Skip to content

ibrl: nontemporal memcpy on bundle path #7580

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
cavemanloverboy wants to merge 1 commit into
base: main
Choose a base branch
from
Open

ibrl: nontemporal memcpy on bundle path #7580

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
124 changes: 61 additions & 63 deletions src/disco/pack/fd_pack.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1128,6 +1128,64 @@ delete_worst( fd_pack_t * pack,
return delete_transaction( pack, worst, 1, 1 );
}

static inline
void fd_pack_copy_txnp( fd_txn_p_t * out, fd_pack_ord_txn_t const * cur ) {
fd_txn_t const * txn = TXN(cur->txn);
if(
#if FD_HAS_AVX512 && FD_PACK_USE_NON_TEMPORAL_MEMCPY
FD_LIKELY( cur->txn->payload_sz>=1024UL )
#else
0
#endif
) {
#if FD_HAS_AVX512 && FD_PACK_USE_NON_TEMPORAL_MEMCPY
_mm512_stream_si512( (void*)(out->payload+ 0UL), _mm512_load_epi64( cur->txn->payload+ 0UL ) );
_mm512_stream_si512( (void*)(out->payload+ 64UL), _mm512_load_epi64( cur->txn->payload+ 64UL ) );
_mm512_stream_si512( (void*)(out->payload+ 128UL), _mm512_load_epi64( cur->txn->payload+ 128UL ) );
_mm512_stream_si512( (void*)(out->payload+ 192UL), _mm512_load_epi64( cur->txn->payload+ 192UL ) );
_mm512_stream_si512( (void*)(out->payload+ 256UL), _mm512_load_epi64( cur->txn->payload+ 256UL ) );
_mm512_stream_si512( (void*)(out->payload+ 320UL), _mm512_load_epi64( cur->txn->payload+ 320UL ) );
_mm512_stream_si512( (void*)(out->payload+ 384UL), _mm512_load_epi64( cur->txn->payload+ 384UL ) );
_mm512_stream_si512( (void*)(out->payload+ 448UL), _mm512_load_epi64( cur->txn->payload+ 448UL ) );
_mm512_stream_si512( (void*)(out->payload+ 512UL), _mm512_load_epi64( cur->txn->payload+ 512UL ) );
_mm512_stream_si512( (void*)(out->payload+ 576UL), _mm512_load_epi64( cur->txn->payload+ 576UL ) );
_mm512_stream_si512( (void*)(out->payload+ 640UL), _mm512_load_epi64( cur->txn->payload+ 640UL ) );
_mm512_stream_si512( (void*)(out->payload+ 704UL), _mm512_load_epi64( cur->txn->payload+ 704UL ) );
_mm512_stream_si512( (void*)(out->payload+ 768UL), _mm512_load_epi64( cur->txn->payload+ 768UL ) );
_mm512_stream_si512( (void*)(out->payload+ 832UL), _mm512_load_epi64( cur->txn->payload+ 832UL ) );
_mm512_stream_si512( (void*)(out->payload+ 896UL), _mm512_load_epi64( cur->txn->payload+ 896UL ) );
_mm512_stream_si512( (void*)(out->payload+ 960UL), _mm512_load_epi64( cur->txn->payload+ 960UL ) );
_mm512_stream_si512( (void*)(out->payload+1024UL), _mm512_load_epi64( cur->txn->payload+1024UL ) );
_mm512_stream_si512( (void*)(out->payload+1088UL), _mm512_load_epi64( cur->txn->payload+1088UL ) );
_mm512_stream_si512( (void*)(out->payload+1152UL), _mm512_load_epi64( cur->txn->payload+1152UL ) );
_mm512_stream_si512( (void*)(out->payload+1216UL), _mm512_load_epi64( cur->txn->payload+1216UL ) );
/* Copied out to 1280 bytes, which copies some other fields we needed to
copy anyway. */
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, payload_sz )+sizeof(((fd_txn_p_t*)NULL)->payload_sz )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, blockhash_slot )+sizeof(((fd_txn_p_t*)NULL)->blockhash_slot)<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, scheduler_arrival_time_nanos )+sizeof(((fd_txn_p_t*)NULL)->scheduler_arrival_time_nanos )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, source_tpu )+sizeof(((fd_txn_p_t*)NULL)->source_tpu )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, source_ipv4 )+sizeof(((fd_txn_p_t*)NULL)->source_ipv4 )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, flags )+sizeof(((fd_txn_p_t*)NULL)->flags )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, _ ) <=1280UL, nt_memcpy );
const ulong offset_into_txn = 1280UL - offsetof(fd_txn_p_t, _ );
fd_memcpy( offset_into_txn+(uchar *)TXN(out), offset_into_txn+(uchar const *)txn,
fd_ulong_max( offset_into_txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) )-offset_into_txn );
#endif
} else {
fd_memcpy( out->payload, cur->txn->payload, cur->txn->payload_sz );
fd_memcpy( TXN(out), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );
out->payload_sz = cur->txn->payload_sz;
out->pack_cu.requested_exec_plus_acct_data_cus = cur->txn->pack_cu.requested_exec_plus_acct_data_cus;
out->pack_cu.non_execution_cus = cur->txn->pack_cu.non_execution_cus;
out->scheduler_arrival_time_nanos = cur->txn->scheduler_arrival_time_nanos;
out->source_tpu = cur->txn->source_tpu;
out->source_ipv4 = cur->txn->source_ipv4;
out->flags = cur->txn->flags;
}
}


