summaryrefslogtreecommitdiff
path: root/tools/proxyclient/experiments/agx_parallel.py
diff options
context:
space:
mode:
Diffstat (limited to 'tools/proxyclient/experiments/agx_parallel.py')
-rw-r--r--tools/proxyclient/experiments/agx_parallel.py341
1 files changed, 341 insertions, 0 deletions
diff --git a/tools/proxyclient/experiments/agx_parallel.py b/tools/proxyclient/experiments/agx_parallel.py
new file mode 100644
index 0000000..163497e
--- /dev/null
+++ b/tools/proxyclient/experiments/agx_parallel.py
@@ -0,0 +1,341 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: MIT
+
+import sys, pathlib, time
+sys.path.append(str(pathlib.Path(__file__).resolve().parents[1]))
+
+import atexit, sys
+
+from m1n1.setup import *
+from m1n1.constructutils import Ver
+from m1n1.utils import *
+
+Ver.set_version(u)
+
+from m1n1.agx import AGX
+from m1n1.agx.render import *
+
+from m1n1 import asm
+
+from m1n1.gpiola import GPIOLogicAnalyzer
+
+analyzer_cpu = 1
+
+p.pmgr_adt_clocks_enable("/arm-io/gfx-asc")
+p.pmgr_adt_clocks_enable("/arm-io/sgx")
+p.smp_start_secondaries()
+p.mmu_init_secondary(analyzer_cpu)
+iface.dev.timeout = 42
+
+## heater code
+if True:
+ code = u.malloc(0x1000)
+
+ util = asm.ARMAsm("""
+ bench:
+ mrs x1, CNTPCT_EL0
+ 1:
+ sub x0, x0, #1
+ cbnz x0, 1b
+
+ mrs x2, CNTPCT_EL0
+ sub x0, x2, x1
+ ret
+ """, code)
+ iface.writemem(code, util.data)
+ p.dc_cvau(code, len(util.data))
+ p.ic_ivau(code, len(util.data))
+
+ LOOPS = 80000000000
+ for idx in range(2, 8):
+ print(f"bench {idx}")
+ p.smp_call(idx, util.bench, LOOPS)
+
+agx = AGX(u)
+
+mon = RegMonitor(u, ascii=True, bufsize=0x8000000)
+agx.mon = mon
+
+sgx = agx.sgx_dev
+#mon.add(sgx.gpu_region_base, sgx.gpu_region_size, "contexts")
+#mon.add(sgx.gfx_shared_region_base, sgx.gfx_shared_region_size, "gfx-shared")
+#mon.add(sgx.gfx_handoff_base, sgx.gfx_handoff_size, "gfx-handoff")
+
+#mon.add(agx.initdasgx.gfx_handoff_base, sgx.gfx_handoff_size, "gfx-handoff")
+
+atexit.register(p.reboot)
+agx.start()
+
+print("==========================================")
+print("## After init")
+print("==========================================")
+mon.poll()
+agx.poll_objects()
+
+ctx = GPUContext(agx)
+ctx.bind(63)
+ctx0 = GPUContext(agx)
+ctx0.bind(62)
+
+f = GPUFrame(ctx, sys.argv[1], track=False)
+f2 = GPUFrame(ctx0, sys.argv[1], track=False)
+
+RENDERERS = 4
+FRAMES = 8
+
+renderers = []
+
+fault_cmdbuf = f.cmdbuf.clone()
+#fault_cmdbuf.depth_buffer = 0xdeadb000
+
+for i in range(RENDERERS):
+ c = ctx0 if i == 0 else ctx
+ r = GPURenderer(c, 32, bm_slot=0x10 + i, queue=1)
+ renderers.append(r)
+
+ for q in (r.wq_3d, r.wq_ta):
+ q.info.set_prio(2)
+ q.info.push()
+
+#for r in renderers[2:4]:
+ #for q in (r.wq_3d, r.wq_ta):
+ #q.info.set_prio(3)
