use vld1q_u32/vst1q_u32 for memcpy

4 files changed, 43 insertions, 5 deletions

diff --git a/src/camera/veyeimx287m.cpp b/src/camera/veyeimx287m.cpp
index c7e519a..d4c19fe 100644
--- a/src/camera/veyeimx287m.cpp
+++ b/src/camera/veyeimx287m.cpp
@@ -16,6 +16,7 @@
 
 #include "constants.h"
 #include "imagealgos.h"
+#include "mem_utils.h"
 #include "pixels.h"
 // #include "rotaryencoder.h"
 
@@ -80,7 +81,8 @@ std::vector<std::shared_ptr<ICamera> > VeyeIMX287m::search()
     if (!cam->init())
         return {};
 
-    if (!cam->setExposureTimeUs(30))
+    // if (!cam->setExposureTimeUs(30))
+    if (!cam->setExposureTimeUs(250))
         return {};
 
     if (!cam->setLaserLevel(1))
@@ -143,7 +145,8 @@ bool VeyeIMX287m::init()
 bool VeyeIMX287m::setExposureTimeUs(int valueUs)
 {
     //return true;
-    std::cout << __func__ << ": " << valueUs << std::endl << std::flush;
+    std::cout << __func__ << ": " << V4L2_CID_EXPOSURE << " - " << valueUs << std::endl
+              << std::flush;
 
     /*
      * Shutter Time. Value is from 8721ns to 8721*885ns, must be integral
@@ -162,6 +165,7 @@ bool VeyeIMX287m::setExposureTimeUs(int valueUs)
     // setLaserLevel(rand() % 100);
     // int exp = rand() % 10;
     // return setCamParam(V4L2_CID_EXPOSURE, exp * exp * exp * exp * exp * exp);
+    // return setCamParam(V4L2_CID_EXPOSURE, valueUs);
     return setCamParam(V4L2_CID_EXPOSURE, valueUs);
 }
 
@@ -218,7 +222,7 @@ bool VeyeIMX287m::setCamParam(unsigned int v4l2controlId, int value)
         return false;
     }
 
-    // std::cout << __func__ << ": new value is " << ctl.value << std::endl;
+    std::cout << __func__ << ": new value is " << ctl.value << std::endl;
 
     return true;
 }
@@ -374,7 +378,13 @@ void VeyeIMX287m::calcFrameLoop(std::stop_token stopToken)
         {
             t.start();
             // std::lock_guard buffer_lock{m_bufferMutexes[bufferIdx]};
-            memcpy(&image.data, m_videoBuffers[bufferIdx], img_size);
+            // get: 4100-4500
+            // memcpy(&image.data, m_videoBuffers[bufferIdx], img_size);
+            // get: 5000-5100
+            // memcpy_1by1<img_size>((std::byte *) &image.data,
+            //                       (std::byte *) m_videoBuffers[bufferIdx]);
+            memcpy_neon<img_size / sizeof(ARRAY_TYPE)>((ARRAY_TYPE *) &image.data,
+                                                       (ARRAY_TYPE *) m_videoBuffers[bufferIdx]);
             get_elapsed_ns += t.nsecsElapsed();
         }
 
diff --git a/src/main.cpp b/src/main.cpp
index ca2da6f..46f8d6d 100644
--- a/src/main.cpp
+++ b/src/main.cpp
@@ -656,9 +656,11 @@ int main(int argc, char* argv[])
                     // qDebug() << "set new exposure time:" << value;
 
                     // requested_params.exposureTime = value;
-                    if (!camera->setExposureTimeUs(value))
+                    if (!camera->setExposureTimeUs(value)) {
+                        qDebug() << "cannot set exp";
                         return QHttpServerResponse::StatusCode::
                             RequestRangeNotSatisfiable;
+                    }
                 }
 
                 if (json.contains(gainKey))
diff --git a/src/mem_utils.cpp b/src/mem_utils.cpp
new file mode 100644
index 0000000..f5fae0f
--- /dev/null
+++ b/src/mem_utils.cpp
@@ -0,0 +1 @@
+#include "mem_utils.h"
diff --git a/src/mem_utils.h b/src/mem_utils.h
new file mode 100644
index 0000000..8601f78
--- /dev/null
+++ b/src/mem_utils.h
@@ -0,0 +1,25 @@
+#pragma once
+
+#include <cstddef>
+
+#include <arm_neon.h>
+
+template<std::size_t S>
+void memcpy_1by1(std::byte *dst, const std::byte *src)
+{
+    for (std::size_t i{0}; i < S; ++i) {
+        dst[i] = src[i];
+    }
+}
+
+using ARRAY_TYPE = uint32_t;
+template<std::size_t S>
+void memcpy_neon(ARRAY_TYPE *dst, const ARRAY_TYPE *src)
+{
+    uint32x4_t tmp;
+
+    for (std::size_t i{0}; i < (S / 4); i += 4) {
+        tmp = vld1q_u32(src + i);
+        vst1q_u32(&dst[i], tmp);
+    }
+}