checkpoint: first pass at wide vmap baking

This commit is contained in:
Ryan Fleury
2025-10-13 17:00:01 -07:00
parent b688c69f13
commit 89be41fda9
+213 -1
View File
@@ -446,7 +446,6 @@ rdim_bake(Arena *arena, RDIM_BakeParams *params)
// * we must process _all_ of the changes that apply at this voff before moving on
//
RDI_U64 voff = key_ptr->key;
for(;key_ptr < key_opl && key_ptr->key == voff; key_ptr += 1)
{
RDIM_VMapMarker *marker = (RDIM_VMapMarker*)key_ptr->val;
@@ -537,6 +536,219 @@ rdim_bake(Arena *arena, RDIM_BakeParams *params)
}
lane_sync();
//////////////////////////////////////////////////////////////
//- rjf: @rdim_bake_stage bake all vmaps (NEW)
//
#if 0
ProfScope("bake all vmaps (NEW)")
{
Temp scratch = scratch_begin(&arena, 1);
RDI_U64 count = rdim_shared->scope_vmap_count;
RDIM_SortKey *keys = rdim_shared->scope_vmap_keys;
//- rjf: do vmap entry generation for portions of the keys array
typedef struct VMapRangeTask VMapRangeTask;
struct VMapRangeTask
{
VMapRangeTask *next;
U32 idx;
};
Rng1U64 range = lane_range(count);
RDI_U64 lane_vmap_count_cap = dim_1u64(range);
RDI_VMapEntry *lane_vmap = push_array(arena, RDI_VMapEntry, lane_vmap_count_cap);
VMapRangeTask *top_range_task = 0;
RDI_U64 lane_vmap_count_actual = 0;
ProfScope("do vmap entry generation for portions of the keys array")
{
VMapRangeTask *free_range_task = 0;
RDI_VMapEntry *lane_vmap_ptr = lane_vmap;
RDIM_SortKey *key_ptr = keys + range.min;
RDIM_SortKey *key_opl = keys + range.max;
for(;key_ptr < key_opl;)
{
// rjf: get initial range index from range task stack
RDI_U32 initial_idx = (RDI_U32)0xffffffff;
if(top_range_task != 0)
{
initial_idx = top_range_task->idx;
}
// rjf: update range task stack
//
// * we must process _all_ of the changes that apply at this voff before moving on
//
RDI_U64 voff = key_ptr->key;
B32 failed_pop = 0;
for(;key_ptr < key_opl && key_ptr->key == voff && !failed_pop; key_ptr += 1)
{
RDIM_VMapMarker *marker = (RDIM_VMapMarker *)key_ptr->val;
RDI_U32 idx = marker->idx;
// rjf: range begin -> push to stack
if(marker->begin_range)
{
VMapRangeTask *task = free_range_task;
if(task != 0)
{
RDIM_SLLStackPop(free_range_task);
}
else
{
task = rdim_push_array(scratch.arena, VMapRangeTask, 1);
}
RDIM_SLLStackPush(top_range_task, task);
task->idx = idx;
}
// rjf: range ending -> pop matching node from stack (not always the top)
else
{
VMapRangeTask **ptr_in = &top_range_task;
VMapRangeTask *match = 0;
for(VMapRangeTask *node = top_range_task; node != 0;)
{
if(node->idx == idx)
{
match = node;
break;
}
ptr_in = &node->next;
node = node->next;
}
if(match != 0)
{
*ptr_in = match->next;
RDIM_SLLStackPush(free_range_task, match);
}
else
{
failed_pop = 1;
}
}
}
// rjf: get final map state from tracker stack
RDI_U32 final_idx = 0;
if(top_range_task != 0)
{
final_idx = top_range_task->idx;
}
