Merge remote-tracking branch 'xilinx/dev' into feature/attention-stre…

…amline

.github/workflows/quicktest-dev-pr.yml

@@ -22,4 +22,5 @@ jobs:
           export FINN_ROOT=$(pwd)
           export FINN_BUILD_DIR=/tmp/finn_gha
           export FINN_INST_NAME=finn_gha
+          export FINN_SKIP_XRT_DOWNLOAD=1
           ./run-docker.sh quicktest

.readthedocs.yaml

@@ -44,3 +44,6 @@ sphinx:
 python:
    install:
         - requirements: docs/requirements.txt
+
+formats:
+  - pdf

AUTHORS.rst

@@ -34,3 +34,6 @@ Contributors
 * Shashwat Khandelwal (@shashwat1198)
 * Ian Colbert (@i-colbert)
 * Rachit Garg (@rstar900)
+* Christoph Berganski (@iksnagreb)
+* Jonas Kuehle (@vopade)
+* Aditya S (@Adityasrinivas24)

docker/Dockerfile.finn

@@ -31,6 +31,8 @@ FROM ubuntu:jammy-20230126
 LABEL maintainer="Jakoba Petri-Koenig <jakoba.petri-koenig@amd.com>, Yaman Umuroglu <yaman.umuroglu@amd.com>"
 
 ARG XRT_DEB_VERSION="xrt_202220.2.14.354_22.04-amd64-xrt"
+ARG SKIP_XRT
+ARG LOCAL_XRT
 
 WORKDIR /workspace
 
@@ -78,15 +80,19 @@ RUN cd verilator && \
     make install
 
 # install XRT
-RUN wget https://www.xilinx.com/bin/public/openDownload?filename=$XRT_DEB_VERSION.deb -O /tmp/$XRT_DEB_VERSION.deb
-RUN apt install -y /tmp/$XRT_DEB_VERSION.deb
-RUN rm /tmp/$XRT_DEB_VERSION.deb
+RUN if [ -z "$LOCAL_XRT" ] && [ -z "$SKIP_XRT" ];then \
+    wget -U 'Mozilla/5.0 (X11; Linux i686) AppleWebKit/537.17 (KHTML, like Gecko) Chrome/24.0.1312.27 Safari/537.17' "https://www.xilinx.com/bin/public/openDownload?filename=$XRT_DEB_VERSION.deb" -O /tmp/$XRT_DEB_VERSION.deb; fi
+
+COPY requirements.txt $XRT_DEB_VERSION.* /tmp/
+
+RUN if [ -z "$SKIP_XRT" ];then \
+    apt install -y /tmp/$XRT_DEB_VERSION.deb && \
+    rm /tmp/$XRT_DEB_VERSION.deb; fi
 
 # versioned Python package requirements for FINN compiler
 # these are given in requirements.txt
-COPY requirements.txt .
-RUN pip install -r requirements.txt
-RUN rm requirements.txt
+RUN pip install -r /tmp/requirements.txt
+RUN rm /tmp/requirements.txt
 
 # install PyTorch
 RUN pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116
@@ -126,6 +132,9 @@ RUN pip install tokenize-rt==4.2.1
 # pyverilator
 RUN pip install tclwrapper==0.0.1
 
+# assure that we have the right setuptools version
+RUN pip install setuptools==68.2.2
+
 # extra environment variables for FINN compiler
 ENV VIVADO_IP_CACHE "/tmp/vivado_ip_cache"
 

docker/finn_entrypoint.sh

@@ -86,7 +86,7 @@ if [ -f "$VITIS_PATH/settings64.sh" ];then
     source $XILINX_XRT/setup.sh
     gecho "Found XRT at $XILINX_XRT"
   else
-    recho "XRT not found on $XILINX_XRT, did the installation fail?"
+    recho "XRT not found on $XILINX_XRT, did you skip the download or did the installation fail?"
     exit -1
   fi
 else

docs/finn/faq.rst

@@ -81,7 +81,7 @@ Which data layout do FINN-generated accelerators use? Big-endian? Little-endian?
     If you need to do this manually, first examine how the `FINN PYNQ Python drivers <https://github.com/Xilinx/finn-examples/blob/main/finn_examples/driver.py#L379>`_ do this – notice how the input data is
     first reshaped to create the “folded input shape” that reflects the word size of the first layer based on how much it
     was parallelized, then data packing is applied to obtain a raw byte array (with some reversals going on) that can be
-    fed directly to the hardware. Another example of this is the `npy_to_rtlsim_input <https://github.com/Xilinx/finn-base/blob/dev/src/finn/util/data_packing.py#L289>`_ function, which converts npy arrays to lists of Python arbitrary-precision integers that we feed into pyverilator for rtl simulation.
+    fed directly to the hardware. Another example of this is the `npy_to_rtlsim_input <https://github.com/Xilinx/finn/blob/dev/src/finn/util/data_packing.py#L284>`_ function, which converts npy arrays to lists of Python arbitrary-precision integers that we feed into pyverilator for rtl simulation.
 