+ #q.info.push()
+
+#for r in renderers[4:6]:
+ #for q in (r.wq_3d, r.wq_ta):
+ #q.info.set_prio(0)
+ #q.info.push()
+
+#for r in renderers[6:8]:
+ #for q in (r.wq_3d, r.wq_ta):
+ #q.info.set_prio(1)
+ #q.info.push()
+
+print("==========================================")
+print("## Submitting")
+print("==========================================")
+
+for i, r in enumerate(renderers):
+ for j in range(FRAMES):
+ if (i, j) in ((1, 0), (2, 1), (3, 1)):
+ r.submit(fault_cmdbuf)
+ elif i == 0:
+ r.submit(f2.cmdbuf)
+ else:
+ r.submit(f.cmdbuf)
+
+print("==========================================")
+print("## Submitted")
+print("==========================================")
+
+def t(addr):
+ paddr = agx.uat.iotranslate(0, addr, 4)[0][0]
+ if paddr is None:
+ raise Exception(f"Failed to iotranslate {addr:#x}")
+ return paddr
+
+regs = {
+ "ta_cmds": t(agx.initdata.regionB.stats_ta.addrof("total_cmds")),
+ "ta_ts": t(agx.initdata.regionB.stats_ta.stats.addrof("unk_timestamp")),
+}
+
+pend_base = agx.initdata.regionC.addrof("pending_stamps")
+for i in range(5):
+ regs[f"st{i}_info"] = t(pend_base + i*8)
+ regs[f"st{i}_val"] = t(pend_base + i*8 + 4)
+
+for i in range(4):
+ regs[f"ta{i}_cq"] = t(agx.initdata.regionB.stats_ta.stats.queues[i].addrof("cur_cmdqueue"))
+
+regs.update({
+ "pwr_status": t(agx.initdata.regionB.hwdata_a.addrof("pwr_status")),
+ "pstate": t(agx.initdata.regionB.hwdata_a.addrof("cur_pstate")),
+ "temp_c": t(agx.initdata.regionB.hwdata_a.addrof("temp_c")),
+ "pwr_mw": t(agx.initdata.regionB.hwdata_a.addrof("avg_power_mw")),
+ "pwr_ts": t(agx.initdata.regionB.hwdata_a.addrof("update_ts")),
+
+ #"unk_10": t(agx.initdata.regionB.hwdata_a.addrof("unk_10")),
+ #"unk_14": t(agx.initdata.regionB.hwdata_a.addrof("unk_14")),
+ "actual_pstate": t(agx.initdata.regionB.hwdata_a.addrof("actual_pstate")),
+ "tgt_pstate": t(agx.initdata.regionB.hwdata_a.addrof("tgt_pstate")),
+ #"unk_40": t(agx.initdata.regionB.hwdata_a.addrof("unk_40")),
+ #"unk_44": t(agx.initdata.regionB.hwdata_a.addrof("unk_44")),
+ #"unk_48": t(agx.initdata.regionB.hwdata_a.addrof("unk_48")),
+ "freq_mhz": t(agx.initdata.regionB.hwdata_a.addrof("freq_mhz")),
+
+ #"unk_748.0": t(agx.initdata.regionB.hwdata_a.addrof("unk_748")),
+ #"unk_748.1": t(agx.initdata.regionB.hwdata_a.addrof("unk_748")+4),
+ #"unk_748.2": t(agx.initdata.regionB.hwdata_a.addrof("unk_748")+8),
+ #"unk_748.3": t(agx.initdata.regionB.hwdata_a.addrof("unk_748")+12),
+ #"use_percent": t(agx.initdata.regionB.hwdata_a.addrof("use_percent")),
+ #"unk_83c": t(agx.initdata.regionB.hwdata_a.addrof("unk_83c")),
+ "freq_with_off": t(agx.initdata.regionB.hwdata_a.addrof("freq_with_off")),
+ #"unk_ba0": t(agx.initdata.regionB.hwdata_a.addrof("unk_ba0")),
