Richard Yan
6eafa2de54
write operands to elf
2024-04-24 22:09:30 -07:00
Richard Yan
4e9855dc33
highly unrolled a/b load
2024-04-16 22:19:30 -07:00
Richard Yan
449d99f0bb
dram gemm kernel
2024-04-16 17:15:22 -07:00
Richard Yan
99621a0df9
Merge branch 'kernels' of https://github.com/hansungk/vortex-private into kernels
2024-04-15 10:22:19 -07:00
Richard Yan
041d49fb58
update gemmini only kernel
2024-04-15 10:22:00 -07:00
Richard Yan
0bb7aeb45b
add gpu+gemmini gemm kernel
2024-04-15 10:13:37 -07:00
Richard Yan
d8eddb21ea
add gemmini dependency
2024-04-15 10:04:54 -07:00
Hansung Kim
37a60b1141
sgemm_wg: Output C result to binary
2024-04-14 12:36:06 -07:00
Hansung Kim
3383b70732
sgemm_wg: Hardcode device address
2024-04-14 12:36:00 -07:00
Richard Yan
7bf72c9568
cycle counting for fence
2024-04-09 19:53:17 -07:00
Hansung Kim
93a00101ae
sgemm_wg: revert to faster params
2024-04-04 21:06:14 -07:00
Richard Yan
84a31f3384
thread parallel data loading for word strided bank
2024-04-01 11:10:32 -07:00
Richard Yan
e6db1a83af
Merge branch 'kernels' of https://github.com/hansungk/vortex-private into kernels
2024-04-01 11:09:43 -07:00
Hansung Kim
b0c1f77388
vx_start.S: Swizzle stack space
...
Striding stack space for threads by power-of-two risks possibilities of bank
conflicts or cache aliasing problems. Add an extra offset of 4 bytes to avoid
this.
2024-03-29 12:26:14 -07:00
Hansung Kim
fa2b6e2ad0
sgemm_wg: Explicitly limit unroll to reduce stack spilling
...
This needs to be done case-by-case for different BK/TM/TN combinations and
examining the assembly.
2024-03-29 02:48:29 -07:00
Hansung Kim
537b97eb20
common.mk: Don't clean all *.elf
2024-03-28 20:17:26 -07:00
Hansung Kim
e4eec8ab4d
vx_spawn.c: Handle num_clusters > 1
...
WIP: still assumes num_tasks is divisible by num_cluster
2024-03-28 20:16:44 -07:00
Hansung Kim
a9b0814211
sgemm_wg: Document tiling parameter constraints
2024-03-28 18:17:00 -07:00
Hansung Kim
9673db4e8c
sgemm_wg: Fix possible divide-by-0
2024-03-28 17:35:47 -07:00
Hansung Kim
9555b790e7
sgemm_wg: ifdef-guard cluster specific code
2024-03-27 22:45:51 -07:00
Hansung Kim
09822764e7
sgemm_wg: Remove software-based barrier implementation
...
Intra-cluster barrier is now implemented in hardware, transparent to the ISA.
2024-03-27 22:43:45 -07:00
Hansung Kim
870846f20f
vx_spawn.c: Create separate vx_spawn_tasks_contiguous
2024-03-27 15:38:52 -07:00
Hansung Kim
fa6adceb7e
vecaddx: Hardcode args/input device address to match chipyard
...
Don't use mem_alloc/mem_free API
2024-03-27 15:15:52 -07:00
Hansung Kim
4e834f2103
vx_spawn.c: Rewrite cluster-based vx_spawn_tasks variant
...
Implements round-robin allocation of warps to cores & maintains contiguous
thread ID allocation to neighboring threads. Also handles partially-enabled
remainder warp logic.
TODO: Hardcodes only 1 cluster in the system.
2024-03-27 15:14:45 -07:00
Hansung Kim
df1f7f242a
vx_spawn.c: Implement spawn_tasks_cluster_rem_stub
2024-03-27 00:00:44 -07:00
Richard Yan
b88dbd7a83
add cycle count and multi core support
2024-03-26 16:43:49 -07:00
Hansung Kim
b545809496
vecaddx: Use -DRADIANCE
2024-03-26 16:42:36 -07:00
Hansung Kim
4d2c0084d1
common.mk: Compile separate cluster ELF
...