static inline int
validate_transaction( fd_pack_t * pack,
fd_pack_ord_txn_t const * ord,
Expand Down Expand Up @@ -1925,58 +1983,7 @@ fd_pack_schedule_impl( fd_pack_t * pack,
FD_PACK_BITSET_OR( bitset_rw_in_use, cur->rw_bitset );
FD_PACK_BITSET_OR( bitset_w_in_use, cur->w_bitset );

if(
#if FD_HAS_AVX512 && FD_PACK_USE_NON_TEMPORAL_MEMCPY
FD_LIKELY( cur->txn->payload_sz>=1024UL )
#else
0
#endif
) {
#if FD_HAS_AVX512 && FD_PACK_USE_NON_TEMPORAL_MEMCPY
_mm512_stream_si512( (void*)(out->payload+ 0UL), _mm512_load_epi64( cur->txn->payload+ 0UL ) );
_mm512_stream_si512( (void*)(out->payload+ 64UL), _mm512_load_epi64( cur->txn->payload+ 64UL ) );
_mm512_stream_si512( (void*)(out->payload+ 128UL), _mm512_load_epi64( cur->txn->payload+ 128UL ) );
_mm512_stream_si512( (void*)(out->payload+ 192UL), _mm512_load_epi64( cur->txn->payload+ 192UL ) );
_mm512_stream_si512( (void*)(out->payload+ 256UL), _mm512_load_epi64( cur->txn->payload+ 256UL ) );
_mm512_stream_si512( (void*)(out->payload+ 320UL), _mm512_load_epi64( cur->txn->payload+ 320UL ) );
_mm512_stream_si512( (void*)(out->payload+ 384UL), _mm512_load_epi64( cur->txn->payload+ 384UL ) );
_mm512_stream_si512( (void*)(out->payload+ 448UL), _mm512_load_epi64( cur->txn->payload+ 448UL ) );
_mm512_stream_si512( (void*)(out->payload+ 512UL), _mm512_load_epi64( cur->txn->payload+ 512UL ) );
_mm512_stream_si512( (void*)(out->payload+ 576UL), _mm512_load_epi64( cur->txn->payload+ 576UL ) );
_mm512_stream_si512( (void*)(out->payload+ 640UL), _mm512_load_epi64( cur->txn->payload+ 640UL ) );
_mm512_stream_si512( (void*)(out->payload+ 704UL), _mm512_load_epi64( cur->txn->payload+ 704UL ) );
_mm512_stream_si512( (void*)(out->payload+ 768UL), _mm512_load_epi64( cur->txn->payload+ 768UL ) );
_mm512_stream_si512( (void*)(out->payload+ 832UL), _mm512_load_epi64( cur->txn->payload+ 832UL ) );
_mm512_stream_si512( (void*)(out->payload+ 896UL), _mm512_load_epi64( cur->txn->payload+ 896UL ) );
_mm512_stream_si512( (void*)(out->payload+ 960UL), _mm512_load_epi64( cur->txn->payload+ 960UL ) );
_mm512_stream_si512( (void*)(out->payload+1024UL), _mm512_load_epi64( cur->txn->payload+1024UL ) );
_mm512_stream_si512( (void*)(out->payload+1088UL), _mm512_load_epi64( cur->txn->payload+1088UL ) );
_mm512_stream_si512( (void*)(out->payload+1152UL), _mm512_load_epi64( cur->txn->payload+1152UL ) );
_mm512_stream_si512( (void*)(out->payload+1216UL), _mm512_load_epi64( cur->txn->payload+1216UL ) );
/* Copied out to 1280 bytes, which copies some other fields we needed to
copy anyway. */
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, payload_sz )+sizeof(((fd_txn_p_t*)NULL)->payload_sz )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, blockhash_slot )+sizeof(((fd_txn_p_t*)NULL)->blockhash_slot)<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, scheduler_arrival_time_nanos )+sizeof(((fd_txn_p_t*)NULL)->scheduler_arrival_time_nanos )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, source_tpu )+sizeof(((fd_txn_p_t*)NULL)->source_tpu )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, source_ipv4 )+sizeof(((fd_txn_p_t*)NULL)->source_ipv4 )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, flags )+sizeof(((fd_txn_p_t*)NULL)->flags )<=1280UL, nt_memcpy );
FD_STATIC_ASSERT( offsetof(fd_txn_p_t, _ ) <=1280UL, nt_memcpy );
const ulong offset_into_txn = 1280UL - offsetof(fd_txn_p_t, _ );
fd_memcpy( offset_into_txn+(uchar *)TXN(out), offset_into_txn+(uchar const *)txn,
fd_ulong_max( offset_into_txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) )-offset_into_txn );
#endif
} else {
fd_memcpy( out->payload, cur->txn->payload, cur->txn->payload_sz );
fd_memcpy( TXN(out), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );
out->payload_sz = cur->txn->payload_sz;
out->pack_cu.requested_exec_plus_acct_data_cus = cur->txn->pack_cu.requested_exec_plus_acct_data_cus;
out->pack_cu.non_execution_cus = cur->txn->pack_cu.non_execution_cus;
out->scheduler_arrival_time_nanos = cur->txn->scheduler_arrival_time_nanos;
out->source_tpu = cur->txn->source_tpu;
out->source_ipv4 = cur->txn->source_ipv4;
out->flags = cur->txn->flags;
}
fd_pack_copy_txnp( out, cur );
out++;

