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
|
Fix build with llvm-23
LLVM 23 assembles gfx11/gfx12 with +real-true16 by default and rejects the
fake16 syntax that Tensile generates (e.g. v_cvt_f16_f32 v1, v2)
--- a/Tensile/Common.py
+++ b/Tensile/Common.py
@@ -2178,6 +2178,10 @@ def tryAssembler(isaVersion, asmString, debug=False, *options):
if isaVersion[0] >= 10 and '-mno-wavefrontsize64' not in options:
options += ['-mwavefrontsize64']
+ # LLVM 23 defaults to true16 syntax on gfx11/gfx12, Tensile emits fake16
+ if isaVersion[0] in (11, 12):
+ options += ['-Xclangas', '-target-feature', '-Xclangas', '-real-true16']
+
assembler = globalParameters['AssemblerPath']
if assembler is None:
raise ValueError('No assembler available; set TENSILE_ROCM_ASSEMBLER_PATH to point to ROCm Clang.')
--- a/Tensile/KernelWriterAssembly.py
+++ b/Tensile/KernelWriterAssembly.py
@@ -285,6 +285,10 @@ class KernelWriterAssembly(KernelWriter):
rv += ['-mcpu=' + gfxName(isa)]
+ # LLVM 23 defaults to true16 syntax on gfx11/gfx12, Tensile emits fake16
+ if isa[0] in (11, 12):
+ rv += ['-Xclangas', '-target-feature', '-Xclangas', '-real-true16']
+
if wavefrontSize == 64:
rv += ['-mwavefrontsize64']
else:
|
