[VTA] Refactor to increase platform coverage (Ultra96 etc.) (#3496)

* hardware refactor for increased FPGA coverage, small optimizations

* fix header

* cleaning up parameters that won't be needed for now

* streamlining makefile, and simplifying tcl scripts

* moving parameter derivation into pkg_config.py, keeping tcl scripts lightweight

* refactoring tcl script to avoid global variables

* deriving AXI signals in pkg_config.py

* unifying address map definition for hardware and software drivers

* single channel design for ultra96 to simplify build

* enable alu by default, no mul opcode for now

* hardware fix

* new bitstream; vta version

* avoid error when env variable is not set

* ultra96 cleanup

* further cleaning up tcl script for bitstream generation

* preliminary rpc server support on ultra96

* rpc server tracker scripts

* ultra96 ldflag

* ultra96 support

* ultra96 support

* cleanup line

* cmake support for ultra96

* simplify memory instantiation

* cleaning up IP parameter initialization

* fix queue instantiation

* 2019.1 transition

* fix macro def

* removing bus width from config

* cleanup

* fix

* turning off testing for now

* cleanup ultra96 ps insantiation

* minor refactor

* adding comments

* upgrading to tophub v0.6

* model used in TVM target now refers to a specific version of VTA for better autoTVM scheduling

* revert change due to bug

* rename driver files to be for zynq-type devices

* streamlining address mapping

* unifying register map offset values between driver and hardware generator

* rely on cma library for cache flush/invalidation

* coherence management

* not make buffer packing depend on data types that can be wider than 64bits

* refactor config derivation to minimize free parameters

* fix environment/pkg config interaction

* adding cfg dump property to pkgconfig:

* fix rpc reconfig

* fix spacing

* cleanup

* fix spacing

* long line fix

* fix spacing and lint

* fix line length

* cmake fix

* environment fix

* renaming after pynq since the driver stack relies on the pynq library - see pynq.io

* update doc

* adding parameterization to  name

* space

* removing reg width

* vta RPC

* update doc on how to edit vta_config.json

* fix path

* fix path

apps/pynq_rpc/start_rpc_server_to_tracker.sh → apps/vta_rpc/start_rpc_server_to_tracker.py

@@ -17,7 +17,10 @@
 # under the License.
 PROJROOT="$( cd "$( dirname "${BASH_SOURCE[0]}" )/../../" && pwd )"
 
+# Derive target specified by vta_config.json
+VTA_CONFIG=${PROJROOT}/vta/config/vta_config.py
+TARGET=$(python ${VTA_CONFIG} --target)
 
 export PYTHONPATH=${PYTHONPATH}:${PROJROOT}/python:${PROJROOT}/vta/python
 export PYTHONPATH=${PYTHONPATH}:/home/xilinx/pynq
-python3 -m vta.exec.rpc_server --tracker fleet:9190 --key pynq
+python3 -m vta.exec.rpc_server --tracker fleet:9190 --key $TARGET

cmake/modules/VTA.cmake

@@ -38,11 +38,16 @@ elseif(PYTHON)
   string(REGEX MATCHALL "(^| )-D[A-Za-z0-9_=.]*" VTA_DEFINITIONS "${__vta_defs}")
 
   file(GLOB VTA_RUNTIME_SRCS vta/src/*.cc)
-  file(GLOB __vta_target_srcs vta/src/${VTA_TARGET}/*.cc)
+  # Add sim driver sources
+  if(${VTA_TARGET} STREQUAL "sim")
+    file(GLOB __vta_target_srcs vta/src/sim/*.cc)
+  endif()
+  # Add pynq driver sources
+  if(${VTA_TARGET} STREQUAL "pynq" OR ${VTA_TARGET} STREQUAL "ultra96")
+    file(GLOB __vta_target_srcs vta/src/pynq/*.cc)
+  endif()
   list(APPEND VTA_RUNTIME_SRCS ${__vta_target_srcs})
-
-  add_library(vta SHARED ${VTA_RUNTIME_SRCS})
-
+  # Add tsim driver sources
   if(${VTA_TARGET} STREQUAL "tsim")
     target_compile_definitions(vta PUBLIC USE_TSIM)
     include_directories("vta/include")
@@ -50,6 +55,8 @@ elseif(PYTHON)
     list(APPEND RUNTIME_SRCS ${RUNTIME_DPI_SRCS})
   endif()
 
