Merge branch 'spinler-feat-cocotb_test_throughput' into 'devel'

feat: add simulation-based throughput test for DMA

See merge request ndk/ndk-fpga!298

Author: jakubcabal

apps/minimal/tests/cocotb/cocotb_grpc.py

@@ -13,7 +13,7 @@
 import scapy.volatile
 import scapy.contrib.mpls
 
-from ndk_core import NFBDevice
+from cocotbext.ndk_core import NFBDevice
 
 import cocotbext.ofm.utils.sim.modelsim as ms
 import cocotb.utils

apps/minimal/tests/cocotb/cocotb_test.py

@@ -8,7 +8,7 @@
 import cocotb
 from cocotb.triggers import Timer
 
-from ndk_core import NFBDevice
+from cocotbext.ndk_core import NFBDevice
 
 import cocotbext.ofm.utils.sim.modelsim as ms
 import cocotb.utils

apps/minimal/tests/cocotb/cocotb_test_throughput.gpi

@@ -0,0 +1,43 @@
+DIR = ARG1
+IFN = ARG2
+CARD = ARG3
+SPEED = ARG4
+
+if (DIR eq "") DIR = "RX"
+if (IFN eq "") IFN = "transcript"
+if (CARD eq "") CARD = "unknown card"
+if (SPEED eq "") CARD = "100G"
+
+
+MIN=64
+MAX=256+64
+
+set xrange [MIN:MAX]
+set xtics 32
+set mxtics 4
+set ytics 10
+
+set style line 42 lc black lw 0.5 dt ".   "
+
+set grid xtics ytics mxtics, ls 42
+
+scl=1
+set terminal pngcairo size 1920*scl,1080*scl noenhanced font 'Verdana,14' linewidth 2
+
+set xlabel "Packet size [B] (without CRC, without NDP header)"
+set ylabel "DMA data throughput [Gbps]"
+set key right bottom
+
+
+set title DIR . " simulation throughput"
+set output "report-sim-" . DIR ."-throughput.png"
+
+cmd_getfn(ifn) = "< cat " . ifn . " | grep " . DIR . "THR | sed 's/# " . DIR . "THR: //' |sed 's/ /, /g'"
+
+
+plot \
+	cmd_getfn(IFN) using ($1):($3)*0.001 * ((($1)+20) / ($1)) with lines title SPEED . " Eth RAW" , \
+	cmd_getfn(IFN) using ($1):($3)*0.001 * ((($1)+24) / ($1)) with lines title SPEED . " Eth with CRC" , \
+	cmd_getfn(IFN) using ($1):($3)*0.001 * ((($1)+20+8) / ($1)) with lines title SPEED . " Eth with 8B NDP header" , \
+	cmd_getfn(IFN) using ($1):($2)*0.001 * (($1+20) / $1) with lines title CARD . " PCIe" , \
+

