IQTLabs · lk-iqt · Oct 31, 2023 · Oct 20, 2023 · Oct 21, 2023 · Oct 21, 2023
diff --git a/.github/workflows/docker-extras.yml b/.github/workflows/docker-extras.yml
@@ -34,7 +34,7 @@ jobs:
     - name: Build and push platforms
       uses: docker/build-push-action@v5
       with:
-        context: docker
+        context: .
         file: docker/Dockerfile.torchserve
         platforms: linux/amd64,linux/arm64
         push: true
@@ -68,7 +68,7 @@ jobs:
     - name: Build and push platforms
       uses: docker/build-push-action@v5
       with:
-        context: docker
+        context: .
         file: docker/Dockerfile.cuda-torchserve
         platforms: linux/amd64
         push: true

diff --git a/.github/workflows/docker-test.yml b/.github/workflows/docker-test.yml
@@ -7,9 +7,7 @@ jobs:
     - uses: actions/checkout@v4
     - name: docker build
       run: |
-        cd docker
-        docker build -f Dockerfile.torchserve . -t iqtlabs/gamutrf-torchserve:latest
-        cd ..
+        docker build -f docker/Dockerfile.torchserve . -t iqtlabs/gamutrf-torchserve:latest
         ./tests/test_torchserve.sh
   test-gamutrf-extra-images:
     runs-on: ubuntu-latest

diff --git a/Dockerfile b/Dockerfile
@@ -31,7 +31,7 @@ RUN apt-get update && apt-get install --no-install-recommends -y -q \
 
 # nosemgrep:github.workflows.config.dockerfile-source-not-pinned
 FROM ubuntu:22.04
-COPY --from=iqtlabs/gnuradio:3.10.7 /usr/share/uhd/images /usr/share/uhd/images
+COPY --from=iqtlabs/gnuradio:3.10.8 /usr/share/uhd/images /usr/share/uhd/images
 COPY --from=installer /usr/local /usr/local
 COPY --from=installer /gamutrf /gamutrf
 COPY --from=installer /root/.local /root/.local
@@ -40,37 +40,46 @@ ENV DEBIAN_FRONTEND noninteractive
 ENV UHD_IMAGES_DIR /usr/share/uhd/images
 ENV PATH="${PATH}:/root/.local/bin"
 RUN mkdir -p /data/gamutrf
-RUN apt-get update && apt-get install --no-install-recommends -y -q \
-    ca-certificates \
-    libblas3 \
-    libboost-iostreams1.74.0 \
-    libboost-program-options1.74.0 \
-    libboost-thread1.74.0 \
-    libcairo2 \
-    libev4 \
-    libfftw3-3 \
-    libgl1 \
-    libglib2.0-0 \
-    liblapack3 \
-    libopencv-core4.5d \
-    libopencv-imgcodecs4.5d \
-    libopencv-imgproc4.5d \
-    librtlsdr0 \
-    libspdlog1 \
-    libuhd4.1.0 \
-    libunwind8 \
-    libvulkan1 \
-    libzmq5 \
-    mesa-vulkan-drivers \
-    python3 \
-    python3-pyqt5 \
-    python3-pyqt5.sip \
-    python3-zmq \
-    sox \
-    sudo \
-    wget \
-    uhd-host \
-    zstd && \
+# install nvidia's vulkan support if x86.
+# hadolint ignore=DL3008
+RUN if [ "$(arch)" = "x86_64" ] ; then \
+        apt-get update && \
+        apt-get install -y --no-install-recommends ca-certificates dirmngr gpg-agent gpg wget && \
+        apt-key adv --fetch-keys "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/$(arch)/3bf863cc.pub" && \
+        echo "deb http://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/$(arch)/ /" | tee /etc/apt/sources.list.d/nvidia.list && \
+        apt-get update && \
+        apt-get install -y --no-install-recommends libnvidia-gl-545 ; \
+    fi && \
+    apt-get update && apt-get install -y --no-install-recommends \
+        ca-certificates \
+        libblas3 \
+        libboost-iostreams1.74.0 \
+        libboost-program-options1.74.0 \
+        libboost-thread1.74.0 \
+        libcairo2 \
+        libev4 \
+        libfftw3-3 \
+        libgl1 \
+        libglib2.0-0 \
+        liblapack3 \
+        libopencv-core4.5d \
+        libopencv-imgcodecs4.5d \
+        libopencv-imgproc4.5d \
+        librtlsdr0 \
+        libspdlog1 \
+        libuhd4.1.0 \
+        libunwind8 \
+        libvulkan1 \
+        libzmq5 \
+        mesa-vulkan-drivers \
+        python3 \
+        python3-pyqt5 \
+        python3-pyqt5.sip \
+        python3-zmq \
+        sox \
+        uhd-host \
+        wget \
+        zstd && \
     apt-get -y -q clean && rm -rf /var/lib/apt/lists/*
 WORKDIR /gamutrf
 RUN echo "$(find /gamutrf/gamutrf -type f -name \*py -print)"|xargs grep -Eh "^(import|from)\s"|grep -Ev "gamutrf"|sort|uniq|python3