+ #"unk_bb0": t(agx.initdata.regionB.hwdata_a.addrof("unk_bb0")),
+ #"unk_c44": t(agx.initdata.regionB.hwdata_a.addrof("unk_c44")),
+ #"unk_c58": t(agx.initdata.regionB.hwdata_a.addrof("unk_c58")),
+
+ #"unk_3ca0": t(agx.initdata.regionB.hwdata_a.addrof("unk_3ca0")),
+ #"unk_3ca8": t(agx.initdata.regionB.hwdata_a.addrof("unk_3ca8")),
+ #"unk_3cb0": t(agx.initdata.regionB.hwdata_a.addrof("unk_3cb0")),
+ #"ts_last_idle": t(agx.initdata.regionB.hwdata_a.addrof("ts_last_idle")),
+ #"ts_last_poweron": t(agx.initdata.regionB.hwdata_a.addrof("ts_last_poweron")),
+ #"ts_last_poweroff": t(agx.initdata.regionB.hwdata_a.addrof("ts_last_poweroff")),
+ #"unk_3cd0": t(agx.initdata.regionB.hwdata_a.addrof("unk_3cd0")),
+
+ "halt_count": t(agx.initdata.fw_status.addrof("halt_count")),
+ "halted": t(agx.initdata.fw_status.addrof("halted")),
+ "resume": t(agx.initdata.fw_status.addrof("resume")),
+ "unk_40": t(agx.initdata.fw_status.addrof("unk_40")),
+ "unk_ctr": t(agx.initdata.fw_status.addrof("unk_ctr")),
+ "unk_60": t(agx.initdata.fw_status.addrof("unk_60")),
+ "unk_70": t(agx.initdata.fw_status.addrof("unk_70")),
+ "c_118c0": t(agx.initdata.regionC._addr + 0x118c0),
+ "c_118c4": t(agx.initdata.regionC._addr + 0x118c4),
+ "c_118c8": t(agx.initdata.regionC._addr + 0x118c8),
+ "c_118cc": t(agx.initdata.regionC._addr + 0x118cc),
+ "c_118d0": t(agx.initdata.regionC._addr + 0x118d0),
+ "c_118d4": t(agx.initdata.regionC._addr + 0x118d4),
+ "c_118d8": t(agx.initdata.regionC._addr + 0x118d8),
+ "c_118dc": t(agx.initdata.regionC._addr + 0x118dc),
+ "3d_cmds": t(agx.initdata.regionB.stats_3d.addrof("total_cmds")),
+ #"3d_tvb_oflws_1": t(agx.initdata.regionB.stats_3d.stats.addrof("tvb_overflows_1")),
+ #"3d_tvb_oflws_2": t(agx.initdata.regionB.stats_3d.stats.addrof("tvb_overflows_2")),
+ "3d_cur_stamp_id": t(agx.initdata.regionB.stats_3d.stats.addrof("cur_stamp_id")),
+ "3d_ts": t(agx.initdata.regionB.stats_3d.stats.addrof("unk_timestamp")),
+ #"3d_cur_stamp_id": t(agx.initdata.regionB.stats_3d.stats.addrof("cur_stamp_id")),
+})
+
+for i in range(4):
+ regs[f"3d{i}_cq"] = t(agx.initdata.regionB.stats_3d.stats.queues[i].addrof("cur_cmdqueue"))
+
+
+for i, r in enumerate(renderers):
+ regs.update({
+ f"r{i}_3d_done": t(r.wq_3d.info.pointers.addrof("gpu_doneptr")),
+ #f"r{i}_3d_rptr": t(r.wq_3d.info.pointers.addrof("gpu_rptr")),
+ f"r{i}_3d_busy": t(r.wq_3d.info.addrof("busy")),
+ #f"r{i}_3d_blk": t(r.wq_3d.info.addrof("blocked_on_barrier")),
+ #f"r{i}_3d_2c": t(r.wq_3d.info.addrof("unk_2c")),
+ #f"r{i}_3d_54": t(r.wq_3d.info.addrof("unk_54")),
+
+ f"r{i}_ta_done": t(r.wq_ta.info.pointers.addrof("gpu_doneptr")),
+ #f"r{i}_ta_rptr": t(r.wq_ta.info.pointers.addrof("gpu_rptr")),
+ f"r{i}_ta_busy": t(r.wq_ta.info.addrof("busy")),
