// lap_filter_axis_dma.cpp
// 2019/12/11 by marsee
//
#include <stdio.h>
#include <string.h>
#include <ap_int.h>
#include <hls_stream.h>
#include <ap_axi_sdata.h>
#include <stdint.h>
static void dma_read(volatile int32_t *inm, hls::stream<ap_axis<32,1,1,1> >& outs, int32_t x_size, int32_t y_size){
ap_axis<32,1,1,1> pix;
LOOP_DRY: for(int y=0; y<y_size; y++){
#pragma HLS LOOP_TRIPCOUNT min=48 max=600
LOOP_DRX: for(int x=0; x<x_size; x++){
#pragma HLS LOOP_TRIPCOUNT min=64 max=800
#pragma HLS PIPELINE II=1
pix.data = inm[x_size*y+x];
if(x==0 || y==0)
pix.user = 1;
else
pix.user = 0;
if(x==(x_size-1))
pix.last = 1;
else
pix.last = 0;
outs << pix;
}
}
}
static void dma_write(hls::stream<ap_axis<32,1,1,1> >& ins, volatile int32_t *outm, int32_t x_size, int32_t y_size){
ap_axis<32,1,1,1> pix;
LOOP_DWY: for(int y=0; y<y_size; y++){
#pragma HLS LOOP_TRIPCOUNT min=48 max=600
LOOP_DWX: for(int x=0; x<x_size; x++){
#pragma HLS LOOP_TRIPCOUNT min=64 max=800
#pragma HLS PIPELINE II=1
ins >> pix;
outm[x_size*y+x] = pix.data;
}
}
}
// RGBからYへの変換
// RGBのフォーマットは、{8'd0, R(8bits), G(8bits), B(8bits)}, 1pixel = 32bits
// 輝度信号Yのみに変換する。変換式は、Y = 0.299R + 0.587G + 0.114B
// "YUVフォーマット及び YUV<->RGB変換"を参考にした。http://vision.kuee.kyoto-u.ac.jp/~hiroaki/firewire/yuv.html
// 2013/09/27 : float を止めて、すべてint にした
int32_t conv_rgb2y(int32_t rgb){
int32_t r, g, b, y_f;
int32_t y;
b = rgb & 0xff;
g = (rgb>>8) & 0xff;
r = (rgb>>16) & 0xff;
y_f = 77*r + 150*g + 29*b; //y_f = 0.299*r + 0.587*g + 0.114*b;の係数に256倍した
y = y_f >> 8; // 256で割る
return(y);
}
// ラプラシアンフィルタ
// x0y0 x1y0 x2y0 -1 -1 -1
// x0y1 x1y1 x2y1 -1 8 -1
// x0y2 x1y2 x2y2 -1 -1 -1
int32_t laplacian_fil(int32_t x0y0, int32_t x1y0, int32_t x2y0, int32_t x0y1,
int32_t x1y1, int32_t x2y1, int32_t x0y2, int32_t x1y2, int32_t x2y2)
{
int32_t y;
y = -x0y0 -x1y0 -x2y0 -x0y1 +8*x1y1 -x2y1 -x0y2 -x1y2 -x2y2;
if (y<0)
y = 0;
else if (y>255)
y = 255;
return(y);
}
static void
lap_filter_axis(hls::stream<ap_axis<32,1,1,1> >& ins, hls::stream<ap_axis<32,1,1,1> >& outs, int32_t x_size, int32_t y_size){
ap_axis<32,1,1,1> pix;
ap_axis<32,1,1,1> lap;
int32_t line_buf[2][1920]; // supported HD resolution
#pragma HLS array_partition variable=line_buf block factor=2 dim=1
#pragma HLS resource variable=line_buf core=RAM_2P
int32_t pix_mat[3][3];
#pragma HLS array_partition variable=pix_mat complete
int32_t lap_fil_val;
Loop1 : do { // user が 1になった時にフレームがスタートする
#pragma HLS LOOP_TRIPCOUNT min=1 max=1 avg=1
ins >> pix;
} while(pix.user == 0);
Loop2 : for (int y=0; y<y_size; y++){
#pragma HLS LOOP_TRIPCOUNT min=48 max=600
Loop3 : for (int x=0; x<x_size; x++){
#pragma HLS LOOP_TRIPCOUNT min=64 max=800
#pragma HLS PIPELINE II=1
if (!(x==0 && y==0)) // 最初の入力はすでに入力されている
ins >> pix; // AXI4-Stream からの入力
Loop4 : for (int k=0; k<3; k++){
Loop5 : for (int m=0; m<2; m++){
#pragma HLS UNROLL