diff --git a/README.md b/README.md
@@ -175,12 +175,22 @@ Run ```echo 0 > /sys/module/usbcore/parameters/usbfs_memory_mb``` as root before
 
 ##### ```[ERROR] [USB] USB open failed: insufficient permissions```
 
-Ettus SDRs download firmware and switch USB identities when first powered up. Restart the affected container to work around this.
+Ettus SDRs download firmware and switch USB identities when first powered up. Restart the affected container to work around this (if run with docker compose, restart will happen automatically).
 
 ##### ```[ERROR] [UHD] An unexpected exception was caught in a task loop.The task loop will now exit, things may not work.boost: mutex lock failed in pthread_mutex_lock: Invalid argument```
 
 UHD driver arguments ```num_recv_frames``` or ```recv_frame_size``` may be too high. The defaults are defined as ETTUS_ARGS in [utils.py](gamutrf/utils.py). Try reducing one or both via ```--sdrargs```. For example, ```--sdrargs num_recv_frames=64,recv_frame_size=8200,type=b200```.
 
+#### ```[ERROR] [UHD] EnvironmentError: IOError: usb rx6 transfer status: LIBUSB_TRANSFER_OVERFLOW```
+
+Stop containers, and reset the Ettus as follows:
+
+```
+$ /usr/lib/uhd/utils/b2xx_fx3_utils -D
+$ /usr/lib/uhd/utils/b2xx_fx3_utils -U
+$ /usr/lib/uhd/utils/b2xx_fx3_utils -S
+```
+
 #### Scanner with Ettus SDR shows implausible low power at approx 100MHz intervals
 
 Ettus radios periodically need extra time to produce good data when being retuned rapidly by the scanner. Increasing the value of ```--db_clamp_floor``` will cause the scanner to discard windows after retuning (effectively waiting for the retune command to be executed and produce good data before proceeding).

diff --git a/docker/Dockerfile.base b/docker/Dockerfile.base
@@ -32,7 +32,7 @@ WORKDIR /root/uhd_sample_recorder/build
 RUN CMAKE_BUILD_TYPE=Release cmake ../lib && make -j $(nproc) && cp uhd_sample_recorder /usr/local/bin
 
 FROM ubuntu:22.04 as driver-builder
-COPY --from=iqtlabs/gnuradio:3.10.7 /usr/local /usr/local
+COPY --from=iqtlabs/gnuradio:3.10.8 /usr/local /usr/local
 ENV DEBIAN_FRONTEND noninteractive
 RUN apt-get update && apt-get install -y --no-install-recommends \
   build-essential \
@@ -72,7 +72,7 @@ WORKDIR /root/lime-tools/build
 RUN cmake .. && make install
 
 FROM ubuntu:22.04 as gr-iqtlabs-builder
-COPY --from=iqtlabs/gnuradio:3.10.7 /usr/local /usr/local
+COPY --from=iqtlabs/gnuradio:3.10.8 /usr/local /usr/local
 ENV DEBIAN_FRONTEND noninteractive
 RUN apt-get update && apt-get install -y --no-install-recommends \
   build-essential \
@@ -87,8 +87,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
   libvulkan-dev \
   python3-numpy
 WORKDIR /root
-RUN git clone https://github.com/iqtlabs/gr-iqtlabs -b 1.0.44
-RUN sed -i /SPIRV-Tools/d gr-iqtlabs/lib/CMakeLists.txt
+RUN git clone https://github.com/iqtlabs/gr-iqtlabs -b 1.0.46
 COPY --from=iqtlabs/gamutrf-vkfft:latest /root /root/gr-iqtlabs
 WORKDIR /root/gr-iqtlabs/build
 COPY --from=sigmf-builder /usr/local /usr/local
@@ -112,7 +111,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     python3 \
     python3-numpy \
     &&  apt-get -y -q clean && rm -rf /var/lib/apt/lists/*
-COPY --from=iqtlabs/gnuradio:3.10.7 /usr/local /usr/local
+COPY --from=iqtlabs/gnuradio:3.10.8 /usr/local /usr/local
 COPY --from=driver-builder /usr/local /usr/local
 COPY --from=gr-iqtlabs-builder /usr/local /usr/local
 COPY --from=uhd_sample_recorder-builder /usr/local /usr/local

