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

Restructure

Browse Source
This commit is contained in:
felsabbagh3
2019-03-22 04:14:52 -04:00
parent 097e0217de
commit 6c64fa35f8
239 changed files with 3839 additions and 819 deletions
Download Patch File Download Diff File Expand all files Collapse all files

150
kernel/vx_os/vx_back/vx_back.c Normal file
View File

@@ -0,0 +1,150 @@
#include "vx_back.h"
#include "../vx_io/vx_io.h"
void vx_before_main()
{
// unsigned num_available_warps = vx_available_warps();
for (int i = 0; i < 8; i++)
{
queue_initialize(q + i);
}
}
void vx_reschedule_warps()
{
register unsigned curr_warp asm("s10");
// vx_printf("Reschedule: ", curr_warp);
if (queue_isEmpty(q+curr_warp))
{
// vx_printf("Done: ", curr_warp);
done[curr_warp] = 1;
if (curr_warp == 0)
{
vx_load_context();
return;
}
ECALL;
}
Job j;
queue_dequeue(q+curr_warp,&j);
// vx_printf("Reschedule -> ", j.wid);
asm __volatile__("mv sp,%0"::"r" (j.base_sp):);
vx_createThreads(j.n_threads, j.wid, j.func_ptr, j.args, j.assigned_warp);
ECALL;
}
void vx_schedule_warps()
{
unsigned num_available_warps = vx_available_warps();
asm __volatile__("mv s3, sp");
for (int curr_warp = 1; curr_warp < num_available_warps; ++curr_warp)
{
if (!queue_isEmpty(q+curr_warp))
{
Job j;
queue_dequeue(q+curr_warp,&j);
asm __volatile__("mv sp,%0"::"r" (j.base_sp):);
vx_wspawn(j.n_threads, j.wid, j.func_ptr, j.args, j.assigned_warp);
}
}
asm __volatile__("mv sp, s3");
vx_save_context();
// vx_print_str("saved context\n");
register unsigned val asm("tp");
if (val)
{
if (!queue_isEmpty(q))
{
// vx_print_str("found something for w0\n");
Job j;
queue_dequeue(q,&j);
// vx_printf("num_threads: ", j.n_threads);
asm __volatile__("mv sp,%0"::"r" (j.base_sp):);
vx_createThreads(j.n_threads, j.wid, j.func_ptr, j.args, j.assigned_warp);
}
}
}
void vx_spawnWarps(unsigned num_Warps, unsigned num_threads, FUNC, void * args)
{
vx_before_main();
unsigned num_available_warps = vx_available_warps();
// vx_printf("Num available warps: ", num_available_warps);
asm __volatile__("addi s2, sp, 0");
int warp = 0;
for (unsigned i = 0; i < num_Warps; i++)
{
asm __volatile__("lui s3, 0xFFFF0");
asm __volatile__("add sp, sp, s3");
register unsigned stack_ptr asm("sp");
Job j;
j.wid = i;
j.n_threads = num_threads;
j.base_sp = stack_ptr;
j.func_ptr = (unsigned) func;
j.args = args;
j.assigned_warp = warp;
queue_enqueue(q + warp,&j);
++warp;
if (warp >= num_available_warps) warp = 0;
}
asm __volatile__("addi sp, s2, 0");
vx_schedule_warps();
}
void vx_wait_for_warps(unsigned num_wait)
{
// vx_printf("wait for: ", num_wait);
unsigned num_available_warps = vx_available_warps();
unsigned num = 0;
while (num != num_wait)
{
num = 0;
for (int i = 0; i < num_available_warps; i++)
{
if (done[i] == 1)
{
num += 1;
}
}
}
// vx_printf("num found: ", num);
for (int i = 0; i < num_available_warps; i++) done[i] = 0;
}
void * vx_get_arg_struct(void)
{
register void *ret asm("s7");
return ret;
}

55
kernel/vx_os/vx_back/vx_back.h Normal file
View File

@@ -0,0 +1,55 @@
#pragma once
#include <stdbool.h>
#include "../vx_util/queue.h"
#define WSPAWN asm __volatile__(".word 0x3006b"::);
#define CLONE asm __volatile__(".word 0x3506b":::);
#define JALRS asm __volatile__(".word 0x1bfe0eb":::"s10");
#define ECALL asm __volatile__(".word 0x00000073");
#define JMPRT asm __volatile__(".word 0x5406b");
#define SPLIT asm __volatile__(".word 0xf206b");
#define P_JUMP asm __volatile__(".word 0x1ff707b");
#define JOIN asm __volatile__(".word 0x306b");
#define __if(val) bool temp = !val; \
register unsigned p asm("t5") = temp; \
register void * e asm("t6") = &&ELSE; \
SPLIT; \
P_JUMP; \
#define __else register void * w asm("t3") = &&AFTER; \
asm __volatile__("jr t3"); \
ELSE: asm __volatile__("nop");
#define __end_if AFTER:\
JOIN;
static int done[] = {0, 0, 0, 0, 0, 0, 0};
static int main_sp[1];
unsigned context[32];
void vx_save_context(void);
void vx_load_context(void);
#define FUNC void (func)(unsigned, unsigned)
unsigned vx_available_warps(void);
unsigned vx_available_threads(void);
void vx_createThreads(unsigned, unsigned, unsigned, void *, unsigned);
void vx_wspawn(unsigned, unsigned, unsigned, void *, unsigned);
void vx_spawnWarps(unsigned num_Warps, unsigned num_threads, FUNC, void *);
void vx_schedule_warps(void);
void vx_reschedule_warps(void);
void vx_wait_for_warps(unsigned);
void * vx_get_arg_struct(void);