for( fd_txn_acct_iter_t iter=fd_txn_acct_iter_init( txn, FD_TXN_ACCT_CAT_WRITABLE );
Expand Down Expand Up @@ -2401,16 +2408,7 @@ fd_pack_try_schedule_bundle( fd_pack_t * pack,
_next = treap_rev_iter_next( _cur, pool );

fd_pack_ord_txn_t * cur = treap_rev_iter_ele( _cur, pool );
fd_txn_t const * txn = TXN(cur->txn);
fd_memcpy( out->payload, cur->txn->payload, cur->txn->payload_sz );
fd_memcpy( TXN(out), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );
out->payload_sz = cur->txn->payload_sz;
out->pack_cu.requested_exec_plus_acct_data_cus = cur->txn->pack_cu.requested_exec_plus_acct_data_cus;
out->pack_cu.non_execution_cus = cur->txn->pack_cu.non_execution_cus;
out->scheduler_arrival_time_nanos = cur->txn->scheduler_arrival_time_nanos;
out->source_tpu = cur->txn->source_tpu;
out->source_ipv4 = cur->txn->source_ipv4;
out->flags = cur->txn->flags;
fd_pack_copy_txnp( out, cur );
out++;

pack->cumulative_block_cost += cur->compute_est;
Expand Down Expand Up @@ -3225,4 +3223,4 @@ fd_pack_verify( fd_pack_t * pack,
}

void * fd_pack_leave ( fd_pack_t * pack ) { FD_COMPILER_MFENCE(); return (void *)pack; }
void * fd_pack_delete( void * mem ) { FD_COMPILER_MFENCE(); return mem; }
void * fd_pack_delete( void * mem ) { FD_COMPILER_MFENCE(); return mem; }