10. 3. Flashing and Validation

This phase documents how to flash the MCU firmware onto the Shikra CS2390 and verify that model inference executed successfully.

10.3.1 Flash the Firmware

The build produces two firmware binaries under build/ms/bin/shikra.lpaicp.test.

adb shell mount -o rw,remount /usr
adb push lpaicp.mbn /usr/lib/firmware/qcom/shikra/
adb push lpaicp_dtb.mbn /usr/lib/firmware/qcom/shikra/
adb shell sync
adb shell reboot

10.3.2. Validate the LPAI Subsystem Is Running

Once the device has booted, check that the remoteproc driver has loaded the LPAI firmware successfully:

cat /sys/class/remoteproc/remoteproc*/{state,name}

Expected output: The state file should read running and the name file should identify the LPAI core (e.g. soccp). Any other state (e.g. offline, crashed) indicates a load failure — check dmesg | grep -i remoteproc for details.

10.3.3. T32 Analysis

The Shikra LPAICP MCU has no UART — all console output goes to a RAM console (ram_console_buf). Debug and result readout is done through Lauterbach T32 over RISC-V JTAG. The TFLM inference harness writes all results to volatile global variables which T32 can read at any time — no need to set breakpoints.

10.3.3.1. Get the RAM console address

The ram_console_buf address changes with every build. Get it from the ELF:

nm LPAICP_IMG_shikra.lpaicp.test.elf | grep ram_console_buf

Example output:

b45a7a8e B ram_console_buf

The address is 0xb45a7a8e. Update the CMM script with the address from your build.

10.3.3.2. T32 CMM script

Update the &elf path and ram_console_buf address to match your build, then run the script in T32.

; ============================================================================
; read_tflite_infer.cmm
; Read TFLite-Micro model latency results from a device that is ALREADY booted
; and running the inference loop.
;
; tflite_infer_run() runs a batch of timed invokes, publishes fresh
; min/avg/max into the g_tfli_* globals, bumps g_tfli_batches LAST (the
; commit marker), sleeps, and repeats. Halting at any time and reading the
; globals yields the latest coherent batch -- as long as g_tfli_batches >= 1
; and g_tfli_status == 0.
;
; How to run (after the target is connected):
;   DO <path-to-this-script>/read_tflite_infer.cmm
; ============================================================================
local &elf
local &ram_console_addr

; ==== EDIT THESE TWO VALUES for your build ====================================
; Full path to the ELF as this Windows machine sees it.
&elf = "C:\path\to\LPAICP_IMG_shikra.lpaicp.test.elf"
; Get this address from: nm <elf> | grep ram_console_buf
&ram_console_addr = 0xb45a7a8e
; ==============================================================================

; ---- load symbols only (do NOT touch target code / do NOT reset) ------------
Data.LOAD.Elf "&elf" /NOCODE

; ---- halt so memory reads are coherent (does not restart anything) ----------
IF STATE.RUN()
(
    PRINT "read_tflite_infer: halting hart to read memory..."
    Break
    WAIT !STATE.RUN()
)

; ---- display the latest published batch -------------------------------------
; status / arena / iters / commit-count
Var.View g_tfli_status g_tfli_batches g_tfli_arena_used g_tfli_arena_size g_tfli_iters

; latency per inference in microseconds (QTMR 19.2 MHz) -- the headline number
Var.View g_tfli_min_us g_tfli_avg_us g_tfli_max_us

; latency in raw QTMR ticks + RISC-V CPU cycles (mcycle) + effective core MHz
Var.View g_tfli_min_ticks g_tfli_avg_ticks g_tfli_max_ticks
Var.View g_tfli_min_mcyc g_tfli_avg_mcyc g_tfli_max_mcyc g_tfli_cpu_mhz

; output tensor of the last invoke : raw int8 and argmax (predicted class)
Var.View g_tfli_out g_tfli_argmax
; Note: Var.View %Float g_tfli_out_f requires an HLL Debugging license.
; If not available, skip it -- all latency results are accessible without HLL.

