START(c_offload_target_descriptor_setup);
// get input/output buffer addresses
if (func->in_datalen > 0 || func->out_datalen > 0) {
if (func->data_offset != 0) {
in_data = (char*) misc_data + func->data_offset;
out_data = (char*) return_data;
}
else {
char *inout_buf = (char*) buffers[--buffer_count];
in_data = inout_buf;
out_data = inout_buf;
}
}
// assign variable descriptors
ofld.m_vars_total = func->vars_num;
if (ofld.m_vars_total > 0) {
uint64_t var_data_len = ofld.m_vars_total * sizeof(VarDesc);
ofld.m_vars = (VarDesc*) malloc(var_data_len);
if (ofld.m_vars == NULL)
LIBOFFLOAD_ERROR(c_malloc);
memcpy(ofld.m_vars, in_data, var_data_len);
in_data += var_data_len;
func->in_datalen -= var_data_len;
}
// timer data
if (func->timer_enabled) {
uint64_t timer_data_len = OFFLOAD_TIMER_DATALEN();
timer_data = out_data;
out_data += timer_data_len;
func->out_datalen -= timer_data_len;
}
// init Marshallers
ofld.m_in.init_buffer(in_data, func->in_datalen);
ofld.m_out.init_buffer(out_data, func->out_datalen);
// copy buffers to offload descriptor
std::copy(buffers, buffers + buffer_count,
std::back_inserter(ofld.m_buffers));
OFFLOAD_TIMER_STOP(c_offload_target_descriptor_setup);
// find offload entry address
OFFLOAD_TIMER_START(c_offload_target_func_lookup);
offload_func_with_parms entry = (offload_func_with_parms)
__offload_entries.find_addr(name);
if (entry == NULL) {
#if OFFLOAD_DEBUG > 0
if (console_enabled > 2) {
__offload_entries.dump();
}
#endif
LIBOFFLOAD_ERROR(c_offload_descriptor_offload, name);
exit(1);
}
OFFLOAD_TIMER_STOP(c_offload_target_func_lookup);
OFFLOAD_TIMER_START(c_offload_target_func_time);
// execute offload entry
entry(&ofld);
OFFLOAD_TIMER_STOP(c_offload_target_func_time);
OFFLOAD_TIMER_STOP(c_offload_target_total_time);
// copy timer data to the buffer
OFFLOAD_TIMER_TARGET_DATA(timer_data);
OFFLOAD_DEBUG_TRACE(2, "Offload \"%s\" finished\n", name);
#ifdef SEP_SUPPORT
if (sep_monitor) {
if (__sync_sub_and_fetch(&sep_counter, 1) == 0) {
OFFLOAD_DEBUG_TRACE(2, "VTPauseSampling\n");
VTPauseSampling();
}
}
#endif // SEP_SUPPORT
}
extern "C" OFFLOAD OFFLOAD_TARGET_ACQUIRE(
TARGET_TYPE target_type,
int target_number,
int is_optional,
_Offload_status* status,
const char* file,
uint64_t line
)
{
bool retval;
OFFLOAD ofld;
// initialize status
if (status != 0) {
status->result = OFFLOAD_UNAVAILABLE;
status->device_number = -1;
status->data_sent = 0;
status->data_received = 0;
}
// make sure libray is initialized
retval = __offload_init_library();
// OFFLOAD_TIMER_INIT must follow call to __offload_init_library
OffloadHostTimerData * timer_data = OFFLOAD_TIMER_INIT(file, line);
OFFLOAD_TIMER_START(timer_data, c_offload_host_total_offload);
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
// initialize all devices is init_type is on_offload_all
if (retval && __offload_init_type == c_init_on_offload_all) {
for (int i = 0; i < mic_engines_total; i++) {
mic_engines[i].init();
}
}
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
OFFLOAD_TIMER_START(timer_data, c_offload_host_target_acquire);
if (target_type == TARGET_HOST) {
// Host always available
retval = true;
}
else if (target_type == TARGET_MIC) {
if (target_number >= -1) {
if (retval) {
if (target_number >= 0) {
// User provided the device number
target_number = target_number % mic_engines_total;
}
else {
// use device 0
target_number = 0;
}
// reserve device in ORSL
if (is_optional) {
if (!ORSL::try_reserve(target_number)) {
target_number = -1;
}
}
else {
