diff --git a/cuburn/device_code.py b/cuburn/device_code.py index 441dd24..2c502aa 100644 --- a/cuburn/device_code.py +++ b/cuburn/device_code.py @@ -106,6 +106,7 @@ class IterThread(PTXEntryPoint): label('iter_loop_choose_xform') with block("Choose the xform for each warp"): + timeout.check_time(5) comment("On subsequent runs, only warp 0 will hit this code") reg.u32('x_addr x_offset') reg.f32('xf_sel') diff --git a/cuburn/variations.py b/cuburn/variations.py index 44715fe..483c772 100644 --- a/cuburn/variations.py +++ b/cuburn/variations.py @@ -110,7 +110,7 @@ class Variations(PTXFragment): @ptx_func def linear(self, xo, yo, xi, yi, wgt): op.fma.rn.ftz.f32(xo, xi, wgt, xo) - op.fma.rn.ftz.f32(yo, yi, wgt, xo) + op.fma.rn.ftz.f32(yo, yi, wgt, yo) @ptx_func def sinusoidal(self, xo, yo, xi, yi, wgt):