PRINT "Dumping ram_console_buf..."
Data.dump EZAXI:&ram_console_addr /DIALOG /NoHex /STRING

PRINT "PASS = g_tfli_status==0, g_tfli_batches>=1, non-zero g_tfli_avg_us"
PRINT "Latency = g_tfli_avg_us microseconds/inference"
enddo

10.3.4. T32 Global Variables Reference

Variable	Type	Meaning
`g_tfli_status`	`int32_t`	`0` = success; `-1` = never ran; `-3` = k_malloc failed; `-4` = AllocateTensors failed; `-5` = input size mismatch; `-6` = warm-up Invoke failed; `-7` = timed Invoke failed
`g_tfli_batches`	`uint32_t`	Batch counter — bumped last after each publish; read when `>= 1` for coherent results
`g_tfli_arena_used`	`uint32_t`	Bytes used by tensor arena — use to tune `TFLI_ARENA_KB` (see Phase 2, section 2.5.2.1)
`g_tfli_arena_size`	`uint32_t`	Configured arena size in bytes (`TFLI_ARENA_KB * 1024`)
`g_tfli_iters`	`uint32_t`	Number of timed Invoke() calls per batch (`TFLI_ITERS`)
`g_tfli_min_us`	`uint32_t`	Minimum inference latency in microseconds
`g_tfli_avg_us`	`uint32_t`	Average inference latency in microseconds — the headline number
`g_tfli_max_us`	`uint32_t`	Maximum inference latency in microseconds
`g_tfli_min_ticks`	`uint32_t`	Minimum latency in QTMR ticks (19.2 MHz clock)
`g_tfli_avg_ticks`	`uint32_t`	Average latency in QTMR ticks
`g_tfli_max_ticks`	`uint32_t`	Maximum latency in QTMR ticks
`g_tfli_min_mcyc`	`uint64_t`	Minimum latency in RISC-V `mcycle` CPU cycles
`g_tfli_avg_mcyc`	`uint64_t`	Average latency in RISC-V `mcycle` CPU cycles
`g_tfli_max_mcyc`	`uint64_t`	Maximum latency in RISC-V `mcycle` CPU cycles
`g_tfli_cpu_mhz`	`uint32_t`	Effective CPU MHz (avg_mcyc / avg_us)
`g_tfli_out[2]`	`int8_t[2]`	Raw output tensor (int8) — 2 values for 2-class model
`g_tfli_out_f[2]`	`float[2]`	Dequantized output tensor (float) — requires HLL license to view in T32
`g_tfli_argmax`	`int32_t`	Predicted class (argmax of output tensor)

10.3.5. Interpreting Results

10.3.5.1. Validity Check

Before reporting results, always verify:

g_tfli_status  == 0          (success)
g_tfli_batches >= 1          (at least one complete batch published)

If g_tfli_status is non-zero, see the error table in Phase 2, section 2.5.2.1 for root cause.

10.3.5.2. Latency

The headline number is g_tfli_avg_us (average inference latency in microseconds over TFLI_ITERS iterations). Cross-check: g_tfli_avg_ticks / 19.2 MHz = g_tfli_avg_us (within rounding). g_tfli_cpu_mhz gives the effective CPU frequency derived from RISC-V mcycle and the QTMR timestamp — useful for confirming the CPU is running at the expected clock rate.

10.3.5.3. Arena Sizing

If g_tfli_arena_used is close to g_tfli_arena_size, increase TFLI_ARENA_KB in inc/tflite_infer.h. For the procedure to find the minimum safe value, see Phase 2, section 2.5.2.1.

10.3.5.4. RAM Console Output

The RAM console captures printk() output written during boot and inference. After a successful run, you should see lines like:

Reached main. Configuration: shikra_lpaicp_test
TFLI: arena_used=XXXX/4096 B, iters/batch=100, out_classes=2
TFLI[1]: lat avg=XXX us (min=XXX max=XXX) | cpu avg=XXXXXX cyc @~XXX MHz | argmax=0

The TFLI[N] line is the per-batch latency summary. argmax is the predicted class for the embedded input sample.

​10. 3. Flashing and Validation

​10.3.1 Flash the Firmware

​10.3.2. Validate the LPAI Subsystem Is Running

​10.3.3. T32 Analysis

​10.3.3.1. Get the RAM console address

​10.3.3.2. T32 CMM script

​10.3.4. T32 Global Variables Reference

​10.3.5. Interpreting Results

​10.3.5.1. Validity Check

​10.3.5.2. Latency

​10.3.5.3. Arena Sizing

​10.3.5.4. RAM Console Output