summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/cuda-sim/cuda-sim.cc12
-rw-r--r--src/cuda-sim/instructions.cc43
2 files changed, 41 insertions, 14 deletions
diff --git a/src/cuda-sim/cuda-sim.cc b/src/cuda-sim/cuda-sim.cc
index 8f684e2..9735a0e 100644
--- a/src/cuda-sim/cuda-sim.cc
+++ b/src/cuda-sim/cuda-sim.cc
@@ -1403,15 +1403,15 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id)
_memory_op_t insn_memory_op = no_memory_op;
unsigned insn_data_size = 0;
if ( (pI->has_memory_read() || pI->has_memory_write()) ) {
- //if(!((inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP)))
+ if(!((inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP)))
//if(!((inst_opcode==MMA_LD_OP||inst_opcode==VP_LD_OP)))
- //{
+ {
insn_memaddr = last_eaddr();
insn_space = last_space();
unsigned to_type = pI->get_type();
insn_data_size = datatype2size(to_type);
insn_memory_op = pI->has_memory_read() ? memory_load : memory_store;
- //}
+ }
}
if ( pI->get_opcode() == BAR_OP && pI->barrier_op() == RED_OPTION) {
@@ -1485,14 +1485,14 @@ void ptx_thread_info::ptx_exec_inst( warp_inst_t &inst, unsigned lane_id)
// "Return values"
if(!skip) {
- //if(!((inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP)))
+ if(!((inst_opcode==MMA_LD_OP||inst_opcode==MMA_ST_OP||inst_opcode==VP_LD_OP||inst_opcode==VP_ST_OP)))
//if(!((inst_opcode==MMA_LD_OP||inst_opcode==VP_LD_OP)))
- //{
+ {
inst.space = insn_space;
inst.set_addr(lane_id, insn_memaddr);
inst.data_size = insn_data_size; // simpleAtomicIntrinsics
assert( inst.memory_op == insn_memory_op );
- //}
+ }
}
} catch ( int x ) {
diff --git a/src/cuda-sim/instructions.cc b/src/cuda-sim/instructions.cc
index 77e8e71..7477b49 100644
--- a/src/cuda-sim/instructions.cc
+++ b/src/cuda-sim/instructions.cc
@@ -3180,7 +3180,7 @@ void vp_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
thread->m_last_memory_space = space;
}
}
-void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
+void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst )
{
size_t size;
unsigned smid;
@@ -3213,6 +3213,9 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
memory_space *mem = NULL;
addr_t addr = addr_reg.u32;
+
+ new_addr_type mem_txn_addr[MAX_ACCESSES_PER_INSN_PER_THREAD];
+ int num_mem_txn=0;
smid = thread->get_hw_sid();
if( whichspace(addr) == shared_space ) {
@@ -3240,7 +3243,8 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
//mem->write(new_addr+4*acc_float_offset(k,wmma_layout,stride),size/8,&v[k].s64,thread,pI);
push_addr=new_addr+4*acc_float_offset(k,wmma_layout,stride);
mem->write(push_addr,size/8,&v[k].s64,thread,pI);
-
+ mem_txn_addr[num_mem_txn++]=push_addr;
+
if(g_debug_instruction){
printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,v[0].s64,v[1].s64,v[2].s64,v[3].s64,v[4].s64,v[5].s64,v[6].s64,v[7].s64);
float temp;
@@ -3252,7 +3256,6 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
}
printf("\n");
}
-
}
else if(type==F16_TYPE){
if(wmma_layout==ROW){
@@ -3265,14 +3268,21 @@ void mma_st_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
push_addr=new_addr+k*2*stride;
mem->write(push_addr,size/8,&nw_v[k].s64,thread,pI);
}
+ if(k%2==0)
+ mem_txn_addr[num_mem_txn++]=push_addr;
+
if(g_debug_instruction)
printf("wmma:store:thread%d=%x,%x,%x,%x,%x,%x,%x,%x\n",thrd,nw_v[0].s64,nw_v[1].s64,nw_v[2].s64,nw_v[3].s64,nw_v[4].s64,nw_v[5].s64,nw_v[6].s64,nw_v[7].s64);
}
}
delete [] v;
- thread->m_last_effective_address = addr;
- thread->m_last_memory_space = space;
+ inst.space = space;
+ inst.set_addr(thrd, (new_addr_type *)mem_txn_addr , num_mem_txn);
+ inst.data_size = 4; // 4 byte transaction
+ assert( inst.memory_op == insn_memory_op );
+ //thread->m_last_effective_address = addr;
+ //thread->m_last_memory_space = space;
}
}
void vp_ld_impl(const ptx_instruction *pI, core_t *core, warp_inst_t inst)
@@ -3423,7 +3433,7 @@ void vp_ld_impl(const ptx_instruction *pI, core_t *core, warp_inst_t inst)
}
-void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
+void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t &inst )
{
size_t size;
int t,i;
@@ -3465,6 +3475,9 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
addr_t new_addr = addr+thread_group_offset(thrd,wmma_type,wmma_layout,type,stride)*size/8;
addr_t fetch_addr;
+ new_addr_type mem_txn_addr[MAX_ACCESSES_PER_INSN_PER_THREAD];
+ int num_mem_txn=0;
+
if(wmma_type==LOAD_A){
for(i=0;i<16;i++){
if(wmma_layout==ROW){
@@ -3480,7 +3493,10 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
else{
printf("mma_ld:wrong_layout_type\n");
abort();
+
}
+ if(i%2==0)
+ mem_txn_addr[num_mem_txn++]=fetch_addr;
}
}
else if(wmma_type==LOAD_B){
@@ -3499,6 +3515,8 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
printf("mma_ld:wrong_layout_type\n");
abort();
}
+ if(i%2==0)
+ mem_txn_addr[num_mem_txn++]=fetch_addr;
}
}
else if(wmma_type==LOAD_C){
@@ -3518,11 +3536,14 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
printf("mma_ld:wrong_type\n");
abort();
}
+ if(i%2==0)
+ mem_txn_addr[num_mem_txn++]=fetch_addr;
}
else if(type==F32_TYPE){
//mem->read(new_addr+4*acc_float_offset(i,wmma_layout,stride),size/8,&data[i].s64);
fetch_addr=new_addr+4*acc_float_offset(i,wmma_layout,stride);
mem->read(fetch_addr,size/8,&data[i].s64);
+ mem_txn_addr[num_mem_txn++]=fetch_addr;
}
else{
printf("wrong type");
@@ -3534,6 +3555,12 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
printf("wrong wmma type\n");;
abort();
}
+ //generate timing memory request
+ inst.space = space;
+ inst.set_addr(thrd, (new_addr_type *)mem_txn_addr , num_mem_txn);
+ inst.data_size = 4; // 4 byte transaction
+ assert( inst.memory_op == insn_memory_op );
+
if(g_debug_instruction){
if(type==F16_TYPE){
printf("\nmma_ld:thread%d= ",thrd);
@@ -3598,8 +3625,8 @@ void mma_ld_impl( const ptx_instruction *pI, core_t *core, warp_inst_t inst )
}
}
- thread->m_last_effective_address = addr;
- thread->m_last_memory_space = space;
+ //thread->m_last_effective_address = addr;
+ //thread->m_last_memory_space = space;
}
}
generated by cgit v1.3 (git 2.53.0) at 2026-07-11 01:29:02 +0000