Skip to content

Commit d02f862

Browse files
Yu-ZhewenYu-Zhewen
committed
WIP
Signed-off-by: Yu-Zhewen <[email protected]>
1 parent 5aa6453 commit d02f862

File tree

7 files changed

+1634
-0
lines changed

7 files changed

+1634
-0
lines changed

‎compiler/plugins/target/ROCM/builtins/mlir_ukernel/TEMPLATE_GUIDE.md‎

Lines changed: 131 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,131 @@
1+
# MLIR Microkernel Template Guide
2+
3+
## Data-Tiled Kernels
4+
5+
### `pingpong_dt_medium_f4E2M1FN`
6+
```bash
7+
python mlir_ukernel_gen.py iree_uk_amdgpu_dt_scaled_matmul_f4E2M1FN.mlir.in \
8+
-D INTRINSICS_M=4 INTRINSICS_N=8 INTRINSICS_K=2 SUBGROUPS_M=2 SUBGROUPS_N=2 ARCH=gfx950 \
9+
-o iree_uk_amdgpu_dt_scaled_matmul_f4E2M1FN.mlir
10+
```
11+
12+
### `pingpong_dt_large_f16`
13+
```bash
14+
python mlir_ukernel_gen.py iree_uk_amdgpu_dt_matmul_large.mlir.in \
15+
-D ELEM_TYPE=f16 INTRINSIC=MFMA_F32_16x16x16_F16 \
16+
INTRINSICS_M=8 INTRINSICS_N=4 INTRINSICS_K=1 SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
17+
SIZE_MIN_0=512 SIZE_DIV_0=64 SIZE_MIN_1=32832 SIZE_DIV_1=64 SIZE_MIN_2=512 SIZE_DIV_2=64 \
18+
-o iree_uk_amdgpu_dt_matmul_f16.mlir
19+
```
20+
21+
### `pingpong_dt_large_f8E4M3FNUZ`
22+
```bash
23+
python mlir_ukernel_gen.py iree_uk_amdgpu_dt_matmul_large.mlir.in \
24+
-D ELEM_TYPE=f8E4M3FNUZ INTRINSIC=MFMA_F32_16x16x32_F8E4M3FNUZ \
25+
INTRINSICS_M=4 INTRINSICS_N=8 INTRINSICS_K=1 SUBGROUPS_M=2 SUBGROUPS_N=2 ARCH=gfx942 \
26+
SIZE_MIN_0=64 SIZE_MIN_1=2048 SIZE_MAX_1=8192 \
27+
-o iree_uk_amdgpu_dt_matmul_f8E4M3FNUZ_large.mlir
28+
```
29+
30+
### `pingpong_dt_medium_f8E4M3FNUZ`
31+
```bash
32+
python mlir_ukernel_gen.py iree_uk_amdgpu_dt_matmul_medium.mlir.in \
33+
-D ELEM_TYPE=f8E4M3FNUZ INTRINSIC=MFMA_F32_16x16x32_F8E4M3FNUZ \
34+
INTRINSICS_M=8 INTRINSICS_N=2 INTRINSICS_K=2 SUBGROUPS_M=1 SUBGROUPS_N=8 ARCH=gfx942 \
35+
SIZE_MIN_0=32 \
36+
-o iree_uk_amdgpu_dt_matmul_f8E4M3FNUZ_medium.mlir
37+
```
38+
39+
### `pingpong_dt_large_f8E4M3FN`
40+
```bash
41+
python mlir_ukernel_gen.py iree_uk_amdgpu_dt_matmul_large.mlir.in \
42+
-D ELEM_TYPE=f8E4M3FN INTRINSIC=MFMA_F32_16x16x32_F8E4M3FN \
43+
INTRINSICS_M=4 INTRINSICS_N=8 INTRINSICS_K=1 SUBGROUPS_M=2 SUBGROUPS_N=2 ARCH=gfx950 \
44+
SIZE_MIN_0=64 SIZE_MIN_1=2048 SIZE_MAX_1=8192 \
45+
-o iree_uk_amdgpu_dt_matmul_f8E4M3FN_large.mlir
46+
```
47+
48+
### `pingpong_dt_medium_f8E4M3FN`
49+
```bash
50+
python mlir_ukernel_gen.py iree_uk_amdgpu_dt_matmul_medium.mlir.in \
51+
-D ELEM_TYPE=f8E4M3FN INTRINSIC=MFMA_F32_16x16x32_F8E4M3FN \
52+
INTRINSICS_M=8 INTRINSICS_N=2 INTRINSICS_K=2 SUBGROUPS_M=1 SUBGROUPS_N=8 ARCH=gfx950 \
53+
SIZE_MIN_0=32 \
54+
-o iree_uk_amdgpu_dt_matmul_f8E4M3FN_medium.mlir
55+
```
56+
57+
---
58+
59+
## Non-Data-Tiled Kernels
60+
61+
### `pingpong_large_f16`
62+
```bash