pix_mat[k][m] = pix_mat[k][m+1];
}
}
pix_mat[0][2] = line_buf[0][x];
pix_mat[1][2] = line_buf[1][x];
int32_t y_val = conv_rgb2y(pix.data);
pix_mat[2][2] = y_val;
line_buf[0][x] = line_buf[1][x]; // 行の入れ替え
line_buf[1][x] = y_val;
lap_fil_val = laplacian_fil( pix_mat[0][0], pix_mat[0][1], pix_mat[0][2],
pix_mat[1][0], pix_mat[1][1], pix_mat[1][2],
pix_mat[2][0], pix_mat[2][1], pix_mat[2][2]);
lap.data = (lap_fil_val<<16)+(lap_fil_val<<8)+lap_fil_val; // RGB同じ値を入れる
if (x<2 || y<2) // 最初の2行とその他の行の最初の2列は無効データなので0とする
lap.data = 0;
if (x==0 && y==0) // 最初のデータでは、TUSERをアサートする
lap.user = 1;
else
lap.user = 0;
if (x == (x_size-1)) // 行の最後で TLAST をアサートする
lap.last = 1;
else
lap.last = 0;
outs << lap; // AXI4-Stream へ出力
}
}
}
extern "C" {
void lap_filter_axis_dma(volatile int32_t *inm, volatile int32_t *outm, int32_t x_size, int32_t y_size){
#pragma HLS INTERFACE m_axi port = inm offset = slave bundle = gmem
#pragma HLS INTERFACE m_axi port = outm offset = slave bundle = gmem
#pragma HLS INTERFACE s_axilite port = x_size bundle = control
#pragma HLS INTERFACE s_axilite port = y_size bundle = control
#pragma HLS INTERFACE s_axilite port = return bundle = control
#pragma HLS dataflow
hls::stream<ap_axis<32,1,1,1> > in_stream;
hls::stream<ap_axis<32,1,1,1> > out_stream;
#pragma HLS STREAM variable = in_stream depth = 32
#pragma HLS STREAM variable = out_stream depth = 32
dma_read(inm, in_stream, x_size, y_size);
lap_filter_axis(in_stream, out_stream, x_size, y_size);
dma_write(out_stream, outm, x_size, y_size);
}
}
// lap_filter_axis_dma_host.cpp
// 2019/12/12 by marsee
// 2019/12/25 : getProfilingInfo()による時間計測を追加
//
// Vitis-Tutorials/docs/mixing-c-rtl-kernels/reference-files/src/host/host_step1.cpp のコードを引用します
// https://github.com/Xilinx/Vitis-Tutorials/blob/master/docs/mixing-c-rtl-kernels/reference-files/src/host/host_step1.cpp
#define CL_HPP_CL_1_2_DEFAULT_BUILD
#define CL_HPP_TARGET_OPENCL_VERSION 120
#define CL_HPP_MINIMUM_OPENCL_VERSION 120
#define CL_HPP_ENABLE_PROGRAM_CONSTRUCTION_FROM_ARRAY_COMPATIBILITY 1
#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <vector>
#include <CL/cl2.hpp>
#include <iostream>
#include <fstream>
#include <CL/cl_ext_xilinx.h>
#include <unistd.h>
#include <limits.h>
#include <sys/stat.h>
#include <ap_int.h>
#include <hls_stream.h>
#include <ap_axi_sdata.h>
#include "bmp_header.h"
int laplacian_fil_soft(int32_t x0y0, int32_t x1y0, int32_t x2y0, int32_t x0y1, int32_t x1y1, int32_t x2y1, int32_t x0y2, int32_t x1y2, int32_t x2y2);
int32_t conv_rgb2y_soft(int32_t rgb);
int32_t lap_filter_axis_soft(hls::stream<ap_axis<32,1,1,1> >& ins, hls::stream<ap_axis<32,1,1,1> >& outs, int32_t width, int32_t height); // software
static const std::string error_message =
"Error: Result mismatch:\n"
"i = %d CPU result = %d Device result = %d\n";
//Some Library functions to be used.
