diff --git a/implot_demo.cpp b/implot_demo.cpp index 03132e9..d17d4fa 100644 --- a/implot_demo.cpp +++ b/implot_demo.cpp @@ -31,11 +31,13 @@ #define sprintf sprintf_s #endif +namespace ImPlot { + /// Choose whether the demo uses double or float versions of the ImPlot API. /// NB: You don't ever need to typdef of define values for ImPlot. This /// is only being done here for the sake of demoing both precision types. -#define IMPLOT_DEMO_USE_DOUBLE +// #define IMPLOT_DEMO_USE_DOUBLE #ifdef IMPLOT_DEMO_USE_DOUBLE typedef double t_float; typedef ImPlotPoint t_float2; @@ -54,8 +56,6 @@ typedef ImVec2 t_float2; #define Fmod fmodf #endif -namespace ImPlot { - t_float RandomRange(t_float min, t_float max) { t_float scale = rand() / (t_float) RAND_MAX; return min + scale * ( max - min );