apps/minimal/tests/cocotb/cocotb_test_throughput.py

@@ -0,0 +1,284 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright (C) 2025 CESNET z. s. p. o.
+# Author(s): Martin Spinler <[email protected]>
+
+# INFO
+######
+# This testbench measures maximal throughput of RX+TX DMA
+# and produces graphs in app/build/card directory
+
+import os
+import sys
+import logging
+import cocotb
+import itertools
+from cocotb.triggers import Timer, Event, First, RisingEdge
+
+from cocotbext.ndk_core import NFBDevice
+
+import cocotbext.ofm.utils.sim.modelsim as ms
+import cocotb.utils
+
+from cocotbext.ofm.utils.sim.bus import MfbBus, MiBus, DmaUpMvbBus, DmaDownMvbBus
+
+
+from ofm.comp.mfb_tools.debug.gen_loop_switch import GenLoopSwitch
+
+# Configuration parameters
+##########################
+
+# used channels per DMA_ENDPOINT
+USED_CHANNELS = {
+    "RX": 16,
+    "TX": 16,
+}
+# Max expected throughput (for graph generation)
+PORT_THROUGHPUT = 100
+PKTLEN = [x for x in range(64, 300) if x % 8 in [0, 1]]
+
+
+logging.basicConfig(stream=sys.stderr, force=True)
+logging.getLogger().setLevel(logging.INFO)
+
+logger = logging.getLogger(__name__)
+
+logger_mi = logging.getLogger("cocotb.nfb.ext.python_servicer")
+#logger_mi.setLevel(logging.DEBUG)
+
+# Shortcuts
+e = cocotb.external
+st = cocotb.utils.get_sim_time
+add_cursor = ms.add_cursor
+print = ms.print
+
+
+def run_gnuplot(d, dev):
+    file_path = os.path.realpath(__file__)[:-3] # crop the .py
+    os.system(f'gnuplot -c "{file_path}.gpi" "{d}" "{os.getcwd()}/transcript" "{dev._card_name}" "{PORT_THROUGHPUT}G"')
+
+
+async def get_dev(dut, init=True, **kwargs):
+    dev = NFBDevice(dut, **kwargs)
+    if init:
+        await dev.init()
+    return dev, dev.nfb
+
+
+async def sendmsgs(txq, msgs):
+    await e(txq.sendmsg)(msgs)
+
+
+async def rx_push_descs(rxq, n, ev):
+    burst = 64
+    for i in range(n):
+        pd = cocotb.start_soon(rxq._push_desc(flush=((i % burst) == (burst - 1))))
+        evw = ev.wait()
+        tr = await First(pd, evw)
+        if tr == evw:
+            return
+
+
+async def recvmsgs(rxq, event):
+    while not event.is_set():
+        m = await e(rxq.recvmsg)()
+        if m:
+            pass
+        else:
+            await Timer(5, units='ns')
+
+
+async def rx_stop_ch(rxq):
+    await e(rxq.stop)()
+
+
+async def wait_for_val(signal, val, clk, n, to=10000):
+    re = RisingEdge(clk)
+    i = 0
+    while i != n:
+        i = i + 1 if signal.value == val else 0
+        await re
+        to -= 1
+        if to <= 0:
+            raise TimeoutError()
+
+
+def get_msg(pktlen, n, ch):
+    return (bytes(itertools.chain([0, ch, (n >> 8) & 0xFF, (n) & 0xFF] * ((pktlen + 3) // 4)))[:pktlen], bytes(), 0)
+
+
+def get_channel_list(tdir):
+    EP_CHANNELS = getattr(core.dma_i, f"{tdir}_CHANNELS").value
+
+    _chnls = min(USED_CHANNELS[tdir], EP_CHANNELS)
+    chnls = []
+    for ch in range(DMA_STREAMS * EP_CHANNELS):
+        if ch % EP_CHANNELS < _chnls:
+            chnls.append(ch)
+    return _chnls, chnls
+
+
+@cocotb.test(timeout_time=80000, timeout_unit='us', skip=False)
+async def test_ndp_sendmsg_burst(dut):
+    tdir = "TX"
+    clk  = core.dma_i.USR_CLK
+    re = RisingEdge(clk)
+    bm = busm[tdir]
+    for m in bm:
+        cocotb.start_soon(m.monitor(clk))
+
+    dev, nfb = await get_dev(dut)
+
+    gls = []
+    for i in range(DMA_STREAMS):
+        gls.append(GenLoopSwitch(nfb, index=i))
+
+    for g in gls:
+        await e(setattr)(g.r2l, 'input', 2)
+
+    stream_chnls, chnls = get_channel_list(tdir)
+    tasks = []
+    for pktlen in PKTLEN:
+        for ch in chnls:
+            await e(nfb.ndp.tx[ch].start)()
+
+        for ch in chnls:
+            npkts = min(256, 262144 // pktlen)
+            msgs = [get_msg(pktlen, n, ch) for n in range(npkts)]
+            t = cocotb.start_soon(sendmsgs(nfb.ndp.tx[ch], msgs))
+            tasks.append((ch, t))
+
+        for m in bm:
+            await wait_for_val(m._get_handle('SRC_RDY'), 1, core.dma_i.USR_CLK, 1, 10000)
+        for i in range(1000):
+            await re
+
+        #add_cursor(f"Measure start {pktlen}")
+        for m in bm:
+            m.clear()
+        for i in range(2000):
+            await re
+        #add_cursor(f"Measure stop {pktlen}")
+
+        mpps = PORT_THROUGHPUT * 1e3 / (8 * (pktlen + 24))
+        eth_raw = (pktlen * mpps * 8)
+        print("TXTHR:", pktlen, sum(m._thr * 1000 for m in bm), eth_raw)
+
+        for ch, t in tasks:
+            await t
+        for ch, t in tasks:
+            await e(nfb.ndp.tx[ch].stop)()
+
+        for m in bm:
+            await wait_for_val(m._get_handle('SRC_RDY'), 0, core.dma_i.USR_CLK, 150, 10000)
+        tasks.clear()
+
+    run_gnuplot(tdir, dev)
+
+
+@cocotb.test(timeout_time=80000, timeout_unit='us', skip=False)
+async def test_ndp_recvmsg_burst(dut):
+    tdir = "RX"
+    clk  = core.dma_i.USR_CLK
+    re = RisingEdge(clk)
+    bm = busm[tdir]
+    for m in bm:
+        cocotb.start_soon(m.monitor(clk))
+
+    dev, nfb = await get_dev(dut)
+
+    gls = []
+    for i in range(DMA_STREAMS):
+        gls.append(GenLoopSwitch(nfb, index=i))
+
+    stop_channels = False
+    channels_running = False
+
+    stream_chnls, chnls = get_channel_list(tdir)
+    tasks = []
+    for pktlen in PKTLEN:
+        if not channels_running:
+            channels_running = True
+            for ch in chnls:
+                await e(nfb.ndp.rx[ch].start)()
+                evd = Event()
+                td = cocotb.start_soon(rx_push_descs(dev.dma.rx[ch], 2**24, evd))
+                await re
+                evr = Event()
+                tr = cocotb.start_soon(recvmsgs(nfb.ndp.rx[ch], evr))
+                tasks.append((ch, td, tr, evd, evr))
+
+            for i in range(300):
+                await re
+        for g in gls:
+            await e(g.l2r.gen_start)(True, pktlen, 0, stream_chnls)
+
+        for m in bm:
+            await wait_for_val(m._get_handle('SRC_RDY'), 1, core.dma_i.USR_CLK, 10, 10000)
+
+        for i in range(1000):
+            await re
+        #add_cursor(f"Measure start {pktlen}")
+        for m in bm:
+            m.clear()
+        for i in range(3000):
+            await re
+        #add_cursor(f"Measure stop {pktlen}")
+
+        mpps = PORT_THROUGHPUT * 1e3 / (8 * (pktlen + 20))
+        eth_raw = (pktlen * mpps * 8)
+        print("RXTHR:", pktlen, sum(m._thr * 1000 for m in bm), eth_raw)
+
+        for g in gls:
+            await e(g.l2r.gen_stop)()
+
+        for i in range(500):
+            await re
+        for m in bm:
+            await wait_for_val(m._get_handle('SRC_RDY'), 0, core.dma_i.USR_CLK, 10000, 200000)
+
+        for i in range(100):
+            await re
+
+        if stop_channels:
+            channels_running = False
+
+            tasks_stop = []
+            for ch, td, tr, evd, evr in tasks:
+                evd.set()
+                await td
+                evr.set()
+                await tr
+                t = cocotb.start_soon(rx_stop_ch(nfb.ndp.rx[ch]))
+                tasks_stop.append(t)
+            for t in tasks_stop:
+                await t
+
+    run_gnuplot(tdir, dev)
+
+
+core = NFBDevice.core_instance_from_top(cocotb.top)
+#ms.cmd(f"log -recursive {ms.cocotb2path(core)}/*")
+
+DMA_STREAMS = core.dma_i.DMA_STREAMS.value
+
+pcic = core.pcie_i.pcie_core_i
+adapter = pcic.pcie_adapter_g[0].pcie_adapter_i
+
+busm = {"RX": [], "TX": []}
+for d in busm:
+    for i in range(DMA_STREAMS):
+        t = MfbBus(core.dma_i.gls_g[i].gls_en_g.gen_loop_switch_i, f'DMA_{d}_MFB', label=f"DMA_{d}_MFB{i}")
+        t.add_wave()
+        busm[d].append(t)
+
+for i in range(core.dma_i.DMA_ENDPOINTS.value):
+    DmaUpMvbBus(core.dma_i, 'PCIE_RQ_MVB', i, label=f"RQ_MVB_{i}").add_wave()
+    MfbBus(core.dma_i, 'PCIE_RQ_MFB', i, label=f"RQ_MFB_{i}").add_wave()
+    DmaDownMvbBus(core.dma_i, 'PCIE_RC_MVB', i, label=f"RC_MVB_{i}").add_wave()
+    MfbBus(core.dma_i, 'PCIE_RC_MFB', i, label=f"RC_MFB_{i}").add_wave()
+
+#for m in ["RC", "RQ", "CC", "CQ"]:
+#    MfbBus(pcic, '{m}_MFB', 0).add_wave()
+
+for i in range(core.pcie_i.PCIE_ENDPOINTS.value):
+    MiBus(core.pcie_i, 'MI', i, label=f'MI_PCIe{i}').add_wave()

