DDR example

examples/app_single_model_on_one_tile_and_DDR/Makefile

@@ -0,0 +1,43 @@
+TARGET = XCORE-AI-EXPLORER
+APP_NAME = app_device
+
+APP_FLAGS = -report \
+-O3 -fcomment-asm -mcmodel=large -fxscope -Wno-xcore-fptrgroup \
+-Wno-unknown-pragmas \
+-Wno-bidirectional-buffered-port \
+-lquadflash \
+-g
+#-Wm,--image-size,8388608,--image-base,8388608
+
+# Comment out the below TFLMC_XCORE_PROFILE definition to disable profiling
+APP_FLAGS += -DTFLMC_XCORE_PROFILE
+
+# Enable the below SHARED_TENSOR_ARENA definition for multiple models 
+# or to manage tensor arena array yourself
+#APP_FLAGS += -DSHARED_TENSOR_ARENA
+
+# *******************************************************
+# Include AI Tools library and headers
+# *******************************************************
+# The following variables are defined for AI Tools library 
+# and should be included when building the app.
+# XMOS_AITOOLSLIB_DEFINITIONS
+# XMOS_AITOOLSLIB_INCLUDES
+# XMOS_AITOOLSLIB_LIBRARIES
+#
+ifeq ($(XMOS_AITOOLSLIB_PATH),)
+$(error Path to XMOS AI Tools library and headers not set correctly!)
+endif
+include ${XMOS_AITOOLSLIB_PATH}/buildfiles/aitoolslib.make
+
+XCC_FLAGS = $(APP_FLAGS) $(XMOS_AITOOLSLIB_DEFINITIONS) $(XMOS_AITOOLSLIB_INCLUDES)
+XCC_CPP_FLAGS = $(APP_FLAGS) -std=c++14 $(XMOS_AITOOLSLIB_DEFINITIONS) $(XMOS_AITOOLSLIB_INCLUDES)
+XCC_MAP_FLAGS = $(APP_FLAGS) $(XMOS_AITOOLSLIB_LIBRARIES)
+
+#=============================================================================
+# The following part of the Makefile includes the common build infrastructure
+# for compiling XMOS applications. You should not need to edit below here.
+
+XMOS_MAKE_PATH ?= ../..
+include $(XMOS_MAKE_PATH)/xcommon/module_xcommon/build/Makefile.common
+

examples/app_single_model_on_one_tile_and_DDR/generate_optimized_cpp_for_xcore.py

@@ -0,0 +1,41 @@
+from xmos_ai_tools import xformer
+
+TFLITE_MODEL_PATH = "mobilenetv2.tflite"
+OPTIMIZED_MODEL_PATH = "src/model.tflite"
+
+OPTIMIZED_MODEL_PATH = "src/model.tflite"
+WEIGHT_PARAMS_PATH = "src/model_weights.c"
+FLASH_IMAGE_PATH = "src/xcore_flash_binary.out"
+print("Generating app cpp files for model...")
+xformer.convert(
+    TFLITE_MODEL_PATH,
+    OPTIMIZED_MODEL_PATH,
+    {
+        "xcore-thread-count": "5",
+        # set conv err threshold
+        "xcore-conv-err-threshold": "3",
+        # operation splitting to reduce tensor arena size
+        "xcore-op-split-tensor-arena": "True",
+        "xcore-op-split-top-op": "0,7",
+        "xcore-op-split-bottom-op": "6,14",
+        "xcore-op-split-num-splits": "8,4",
+        # write weights as an array to be 
+        # placed on second tile
+        "xcore-load-tile" : "True",
+        # roughly(tensors are not split) specifies 
+        # size of weights to move out to header file
+        # "xcore-max-load-external-size" : "270000",
+        # move weights to this file
+        "xcore-weights-file" : WEIGHT_PARAMS_PATH,
+    },
+)
+xformer.print_optimization_report()
+
+# # Generate flash image to be flashed using xflash
+# xformer.generate_flash(
+#     output_file=FLASH_IMAGE_PATH,
+#     model_files=[OPTIMIZED_MODEL_PATH],
+#     param_files=[WEIGHT_PARAMS_PATH],
+# )
+
+print("Done!")

examples/app_single_model_on_one_tile_and_DDR/src/config.xscope

@@ -0,0 +1,21 @@
+<?xml version="1.0" encoding="UTF-8"?>
+
+<!-- ======================================================= -->
+<!-- The 'ioMode' attribute on the xSCOPEconfig              -->
+<!-- element can take the following values:                  -->
+<!--   "none", "basic", "timed"                              -->
+<!--                                                         -->
+<!-- The 'type' attribute on Probe                           -->
+<!-- elements can take the following values:                 -->
+<!--   "STARTSTOP", "CONTINUOUS", "DISCRETE", "STATEMACHINE" -->
+<!--                                                         -->
+<!-- The 'datatype' attribute on Probe                       -->
+<!-- elements can take the following values:                 -->
+<!--   "NONE", "UINT", "INT", "FLOAT"                        -->
+<!-- ======================================================= -->
+
+<xSCOPEconfig ioMode="basic" enabled="true">
+    <!-- For example: -->
+    <!-- <Probe name="Probe Name" type="CONTINUOUS" datatype="UINT" units="Value" enabled="true"/> -->
+    <!-- From the target code, call: xscope_int(PROBE_NAME, value); -->
+</xSCOPEconfig>

examples/app_single_model_on_one_tile_and_DDR/src/main.xc

@@ -0,0 +1,52 @@
+#include <stdint.h>
+#include <stdio.h>
+#include <platform.h>
+#include "tile_ram_server.h"
+// #include "flash_server.h"
+
+// #define NFLASH_SPECS 1
+
+// fl_QuadDeviceSpec flash_spec[NFLASH_SPECS] = {
+//     FL_QUADDEVICE_DEFAULT //FL_QUADDEVICE_MACRONIX_MX25R6435FM2IH0
+// };
+
+// on tile[0]: fl_QSPIPorts qspi = {
+//     PORT_SQI_CS,
+//     PORT_SQI_SCLK,
+//     PORT_SQI_SIO,
+//     XS1_CLKBLK_2
+// };
+
+#define NUMBER_OF_MODELS 1
+extern void model_init(chanend f);
+extern void inference();
+
+// Hack for xc to be happy with EXTMEM
+// Have to specify the extern array size
+extern const int8_t tile_server_weights[3587236];
+const int8_t* weights = (int8_t*)&tile_server_weights;
+
+int main(void) {
+    chan c_flash_or_tile[NUMBER_OF_MODELS];
+
+    par {
+        on tile[0]: {
+            flash_t headers[NUMBER_OF_MODELS];
+            tile_ram_server(c_flash_or_tile, headers, NUMBER_OF_MODELS, weights);
+            // flash_server(c_flash_or_tile, headers, NUMBER_OF_MODELS, qspi, flash_spec, 1);
+        }
+
+        on tile[1]: {
+            unsafe {
+            c_flash_or_tile[0] <: FLASH_SERVER_INIT;
+
+            model_init(c_flash_or_tile[0]);
+
+            inference();
+
+            c_flash_or_tile[0] <: FLASH_SERVER_QUIT;
+            }
+        }
+    }
+    return 0;
+}