template <typename T>
struct aligned_allocator
{
using value_type = T;
T* allocate(std::size_t num)
{
void* ptr = nullptr;
if (posix_memalign(&ptr,4096,num*sizeof(T)))
throw std::bad_alloc();
return reinterpret_cast<T*>(ptr);
}
void deallocate(T* p, std::size_t num)
{
free(p);
}
};
#define OCL_CHECK(error,call) \
call; \
if (error != CL_SUCCESS) { \
printf("%s:%d Error calling " #call ", error code is: %d\n", \
__FILE__,__LINE__, error); \
exit(EXIT_FAILURE); \
}
namespace xcl {
std::vector<cl::Device> get_devices(const std::string& vendor_name) {
size_t i;
cl_int err;
std::vector<cl::Platform> platforms;
OCL_CHECK(err, err = cl::Platform::get(&platforms));
cl::Platform platform;
for (i = 0 ; i < platforms.size(); i++){
platform = platforms[i];
OCL_CHECK(err, std::string platformName = platform.getInfo<CL_PLATFORM_NAME>(&err));
if (platformName == vendor_name){
std::cout << "Found Platform" << std::endl;
std::cout << "Platform Name: " << platformName.c_str() << std::endl;
break;
}
}
if (i == platforms.size()) {
std::cout << "Error: Failed to find Xilinx platform" << std::endl;
exit(EXIT_FAILURE);
}
//Getting ACCELERATOR Devices and selecting 1st such device
std::vector<cl::Device> devices;
OCL_CHECK(err, err = platform.getDevices(CL_DEVICE_TYPE_ACCELERATOR, &devices));
return devices;
}
std::vector<cl::Device> get_xil_devices() {
return get_devices("Xilinx");
}
char* read_binary_file(const std::string &xclbin_file_name, unsigned &nb)
{
std::cout << "INFO: Reading " << xclbin_file_name << std::endl;
if(access(xclbin_file_name.c_str(), R_OK) != 0) {
printf("ERROR: %s xclbin not available please build\n", xclbin_file_name.c_str());
exit(EXIT_FAILURE);
}
//Loading XCL Bin into char buffer
std::cout << "Loading: '" << xclbin_file_name.c_str() << "'\n";
std::ifstream bin_file(xclbin_file_name.c_str(), std::ifstream::binary);
bin_file.seekg (0, bin_file.end);
nb = bin_file.tellg();
bin_file.seekg (0, bin_file.beg);
char *buf = new char [nb];
bin_file.read(buf, nb);
return buf;
}
};
int main(int argc, char* argv[])
{
long x, y;
BITMAPFILEHEADER bmpfhr; // BMPファイルのファイルヘッダ(for Read)
BITMAPINFOHEADER bmpihr; // BMPファイルのINFOヘッダ(for Read)
FILE *fbmpr, *fbmpw;
int32_t blue, green, red;
const char* xclbinFilename;
hls::stream<ap_axis<32,1,1,1> > ins_soft;
hls::stream<ap_axis<32,1,1,1> > outs_soft;
ap_axis<32,1,1,1> pix;
ap_axis<32,1,1,1> vals_soft;
if (argc==2) {
xclbinFilename = argv[1];
std::cout <<"Using FPGA binary file specfied through the command line: " << xclbinFilename << std::endl;
}
else {