+  add_library(vta SHARED ${VTA_RUNTIME_SRCS})
+
   target_include_directories(vta PUBLIC vta/include)
 
   foreach(__def ${VTA_DEFINITIONS})
@@ -62,7 +69,7 @@ elseif(PYTHON)
   endif(APPLE)
 
   # PYNQ rules for Pynq v2.4
-  if(${VTA_TARGET} STREQUAL "pynq")
+  if(${VTA_TARGET} STREQUAL "pynq" OR ${VTA_TARGET} STREQUAL "ultra96")
     find_library(__cma_lib NAMES cma PATH /usr/lib)
     target_link_libraries(vta ${__cma_lib})
   endif()

docs/vta/dev/config.rst

@@ -36,10 +36,6 @@ below.
 +=======================+============+========================================================+
 | ``TARGET``            | String     | The TVM device target.                                 |
 +-----------------------+------------+--------------------------------------------------------+
-| ``HW_TARGET``         | Int        | FPGA frequency in MHz.                                 |
-+-----------------------+------------+--------------------------------------------------------+
-| ``HW_CLK_TARGET``     | Int        | FPGA clock period in ns target for HLS tool.           |
-+-----------------------+------------+--------------------------------------------------------+
 | ``HW_VER``            | String     | VTA hardware version number.                           |
 +-----------------------+------------+--------------------------------------------------------+
 | ``LOG_INP_WIDTH``     | Int (log2) | Input data type signed integer width.                  |
@@ -48,13 +44,9 @@ below.
 +-----------------------+------------+--------------------------------------------------------+
 | ``LOG_ACC_WIDTH``     | Int (log2) | Accumulator data type signed integer width.            |
 +-----------------------+------------+--------------------------------------------------------+
-| ``LOG_OUT_WIDTH``     | Int (log2) | Output data type signed integer width.                 |
-+-----------------------+------------+--------------------------------------------------------+
-| ``LOG_BATCH``         | Int (log2) | VTA matrix multiply intrinsic output dimension 0.      |
-+-----------------------+------------+--------------------------------------------------------+
-| ``LOG_BLOCK_IN``      | Int (log2) | VTA matrix multiply reduction dimension.               |
+| ``LOG_BATCH``         | Int (log2) | VTA matrix multiply intrinsic input/output dimension 0.|
 +-----------------------+------------+--------------------------------------------------------+
-| ``LOG_BLOCK_OUT``     | Int (log2) | VTA matrix multiply intrinsic output dimension 1.      |
+| ``LOG_BLOCK``         | Int (log2) | VTA matrix multiply inner dimensions.                  |
 +-----------------------+------------+--------------------------------------------------------+
 | ``LOG_UOP_BUFF_SIZE`` | Int (log2) | Micro-op on-chip buffer in Bytes.                      |
 +-----------------------+------------+--------------------------------------------------------+
@@ -75,13 +67,8 @@ below.
 
 We provide additional detail below regarding each parameter:
 
- - ``TARGET``: Can be set to ``"pynq"`` or ``"sim"``.
- - ``HW_TARGET``: In pynq mode, can be set to ``100``, ``142``, ``167``, or ``200`` MHz.
- - ``HW_CLK_TARGET``: The lower the target, the more pipeline stages HLS will insert to achieve timing closure during place and route (this can also slightly decrease performance).
+ - ``TARGET``: Can be set to ``"pynq"``, ``"ultra96"``, ``"sim"`` (fast simulator), or ``"tsim"`` (cycle accurate sim with verilator).
  - ``HW_VER``: Hardware version which increments everytime the VTA hardware design changes. This parameter is used to uniquely idenfity hardware bitstreams.
- - ``LOG_OUT_WIDTH``: We recommend matching ``LOG_OUT_WIDTH`` to ``LOG_INP_WIDTH``.
- - ``LOG_BATCH``: Equivalent to A in multiplication of shape (A, B) x (B, C), or typically, the batch dimension.
- - ``LOG_BATCH``: Equivalent to A in multiplication of shape (A, B) x (B, C), or typically, the batch dimension.
- - ``LOG_BLOCK_IN``: Equivalent to B in multiplication of shape (A, B) x (B, C), or typically, the input channel dimension.
- - ``LOG_BLOCK_OUT``: Equivalent to C in multiplication of shape (A, B) x (B, C), or typically, the output channel dimension.
+ - ``LOG_BATCH``: Equivalent to A in multiplication of shape (A, B) x (B, C), or typically, the batch dimension of inner tensor computation.
+ - ``LOG_BLOCK``: Equivalent to B and C in multiplication of shape (A, B) x (B, C), or typically, the input/output channel dimensions of the innter tensor computation.
 