apps/minimal/tests/cocotb/issues/issue1.py

@@ -11,7 +11,7 @@
 from cocotbext.ofm.utils.sim.bus import MfbBus, DmaDownMvbBus
 from cocotbext.ofm.utils.scapy import simple_tcp_bytes
 
-from ndk_core import NFBDevice
+from cocotbext.ndk_core import NFBDevice
 
 print = ms.print
 

apps/minimal/tests/cocotb/issues/issue2.py

@@ -9,7 +9,7 @@
 
 from cocotbext.ofm.utils.sim.bus import MiBus
 
-from ndk_core import NFBDevice
+from cocotbext.ndk_core import NFBDevice
 
 print = ms.print
 

comp/debug/mem_tester/amm_gen/amm_gen.vhd

@@ -207,7 +207,7 @@ architecture FULL of AMM_GEN is
     signal amm_write_delayed        : std_logic_vector(AMM_WRITE_DELAY downto 0);
     signal amm_read_delayed         : std_logic_vector(AMM_READ_DELAY downto 0);
 
-    signal target_burst_cnt         : std_logic_vector(BURST_BITS - 1 downto 0);
+    signal target_burst_cnt         : std_logic_vector(BURST_BITS - 1 downto 0) := std_logic_vector(to_unsigned(INIT_BURST_CNT, BURST_BITS));
     signal target_burst_cnt_lim     : std_logic_vector(BURST_BITS - 1 downto 0);    -- Indexed from 0 to match curr_burst
     signal curr_burst_cnt           : std_logic_vector(BURST_BITS - 1 downto 0);
     -- To restore burst cnt when amm_ready occurs