 Why does FIFO sizing take so long for my network? Is something wrong?
     The automatic FIFO sizing in FINN can take quite long. It unfortunately doesn’t really parallelize on multiple cores since

fetch-repos.sh

@@ -28,8 +28,8 @@
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 QONNX_COMMIT="1a4957ebf2aaf139217fd56109386d4518dd6127"
-FINN_EXP_COMMIT="de99347e936d51715f5356a1b6c64e37b91c23c2"
-BREVITAS_COMMIT="84f42259ec869eb151af4cb8a8b23ad925f493db"
+FINN_EXP_COMMIT="0724be21111a21f0d81a072fccc1c446e053f851"
+BREVITAS_COMMIT="d4834bd2a0fad3c1fbc0ff7e1346d5dcb3797ea4"
 PYVERILATOR_COMMIT="ce0a08c20cb8c1d1e84181d6f392390f846adbd1"
 CNPY_COMMIT="4e8810b1a8637695171ed346ce68f6984e585ef4"
 HLSLIB_COMMIT="16e5847a5e3ef76cffe84c8fad2f010d593457d3"

finn-rtllib/fifo/hdl/Q_srl.v

@@ -74,7 +74,8 @@ module Q_srl (clock, reset, i_d, i_v, i_r, o_d, o_v, o_r, count, maxcount);
    parameter depth = 16;   // - greatest #items in queue  (2 <= depth <= 256)
    parameter width = 16;   // - width of data (i_d, o_d)
 
-   parameter addrwidth = $clog2(depth);
+   localparam countwidth = $clog2(depth + 1);
+   localparam addrwidth = $clog2(depth);
 
    input     clock;
    input     reset;
@@ -89,10 +90,10 @@ module Q_srl (clock, reset, i_d, i_v, i_r, o_d, o_v, o_r, count, maxcount);
    input              o_r;	// - output stream ready
    wire               o_b;	// - output stream back-pressure
 
-   output [addrwidth:0] count;  // - output number of elems in queue
-   output [addrwidth:0] maxcount;  // - maximum observed count since reset
+   output [countwidth-1:0] count;  // - output number of elems in queue
+   output [countwidth-1:0] maxcount;  // - maximum observed count since reset
 
-   reg [addrwidth:0] maxcount_reg;  // - maximum count seen until now
+   reg [countwidth-1:0] maxcount_reg;  // - maximum count seen until now
    reg    [addrwidth-1:0] addr, addr_, a_;		// - SRL16 address
 							//     for data output
    reg 			  shift_en_;			// - SRL16 shift enable
@@ -183,58 +184,58 @@ module Q_srl (clock, reset, i_d, i_v, i_r, o_d, o_v, o_r, count, maxcount);
    end // always @ (posedge clock or negedge reset)
 
    always @* begin					// - combi always
-        srlo_       <=  'bx;
-        shift_en_o_ <= 1'bx;
-        shift_en_   <= 1'bx;
-        addr_       <=  'bx;
-        state_      <= 2'bx;
+        srlo_       =  'bx;
+        shift_en_o_ = 1'bx;
+        shift_en_   = 1'bx;
+        addr_       =  'bx;
+        state_      = 2'bx;
       case (state)
 
 	state_empty: begin		    // - (empty, will not produce)
 	      if (i_v) begin		    // - empty & i_v => consume
-		 srlo_       <= i_d;
-		 shift_en_o_ <= 1;
-		 shift_en_   <= 1'bx;
-		 addr_       <= 0;
-		 state_      <= state_one;
+		 srlo_       = i_d;
+		 shift_en_o_ = 1;
+		 shift_en_   = 1'bx;
+		 addr_       = 0;
+		 state_      = state_one;
 	      end
 	      else	begin		    // - empty & !i_v => idle
-		 srlo_       <= 'bx;
-		 shift_en_o_ <= 0;
-		 shift_en_   <= 1'bx;
-		 addr_       <= 0;
-		 state_      <= state_empty;
+		 srlo_       = 'bx;
+		 shift_en_o_ = 0;
+		 shift_en_   = 1'bx;
+		 addr_       = 0;
+		 state_      = state_empty;
 	      end
 	end
 