... using -DRADIANCE, which the kernel C code use explicitly to switch between
vx_spawn_tasks and vx_spawn_tasks_cluster. This is to ease running both simX
and Chipyard simulations without mixing up binaries.
2024-03-26 16:37:44 -07:00
Hansung Kim
3729a05adc
vx_spawn.c: Separate cluster-based scheduling code from original
2024-03-26 16:36:57 -07:00
Hansung Kim
f050a08d77
Write vx_spawn_tasks_cluster
...
This scheduling logic tries to evenly distribute warps across *all* cores,
instead of trying to fill up the first cores as much as possible. This scheme
is necessary for the intra-cluster cores which are assumed to have equal
workloads distributed.
2024-03-26 10:45:14 -07:00
Hansung Kim
7f00e6c376
vecaddx: Change arg device address to 7fff0000
2024-03-26 10:44:33 -07:00
Hansung Kim
cc7b34ec5b
vecaddx: Write args.bin and input.bin
2024-03-26 10:44:02 -07:00
Hansung Kim
ff401bdec0
Cleanup tests/.gitignore
2024-03-24 01:47:00 -07:00
Hansung Kim
7d177492b2
Move CORES_PER_CLUSTER to vx_spawn.h
2024-03-24 01:45:30 -07:00
Hansung Kim
8f3474b151
Don't clean *.bin
2024-03-24 01:45:08 -07:00
Hansung Kim
f590c4b417
Add vx_spawn.h as dependency to kernel/Makefile
2024-03-24 01:44:49 -07:00
Richard Yan
c18267443f
matmul kernel switch to proper fence and fsm
2024-03-20 15:22:25 -07:00
Richard Yan
94ad1850a9
implement correct gemmini fence and loop fsm support
2024-03-20 15:18:31 -07:00
Hansung Kim
12ee2a3a0f
Write cluster-aware thread scheduling
...
NOTE: cores per cluster is hardcoded as a constant
2024-03-18 16:40:02 -07:00
Hansung Kim
3e6771237f
Merge remote-tracking branch 'sungwoong/master' into kernels
2024-03-14 09:48:31 -07:00
Hansung Kim
2036d37840
sgemm_wg: Prevent run-ahead using ternary flags; reduce mem accesses
2024-03-13 21:35:24 -07:00
Hansung Kim
510a834db5
sgemm_wg: Implement software barrier for inter-core synchronization
2024-03-12 15:34:42 -07:00
Hansung Kim
fbe872c831
sgemm_wg: Add missing makefile dep to common.h
2024-03-12 15:34:17 -07:00
Sungwoong Ha
3c2a266d37
second pass
2024-03-01 21:27:26 -08:00
Sungwoong Ha
a9709edae2
first pass
2024-03-01 21:05:52 -08:00
Hansung Kim
6f4dfe5a0e
sgemm_wg: Implement 2D threadtiling
2024-02-29 14:40:54 -08:00
Hansung Kim
a06b2dd20e
sgemm_wg: Cleanup & proper unroll
2024-02-28 21:17:42 -08:00
Hansung Kim
46f242e520
sgemm_wg: Constantify BM/BN/BK/TM, computationally set gridsize and TB/core
2024-02-27 22:23:25 -08:00
Hansung Kim
27646bb507
sgemm_wg: Implement multiple C per thread with sliding A/B blocks
2024-02-27 22:06:01 -08:00
Hansung Kim
a2ea27b2b5
vx_spawn: Add spawn_tasks_contiguous_all_stub
...
Spawns tasks in a way that the threads in a warp see contiguous
thread_id, unlike the original variant where each thread were allocated
a range of thread_id that spans the number of batches.
E.g. in a 4-thread config, instead of mapping IDs (0,2,4,6)->(1,3,5,7),
map (0,1,2,3)->(4,5,6,7).
TODO remaining logic not implemented.
2024-02-27 15:46:02 -08:00
