Philippe Tillet
|
b8cd63e0da
|
[codegen] separated lower_dot_inst into lower_outer_dot ||
lower_hmma_dot || lower_scanline_dot
|
2019-08-12 21:48:30 -07:00 |
|
Philippe Tillet
|
4bc5758a22
|
[general] some cleaning:
* trans/dot -> peephole
* isel -> added function for tile-level lowering
|
2019-08-12 21:15:21 -07:00 |
|
Philippe Tillet
|
1400d960a6
|
[auto-tuning] much smaller parameters space
|
2019-08-12 21:15:21 -07:00 |
|
Philippe Tillet
|
fd49cdc92b
|
[dnn][blocksparse] added dw code
|
2019-08-08 19:15:35 -07:00 |
|
Philippe Tillet
|
f93099bda1
|
[codegen][transform][trans] fixed incorrect replace_all_uses_with
|
2019-08-07 21:50:16 -07:00 |
|
Philippe Tillet
|
7578c27d3d
|
[general][filesystem] added structure and namespace to code generation files
|
2019-08-07 21:17:17 -07:00 |
|
Philippe Tillet
|
392b55280d
|
[codegen] some cleaning for batched matmul
|
2019-08-07 21:17:17 -07:00 |
|
Philippe Tillet
|
7b75b68edc
|
dirty but working warp-splitting
|
2019-08-06 21:07:13 -07:00 |
|
Philippe Tillet
|
494bfa7671
|
didn't break correctness of existing HMMA
|
2019-08-06 17:34:00 -07:00 |
|
Philippe Tillet
|
46e9863ebe
|
better fp16 support for dot
|
2019-08-06 17:19:13 -07:00 |
|
Philippe Tillet
|
0e201e18ff
|
fixed simple FP16 test
|
2019-08-06 17:14:16 -07:00 |
|
Philippe Tillet
|
6c39cdbace
|
making sure changes didn't break HMMA
|
2019-08-06 16:48:53 -07:00 |
|
Philippe Tillet
|
cf256a636c
|
fixup
|
2019-08-06 16:44:16 -07:00 |
|
Philippe Tillet
|
5efdb7978e
|
more improvements and regressions
|
2019-08-06 16:21:20 -07:00 |
|
Philippe Tillet
|
26c9849462
|
[ir][instructions] added permutations option for trans
|
2019-08-05 21:19:13 -07:00 |
|
Philippe Tillet
|
d62e581ab3
|
basic split-k across warps working for GEMM
|
2019-08-05 19:33:28 -07:00 |
|
Philippe Tillet
|
899b2b72e1
|
simple constexpr
|
2019-08-05 13:06:56 -07:00 |
|
Philippe Tillet
|
d869d9a924
|
[codegen][selection] more flexible instruction selection for reduce_inst
|
2019-08-04 16:34:36 -07:00 |
|
Philippe Tillet
|
6be532c6a2
|
[codegen][selection] adding support for reduction along arbitrary axis
|
2019-08-02 21:29:36 -07:00 |
|
Philippe Tillet
|
d9945692a9
|
[dnn] better specification of recompilation key
|
2019-08-02 17:42:48 -07:00 |
|
Philippe Tillet
|
3b92ddf7e6
|
[codegen/reassociation] now recursively takes pointer arguments into account as well
|
2019-07-31 18:41:56 -07:00 |
|
Philippe Tillet
|
f7bd976fc7
|
[dnn/blocksparse] added heuristics for block-sparse dot
|
2019-07-31 17:12:36 -07:00 |
|
Philippe Tillet
|
bb32ac56c9
|
[codegen/optimize_dce.cpp] fixed bugs whereby barriers were removed by DCE
|
2019-07-31 15:11:10 -07:00 |
|
Philippe Tillet
|
5af7e5adac
|
Made sure it works for FP16
|
2019-07-30 20:02:16 -07:00 |
|
Philippe Tillet
|
080bf1af88
|
[dnn/blocksparse/dot]: BlocksparseDx also working
|
2019-07-30 11:42:31 -07:00 |
|
Philippe Tillet
|
dc11f70fad
|
[dnn/blocksparse] FPROP test passes!
