fix: correct gfx10 ctl stack size (tinygrad#9384)

Author: wozeparrot

tinygrad/runtime/ops_amd.py

@@ -584,7 +584,8 @@ def __init__(self, device:str=""):
     sgrp_size_per_cu, lds_size_per_cu, hwreg_size_per_cu = 0x4000, 0x10000, 0x1000
     vgpr_size_per_cu = 0x60000 if self.target in {110000, 110001, 120000, 120001} else 0x40000
     wg_data_size = round_up((vgpr_size_per_cu + sgrp_size_per_cu + lds_size_per_cu + hwreg_size_per_cu) * (self.max_cu_id + 1), mmap.PAGESIZE)
-    ctl_stack_size = round_up(12 * (self.max_cu_id + 1) * (self.max_wave_id + 1) + 8 + 40, mmap.PAGESIZE) if self.target//10000 != 10 else 0x7000
+    ctl_stack_size = round_up(12 * (self.max_cu_id + 1) * (self.max_wave_id + 1) + 8 + 40, mmap.PAGESIZE)
+    if self.target//10000 == 10: ctl_stack_size = min(ctl_stack_size, 0x7000)
     debug_memory_size = round_up((self.max_cu_id + 1) * (self.max_wave_id + 1) * 32, 64)
 
     self.compute_queue = self.create_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE, 0x800000, ctx_save_restore_size=wg_data_size + ctl_stack_size,