void __dv_desc_dump(const char *name, const ArrDesc *dvp)
{
OFFLOAD_TRACE(3, "%s DV %p\n", name, dvp);
if (dvp != 0) {
OFFLOAD_TRACE(3,
" dv->Base = 0x%lx\n"
" dv->Len = 0x%lx\n"
" dv->Offset = 0x%lx\n"
" dv->Flags = 0x%lx\n"
" dv->Rank = 0x%lx\n"
" dv->Resrvd = 0x%lx\n",
dvp->Base,
dvp->Len,
dvp->Offset,
dvp->Flags,
dvp->Rank,
dvp->Reserved);
for (int i = 0 ; i < dvp->Rank; i++) {
OFFLOAD_TRACE(3,
" (%d) Extent=%ld, Multiplier=%ld, LowerBound=%ld\n",
i,
dvp->Dim[i].Extent,
dvp->Dim[i].Mult,
dvp->Dim[i].LowerBound);
}
}
}
void OffloadDescriptor::merge_var_descs(
VarDesc *vars,
VarDesc2 *vars2,
int vars_total
)
{
// number of variable descriptors received from host and generated
// locally should match
if (m_vars_total < vars_total) {
LIBOFFLOAD_ERROR(c_merge_var_descs1);
exit(1);
}
for (int i = 0; i < m_vars_total; i++) {
if (i < vars_total) {
// variable type must match
if (m_vars[i].type.bits != vars[i].type.bits) {
LIBOFFLOAD_ERROR(c_merge_var_descs2);
exit(1);
}
m_vars[i].ptr = vars[i].ptr;
m_vars[i].into = vars[i].into;
const char *var_sname = "";
if (vars2 != NULL) {
if (vars2[i].sname != NULL) {
var_sname = vars2[i].sname;
}
}
OFFLOAD_DEBUG_TRACE_1(2, get_offload_number(), c_offload_var,
" VarDesc %d, var=%s, %s, %s\n",
i, var_sname,
vardesc_direction_as_string[m_vars[i].direction.bits],
vardesc_type_as_string[m_vars[i].type.src]);
if (vars2 != NULL && vars2[i].dname != NULL) {
OFFLOAD_TRACE(2, " into=%s, %s\n", vars2[i].dname,
vardesc_type_as_string[m_vars[i].type.dst]);
}
}
OFFLOAD_TRACE(2,
" type_src=%d, type_dstn=%d, direction=%d, "
"alloc_if=%d, free_if=%d, align=%d, mic_offset=%d, flags=0x%x, "
"offset=%lld, size=%lld, count/disp=%lld, ptr=%p into=%p\n",
m_vars[i].type.src,
m_vars[i].type.dst,
m_vars[i].direction.bits,
m_vars[i].alloc_if,
m_vars[i].free_if,
m_vars[i].align,
m_vars[i].mic_offset,
m_vars[i].flags.bits,
m_vars[i].offset,
m_vars[i].size,
m_vars[i].count,
m_vars[i].ptr,
m_vars[i].into);
}
}
static void BufReleaseRef(void * buf)
{
mutex_locker_t locker(add_ref_lock);
RefInfo * info = ref_data[buf];
if (info) {
--info->count;
if (info->count == 0 && info->is_added) {
OFFLOAD_TRACE(1, "Calling COIBufferReleaseRef AddRef count = %d\n",
((RefInfo *) ref_data[buf])->count);
BufferReleaseRef(buf);
info->is_added = 0;
}
}
}
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::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);
}
void OffloadDescriptor::merge_var_descs(
VarDesc *vars,
VarDesc2 *vars2,
int vars_total
)
{
// number of variable descriptors received from host and generated
// locally should match
if (m_vars_total < vars_total) {
LIBOFFLOAD_ERROR(c_merge_var_descs1);
exit(1);
}
for (int i = 0; i < m_vars_total; i++) {
// instead of m_vars[i].type.src we will use m_vars_extra[i].type_src
if (i < vars_total) {
// variable type must match
if (m_vars[i].type.bits != vars[i].type.bits) {
OFFLOAD_TRACE(2,
"m_vars[%d].type.bits=%08x, vars[%d].type.bits=%08x\n",
i, m_vars[i].type.bits, i, vars[i].type.bits);
LIBOFFLOAD_ERROR(c_merge_var_descs2);
exit(1);
}
if (m_vars[i].type.src == c_extended_type) {
VarDescExtendedType *etype =
reinterpret_cast<VarDescExtendedType*>(vars[i].ptr);
m_vars_extra[i].type_src = etype->extended_type;
m_vars[i].ptr = etype->ptr;
}
else {
m_vars_extra[i].type_src = m_vars[i].type.src;
if (!(m_vars[i].flags.use_device_ptr &&
m_vars[i].type.src == c_dv)) {
m_vars[i].ptr = vars[i].ptr;
}
}
// instead of m_vars[i].type.dst we will use m_vars_extra[i].type_dst
if (m_vars[i].type.dst == c_extended_type && i < vars_total) {
VarDescExtendedType *etype =
reinterpret_cast<VarDescExtendedType*>(vars[i].into);
m_vars_extra[i].type_dst = etype->extended_type;
m_vars[i].into = etype->ptr;
}
else {
m_vars_extra[i].type_dst = m_vars[i].type.dst;
m_vars[i].into = vars[i].into;
}
const char *var_sname = "";
if (vars2 != NULL) {
if (vars2[i].sname != NULL) {
var_sname = vars2[i].sname;
}
}
OFFLOAD_DEBUG_TRACE_1(2, get_offload_number(), c_offload_var,
" VarDesc %d, var=%s, %s, %s\n",
i, var_sname,
vardesc_direction_as_string[m_vars[i].direction.bits],
vardesc_type_as_string[m_vars_extra[i].type_src]);
if (vars2 != NULL && vars2[i].dname != NULL) {
OFFLOAD_TRACE(2, " into=%s, %s\n", vars2[i].dname,
vardesc_type_as_string[m_vars_extra[i].type_dst]);
}
}
else {
m_vars_extra[i].type_src = m_vars[i].type.src;
m_vars_extra[i].type_dst = m_vars[i].type.dst;
}
OFFLOAD_TRACE(2,
" type_src=%d, type_dstn=%d, direction=%d, "
"alloc_if=%d, free_if=%d, align=%d, mic_offset=%d, flags=0x%x, "
"offset=%lld, size=%lld, count/disp=%lld, ptr=%p into=%p\n",
m_vars_extra[i].type_src,
m_vars_extra[i].type_dst,
m_vars[i].direction.bits,
m_vars[i].alloc_if,
m_vars[i].free_if,
m_vars[i].align,
m_vars[i].mic_offset,
m_vars[i].flags.bits,
m_vars[i].offset,
m_vars[i].size,
m_vars[i].count,
m_vars[i].ptr,
m_vars[i].into);
}
}
