Add XDMA Software (#61)

target/fpga/sw/send_uart.sh

@@ -11,10 +11,10 @@ else
 		FILE=$1
 	fi
 
-	stty -F /dev/ttyUSB1 cs8 1000000 ignbrk -brkint -imaxbel -opost -onlcr -isig -icanon -iexten -echo -echoe -echok -echoctl -echoke noflsh -ixon crtscts
+	stty -F /dev/ttyUSB2 cs8 1000000 ignbrk -brkint -imaxbel -opost -onlcr -isig -icanon -iexten -echo -echoe -echok -echoctl -echoke noflsh -ixon crtscts
 
-	echo -n 2 > /dev/ttyUSB1
+	echo -n 2 > /dev/ttyUSB2
 
-	sx -k "$FILE" < /dev/ttyUSB1 > /dev/ttyUSB1
+	sx -k "$FILE" < /dev/ttyUSB2 > /dev/ttyUSB2
 
 fi

target/sim/sw/device/apps/snax/snax-test-integration/Makefile

@@ -14,4 +14,4 @@ include ./data/Makefile
 include ../../common.mk
 
 
-$(DEP): $(DATA_H)
+$(DEP): $(DATA_H)

target/sim/sw/device/apps/snax/snax-xdma-maxpool/Makefile

@@ -0,0 +1,20 @@
+# Copyright 2023 ETH Zurich and University of Bologna.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Yunhao Deng <[email protected]>
+
+APP     = snax-xdma-maxpool
+
+INCDIRS += data
+INCDIRS += INCDIRS += ../../../snax/xdma/include/
+
+# Add binary to final build
+RISCV_LDFLAGS += ../../../snax/xdma/build/snax-xdma-lib.o
+
+SRCS    = src/snax-xdma-maxpool.c
+
+include ./data/Makefile
+include ../../common.mk
+
+$(DEP): $(DATA_H)

target/sim/sw/device/apps/snax/snax-xdma-maxpool/data/Makefile

@@ -0,0 +1,23 @@
+# Copyright 2023 KU Leuven.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Xiaoling Yi <[email protected]>
+
+# Usage of absolute paths is required to externally include this Makefile
+MK_DIR   := $(dir $(realpath $(lastword $(MAKEFILE_LIST))))
+DATA_DIR := $(realpath $(MK_DIR))
+
+DATA_CFG ?= $(DATA_DIR)/params.hjson
+
+DATA_H = $(DATA_DIR)/data.h
+
+$(DATA_H): $(DATA_DIR)/datagen.py $(DATA_CFG)
+	$< -c $(DATA_CFG) > $@
+
+.PHONY: clean-data clean
+
+clean-data:
+	rm -f $(DATA_H)
+
+clean: clean-data