docs/vta/install.md

@@ -61,7 +61,7 @@ To do so,
 
 ```bash
 cd <tvm root>
-cp vta/config/vta_config.json vta_config.json
+vim vta/config/vta_config.json
 # edit vta_config.json
 make vta
 ```
@@ -118,7 +118,7 @@ cd /home/xilinx/tvm
 mkdir build
 cp cmake/config.cmake build/.
 # Copy pynq specific configuration
-cp vta/config/pynq_sample.json build/vta_config.json
+cp vta/config/pynq_sample.json vta/config/vta_config.json
 cd build
 cmake ..
 make runtime vta -j2
@@ -147,13 +147,12 @@ export VTA_PYNQ_RPC_PORT=9091
 ```
 
 In addition, you'll need to edit the `vta_config.json` file on the host to indicate that we are targeting the Pynq platform, by setting the `TARGET` field to `"pynq"`.
-Alternatively, you can copy the default `vta/config/pynq_sample.json` into the TVM root as `vta_config.json`.
 > Note: in contrast to our simulation setup, there are no libraries to compile on the host side since the host offloads all of the computation to the Pynq board.
 
 ```bash
 # On the Host-side
 cd <tvm root>
-cp vta/config/pynq_sample.json vta_config.json
+cp vta/config/pynq_sample.json vta/config/vta_config.json
 ```
 
 This time again, we will run the 2D convolution testbench.
@@ -187,28 +186,28 @@ This third and last guide allows users to generate custom VTA bitstreams using f
 
 ### Xilinx Toolchain Installation
 
-We recommend using `Vivado 2018.2` since our scripts have been tested to work on this version of the Xilinx toolchains.
+We recommend using `Vivado 2019.1` since our scripts have been tested to work on this version of the Xilinx toolchains.
 Our guide is written for Linux (Ubuntu) installation.
 
-You’ll need to install Xilinx’ FPGA compilation toolchain, [Vivado HL WebPACK 2018.2](https://www.xilinx.com/products/design-tools/vivado.html), which a license-free version of the Vivado HLx toolchain.
+You’ll need to install Xilinx’ FPGA compilation toolchain, [Vivado HL WebPACK 2019.1](https://www.xilinx.com/products/design-tools/vivado.html), which a license-free version of the Vivado HLx toolchain.
 
 #### Obtaining and Launching the Vivado GUI Installer
 
-1. Go to the [download webpage](https://www.xilinx.com/support/download/index.html/content/xilinx/en/downloadNav/vivado-design-tools/2018-2.html), and download the Linux Self Extracting Web Installer for Vivado HLx 2018.2: WebPACK and Editions.
+1. Go to the [download webpage](https://www.xilinx.com/support/download/index.html/content/xilinx/en/downloadNav/vivado-design-tools/2019-1.html), and download the Linux Self Extracting Web Installer for Vivado HLx 2019.1: WebPACK and Editions.
 2. You’ll have to sign in with a Xilinx account. This requires a Xilinx account creation that will take 2 minutes.
-3. Complete the Name and Address Verification by clicking “Next”, and you will get the opportunity to download a binary file, called `Xilinx_Vivado_SDK_Web_2018.2_0614_1954_Lin64.bin`.
+3. Complete the Name and Address Verification by clicking “Next”, and you will get the opportunity to download a binary file, called `Xilinx_Vivado_SDK_Web_2019.1_0524_1430_Lin64.bin`.