xclbinFilename = "../lap_filter_axim.xclbin";
std::cout << "No FPGA binary file specified through the command line, using:" << xclbinFilename <<std::endl;
}
if ((fbmpr = fopen("test.bmp", "rb")) == NULL){ // test.bmp をオープン
fprintf(stderr, "Can't open test.bmp by binary read mode\n");
exit(1);
}
// bmpヘッダの読み出し
fread(&bmpfhr.bfType, sizeof(uint16_t), 1, fbmpr);
fread(&bmpfhr.bfSize, sizeof(uint32_t), 1, fbmpr);
fread(&bmpfhr.bfReserved1, sizeof(uint16_t), 1, fbmpr);
fread(&bmpfhr.bfReserved2, sizeof(uint16_t), 1, fbmpr);
fread(&bmpfhr.bfOffBits, sizeof(uint32_t), 1, fbmpr);
fread(&bmpihr, sizeof(BITMAPINFOHEADER), 1, fbmpr);
// ピクセルを入れるメモリをアロケートする
std::vector<int32_t,aligned_allocator<int32_t>> rd_bmp(bmpihr.biWidth * bmpihr.biHeight);
std::vector<int32_t,aligned_allocator<int32_t>> hw_lapd(bmpihr.biWidth * bmpihr.biHeight);
size_t size_in_bytes = (bmpihr.biWidth * bmpihr.biHeight) * sizeof(int32_t);
// rd_bmp にBMPのピクセルを代入。その際に、行を逆転する必要がある
for (y=0; y<bmpihr.biHeight; y++){
for (x=0; x<bmpihr.biWidth; x++){
blue = fgetc(fbmpr);
green = fgetc(fbmpr);
red = fgetc(fbmpr);
rd_bmp[((bmpihr.biHeight-1)-y)*bmpihr.biWidth+x] = (blue & 0xff) | ((green & 0xff)<<8) | ((red & 0xff)<<16);
}
}
fclose(fbmpr);
std::vector<cl::Device> devices = xcl::get_xil_devices();
cl::Device device = devices[0];
devices.resize(1);
// Creating Context and Command Queue for selected device
cl::Context context(device);
cl::CommandQueue q(context, device, CL_QUEUE_PROFILING_ENABLE);
// Load xclbin
std::cout << "Loading: '" << xclbinFilename << "'\n";
std::ifstream bin_file(xclbinFilename, std::ifstream::binary);
bin_file.seekg (0, bin_file.end);
unsigned nb = bin_file.tellg();
bin_file.seekg (0, bin_file.beg);
char *buf = new char [nb];
bin_file.read(buf, nb);
// Creating Program from Binary File
cl::Program::Binaries bins;
bins.push_back({buf,nb});
cl::Program program(context, devices, bins);
// This call will get the kernel object from program. A kernel is an
// OpenCL function that is executed on the FPGA.
cl::Kernel krnl_lap_filter(program,"lap_filter_axis_dma");
// These commands will allocate memory on the Device. The cl::Buffer objects can
// be used to reference the memory locations on the device.
cl::Buffer rd_bmp_buf(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY,
size_in_bytes, rd_bmp.data());
cl::Buffer hw_lapd_buf(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_WRITE,
size_in_bytes, hw_lapd.data());
// Data will be transferred from system memory over PCIe to the FPGA on-board
// DDR memory.
