From fb10ea563e2bb982cec952b79defec7480fc5cd3 Mon Sep 17 00:00:00 2001 From: ignaciosica Date: Fri, 22 Nov 2024 16:29:01 -0300 Subject: [PATCH] typedef bf16 amd (#7850) --- tinygrad/renderer/cstyle.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tinygrad/renderer/cstyle.py b/tinygrad/renderer/cstyle.py index 007190a6b4..26d2818488 100644 --- a/tinygrad/renderer/cstyle.py +++ b/tinygrad/renderer/cstyle.py @@ -412,7 +412,7 @@ class AMDRenderer(CStyleLanguage): prefix = ["#define INFINITY (__builtin_inff())","#define NAN (__builtin_nanf(\"\"))","typedef long unsigned int size_t;","#define half _Float16"] used_dtypes = uops_to_dtypes(uops) - if any(dt.scalar() == dtypes.bfloat16 for dt in used_dtypes): prefix.append("struct hip_bfloat16 { unsigned short data; };") + if any(dt.scalar() == dtypes.bfloat16 for dt in used_dtypes): prefix.append("typedef unsigned short hip_bfloat16;") prefix += [self.render_vector_prefix(dt) for dt in used_dtypes if dt.count > 1] for arg in dedup([uop.arg for uop in uops if uop.op is Ops.WMMA]): # TODO: handle TCs f32_bf16 and bf16_bf16 w/ wrapper