wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update Wu architecture kernel implementations and runtime library

Browse Source
This commit is contained in:
Zhongdi LUO
2026-05-26 12:59:35 +00:00
parent e7229dae27
commit 9f4be1b8f7
8 changed files with 142 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
lib/src/vx_spawn.c
View File

@@ -76,7 +76,7 @@ static void __attribute__ ((noinline)) spawn_tasks_all_stub() {
static void __attribute__ ((noinline)) spawn_tasks_contiguous_all_stub() {
int NT = vx_num_threads();
int NW = vx_num_warps();
int NW = NUM_SCALAR_WARPS;
int cid = vx_core_id();
int wid = vx_warp_id();
int tid = vx_thread_id();
@@ -96,7 +96,7 @@ static void __attribute__ ((noinline)) spawn_tasks_contiguous_all_stub() {
static void __attribute__ ((noinline)) spawn_tasks_cluster_all_stub() {
int NT = vx_num_threads();
int NW = vx_num_warps();
int NW = NUM_SCALAR_WARPS;
int cid = vx_core_id();
int wid = vx_warp_id();
int tid = vx_thread_id();
@@ -187,7 +187,7 @@ static void __attribute__ ((noinline)) spawn_tasks_all_cb() {
void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg) {
// device specs
const int NC = vx_num_cores();
const int NW = vx_num_warps();
const int NW = NUM_SCALAR_WARPS;
const int NT = vx_num_threads();
// NOTE: assumes divisible
const int num_cluster = NC / CORES_PER_CLUSTER;
@@ -243,7 +243,7 @@ void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg
const int num_full_waves = num_warps_this_core / NW;
const int rem_full_warps_in_last_wave = num_warps_this_core % NW;
const const int offset = cluster_id * num_tasks_this_cluster;
const int offset = cluster_id * num_tasks_this_cluster;
wspawn_tasks_args_t wspawn_args = {callback, arg, offset, num_full_waves,
rem_full_warps_in_last_wave};
g_wspawn_args[core_id] = &wspawn_args;
@@ -289,7 +289,7 @@ void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg
void vx_spawn_tasks_contiguous(int num_tasks, vx_spawn_tasks_cb callback , void * arg) {
// device specs
int NC = vx_num_cores();
int NW = vx_num_warps();
int NW = NUM_SCALAR_WARPS;
int NT = vx_num_threads();
// current core id
@@ -361,7 +361,7 @@ void vx_spawn_tasks_contiguous(int num_tasks, vx_spawn_tasks_cb callback , void
void vx_spawn_tasks(int num_tasks, vx_spawn_tasks_cb callback , void * arg) {
// device specs
int NC = vx_num_cores();
int NW = vx_num_warps();
int NW = NUM_SCALAR_WARPS;
int NT = vx_num_threads();
// current core id
@@ -515,7 +515,7 @@ void vx_spawn_kernel(context_t * ctx, vx_spawn_kernel_cb callback, void * arg) {
// device specs
int NC = vx_num_cores();
int NW = vx_num_warps();
int NW = NUM_SCALAR_WARPS;
int NT = vx_num_threads();
// current core id

9
lib/src/vx_start.S
View File

@@ -22,9 +22,9 @@
_start:
# initialize per-thread registers
csrr t0, VX_CSR_NUM_WARPS # get num warps
li t0, ((1 << NUM_SCALAR_WARPS) - 1) # scalar warp mask
la t1, init_regs_all
.insn r RISCV_CUSTOM0, 1, 0, x0, t0, t1 # wspawn t0, t1
.insn r RISCV_CUSTOM0, 6, 0, x0, t0, t1 # wspawn_mask t0, t1
li t0, -1
.insn r RISCV_CUSTOM0, 0, 0, x0, t0, x0 # tmc t0
jal init_regs
@@ -35,9 +35,9 @@ _start:
jal vx_wspawn_wait
# initialize TLS for all warps
csrr t0, VX_CSR_NUM_WARPS # get num warps
li t0, ((1 << NUM_SCALAR_WARPS) - 1) # scalar warp mask
la t1, init_tls_all
.insn r RISCV_CUSTOM0, 1, 0, x0, t0, t1 # wspawn t0, t1
.insn r RISCV_CUSTOM0, 6, 0, x0, t0, t1 # wspawn_mask t0, t1
li t0, -1
.insn r RISCV_CUSTOM0, 0, 0, x0, t0, x0 # tmc t0
call __init_tls
@@ -150,4 +150,3 @@ vx_wspawn_wait:
.weak __dso_handle
__dso_handle:
.long 0