q.enqueueMigrateMemObjects({rd_bmp_buf},0/* 0 means from host*/);
//set the kernel Arguments
krnl_lap_filter.setArg(0,rd_bmp_buf);
krnl_lap_filter.setArg(1,hw_lapd_buf);
krnl_lap_filter.setArg(2,bmpihr.biWidth);
krnl_lap_filter.setArg(3,bmpihr.biHeight);
cl::Event event;
uint64_t lapf_start, lapf_end;
//Launch the Kernel
q.enqueueTask(krnl_lap_filter, NULL, &event);
// The result of the previous kernel execution will need to be retrieved in
// order to view the results. This call will transfer the data from FPGA to
// source_results vector
q.enqueueMigrateMemObjects({hw_lapd_buf},CL_MIGRATE_MEM_OBJECT_HOST);
q.finish();
// 時間計測
event.getProfilingInfo<uint64_t>(CL_PROFILING_COMMAND_START, &lapf_start);
event.getProfilingInfo<uint64_t>(CL_PROFILING_COMMAND_END, &lapf_end);
auto lapf_time = lapf_end - lapf_start;
printf("lap_filter_axis_dma: %lu ns\n", lapf_time);
// ソフトウェアとハードウェアのチェック
// ins_soft に入力データを用意する
for(int i=0; i<5; i++){ // dummy data
pix.user = 0;
pix.data = int32_t(i);
ins_soft << pix;
}
for(int j=0; j < bmpihr.biHeight; j++){
for(int i=0; i < bmpihr.biWidth; i++){
pix.data = rd_bmp[(j*bmpihr.biWidth)+i];
if (j==0 && i==0) // 最初のデータの時に TUSER を 1 にする
pix.user = 1;
else
pix.user = 0;
if (i == bmpihr.biWidth-1) // 行の最後でTLASTをアサートする
pix.last = 1;
else
pix.last = 0;
ins_soft << pix;
}
}
lap_filter_axis_soft(ins_soft, outs_soft, bmpihr.biWidth, bmpihr.biHeight); // ソフトウェアのラプラシアン・フィルタ
// ハードウェアとソフトウェアのラプラシアン・フィルタの値のチェック
for (y=0; y<bmpihr.biHeight; y++){
for (x=0; x<bmpihr.biWidth; x++){
outs_soft >> vals_soft;
if (hw_lapd[y*bmpihr.biWidth+x] != vals_soft.data){
printf("ERROR HW and SW results mismatch x = %ld, y = %ld, HW = %d, SW = %d\n", x, y, int(hw_lapd[y*bmpihr.biWidth+x]), int(vals_soft.data));
//return(1);
}
}
}
printf("Success HW and SW results match\n");
// ハードウェアのラプラシアンフィルタの結果を temp_lap.bmp へ出力する
if ((fbmpw=fopen("temp_lap.bmp", "wb")) == NULL){
fprintf(stderr, "Can't open temp_lap.bmp by binary write mode\n");
exit(1);
}
// BMPファイルヘッダの書き込み
fwrite(&bmpfhr.bfType, sizeof(uint16_t), 1, fbmpw);
fwrite(&bmpfhr.bfSize, sizeof(uint32_t), 1, fbmpw);
fwrite(&bmpfhr.bfReserved1, sizeof(uint16_t), 1, fbmpw);
fwrite(&bmpfhr.bfReserved2, sizeof(uint16_t), 1, fbmpw);
fwrite(&bmpfhr.bfOffBits, sizeof(uint32_t), 1, fbmpw);
fwrite(&bmpihr, sizeof(BITMAPINFOHEADER), 1, fbmpw);
// RGB データの書き込み、逆順にする
for (y=0; y<bmpihr.biHeight; y++){
for (x=0; x<bmpihr.biWidth; x++){
blue = hw_lapd[((bmpihr.biHeight-1)-y)*bmpihr.biWidth+x] & 0xff;
green = (hw_lapd[((bmpihr.biHeight-1)-y)*bmpihr.biWidth+x] >> 8) & 0xff;
red = (hw_lapd[((bmpihr.biHeight-1)-y)*bmpihr.biWidth+x]>>16) & 0xff;
fputc(blue, fbmpw);
fputc(green, fbmpw);
fputc(red, fbmpw);
}
}
fclose(fbmpw);
return(0);
}