target/sim/sw/device/apps/snax/snax-xdma-maxpool/data/datagen.py

@@ -0,0 +1,133 @@
+#!/usr/bin/env python3
+
+# Copyright 2024 KU Leuven.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Fanchen Kong <[email protected]>
+
+import numpy as np
+import argparse
+import pathlib
+import hjson
+import sys
+import os
+import subprocess
+
+# Add data utility path
+sys.path.append(os.path.join(os.path.dirname(__file__), "../../../../../../../../util/sim/"))
+from data_utils import format_scalar_definition, format_vector_definition  # noqa E402
+
+bender_command = subprocess.run(['bender', 'path', 'snitch_cluster'],
+                                capture_output=True, text=True)
+snax_utils_path = bender_command.stdout.strip()
+
+sys.path.append(snax_utils_path + "/util/sim/")
+
+from snax_utils import max_pooling  # noqa E402
+np.random.seed(42)
+
+
+# Add stdint.h header
+def emit_header_file(**kwargs):
+    emit_str = "#include <stdint.h>\n\n"
+    emit_str += "#include <stdbool.h> \n\n"
+    emit_str += emit_data(**kwargs)
+    return emit_str
+
+
+def emit_data(**kwargs):
+    MIN = -128
+    MAX = 127
+
+    data_str = ""
+    data_str += format_scalar_definition("int8_t",
+                                         "H",
+                                         kwargs["H"]) + "\n"
+    data_str += format_scalar_definition("int8_t",
+                                         "W",
+                                         kwargs["W"]) + "\n"
+    data_str += format_scalar_definition("int8_t",
+                                         "Cin",
+                                         kwargs["Cin"]) + "\n"
+    data_str += format_scalar_definition("int8_t",
+                                         "Kh",
+                                         kwargs["Kh"]) + "\n"
+    data_str += format_scalar_definition("int8_t", "Kw", kwargs["Kw"]) + "\n"
+    data_str += format_scalar_definition("int8_t",
+                                         "pad_h", kwargs["pad_h"]) + "\n"
+    data_str += format_scalar_definition("int8_t",
+                                         "pad_w", kwargs["pad_w"]) + "\n"
+    data_str += format_scalar_definition("int8_t",
+                                         "stride_h", kwargs["stride_h"]) + "\n"
+    data_str += format_scalar_definition("int8_t",
+                                         "stride_w", kwargs["stride_w"]) + "\n"
+    padded_h = kwargs["H"] + 2 * kwargs["pad_h"]
+    padded_w = kwargs["W"] + 2 * kwargs["pad_w"]
+    out_h = (kwargs["H"] + 2 * kwargs["pad_h"] -
+             kwargs["Kh"]) // kwargs["stride_h"] + 1
+    out_w = (kwargs["W"] + 2 * kwargs["pad_w"] -
+             kwargs["Kw"]) // kwargs["stride_w"] + 1
+
+    data_str += format_scalar_definition("int8_t", "out_H", out_h) + "\n"
+    data_str += format_scalar_definition("int8_t", "out_W", out_w) + "\n"
+    data_str += format_scalar_definition("int8_t", "padded_H", padded_h) + "\n"
+    data_str += format_scalar_definition("int8_t", "padded_W", padded_w) + "\n"
+
+    # Generating random input data vector
+    data_in = np.random.randint(
+        MIN, MAX, (1, kwargs["H"], kwargs["W"], kwargs["Cin"])
+    )
+    padded_data_in = np.pad(
+        data_in,
+        (
+            (0, 0),
+            (kwargs["pad_h"], kwargs["pad_h"]),
+            (kwargs["pad_w"], kwargs["pad_w"]),
+            (0, 0),
+        ),
+        "constant",
+    )
+    # Generating golden data
+    c_golden = max_pooling(
+        data_in,
+        kwargs["Kw"],
+        kwargs["Kh"],
+        kwargs["stride_w"],
+        kwargs["stride_h"],
+        kwargs["pad_w"],
+        kwargs["pad_h"],
+        "HWC",
+    )
+    data_str += format_vector_definition("int8_t",
+                                         "padded_data_in",
+                                         padded_data_in.reshape(-1)) + "\n"
+    data_str += format_vector_definition("int8_t",
+                                         "golden_data_out",
+                                         c_golden.reshape(-1)) + "\n"
+
+    return data_str
+
+
+def main():
+    # Parsing cmd args
+    parser = argparse.ArgumentParser(description="Generating data for kernels")
+    parser.add_argument(
+        "-c",
+        "--cfg",
+        type=pathlib.Path,
+        required=True,
+        help="Select param config file kernel",
+    )
+    args = parser.parse_args()
+
+    # Load param config file
+    with args.cfg.open() as f:
+        param = hjson.loads(f.read())
+
+    # Emit header file
+    print(emit_header_file(**param))
+
+
+if __name__ == "__main__":
+    main()

target/sim/sw/device/apps/snax/snax-xdma-maxpool/data/params.hjson

@@ -0,0 +1,16 @@
+// Copyright 2023 KU Leuven.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Fanchen Kong <[email protected]>
+{
+    H: 32,
+    W: 32,
+    Cin: 8,
+    Kh: 3,
+    Kw: 3,
+    pad_h: 1,
+    pad_w: 1,
+    stride_h: 1,
+    stride_w: 1
+}