if (!ORSL::reserve(target_number)) {
target_number = -1;
}
}
// initialize device
if (target_number >= 0 &&
__offload_init_type == c_init_on_offload) {
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
mic_engines[target_number].init();
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
}
}
else {
// fallback to CPU
target_number = -1;
}
if (target_number < 0 || !retval) {
if (!is_optional && status == 0) {
LIBOFFLOAD_ERROR(c_device_is_not_available);
exit(1);
}
retval = false;
}
}
else {
LIBOFFLOAD_ERROR(c_invalid_device_number);
exit(1);
}
}
if (retval) {
ofld = new OffloadDescriptor(target_number, status,
!is_optional, false, timer_data);
OFFLOAD_TIMER_HOST_MIC_NUM(timer_data, target_number);
Offload_Report_Prolog(timer_data);
OFFLOAD_DEBUG_TRACE_1(2, timer_data->offload_number, c_offload_start,
"Starting offload: target_type = %d, "
"number = %d, is_optional = %d\n",
target_type, target_number, is_optional);
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire);
}
else {
ofld = NULL;
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire);
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_total_offload);
offload_report_free_data(timer_data);
}
return ofld;
}
extern "C" OFFLOAD OFFLOAD_TARGET_ACQUIRE1(
const int* device_num,
const char* file,
uint64_t line
)
{
int target_number;
// make sure libray is initialized and at least one device is available
if (!__offload_init_library()) {
LIBOFFLOAD_ERROR(c_device_is_not_available);
exit(1);
}
// OFFLOAD_TIMER_INIT must follow call to __offload_init_library
OffloadHostTimerData * timer_data = OFFLOAD_TIMER_INIT(file, line);
OFFLOAD_TIMER_START(timer_data, c_offload_host_total_offload);
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
if (__offload_init_type == c_init_on_offload_all) {
for (int i = 0; i < mic_engines_total; i++) {
mic_engines[i].init();
}
}
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
OFFLOAD_TIMER_START(timer_data, c_offload_host_target_acquire);
// use default device number if it is not provided
if (device_num != 0) {
target_number = *device_num;
}
else {
target_number = __omp_device_num;
}
// device number should be a non-negative integer value
if (target_number < 0) {
LIBOFFLOAD_ERROR(c_omp_invalid_device_num);
exit(1);
}
// should we do this for OpenMP?
target_number %= mic_engines_total;
// reserve device in ORSL
if (!ORSL::reserve(target_number)) {
LIBOFFLOAD_ERROR(c_device_is_not_available);
exit(1);
}
// initialize device(s)
OFFLOAD_TIMER_START(timer_data, c_offload_host_initialize);
if (__offload_init_type == c_init_on_offload) {
mic_engines[target_number].init();
}
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_initialize);
OFFLOAD ofld =
new OffloadDescriptor(target_number, 0, true, true, timer_data);
OFFLOAD_TIMER_HOST_MIC_NUM(timer_data, target_number);
Offload_Report_Prolog(timer_data);
OFFLOAD_DEBUG_TRACE_1(2, timer_data->offload_number, c_offload_start,
"Starting OpenMP offload, device = %d\n",
target_number);
OFFLOAD_TIMER_STOP(timer_data, c_offload_host_target_acquire);
return ofld;
}
void OffloadDescriptor::scatter_copyin_data()
{
OFFLOAD_TIMER_START(c_offload_target_scatter_inputs);
OFFLOAD_DEBUG_TRACE(2, "IN buffer @ %p size %lld\n",
m_in.get_buffer_start(),
m_in.get_buffer_size());
OFFLOAD_DEBUG_DUMP_BYTES(2, m_in.get_buffer_start(),
m_in.get_buffer_size());
// receive data
for (int i = 0; i < m_vars_total; i++) {
bool src_is_for_mic = (m_vars[i].direction.out ||
m_vars[i].into == NULL);
void** ptr_addr = src_is_for_mic ?