 	state_one: begin		    // - (contains one)
 	      if (i_v && o_b) begin	    // - one & i_v & o_b => consume
-		 srlo_       <= 'bx;
-		 shift_en_o_ <= 0;
-		 shift_en_   <= 1;
-		 addr_       <= 0;
-		 state_      <= state_more;
+		 srlo_       = 'bx;
+		 shift_en_o_ = 0;
+		 shift_en_   = 1;
+		 addr_       = 0;
+		 state_      = state_more;
 	      end
 	      else if (i_v && !o_b) begin   // - one & i_v & !o_b => cons+prod
-		 srlo_       <= i_d;
-		 shift_en_o_ <= 1;
-		 shift_en_   <= 1;
-		 addr_       <= 0;
-		 state_      <= state_one;
+		 srlo_       = i_d;
+		 shift_en_o_ = 1;
+		 shift_en_   = 1;
+		 addr_       = 0;
+		 state_      = state_one;
 	      end
 	      else if (!i_v && o_b) begin   // - one & !i_v & o_b => idle
-		 srlo_       <= 'bx;
-		 shift_en_o_ <= 0;
-		 shift_en_   <= 1'bx;
-		 addr_       <= 0;
-		 state_      <= state_one;
+		 srlo_       = 'bx;
+		 shift_en_o_ = 0;
+		 shift_en_   = 1'bx;
+		 addr_       = 0;
+		 state_      = state_one;
 	      end
 	      else if (!i_v && !o_b) begin  // - one & !i_v & !o_b => produce
-		 srlo_       <= 'bx;
-		 shift_en_o_ <= 0;
-		 shift_en_   <= 1'bx;
-		 addr_       <= 0;
-		 state_      <= state_empty;
+		 srlo_       = 'bx;
+		 shift_en_o_ = 0;
+		 shift_en_   = 1'bx;
+		 addr_       = 0;
+		 state_      = state_empty;
 	      end
 	end // case: state_one
 
@@ -243,60 +244,60 @@ module Q_srl (clock, reset, i_d, i_v, i_r, o_d, o_v, o_r, count, maxcount);
 					    // - (full, will not consume)
 					    // - (full here if depth==2)
 	      if (o_b) begin		    // - full & o_b => idle
-		 srlo_       <= 'bx;
-		 shift_en_o_ <= 0;
-		 shift_en_   <= 0;
-		 addr_       <= addr;
-		 state_      <= state_more;
+		 srlo_       = 'bx;
+		 shift_en_o_ = 0;
+		 shift_en_   = 0;
+		 addr_       = addr;
+		 state_      = state_more;
 	      end
 	      else begin		    // - full & !o_b => produce
-		 srlo_       <= srl[addr];
-		 shift_en_o_ <= 1;
-		 shift_en_   <= 0;
-//		 addr_       <= addr-1;
-//		 state_      <= state_more;
-		 addr_       <= addr_zero_ ? 0         : addr-1;
-		 state_      <= addr_zero_ ? state_one : state_more;
+		 srlo_       = srl[addr];
+		 shift_en_o_ = 1;
+		 shift_en_   = 0;
+//		 addr_       = addr-1;
+//		 state_      = state_more;
+		 addr_       = addr_zero_ ? 0         : addr-1;
+		 state_      = addr_zero_ ? state_one : state_more;
 	      end
 	   end
 	   else begin			    // - (mid: neither empty nor full)
 	      if (i_v && o_b) begin	    // - mid & i_v & o_b => consume
-		 srlo_       <= 'bx;
-		 shift_en_o_ <= 0;
-		 shift_en_   <= 1;
-		 addr_       <= addr+1;
-		 state_      <= state_more;
+		 srlo_       = 'bx;
+		 shift_en_o_ = 0;
+		 shift_en_   = 1;
+		 addr_       = addr+1;
+		 state_      = state_more;
 	      end
 	      else if (i_v && !o_b) begin   // - mid & i_v & !o_b => cons+prod
-		 srlo_       <= srl[addr];
-		 shift_en_o_ <= 1;
-		 shift_en_   <= 1;
-		 addr_       <= addr;
-		 state_      <= state_more;
+		 srlo_       = srl[addr];
+		 shift_en_o_ = 1;
+		 shift_en_   = 1;
+		 addr_       = addr;
+		 state_      = state_more;
 	      end
 	      else if (!i_v && o_b) begin   // - mid & !i_v & o_b => idle
-		 srlo_       <= 'bx;
-		 shift_en_o_ <= 0;
-		 shift_en_   <= 0;
-		 addr_       <= addr;
-		 state_      <= state_more;
+		 srlo_       = 'bx;
+		 shift_en_o_ = 0;
+		 shift_en_   = 0;
+		 addr_       = addr;
+		 state_      = state_more;
 	      end
 	      else if (!i_v && !o_b) begin  // - mid & !i_v & !o_b => produce
-		 srlo_       <= srl[addr];
-		 shift_en_o_ <= 1;
-		 shift_en_   <= 0;
-		 addr_       <= addr_zero_ ? 0         : addr-1;
-		 state_      <= addr_zero_ ? state_one : state_more;
+		 srlo_       = srl[addr];
+		 shift_en_o_ = 1;
+		 shift_en_   = 0;
+		 addr_       = addr_zero_ ? 0         : addr-1;
+		 state_      = addr_zero_ ? state_one : state_more;
 	      end
 	   end // else: !if(addr_full)
 	end // case: state_more
 
 	default: begin
-		 srlo_       <=  'bx;
-		 shift_en_o_ <= 1'bx;
-		 shift_en_   <= 1'bx;
-		 addr_       <=  'bx;
-		 state_      <= 2'bx;
+		 srlo_       =  'bx;
+		 shift_en_o_ = 1'bx;
+		 shift_en_   = 1'bx;
+		 addr_       =  'bx;
+		 state_      = 2'bx;
 	end // case: default
 
       endcase // case(state)