target/sim/sw/device/apps/snax/snax-xdma-maxpool/src/snax-xdma-maxpool.c

@@ -0,0 +1,106 @@
+// Copyright 2024 KU Leuven.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Fanchen Kong <[email protected]>
+
+#include "data.h"
+#include "snax-xdma-lib.h"
+#include "snrt.h"
+
+int main() {
+    // Set err value for checking
+    int err = 0;
+    // Obtain the start address of the TCDM memory
+    uint32_t dma_load_input_start;
+    uint32_t dma_load_input_end;
+    uint32_t *tcdm_baseaddress = (uint32_t *)snrt_l1_next();
+    // Put the input at the starting of tcdm
+    uint8_t *tcdm_in = tcdm_baseaddress;
+    // Put the output at the middle of tcdm
+    uint8_t *tcdm_out = tcdm_in + 0x10000 * sizeof(uint8_t);
+
+    if (snrt_is_dm_core()) {
+        // The xdma core is the last compute core in the cluster
+        uint32_t sstride_src[1] = {8};
+        uint32_t sstride_dst[1] = {8};
+        uint32_t tstride_src[2] = {8, 512};
+        uint32_t tbound_src[2] = {3, 3};
+
+        // First we need to transfer the input data from L3->TCDM
+        // Here we use the 2d iDMA transfer
+        dma_load_input_start = snrt_mcycle();
+        snrt_dma_start_2d(
+            tcdm_in, padded_data_in, padded_W * Cin * sizeof(uint8_t),
+            512 * sizeof(uint8_t), padded_W * Cin * sizeof(uint8_t),
+            padded_H * sizeof(uint8_t));
+        snrt_dma_wait_all();
+        dma_load_input_end = snrt_mcycle();
+
+        // --------------------- Configure the Ext --------------------- //
+
+        // There are three extensions in xdma
+        // VerilogMemset, Maxpool, Transposer
+        // 0            , 1      , 2
+        // We want to only use Maxpool
+        // Hence we need to disable the 0 and 2
+        // and we set the maxpool csr to 9 since we need 3x3 pooling
+        if (xdma_disable_dst_ext(0) != 0) {
+            printf("Error in disabling xdma extension 0 \r\n");
+            err++;
+        } else {
+            printf("The xdma extension 0 is disabled \r\n");
+        }
+
+        uint32_t ext_param_maxpool_size[1] = {9};
+        if (xdma_enable_dst_ext(1, ext_param_maxpool_size) != 0) {
+            printf("Error in enabling xdma extension 1 \r\n");
+            err++;
+        } else {
+            printf("The xdma extension 1 is enabled \r\n");
+        }
+
+        if (xdma_disable_dst_ext(2) != 0) {
+            printf("Error in disabling xdma extension 2 \r\n");
+            err++;
+        } else {
+            printf("The xdma extension 2 is disabled \r\n");
+        }
+
+        // --------------------- Configure the AGU --------------------- //
+        uint8_t *local_src_pointer;
+        uint8_t *local_dst_pointer;
+        int task_id;
+        for (int i = 0; i < out_H; i++) {
+            for (int j = 0; j < out_W / 8; j++) {
+                local_src_pointer = tcdm_in + j * 64 + i * 512;
+                local_dst_pointer = tcdm_out + j * 64 + i * 256;
+                if (xdma_memcpy_nd(local_src_pointer, local_dst_pointer,
+                                   sstride_src, sstride_dst, 2, tstride_src,
+                                   tbound_src, 0, NULL, NULL, 0xFFFFFFFF,
+                                   0xFFFFFFFF, 0xFFFFFFFF) != 0) {
+                    printf("Error in xdma agu configuration \r\n");
+                    err++;
+                } else {
+                    printf("The xdma agu is configured \r\n");
+                }
+                int task_id = xdma_start();
+                xdma_wait(task_id);
+                printf("i = %d, j = %d is done \r\n", i, j);
+            }
+        }
+
+        // --------------------- Checking the Results --------------------- //
+        printf("Checking the results \r\n");
+        for (int i = 0; i < out_H * out_W * Cin; i++) {
+            if ((int8_t)tcdm_out[i] != golden_data_out[i]) {
+                printf("The maxpool is incorrect! \r\n");
+                printf("tcdm_out[%d]=%d, golden_data_out[%d]=%d", i,
+                       (int8_t)tcdm_out[i], i, golden_data_out[i]);
+            }
+        }
+        printf("Checking is done. All values are right \r\n");
+    }
+
+    return 0;
+}

target/sim/sw/device/apps/snax/snax-xdma-memset/Makefile

@@ -0,0 +1,16 @@
+# Copyright 2023 ETH Zurich and University of Bologna.
+# Licensed under the Apache License, Version 2.0, see LICENSE for details.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Yunhao Deng <[email protected]>
+
+APP     = snax-xdma-memset
+
+INCDIRS += ../../../snax/xdma/include/
+
+# Add binary to final build
+RISCV_LDFLAGS += ../../../snax/xdma/build/snax-xdma-lib.o
+
+SRCS    = src/snax-xdma-memset.c
+
+include ../../common.mk