 4. Now that the file is downloaded, go to your `Downloads` directory, and change the file permissions so it can be executed:
 ```bash
-chmod u+x Xilinx_Vivado_SDK_Web_2018.2_0614_1954_Lin64.bin
+chmod u+x Xilinx_Vivado_SDK_Web_2019.1_0524_1430_Lin64.bin
 ```
 5. Now you can execute the binary:
 ```bash
-./Xilinx_Vivado_SDK_Web_2018.2_0614_1954_Lin64.bin
+./Xilinx_Vivado_SDK_Web_2019.1_0524_1430_Lin64.bin
 ```
 
 #### Xilinx Vivado GUI Installer Steps
 
-At this point you've launched the Vivado 2018.2 Installer GUI program.
+At this point you've launched the Vivado 2019.1 Installer GUI program.
 
 1. Click “Next” on the *Welcome* screen.
 2. On the *Select Install Type* screen, enter your Xilinx user credentials under the “User Authentication” box and select the “Download and Install Now” option before clicking “Next” .
@@ -230,8 +229,8 @@ At this point you've launched the Vivado 2018.2 Installer GUI program.
 
 The last step is to update your `~/.bashrc` with the following lines. This will include all of the Xilinx binary paths so you can launch compilation scripts from the command line.
 ```bash
-# Xilinx Vivado 2018.2 environment
-export XILINX_VIVADO=${XILINX_PATH}/Vivado/2018.2
+# Xilinx Vivado 2019.1 environment
+export XILINX_VIVADO=${XILINX_PATH}/Vivado/2019.1
 export PATH=${XILINX_VIVADO}/bin:${PATH}
 ```
 

python/tvm/autotvm/tophub.py

@@ -44,7 +44,7 @@ PACKAGE_VERSION = {
     'opencl':  "v0.02",
     'mali':    "v0.05",
 
-    'vta':     "v0.05",
+    'vta':     "v0.06",
 }
 
 logger = logging.getLogger('autotvm')

vta/config/pynq_sample.json

 {
   "TARGET" : "pynq",
-  "HW_FREQ" : 100,
-  "HW_CLK_TARGET" : 8,
-  "HW_VER" : "0.0.0",
+  "HW_VER" : "0.0.1",
   "LOG_INP_WIDTH" : 3,
   "LOG_WGT_WIDTH" : 3,
   "LOG_ACC_WIDTH" : 5,
-  "LOG_OUT_WIDTH" : 3,
   "LOG_BATCH" : 0,
-  "LOG_BLOCK_IN" : 4,
-  "LOG_BLOCK_OUT" : 4,
+  "LOG_BLOCK" : 4,
   "LOG_UOP_BUFF_SIZE" : 15,
-  "LOG_INP_BUFF_SIZE" : 15,
+  "LOG_INP_BUFF_SIZE" :15,
   "LOG_WGT_BUFF_SIZE" : 18,
   "LOG_ACC_BUFF_SIZE" : 17
 }

vta/config/ultra96_sample.json

+{
+  "TARGET" : "ultra96",
+  "HW_VER" : "0.0.1",
+  "LOG_INP_WIDTH" : 3,
+  "LOG_WGT_WIDTH" : 3,
+  "LOG_ACC_WIDTH" : 5,
+  "LOG_BATCH" : 0,
+  "LOG_BLOCK" : 4,
+  "LOG_UOP_BUFF_SIZE" : 15,
+  "LOG_INP_BUFF_SIZE" :15,
+  "LOG_WGT_BUFF_SIZE" : 18,
+  "LOG_ACC_BUFF_SIZE" : 17
+}

vta/config/vta_config.json

 {
   "TARGET" : "sim",
-  "HW_FREQ" : 100,
-  "HW_CLK_TARGET" : 7,
-  "HW_VER" : "0.0.0",
+  "HW_VER" : "0.0.1",
   "LOG_INP_WIDTH" : 3,
   "LOG_WGT_WIDTH" : 3,
   "LOG_ACC_WIDTH" : 5,
-  "LOG_OUT_WIDTH" : 3,
   "LOG_BATCH" : 0,
-  "LOG_BLOCK_IN" : 4,
-  "LOG_BLOCK_OUT" : 4,
+  "LOG_BLOCK" : 4,
   "LOG_UOP_BUFF_SIZE" : 15,
   "LOG_INP_BUFF_SIZE" : 15,
   "LOG_WGT_BUFF_SIZE" : 18,

vta/hardware/xilinx/Makefile

@@ -17,81 +17,30 @@
 
 # Directories
 ROOTDIR = $(CURDIR)
-BUILD_NAME = build
-BUILD_DIR = $(ROOTDIR)/../../$(BUILD_NAME)/hardware/xilinx
-SCRIPT_DIR = $(ROOTDIR)/scripts