static_cast<void**>(m_vars[i].ptr) :
static_cast<void**>(m_vars[i].into);
int type = src_is_for_mic ? m_vars_extra[i].type_src :
m_vars_extra[i].type_dst;
bool is_static = src_is_for_mic ?
m_vars[i].flags.is_static :
m_vars[i].flags.is_static_dstn;
void *ptr = NULL;
if (m_vars[i].flags.alloc_disp) {
int64_t offset = 0;
m_in.receive_data(&offset, sizeof(offset));
}
if (VAR_TYPE_IS_DV_DATA_SLICE(type) ||
VAR_TYPE_IS_DV_DATA(type)) {
ArrDesc *dvp = (type == c_dv_data_slice || type == c_dv_data)?
reinterpret_cast<ArrDesc*>(ptr_addr) :
*reinterpret_cast<ArrDesc**>(ptr_addr);
ptr_addr = reinterpret_cast<void**>(&dvp->Base);
}
// Set pointer values
switch (type) {
case c_data_ptr_array:
{
int j = m_vars[i].ptr_arr_offset;
int max_el = j + m_vars[i].count;
char *dst_arr_ptr = (src_is_for_mic)?
*(reinterpret_cast<char**>(m_vars[i].ptr)) :
reinterpret_cast<char*>(m_vars[i].into);
// if is_pointer is 1 it means that pointer array itself
// is defined either via pointer or as class member.
// i.e. arr_ptr[0:5] or this->ARR[0:5]
if (m_vars[i].flags.is_pointer) {
int64_t offset = 0;
m_in.receive_data(&offset, sizeof(offset));
dst_arr_ptr = *((char**)dst_arr_ptr) + offset;
}
for (; j < max_el; j++) {
if (src_is_for_mic) {
m_vars[j].ptr =
dst_arr_ptr + m_vars[j].ptr_arr_offset;
}
else {
m_vars[j].into =
dst_arr_ptr + m_vars[j].ptr_arr_offset;
}
}
}
break;
case c_data:
case c_void_ptr:
case c_void_ptr_ptr:
case c_cean_var:
case c_dv:
break;
case c_string_ptr:
case c_data_ptr:
case c_string_ptr_ptr:
case c_data_ptr_ptr:
case c_cean_var_ptr:
case c_cean_var_ptr_ptr:
case c_dv_ptr:
// Don't need ptr_addr value for variables from stack buffer.
// Stack buffer address is set at var_desc with #0.
if (i != 0 && m_vars[i].flags.is_stack_buf) {
break;
}
if (TYPE_IS_PTR_TO_PTR(m_vars_extra[i].type_src) ||
TYPE_IS_PTR_TO_PTR(m_vars_extra[i].type_dst)) {
int64_t offset;
m_in.receive_data(&offset, sizeof(offset));
ptr_addr = reinterpret_cast<void**>(
reinterpret_cast<char*>(*ptr_addr) + offset);
}
if (m_vars[i].alloc_if && !m_vars[i].flags.preallocated) {
void *buf = NULL;
if (m_vars[i].flags.sink_addr) {
m_in.receive_data(&buf, sizeof(buf));
}
else {
buf = m_buffers.front();
m_buffers.pop_front();
}
if (buf) {
if (!is_static) {
if (!m_vars[i].flags.sink_addr) {
// increment buffer reference
OFFLOAD_TIMER_START(c_offload_target_add_buffer_refs);
BufferAddRef(buf);
OFFLOAD_TRACE(1, "Calling COIBufferAddRef %p\n", buf);
OFFLOAD_TIMER_STOP(c_offload_target_add_buffer_refs);
}
add_ref_count(buf, 0 == m_vars[i].flags.sink_addr);
OFFLOAD_TRACE(1, " AddRef count = %d\n",
((RefInfo *) ref_data[buf])->count);
}
ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset +
(m_vars[i].flags.is_stack_buf ?
0 : m_vars[i].offset);
}
*ptr_addr = ptr;
}
else if (m_vars[i].flags.sink_addr) {
void *buf;
m_in.receive_data(&buf, sizeof(buf));
void *ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset +
(m_vars[i].flags.is_stack_buf ?
