github-mirrors/vortex
1
0
Fork 0
mirror of https://github.com/vortexgpgpu/vortex.git synced 2025-04-23 21:39:10 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

Script checkin and code cleanup

Browse source
This commit is contained in:
Varsha Singhania 2024-06-18 02:19:57 -04:00
parent 99c6a1af5a
commit 0e3badf723
2 changed files with 31 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

22
run_final.sh Executable file
View file

@ -0,0 +1,22 @@
# Define arrays for threads, warps, and matrix sizes
matrix_sizes=(16 32 64 128 256 512)
tcsizes=(8 16 32)
tcnums=(4 8 16 32)
#lsulanes=(4 16)
#cores=(32)
# Loop through each combination of threads and warps
for size in "${matrix_sizes[@]}"; do
sed -i "s/OPTS ?= -n[0-9]\+/OPTS ?= -n${size}/" ../tests/regression/matmul/Makefile
sed -i "s/OPTS ?= -n[0-9]\+/OPTS ?= -n${size}/" tests/regression/matmul/Makefile
echo "Matrix size changed to ${size} in Makefile"
for tcsize in "${tcsizes[@]}"; do
for tcnum in "${tcnums[@]}"; do
log_name="sim_final/mat${size}/tcsize${tcsize}_tcnum${tcnum}_32w32t"
command="./ci/blackbox.sh --cores=4 --app=matmul --driver=simx --threads=32 --warps=32 --tc_size=${tcsize} --tc_num=${tcnum} --rebuild=1 --perf=1 > ${log_name} 2>&1"
echo "$command"
eval "$command"
done
done
done

28
sim/simx/execute.cpp
View file

@ -1432,8 +1432,6 @@ void Emulator::execute(const Instr &instr, uint32_t wid, instr_trace_t *trace) {
uint32_t data_bytes_store;
uint32_t num_threads_per_tc = MAX (1, num_threads/TC_per_warp);
//int num_warps = MIN()
//int active_tcs = MIN (TC_per_warp, num_output_tiles/num_warps)
//LOAD
if(num_threads > tc_size*tc_size*n_tiles*TC_per_warp)
{
@ -1448,11 +1446,6 @@ void Emulator::execute(const Instr &instr, uint32_t wid, instr_trace_t *trace) {
data_bytes_load = mem_bytes*num_data_per_thread;
//STORE
// DP(3, "DEBUG :: num_threads = " << num_threads);
// DP(3, "DEBUG :: tc_size*tc_size = " << tc_size*tc_size);
//DP(3, "imm = " << immsrc);
if(num_threads > tc_size*tc_size*TC_per_warp)
{
num_threads_actv_st = tc_size*tc_size*TC_per_warp;
@ -1499,8 +1492,6 @@ void Emulator::execute(const Instr &instr, uint32_t wid, instr_trace_t *trace) {
scratchpad[loop_offset + (immsrc*(n_tiles)*tc_size*tc_size) + (t*num_data_per_thread) + n] = *temp_ref;
DP(3, "Scratchpad Index: " << loop_offset + (immsrc*(n_tiles)*tc_size*tc_size) + (t*num_data_per_thread) + n << ", Value: " << scratchpad[loop_offset + (immsrc*(n_tiles)*tc_size*tc_size) + (t*num_data_per_thread) + n]);
}
//loop_offset += tc_size*tc_size;
//}
}
rd_write = true;
} break;
@ -1531,7 +1522,7 @@ void Emulator::execute(const Instr &instr, uint32_t wid, instr_trace_t *trace) {
uint32_t csr_index = (2*num_data_per_thread_st) + n;
uint32_t scratchpad_index = (tc_size*tc_size*2) + (t*num_data_per_thread) + n;
//scratchpad -> csr (TODO :: can intermediate step of moving to CSR be skipped?)
//scratchpad -> csr (TODO :: removed intermediate CSR stage ; incorporate limited scratchmad implementation)
//core_->set_csr(csr_addr[(2*num_data_per_thread) + n], scratchpad[(n_tiles*tc_size*tc_size*2) + (t*num_data_per_thread) + n], t, warp_id_);
Word* temp_ref = &(warp.ireg_file.at(t).at(rsrc0));
*temp_ref = scratchpad[(n_tiles*tc_size*tc_size*2) + (t*num_data_per_thread_st) + n];
@ -1562,14 +1553,14 @@ void Emulator::execute(const Instr &instr, uint32_t wid, instr_trace_t *trace) {
//TC operation [only 1 thread in 1 warp needs to do this]
if (t%threads_per_tc == 0)
{
//TODO - change to systolic array implementation
//TODO : change to systolic array implementation
uint32_t thread_offset = t*(tc_size*tc_size);
int loop_offset = 0;
int offset_b = n_tiles*n_tiles*n_tiles*tc_size*tc_size;
// Loop over all tiles - output stationary
//for(int tiles = 0 ; tiles < n_tiles ; tiles++) //What's the HW implication of this?? A counter implementation?
//{
/*
/*
// TODO : Fix needed for functional correctness
for(int tiles = 0 ; tiles < n_tiles ; tiles++) //What's the HW implication of this?? A counter implementation?
{
for (int i = 0; i < tc_size; i++) { //ROW-1
for (int j = 0; j < tc_size; j++) { //COL-2
int sum = 0;
@ -1579,12 +1570,11 @@ void Emulator::execute(const Instr &instr, uint32_t wid, instr_trace_t *trace) {
}
scratchpad[accu_offset + thread_offset +(i * tc_size + j)] += sum; //[i * col2 + j] = sum
DP(3, "Scratchpad Index: " << accu_offset + (i * tc_size + j) << " , Value=" << scratchpad[accu_offset + (i * tc_size + j)]);
}
}
*/
//loop_offset += tc_size*tc_size; //Move to the next tiled matmul fragment
//}
loop_offset += tc_size*tc_size; //Move to the next tiled matmul fragment
}
*/
}
}