63+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_large.mlir.in \
64+
-D ELEM_TYPE=f16 INTRINSIC=MFMA_F32_16x16x16_F16 \
65+
INTRINSICS_M=8 INTRINSICS_N=4 INTRINSICS_K=4 \
66+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
67+
-o pingpong_large_f16.mlir
68+
```
69+
70+
### `pingpong_medium_f16_expanded`
71+
```bash
72+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_medium.mlir.in \
73+
-D ELEM_TYPE=f16 INTRINSIC=MFMA_F32_16x16x16_F16 \
74+
INTRINSICS_M=4 INTRINSICS_N=4 INTRINSICS_K=8 \
75+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
76+
-o pingpong_medium_f16_expanded.mlir
77+
```
78+
79+
### `pingpong_large_bf16`
80+
```bash
81+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_large.mlir.in \
82+
-D ELEM_TYPE=bf16 INTRINSIC=MFMA_F32_16x16x16_BF16 \
83+
INTRINSICS_M=8 INTRINSICS_N=4 INTRINSICS_K=4 \
84+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
85+
-o pingpong_large_bf16.mlir
86+
```
87+
88+
### `pingpong_medium_bf16_expanded`
89+
```bash
90+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_medium.mlir.in \
91+
-D ELEM_TYPE=bf16 INTRINSIC=MFMA_F32_16x16x16_BF16 \
92+
INTRINSICS_M=4 INTRINSICS_N=4 INTRINSICS_K=8 \
93+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
94+
-o pingpong_medium_bf16_expanded.mlir
95+
```
96+
97+
### `pingpong_medium_f8E4M3FNUZ_expanded`
98+
```bash
99+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_medium.mlir.in \
100+
-D ELEM_TYPE=f8E4M3FNUZ INTRINSIC=MFMA_F32_16x16x32_F8E4M3FNUZ \
101+
INTRINSICS_M=4 INTRINSICS_N=4 INTRINSICS_K=16 \
102+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
103+
-o pingpong_medium_f8E4M3FNUZ_expanded.mlir
104+
```
105+
106+
### `pingpong_large_f8E4M3FNUZ_expanded`
107+
```bash
108+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_large.mlir.in \
109+
-D ELEM_TYPE=f8E4M3FNUZ INTRINSIC=MFMA_F32_16x16x32_F8E4M3FNUZ \
110+
INTRINSICS_M=8 INTRINSICS_N=4 INTRINSICS_K=8 \
111+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
112+
-o pingpong_large_f8E4M3FNUZ_expanded.mlir
113+
```
114+
115+
### `pingpong_medium_f8E4M3FN_expanded`
116+
```bash
117+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_medium.mlir.in \
118+
-D ELEM_TYPE=f8E4M3FN INTRINSIC=MFMA_F32_16x16x32_F8E4M3FN \
119+
INTRINSICS_M=4 INTRINSICS_N=4 INTRINSICS_K=16 \
120+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
121+
-o pingpong_medium_f8E4M3FN_expanded.mlir
122+
```
123+
124+
### `pingpong_large_f8E4M3FN_expanded`
125+
```bash
126+
python mlir_ukernel_gen.py iree_uk_amdgpu_matmul_large.mlir.in \
127+
-D ELEM_TYPE=f8E4M3FN INTRINSIC=MFMA_F32_16x16x32_F8E4M3FN \
128+
INTRINSICS_M=8 INTRINSICS_N=4 INTRINSICS_K=8 \
129+
SUBGROUPS_M=2 SUBGROUPS_N=4 ARCH=gfx942 \
130+
-o pingpong_large_f8E4M3FN_expanded.mlir
131+
```

0 commit comments

Comments
 (0)