// rjf: if final is different from initial - emit new vmap entry
if(failed_pop)
{
lane_vmap_ptr->voff = voff;
lane_vmap_ptr->idx = 0xffffffff;
lane_vmap_ptr += 1;
}
else if(final_idx != initial_idx)
{
lane_vmap_ptr->voff = voff;
lane_vmap_ptr->idx = final_idx;
lane_vmap_ptr += 1;
}
}
lane_vmap_count_actual = (lane_vmap_ptr - lane_vmap);
}
//- rjf: collect all per-lane artifacts
RDI_U64 *lane_vmaps_counts = 0;
VMapRangeTask **lane_leftover_tasks = 0;
ProfScope("collect all lane vmap counts")
{
if(lane_idx() == 0)
{
lane_vmaps_counts = push_array(scratch.arena, RDI_U64, lane_count());
lane_leftover_tasks = push_array(arena, VMapRangeTask *, lane_count());
}
lane_sync_u64(&lane_vmaps_counts, 0);
lane_sync_u64(&lane_leftover_tasks, 0);
lane_vmaps_counts[lane_idx()] = lane_vmap_count_actual;
lane_leftover_tasks[lane_idx()] = top_range_task;
}
lane_sync();
//- rjf: apply leftovers to per-lane vmaps
if(lane_idx() > 0)
{
VMapRangeTask *top_leftover_task = lane_leftover_tasks[lane_idx()-1];
for EachIndex(idx, lane_vmap_count_actual)
{
if(!top_leftover_task)
{
break;
}
if(lane_vmap[idx].idx == 0xffffffff)
{
SLLStackPop(top_leftover_task);
if(top_leftover_task)
{
lane_vmap[idx].idx = top_leftover_task->idx;
}
}
}
}
lane_idx();
//- rjf: lay out all lane vmaps into single range
RDI_U64 *lane_vmaps_offs = 0;
RDI_U64 lane_vmap_count_total = 0;
ProfScope("lay out all lane vmaps into single range")
{
if(lane_idx() == 0)
{
lane_vmaps_offs = push_array(scratch.arena, RDI_U64, lane_count());
RDI_U64 off = 0;
for EachIndex(lidx, lane_count())
{
lane_vmaps_offs[lidx] = off;
off += lane_vmaps_counts[lidx];
}
lane_vmap_count_total = off;
}
}
lane_sync_u64(&lane_vmaps_offs, 0);
lane_sync_u64(&lane_vmap_count_total, 0);
//- rjf: join all lane vmaps
RDI_U64 vmap_count = lane_vmap_count_total + 1;
RDI_VMapEntry *vmap = 0;
ProfScope("join all lane vmaps")
{
if(lane_idx() == 0)
{
vmap = push_array_no_zero(arena, RDI_VMapEntry, vmap_count);
}
lane_sync_u64(&vmap, 0);
MemoryCopy(vmap + lane_vmaps_offs[lane_idx()], lane_vmap, sizeof(lane_vmap[0])*lane_vmaps_counts[lane_idx()]);
}
lane_sync();
//- rjf: combine duplicate neighbors
RDI_U64 vmap_count__deduplicated = 0;
ProfScope("combine duplicate neighbors") if(lane_idx() == 0)
{
RDI_VMapEntry *vmap_ptr = vmap;
RDI_VMapEntry *vmap_opl = vmap + vmap_count;
RDI_VMapEntry *vmap_out = vmap;
for(;vmap_ptr < vmap_opl;)
{
RDI_VMapEntry *vmap_range_first = vmap_ptr;
RDI_U64 idx = vmap_ptr->idx;
vmap_ptr += 1;
for(;vmap_ptr < vmap_opl && vmap_ptr->idx == idx;) vmap_ptr += 1;
rdim_memcpy_struct(vmap_out, vmap_range_first);
vmap_out += 1;
}
vmap_count__deduplicated = (RDI_U64)(vmap_out - vmap);
}
lane_sync_u64(&vmap_count__deduplicated, 0);
scratch_end(scratch);
}
lane_sync();
#endif
//////////////////////////////////////////////////////////////
//- rjf: @rdim_bake_stage build interned path tree
//