diff --git a/docker/Dockerfile.cuda-torchserve b/docker/Dockerfile.cuda-torchserve
@@ -1,19 +1,12 @@
 FROM nvidia/cuda:11.8.0-runtime-ubuntu22.04
 ENV DEBIAN_FRONTEND=noninteractive
 WORKDIR /root
-RUN apt-get update && \
-    apt-get install -y \
-      git \
-      python3-pip
-RUN pip config set global.no-cache-dir false && \
-  git clone https://github.com/pytorch/serve -b v0.8.2 && \
-  cd serve && \
-  python3 ./ts_scripts/install_dependencies.py --cuda cu118 --environment prod && \
-  pip3 install . && \
-  pip3 install -r examples/object_detector/yolo/yolov8/requirements.txt && \
-  cd .. && \
-  rm -rf serve
+COPY torchserve/install-torchserve.sh /torchserve/install-torchserve.sh
+RUN /torchserve/install-torchserve.sh --cuda cu118
 RUN /usr/local/bin/torchserve --help
+COPY torchserve/config.properties /torchserve/config.properties
+COPY torchserve/torchserve-entrypoint.sh /torchserve/torchserve-entrypoint.sh
+ENTRYPOINT ["/torchserve/torchserve-entrypoint.sh"]
 
 # see Dockerfile.torchserve for example, but use
 # docker run --gpus all -ti iqtlabs/gamutrf-cuda-torchserve:latest bash

diff --git a/docker/Dockerfile.torchserve b/docker/Dockerfile.torchserve
@@ -1,16 +1,9 @@
 FROM ubuntu:22.04
 ENV DEBIAN_FRONTEND=noninteractive
 WORKDIR /root
-RUN apt-get update && \
-    apt-get install -y \
-      git \
-      python3-pip
-RUN pip config set global.no-cache-dir false && \
-  git clone https://github.com/pytorch/serve -b v0.8.2 && \
-  cd serve && \
-  python3 ./ts_scripts/install_dependencies.py --environment prod && \
-  pip3 install . && \
-  pip3 install -r examples/object_detector/yolo/yolov8/requirements.txt && \
-  cd .. && \
-  rm -rf serve
+COPY torchserve/install-torchserve.sh /torchserve/install-torchserve.sh
+RUN /torchserve/install-torchserve.sh
 RUN /usr/local/bin/torchserve --help
+COPY torchserve/config.properties /torchserve/config.properties
+COPY torchserve/torchserve-entrypoint.sh /torchserve/torchserve-entrypoint.sh
+CMD ["/torchserve/torchserve-entrypoint.sh"]
diff --git a/docker/Dockerfile.vkfft b/docker/Dockerfile.vkfft
@@ -8,11 +8,26 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
   libvulkan-dev \
   python3-dev
 WORKDIR /root
-RUN git clone https://github.com/DTolm/VkFFT -b v1.3.1
-RUN sed -i -E 's/GIT_TAG\s+"origin.main"/GIT_TAG "13.0.0"/g' VkFFT/CMakeLists.txt
+RUN git clone https://github.com/DTolm/VkFFT -b v1.3.2
 WORKDIR /root/VkFFT/build
-RUN CMAKE_BUILD_TYPE=Release cmake .. && make -j "$(nproc)"
+RUN CMAKE_BUILD_TYPE=Release cmake -DGLSLANG_GIT_TAG=13.0.0 .. && make -j "$(nproc)"
 