int lap_filter_axis_soft(hls::stream<ap_axis<32,1,1,1> >& ins, hls::stream<ap_axis<32,1,1,1> >& outs, int32_t width, int32_t height){
ap_axis<32,1,1,1> pix;
ap_axis<32,1,1,1> lap;
int32_t **line_buf;
int32_t pix_mat[3][3];
int32_t lap_fil_val;
int32_t i;
// line_buf の1次元目の配列をアロケートする
if ((line_buf =(int32_t **)malloc(sizeof(int32_t *) * 2)) == NULL){
fprintf(stderr, "Can't allocate line_buf[3][]\n");
exit(1);
}
// メモリをアロケートする
for (i=0; i<2; i++){
if ((line_buf[i]=(int32_t *)malloc(sizeof(int32_t) * width)) == NULL){
fprintf(stderr, "Can't allocate line_buf[%d]\n", i);
exit(1);
}
}
do { // user が 1になった時にフレームがスタートする
ins >> pix;
} while(pix.user == 0);
for (int y=0; y<height; y++){
for (int x=0; x<width; x++){
if (!(x==0 && y==0)) // 最初の入力はすでに入力されている
ins >> pix; // AXI4-Stream からの入力
for (int k=0; k<3; k++){
for (int m=0; m<2; m++){
pix_mat[k][m] = pix_mat[k][m+1];
}
}
pix_mat[0][2] = line_buf[0][x];
pix_mat[1][2] = line_buf[1][x];
int32_t y_val = conv_rgb2y_soft(pix.data);
pix_mat[2][2] = y_val;
line_buf[0][x] = line_buf[1][x]; // 行の入れ替え
line_buf[1][x] = y_val;
lap_fil_val = laplacian_fil_soft( pix_mat[0][0], pix_mat[0][1], pix_mat[0][2],
pix_mat[1][0], pix_mat[1][1], pix_mat[1][2],
pix_mat[2][0], pix_mat[2][1], pix_mat[2][2]);
lap.data = (lap_fil_val<<16)+(lap_fil_val<<8)+lap_fil_val; // RGB同じ値を入れる
if (x<2 || y<2) // 最初の2行とその他の行の最初の2列は無効データなので0とする
lap.data = 0;
if (x==0 && y==0) // 最初のデータでは、TUSERをアサートする
lap.user = 1;
else
lap.user = 0;
if (x == (width-1)) // 行の最後で TLAST をアサートする
lap.last = 1;
else
lap.last = 0;
outs << lap; // AXI4-Stream へ出力
}
}
for (i=0; i<2; i++)
free(line_buf[i]);
free(line_buf);
return 0;
}
// RGBからYへの変換
// RGBのフォーマットは、{8'd0, R(8bits), G(8bits), B(8bits)}, 1pixel = 32bits
// 輝度信号Yのみに変換する。変換式は、Y = 0.299R + 0.587G + 0.114B
// "YUVフォーマット及び YUV<->RGB変換"を参考にした。http://vision.kuee.kyoto-u.ac.jp/~hiroaki/firewire/yuv.html
// 2013/09/27 : float を止めて、すべてint にした
int32_t conv_rgb2y_soft(int32_t rgb){
int32_t r, g, b, y_f;
int32_t y;
b = rgb & 0xff;
g = (rgb>>8) & 0xff;
r = (rgb>>16) & 0xff;
y_f = 77*r + 150*g + 29*b; //y_f = 0.299*r + 0.587*g + 0.114*b;の係数に256倍した
y = y_f >> 8; // 256で割る
return(y);
}
// ラプラシアンフィルタ
// x0y0 x1y0 x2y0 -1 -1 -1
// x0y1 x1y1 x2y1 -1 8 -1
// x0y2 x1y2 x2y2 -1 -1 -1
int32_t laplacian_fil_soft(int32_t x0y0, int32_t x1y0, int32_t x2y0, int32_t x0y1, int32_t x1y1,
int32_t x2y1, int32_t x0y2, int32_t x1y2, int32_t x2y2)
{
int32_t y;
y = -x0y0 -x1y0 -x2y0 -x0y1 +8*x1y1 -x2y1 -x0y2 -x1y2 -x2y2;
if (y<0)
y = 0;
else if (y>255)
y = 255;
return(y);
}
| 日 | 月 | 火 | 水 | 木 | 金 | 土 |
|---|---|---|---|---|---|---|
| - | - | - | - | - | 1 | 2 |
| 3 | 4 | 5 | 6 | 7 | 8 | 9 |
| 10 | 11 | 12 | 13 | 14 | 15 | 16 |
| 17 | 18 | 19 | 20 | 21 | 22 | 23 |
| 24 | 25 | 26 | 27 | 28 | 29 | 30 |