|
2019-07-29 17:06:20 -07:00 |
|
Philippe Tillet
|
17cb2db356
|
[dnn/blocksparse/dot] prototype version seems to pass basic test
|
2019-07-27 21:21:36 -07:00 |
|
Philippe Tillet
|
2a377bc8b1
|
[ir] deleted mask/merge instructions; will be replaced by masked_load/store and select
|
2019-07-25 15:06:15 -07:00 |
|
Philippe Tillet
|
6ce82dfcdb
|
FINALLY
|
2019-07-23 22:19:57 -07:00 |
|
Philippe Tillet
|
b7fadb9986
|
more stuff
|
2019-07-23 21:22:47 -07:00 |
|
Philippe Tillet
|
397d76156b
|
progress on re-association
|
2019-07-23 17:21:24 -07:00 |
|
Philippe Tillet
|
38b3771c26
|
some reassociation
|
2019-07-23 14:43:18 -07:00 |
|
Philippe Tillet
|
c448876178
|
better benchmarking
|
2019-07-22 19:26:12 -07:00 |
|
Philippe Tillet
|
ead368d1ed
|
[general] a bunch of fixes in anticipation of proper triton vs cudnn
benchmarks
* DNN: Added partial auto-tuning mode and skeleton for heuristics
* Examples: Moduralized benchmarking and now evaluating ResNet-18 shapes
|
2019-07-21 20:17:56 -07:00 |
|
Philippe Tillet
|
b1d81a5802
|
more work on heuristics
|
2019-07-21 18:11:54 -07:00 |
|
Philippe Tillet
|
484e3871cf
|
[dnn/shift] added base pointer for a, b
|
2019-07-20 23:00:27 -07:00 |
|
Philippe Tillet
|
d159455f7b
|
[codegen/alignment_info] better alignment information
|
2019-07-20 21:44:18 -07:00 |
|
Philippe Tillet
|
28c250216c
|
[dnn/gemm] added some bounds checking
|
2019-07-19 21:32:55 -07:00 |
|
Philippe Tillet
|
5215fb0424
|
[codegen] some more optimizations
|
2019-07-19 20:29:03 -07:00 |
|
Philippe Tillet
|
71594da66f
|
[dnn/gemm]: fixed leading dimension in transposed variants
|
2019-07-18 16:35:48 -07:00 |
|
Philippe Tillet
|
f0d8306437
|
[codegen/alignment_info] better handling of constants
|
2019-07-18 16:12:06 -07:00 |
|
Philippe Tillet
|
86f70f8224
|
[codegen/selection] performance fix-up when A is transposed for hmma
|
2019-07-17 21:46:23 -07:00 |
|
Philippe Tillet
|
2f0817b2cd
|
[codegen/selection] tensor cores now used for transposed layotus
|
2019-07-17 17:20:38 -07:00 |
|
Philippe Tillet
|
bfa39b8992
|
preparing the field for tensor cores transposes
|
2019-07-17 13:20:33 -07:00 |
|
Philippe Tillet
|
d2e116d057
|
testing GEMM
|
2019-07-17 12:38:30 -07:00 |
|
Philippe Tillet
|
791c91ee63
|
[dnn/shift] bugfix in static shape division
|
2019-07-17 11:39:17 -07:00 |
|
Philippe Tillet
|
a55b098e88
|
[dnn/shift] now using constant divisions
|
2019-07-16 21:05:21 -07:00 |
|
Philippe Tillet
|
07c964919c
|
[dnn/shift] now strictly only shifting the interior
|
2019-07-16 20:18:48 -07:00 |
|
Philippe Tillet
|
ec24e1e7df
|
trying to remove interior logic
|
2019-07-16 18:47:50 -07:00 |
|
Philippe Tillet
|
5f6dd23fc2
|
[dnn/dot] reverted back to peak tensorcores performance
|
2019-07-16 16:14:58 -07:00 |
|