// CU mode: the waves of a window share one L0 cache (1-6% faster here; slower in the wide layers) // d4r-build-flags: -mcumode #define PWIN_F32ACC #ifndef D4R_K_POS_MT #define D4R_K_POS_MT 2 #elif !defined(__GFX12__) || !defined(D4R_K_FP16_BASELINE) #error "experimental position token tiling requires gfx12 strict K baseline" #endif #include "pwin_pos.h" POS_DECODER(dltss_pwin_dec0_layer, 2, 32, 64, 40, 48, D4R_K_POS_MT)