151
kernel/vx_os/vx_back/vx_back.s Normal file
View File

@@ -0,0 +1,151 @@
.section .text
.type _start, @function
.global _start
_start:
li a0, 4 # Num Warps
csrw 0x20, a0 # Setting the number of available warps
li a0, 8 # Num Threads
csrw 0x21, a0 # Setting the number of available threads
csrw mhartid,zero
csrw misa,zero
lui sp, 0x7ffff
jal vx_before_main
jal main
ecall
.type vx_createThreads, @function
.global vx_createThreads
vx_createThreads:
mv s7 ,a3 # Moving args to s7
mv s10,a4 # Moving assigned_warp to s10
mv t5 ,sp # Saving the current stack pointer to t5
mv t2 , a0 # t2 = num_threads
loop_init:
li a0,1 # i = 0
loop_cond:
bge a0, t2, loop_done # i < num_threads
loop_body:
addi sp,sp,-2048 # Allocate 2k stack for new thread
mv t1, a0 # #lane = i
.word 0x3506b # clone register state
loop_inc:
addi a0, a0, 1
j loop_cond
loop_done:
mv sp,t5 # Restoring the stack
li a0,0 # setting tid = 0 for main thread
mv t6,a2 # setting func_addr
mv s11,t2 # setting num_threads to spawn
.word 0x1bfe0eb
la a0, vx_reschedule_warps
.word 0x5406b
.type vx_wspawn, @function
.global vx_wspawn
vx_wspawn:
la t1, vx_createThreads
.word 0x3006b # WSPAWN instruction
ret
.global context
.type vx_save_context, @function
.global vx_save_context
vx_save_context:
la tp, context
sw x0 , 0 (tp)
sw x1 , 4 (tp)
sw x2 , 8 (tp)
sw x3 , 12(tp)
sw x4 , 16(tp)
sw x5 , 20(tp)
sw x6 , 24(tp)
sw x7 , 28(tp)
sw x8 , 32(tp)
sw x9 , 36(tp)
sw x10, 40(tp)
sw x11, 44(tp)
sw x12, 48(tp)
sw x13, 52(tp)
sw x14, 56(tp)
sw x15, 60(tp)
sw x16, 64(tp)
sw x17, 68(tp)
sw x18, 72(tp)
sw x19, 76(tp)
sw x20, 80(tp)
sw x21, 84(tp)
sw x22, 88(tp)
sw x23, 92(tp)
sw x24, 96(tp)
sw x25, 100(tp)
sw x26, 104(tp)
sw x27, 108(tp)
sw x28, 112(tp)
sw x29, 116(tp)
sw x30, 120(tp)
sw x31, 124(tp)
li tp, 1
ret
.type vx_load_context, @function
.global vx_load_context
vx_load_context:
la tp, context
lw x0 , 0 (tp)
lw x1 , 4 (tp)
lw x2 , 8 (tp)
lw x3 , 12(tp)
lw x4 , 16(tp)
lw x5 , 20(tp)
lw x6 , 24(tp)
lw x7 , 28(tp)
lw x8 , 32(tp)
lw x9 , 36(tp)
lw x10, 40(tp)
lw x11, 44(tp)
lw x12, 48(tp)
lw x13, 52(tp)
lw x14, 56(tp)
lw x15, 60(tp)
lw x16, 64(tp)
lw x17, 68(tp)
lw x18, 72(tp)
lw x19, 76(tp)
lw x20, 80(tp)
lw x21, 84(tp)
lw x22, 88(tp)
lw x23, 92(tp)
lw x24, 96(tp)
lw x25, 100(tp)
lw x26, 104(tp)
lw x27, 108(tp)
lw x28, 112(tp)
lw x29, 116(tp)
lw x30, 120(tp)
lw x31, 124(tp)
li tp, 0
ret
.type vx_available_warps, @function
.global vx_available_warps
vx_available_warps:
csrr a0, 0x20
ret
.type vx_available_threads, @function
.global vx_available_threads
vx_available_threads:
csrr a0, 0x21
ret