 FROM ubuntu:22.04
 # TODO: ideally, should be packaged such that cmake can find it.
+# hadolint ignore=DL3008
+RUN if [ "$(arch)" = "x86_64" ] ; then \
+        apt-get update && \
+        apt-get install -y --no-install-recommends ca-certificates dirmngr gpg-agent gpg wget && \
+        apt-key adv --fetch-keys "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/$(arch)/3bf863cc.pub" && \
+        echo "deb http://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/$(arch)/ /" | tee /etc/apt/sources.list.d/nvidia.list && \
+        apt-get update && \
+        apt-get install -y --no-install-recommends libnvidia-gl-545 ; \
+    fi && \
+    apt-get update && apt-get install -y --no-install-recommends \
+        libvulkan1
 COPY --from=vkfft-builder /root/VkFFT /root/VkFFT
+CMD ["/root/VkFFT/build/VkFFT_TestSuite", "-devices"]
+
+# Test that GPU can be accessed by VkFFT:
+# $ docker run --gpus all --device /dev/dri/renderD128:/dev/dri/renderD128 -ti iqtlabs/gamutrf-vkfft
+# Device id: 0 name: NVIDIA GeForce RTX 4070 Ti API:1.3.260
diff --git a/docs/README-airt.md b/docs/README-airt.md
@@ -85,7 +85,7 @@ install gr-iqtlabs
 $ git clone https://github.com/google/flatbuffers -b v23.5.26
 $ git clone https://github.com/nlohmann/json -b v3.11.2
 $ git clone https://github.com/deepsig/libsigmf -b v1.0.2
-$ git clone https://github.com/iqtlabs/gr-iqtlabs -b 1.0.44
+$ git clone https://github.com/iqtlabs/gr-iqtlabs -b 1.0.46
 $ mkdir -p flatbuffers/build && cd flatbuffers/build && cmake -DCMAKE_INSTALL_PREFIX=~/.conda/envs/$CONDA_DEFAULT_ENV .. && make -j $(nproc) && make install && cd ../..
 $ mkdir -p json/build && cd json/build && cmake -DCMAKE_INSTALL_PREFIX=~/.conda/envs/$CONDA_DEFAULT_ENV .. && make -j $(nproc) && make install && cd ../..
 $ mkdir -p libsigmf/build && cd libsigmf/build && cmake -DUSE_SYSTEM_JSON=ON -DUSE_SYSTEM_FLATBUFFERS=ON -DCMAKE_INSTALL_PREFIX=~/.conda/envs/$CONDA_DEFAULT_ENV -DCMAKE_CXX_FLAGS="-I $HOME/.conda/envs/$CONDA_DEFAULT_ENV/include" .. && make -j $(nproc) && make install && cd ../..
@@ -132,24 +132,22 @@ On a non-AIRT machine that the AIRT can reach over the network, that has an nvid
 
 See https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html
 
-# start torchserve
+# create model archive
 
-From gamutRF's source directory:
+From gamutRF's source directory, and having obtained mini2_snr.pt:
 
 ```
-$ mkdir /tmp/torchserve
-$ cp torchserve/config.properities /tmp/torchserve
-$ docker run --gpus all -p 8081:8081 -p 8080:8080 -v /tmp/torchserve:/torchserve -d iqtlabs/gamutrf-cuda-torchserve torchserve --start --model-store /torchserve --ts-config /torchserve/config.properties --ncs --foreground
+$ pip3 install torch-model-archiver
+$ mkdir /tmp/model_store
+$ torch-model-archiver --force --model-name mini2_snr --version 1.0 --serialized-file /PATH/TO/mini2_snr.pt --handler torchserve/custom_handler.py --export-path /tmp/model_store
 ```
 
-# create and register model
+# start torchserve
 
-From gamutRF's source directory, and having obtained mini2_snr.pt:
+From gamutRF's source directory (mini2_snr is the default model name in torchserve-cuda.yml):
 
 ```
-$ pip3 install torch-model-archiver
-$ torch-model-archiver --force --model-name mini2_snr --version 1.0 --serialized-file /PATH/TO/mini2_snr.pt --handler torchserve/custom_handler.py --export-path /tmp/torchserve
-$ curl -X POST "localhost:8081/models?model_name=mini2_snr&url=mini2_snr.mar&initial_workers=4&batch_size=2"
+$ VOL_PREFIX=/tmp/model_store docker compose -f orchestrator.yml -f torchserve-cuda.yml up -d torchserve
 ```
 
