#!/usr/bin/env python3 # SPDX-License-Identifier: MIT import sys, pathlib, time sys.path.append(str(pathlib.Path(__file__).resolve().parents[1])) from m1n1.setup import * from m1n1 import asm # disable disp0 p.write32(0x23b700360, 0) p.write32(0x23b710018, 0) # mcc init configs = [ (0x3ff001a, 0x33, 0), (0x3ff0039, 0x32, 0), (0x3ff003a, 0x32, 0), (0x3ff0013, 0x00, 0x200), (0x3ff0016, 0x00, 0x200), (0x3ff0012, 0x33, 0x200), (0x3ff0014, 0x33, 0x200), (0x3ff001a, 0x31, 0), (0x3ff001a, 0x32, 0), ] for i, (v0, v1, v2) in enumerate(configs): p.write32(0x20020058c + i * 0x14, v0) p.write32(0x200200590 + i * 0x14, 0) p.write32(0x200200594 + i * 0x14, v1) p.write32(0x200200598 + i * 0x14, 0) p.write32(0x20020059c + i * 0x14, 0) p.write32(0x2002005a0 + i * 0x14, v2) print(p.read64(0x200204000 + 8*i)) p.write32(0x200200580, 0x71) p.set32(0x200200584, 1 << i) p.write32(0x23d2b000c, 0) p.write64(0x210e440f8, 1) p.write64(0x211e440f8, 1) p.write64(0x210e200f8, 0x10001900000) p.write64(0x210e70210, 0x96b4026c05090ac0) p.write64(0x210e70218, 0xa20201f0d000303) p.write64(0x211e200f8, 0x10001900000) p.write64(0x211e70210, 0x3d7a01c201204fa4) p.write64(0x211e70218, 0x408101f18020202) for i in range(8): p.set32(0x20020000c + i * 0x40000, 0x1000) freq = u.mrs(CNTFRQ_EL0) BYTES = (256 * 1024 * 1024) data = u.malloc(BYTES * 2) p.memset64(data, 0x1000, BYTES * 2) 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 copy: mrs x6, CNTPCT_EL0 mov x7, x0 1: ldp x15, x14, [x1, #0x0] stp x15, x14, [x2, #0x0] ldp x13, x12, [x1, #0x10] stp x13, x12, [x2, #0x10] ldp x11, x10, [x1, #0x20] stp x11, x10, [x2, #0x20] ldp x9, x8, [x1, #0x30] stp x9, x8, [x2, #0x30] add x1, x1, #0x40 add x2, x2, #0x40 sub x0, x0, #0x40 cbnz x0, 1b mov x0, x7 sub x1, x1, x7 sub x2, x2, x7 sub x3, x3, #1 cbnz x3, 1b mrs x7, CNTPCT_EL0 sub x0, x7, x6 ret latency: mrs x2, s3_1_c15_c0_0 /* SYS_IMP_APL_PMCR0 */ orr x2, x2, #1 msr s3_1_c15_c0_0, x2 mov x2, #0xffffffffffffffff msr s3_1_c15_c1_0, x2 isb mov x3, #1 1: ldp x2, x7, [x1] add x1, x1, #0x8 isb sub x3, x3, #1 cbnz x3, 1b add x1, x1, #0x80000 mrs x6, s3_2_c15_c0_0 add x7, x6, x0 1: mrs x5, s3_2_c15_c0_0 cmp x5, x7 bmi 1b isb dsb sy mov x3, #10 1: ldr x0, [x1, #0x4000] add x1, x1, x0 add x1, x1, #0x80000 sub x3, x3, #1 cbnz x3, 1b dsb sy isb mrs x3, s3_2_c15_c0_0 isb sub x0, x3, x5 ret """, code) iface.writemem(code, util.data) p.dc_cvau(code, len(util.data)) p.ic_ivau(code, len(util.data)) CREG = [ 0x210e00000, 0x211e00000, ] CLUSTER_PSTATE = 0x20020 LOOPS = 10000000 CLOOPS = 1 def bench_cpu(idx): if idx == 0: elapsed = p.call(util.bench, LOOPS) / freq else: elapsed = p.smp_call_sync(idx, util.bench, LOOPS) / freq if elapsed == 0: return 0 mhz = (LOOPS / elapsed) / 1000000 return mhz def bench_bw(idx): if idx == 0: elapsed = p.call(util.copy | REGION_RX_EL1, BYTES, data, data + BYTES, CLOOPS) / freq else: elapsed = p.smp_call_sync(idx, util.copy | REGION_RX_EL1, BYTES, data, data + BYTES, CLOOPS) / freq if elapsed == 0: return 0 mbs = (BYTES * CLOOPS / elapsed) / 1000000 return mbs def bench_lat(idx, delay, addr): if idx == 0: elapsed = p.call(util.latency | REGION_RX_EL1, delay, addr) else: p.smp_call(idx, util.latency | REGION_RX_EL1, delay, addr) time.sleep(0.1) elapsed = p.smp_wait(idx) if elapsed == 0: return 0 return elapsed def do_bench(): idx = 0 print(f"CPU: {bench_cpu(idx):.2f} MHz") mbs = bench_bw(idx) print(f"CPU: {mbs:.2f} MB/s") p = data + 13 * 1024 * 1024 for i in (0, 1, 2, 3, 4, 5, 10, 20, 100, 1000, 10000, 100000, 1000000, 10000000, 20000000, 50000000): c = 0 runs = 16 for j in range(runs): p = p + 0x14000 c += bench_lat(idx, i, p) print(f"Latency @{i}: {c / runs} cycles") print() print(f"Data buffer @ {data:#x}") p.smp_start_secondaries() for i in range(1, 8): p.mmu_init_secondary(i) e_pstate = p.read64(CREG[0] + CLUSTER_PSTATE) p_pstate = p.read64(CREG[1] + CLUSTER_PSTATE) print(f"E-Core pstate: {e_pstate:x}") print(f"P-Core pstate: {p_pstate:x}") print("== Initial CPU frequencies ==") mcc_configs = [ (0x813057f, 0x1800180), (0x133, 0x55555340), (0, 0x55555340), ] def set_pstate(cluster, pstate, mcc=True): # This really seems to be all that's needed ps1 = pstate ps2 = pstate val = (1<<25) | ps1 | (ps2 << 12) val = p.writeread64(CREG[cluster] + CLUSTER_PSTATE, val) print(f"pstate = {val:#x}") print(f"pstate = {p.read64(CREG[cluster] + CLUSTER_PSTATE):#x}") #p.mask64(CREG[cluster] + CLUSTER_PSTATE, 0xffff0f, val) # Optionally, adjust MCC performance in higher p-core pstates p0, p1 = mcc_configs[mcc] for lane in range(8): p.write32(0x200200dc4 + lane * 0x40000, p0) p.write32(0x200200dbc + lane * 0x40000, p1) # This seems to be about notifying PMP #p.write32(0x23b738004 + cluster*4, pstate) #p.write32(0x23bc34000, 1 << cluster) set_pstate(0, 5, mcc=0) set_pstate(1, 10, mcc=0) do_bench() set_pstate(1, 10, mcc=1) do_bench() set_pstate(1, 10, mcc=2) do_bench()