+ #f"r{i}_ta_blk": t(r.wq_ta.info.addrof("blocked_on_barrier")),
+ #f"r{i}_ta_2c": t(r.wq_ta.info.addrof("unk_2c")),
+ #f"r{i}_ta_54": t(r.wq_ta.info.addrof("unk_54")),
+ f"r{i}_f{j}_ta_stamp1": t(r.stamp_ta1._addr),
+ f"r{i}_ta_stamp2":t(r.stamp_ta2._addr),
+ f"r{i}_f{j}_3d_stamp1": t(r.stamp_3d1._addr),
+ f"r{i}_3d_stamp2":t(r.stamp_3d2._addr),
+ })
+
+ for j in range(FRAMES):
+ work = r.work[j]
+ regs.update({
+ f"r{i}_f{j}_3d_ts": t(work.wc_3d.ts1._addr),
+ f"r{i}_f{j}_ta_ts": t(work.wc_ta.ts1._addr),
+ })
+
+div=4
+ticks = 24000000 // div * 25
+
+la = GPIOLogicAnalyzer(u, regs=regs, cpu=analyzer_cpu, div=div)
+
+
+print("==========================================")
+print("## Poll prior to job start")
+print("==========================================")
+
+#mon.poll()
+#agx.poll_objects()
+
+
+print("==========================================")
+print("## After start")
+print("==========================================")
+#agx.poll_objects()
+
+#mon.poll()
+print("==========================================")
+print("## Waiting")
+print("==========================================")
+
+print("Queues:")
+for i, r in enumerate(renderers):
+ print(f" Renderer {i}")
+ print(f" TA: {r.wq_ta.info._addr:#x} (stamp {r.work[0].ev_ta.id})")
+ #print(r.wq_ta.info)
+ print(f" 3D: {r.wq_3d.info._addr:#x} (stamp {r.work[0].ev_3d.id})")
+ #print(r.wq_3d.info)
+
+print("==========================================")
+print("## Run")
+print("==========================================")
+
+la.start(ticks, bufsize=0x8000000)
+
+try:
+ for r in renderers[:RENDERERS]:
+ r.run()
+
+ for r in renderers[:RENDERERS]:
+ while not r.ev_3d.fired:
+ agx.asc.work()
+ agx.poll_channels()
+ print("==========================================")
+ agx.poll_objects()
+ mon.poll()
+
+ r.wait()
+
+ #agx.poll_objects()
+
+ #print("==========================================")
+ #print("## Stop ASC")
+ #print("==========================================")
+
+ #agx.asc.stop()
+
+ ##time.sleep(0.1)
+
+ ##agx.poll_objects()
+
+ #print("==========================================")
+ #print("## Start ASC")
+ #print("==========================================")
+
+ #agx.asc.start()
+
+ ##agx.poll_objects()
+
+ #print("==========================================")
+ #print("## Run 2")
+ #print("==========================================")
+
+ #for r in renderers[RENDERERS//2:]:
+ #r.run()
+
+ #for r in renderers[RENDERERS//2:]:
+ #while not r.ev_3d.fired:
+ #agx.asc.work()
+ #agx.poll_channels()
+ #print("==========================================")
+
+ #r.wait()
+
+ #agx.poll_objects()
+
+ #mon.poll()
+
+finally:
+ #agx.poll_objects()
+ #mon.poll()
+
+ la.complete()
+ la.show()
+
+time.sleep(2)
+
generated by cgit v1.2.1 (git 2.18.0) at 2026-01-10 20:31:16 +0000