 Now, when starting the scanner, on the AIRT:

diff --git a/gamutrf/grscan.py b/gamutrf/grscan.py
@@ -131,6 +131,9 @@ def __init__(
                 logging.info(
                     f"retuning across {freq_range/1e6}MHz in {self.sweep_sec}s, requires retuning at {target_retune_hz}Hz in {tune_step_hz/1e6}MHz steps ({tune_step_fft} FFTs)"
                 )
+        if not tune_step_fft:
+            logging.info("tune_step_fft cannot be 0 - defaulting to nfft")
+            tune_step_fft = nfft
         tune_dwell_ms = tune_step_fft / fft_rate * 1e3
         logging.info(
             f"requested retuning across {freq_range/1e6}MHz every {tune_step_fft} FFTs, dwell time {tune_dwell_ms}ms"
@@ -151,7 +154,6 @@ def __init__(
             pretune,
         )
         self.fft_blocks = fft_blocks + self.get_db_blocks(nfft, samp_rate, scaling)
-        self.fft_to_inference_block = self.fft_blocks[-1]
 
         retune_fft = self.iqtlabs.retune_fft(
             "rx_freq",
@@ -179,7 +181,7 @@ def __init__(
         logging.info("serving FFT on %s", zmq_addr)
         self.fft_blocks.append((zeromq.pub_sink(1, 1, zmq_addr, 100, False, 65536, "")))
 
-        self.inference_blocks = []
+        self.inference_blocks = [blocks.null_sink(gr.sizeof_float * nfft)]
         if inference_output_dir:
             x = 640
             y = 640
@@ -190,26 +192,24 @@ def __init__(
             image_dir = Path(inference_output_dir, "images")
             Path(inference_output_dir).mkdir(parents=True, exist_ok=True)
             image_dir.mkdir(parents=True, exist_ok=True)
-            self.image_inference_block = self.iqtlabs.image_inference(
-                tag="rx_freq",
-                vlen=nfft,
-                x=x,
-                y=y,
-                image_dir=str(image_dir),
-                convert_alpha=255,
-                norm_alpha=0,
-                norm_beta=1,
-                norm_type=32,  # cv::NORM_MINMAX = 32
-                colormap=16,  # cv::COLORMAP_VIRIDIS = 16, cv::COLORMAP_TURBO = 20,
-                interpolation=1,  # cv::INTER_LINEAR = 1,
-                flip=0,
-                min_peak_points=inference_min_db,
-                model_server=inference_model_server,
-                model_name=inference_model_name,
-            )
             self.inference_blocks = [
-                blocks.stream_to_vector(gr.sizeof_float * nfft, 1),
-                self.image_inference_block,
+                self.iqtlabs.image_inference(
+                    tag="rx_freq",
+                    vlen=nfft,
+                    x=x,
+                    y=y,
+                    image_dir=str(image_dir),
+                    convert_alpha=255,
+                    norm_alpha=0,
+                    norm_beta=1,
+                    norm_type=32,  # cv::NORM_MINMAX = 32
+                    colormap=16,  # cv::COLORMAP_VIRIDIS = 16, cv::COLORMAP_TURBO = 20,
+                    interpolation=1,  # cv::INTER_LINEAR = 1,
+                    flip=0,
+                    min_peak_points=inference_min_db,
+                    model_server=inference_model_server,
+                    model_name=inference_model_name,
+                ),
                 yolo_bbox(
                     str(Path(inference_output_dir, "predictions")),
                     inference_min_confidence,
@@ -223,17 +223,18 @@ def __init__(
         else:
             self.msg_connect((retune_fft, "tune"), (self.sources[0], cmd_port))
         self.connect_blocks(self.sources[0], self.sources[1:])
-        self.connect_blocks(self.fft_to_inference_block, self.inference_blocks)
+        self.connect((retune_fft, 1), (self.inference_blocks[0], 0))
+        self.connect_blocks(self.inference_blocks[0], self.inference_blocks[1:])
         for pipeline_blocks in (
             self.fft_blocks,
             self.samples_blocks,
         ):
             self.connect_blocks(self.sources[-1], pipeline_blocks)
 
-    def connect_blocks(self, source, other_blocks):
+    def connect_blocks(self, source, other_blocks, last_block_port=0):
         last_block = source
         for block in other_blocks:
-            self.connect((last_block, 0), (block, 0))
+            self.connect((last_block, last_block_port), (block, 0))
             last_block = block
 
     def get_db_blocks(self, nfft, samp_rate, scaling):