0 : m_vars[i].offset);
*ptr_addr = ptr;
}
break;
case c_func_ptr:
case c_func_ptr_ptr:
break;
case c_dv_data:
case c_dv_ptr_data:
case c_dv_data_slice:
case c_dv_ptr_data_slice:
if (m_vars[i].alloc_if) {
void *buf;
if (m_vars[i].flags.sink_addr) {
m_in.receive_data(&buf, sizeof(buf));
}
else {
buf = m_buffers.front();
m_buffers.pop_front();
}
if (buf) {
if (!is_static) {
if (!m_vars[i].flags.sink_addr) {
// increment buffer reference
OFFLOAD_TIMER_START(c_offload_target_add_buffer_refs);
BufferAddRef(buf);
OFFLOAD_TIMER_STOP(c_offload_target_add_buffer_refs);
}
add_ref_count(buf, 0 == m_vars[i].flags.sink_addr);
}
ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset + m_vars[i].offset;
}
*ptr_addr = ptr;
}
else if (m_vars[i].flags.sink_addr) {
void *buf;
m_in.receive_data(&buf, sizeof(buf));
ptr = static_cast<char*>(buf) +
m_vars[i].mic_offset + m_vars[i].offset;
*ptr_addr = ptr;
}
break;
default:
LIBOFFLOAD_ERROR(c_unknown_var_type, type);
abort();
}
// Release obsolete buffers for stack of persistent objects.
// The vardesc with i==0 and flags.is_stack_buf==TRUE is always for
// stack buffer pointer.
if (i == 0 &&
m_vars[i].flags.is_stack_buf &&
!m_vars[i].direction.bits &&
m_vars[i].alloc_if &&
m_vars[i].size != 0) {
for (int j=0; j < m_vars[i].size; j++) {
void *buf;
m_in.receive_data(&buf, sizeof(buf));
OFFLOAD_TRACE(4, "Releasing stack buffer %p\n", buf);
BufferReleaseRef(buf);
ref_data.erase(buf);
}
}
// Do copyin
switch (m_vars_extra[i].type_dst) {
case c_data_ptr_array:
break;
case c_data:
case c_void_ptr:
case c_void_ptr_ptr:
case c_cean_var:
if (m_vars[i].direction.in &&
!m_vars[i].flags.is_static_dstn) {
int64_t size;
int64_t disp;
char* ptr = m_vars[i].into ?
static_cast<char*>(m_vars[i].into) :
static_cast<char*>(m_vars[i].ptr);
if (m_vars_extra[i].type_dst == c_cean_var) {
m_in.receive_data((&size), sizeof(int64_t));
m_in.receive_data((&disp), sizeof(int64_t));
}
else {
size = m_vars[i].size;
disp = 0;
}
m_in.receive_data(ptr + disp, size);
}
break;
case c_dv:
if (m_vars[i].direction.bits ||
m_vars[i].alloc_if ||
m_vars[i].free_if) {
char* ptr = m_vars[i].into ?
static_cast<char*>(m_vars[i].into) :
static_cast<char*>(m_vars[i].ptr);
m_in.receive_data(ptr + sizeof(uint64_t),
m_vars[i].size - sizeof(uint64_t));
}
break;
case c_string_ptr:
case c_data_ptr:
case c_string_ptr_ptr:
case c_data_ptr_ptr:
case c_cean_var_ptr:
case c_cean_var_ptr_ptr:
case c_dv_ptr:
case c_dv_data:
case c_dv_ptr_data:
case c_dv_data_slice:
case c_dv_ptr_data_slice:
break;
case c_func_ptr:
case c_func_ptr_ptr:
if (m_vars[i].direction.in) {
m_in.receive_func_ptr((const void**) m_vars[i].ptr);
}
break;
default:
LIBOFFLOAD_ERROR(c_unknown_var_type, m_vars_extra[i].type_dst);
abort();
}
}
OFFLOAD_TRACE(1, "Total copyin data received from host: [%lld] bytes\n",
m_in.get_tfr_size());
OFFLOAD_TIMER_STOP(c_offload_target_scatter_inputs);
OFFLOAD_TIMER_START(c_offload_target_compute);
}