1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
|
---
include: ../../../../clients/include/rocblas_common.yaml
Definitions:
- &gemvN_matrix_sizes
# skinny n large m
- { scan: [ 65536..131072..4096 ], M: *c_scan_value, N: 32, lda: *c_scan_value }
- { scan: [ 16384..65536..2048], M: *c_scan_value, N: 8, lda: *c_scan_value }
# double buffered loads on gfx90a w/ float/double
- { scan: [ 512..8192..512], M: *c_scan_value, N: *c_scan_value, lda: *c_scan_value }
# optimized sizes for gfx906 and gfx908, generic covered by gfx90a
- { scan: [ 1538..12304..1538 ], M: *c_scan_value, N: *c_scan_value, lda: *c_scan_value } # non-multiples of 128
- { M: 4011, N: 4011, lda: 4011, stride_a: 16088200 }
- { M: 8000, N: 8000, lda: 8000, stride_a: 64000000 }
- &gemvT_matrix_sizes
# skinny n large m
- { scan: [ 65536..131072..4096 ], M: *c_scan_value, N: 32, lda: *c_scan_value }
- { scan: [ 16384..65536..2048], M: *c_scan_value, N: 8, lda: *c_scan_value }
# double bufferred loads, special kernel only for gfx908
- { scan: [ 8192..1024..16384 ], M: *c_scan_value, N: *c_scan_value, lda: *c_scan_value }
# generic square range - uses gfx10/gfx11-specific kernels, or generic kernels on other hardware
- { scan: [ 1024..8192..1024], M: *c_scan_value, N: *c_scan_value, lda: *c_scan_value }
- &gemv_batched_matrix_sizes
# m && n <= 32 && batch_count >= 256 (gfx90a only) watch for overlap with qmcpack
- { M: 3 , N: 32, lda: 3, batch_count: [ 256..1024..256 ], stride_a: 96 }
- { M: 24, N: 24, lda: 24, batch_count: [ 256..1024..256 ], stride_a: 1024 }
# qmcpack few rows
- { M: 32, N: [ 2048..8192..1024 ], lda: 32, batch_count: [ 8..32..8 ], stride_a: 49152 }
Tests:
- name: gemvn_bench
category: bench
function: gemv
precision: *single_double_precisions_complex_real
transA: [ N ]
alpha: 1
incx: 1
incy: 1
matrix_size: *gemvN_matrix_sizes
- name: gemvt_bench
category: bench
function: gemv
precision: *single_double_precisions_complex_real
transA: [ T ]
alpha: 1
incx: 1
incy: 1
matrix_size: *gemvT_matrix_sizes
- name: gemv_batched_bench
category: bench
function: gemv_batched
precision: *single_double_precisions_complex_real
transA: [ N, T ]
alpha: 1
incx: 1
incy: 1
matrix_size: *gemv_batched_matrix_sizes
...
|