-SRC_DIR = $(ROOTDIR)/src
-SIM_DIR = $(ROOTDIR)/sim
-TEST_DIR = $(ROOTDIR)/../../tests/hardware/common
-INCLUDE_DIR = $(ROOTDIR)/../../include
+VTA_DIR = $(CURDIR)/../..
+BUILD_DIR = $(VTA_DIR)/build/hardware/xilinx
+SCRIPT_DIR = $(CURDIR)/scripts
+SRC_DIR = $(CURDIR)/src
 
 # Executables
 VIVADO_HLS = vivado_hls
 VIVADO = vivado
-HSI = hsi
-
-# HLS mode
-MODE = skip_sim
-# Debug flag
-DEBUG = false
-# SLURM
-SLURM = false
-# Prevent generation of DSP
-NO_DSP = false
-# Prevent generation of ALU
-NO_ALU = false
 
 # Process VTA JSON config
-VTA_CONFIG = python $(CURDIR)/../../config/vta_config.py
-CFLAGS := $(shell ${VTA_CONFIG} --cflags)
-VTA_TARGET := $(shell ${VTA_CONFIG} --target)
-
-#---------------------
-# VTA Parameters
-#--------------------
-VTA_INP_WIDTH := $(shell ${VTA_CONFIG} --get-inpwidth)
-VTA_WGT_WIDTH := $(shell ${VTA_CONFIG} --get-wgtwidth)
-VTA_ACC_WIDTH := $(shell ${VTA_CONFIG} --get-accwidth)
-VTA_OUT_WIDTH := $(shell ${VTA_CONFIG} --get-outwidth)
-VTA_BATCH := $(shell ${VTA_CONFIG} --get-batch)
-VTA_IN_BLOCK := $(shell ${VTA_CONFIG} --get-blockin)
-VTA_OUT_BLOCK := $(shell ${VTA_CONFIG} --get-blockout)
-VTA_UOP_BUFF_SIZE := $(shell ${VTA_CONFIG} --get-uopbuffsize)
-VTA_INP_BUFF_SIZE := $(shell ${VTA_CONFIG} --get-inpbuffsize)
-VTA_WGT_BUFF_SIZE := $(shell ${VTA_CONFIG} --get-wgtbuffsize)
-VTA_ACC_BUFF_SIZE := $(shell ${VTA_CONFIG} --get-accbuffsize)
-VTA_OUT_BUFF_SIZE := $(shell ${VTA_CONFIG} --get-outbuffsize)
-
-#---------------------
-# FPGA Parameters
-#--------------------
-VTA_CLOCK_FREQ = $(shell ${VTA_CONFIG} --get-fpgafreq)
-VTA_TARGET_PER = $(shell ${VTA_CONFIG} --get-fpgaper)
-
-#---------------------
-# Compilation parameters
-#--------------------
-
-#  Number of threads during compilation
-VTA_HW_COMP_THREADS = 8
+VTA_CONFIG := $(CURDIR)/../../config/vta_config.py
 
 # Derive config name
-CONF = $(shell ${VTA_CONFIG} --cfg-str)
-IP_BUILD_PATH = $(BUILD_DIR)/hls/$(CONF)
-HW_BUILD_PATH = $(BUILD_DIR)/vivado/$(CONF)
-
-ifeq ($(SLURM), true)
-	IP_BUILD_PATH = /scratch/hls/$(CONF)
-	HW_BUILD_PATH = /scratch/vivado/$(CONF)
-endif
+CONF := $(shell python ${VTA_CONFIG} --cfg-str)
+IP_BUILD_PATH := $(BUILD_DIR)/hls/$(CONF)
+HW_BUILD_PATH := $(BUILD_DIR)/vivado/$(CONF)
 
 # IP file path
-IP_PATH = $(BUILD_DIR)/hls/$(CONF)/solution0/impl/ip/xilinx_com_hls_vta_1_0.zip
+IP_PATH := $(BUILD_DIR)/hls/$(CONF)/vta_compute/soln/impl/ip/xilinx_com_hls_compute_1_0.zip
 
 # Bitstream file path
-BIT_PATH = $(BUILD_DIR)/vivado/$(CONF)/export/$(CONF).bit
+BIT_PATH := $(BUILD_DIR)/vivado/$(CONF)/export/$(CONF).bit
 
-.PHONY: all ip bit bsp clean clean_all
+.PHONY: all ip bit clean clean_all
 
 all: bit
 ip: $(IP_PATH)
@@ -100,37 +49,24 @@ bit: $(BIT_PATH)
 $(IP_PATH): $(SRC_DIR)/*
 	mkdir -p $(IP_BUILD_PATH)
 	cd $(IP_BUILD_PATH) && \
-		$(VIVADO_HLS) -f $(SCRIPT_DIR)/hls.tcl \
-		-tclargs $(SRC_DIR) $(SIM_DIR) $(TEST_DIR) $(INCLUDE_DIR) \
-		$(MODE) $(DEBUG) $(NO_DSP) $(NO_ALU) $(VTA_TARGET_PER) \
-		$(VTA_INP_WIDTH) $(VTA_WGT_WIDTH) $(VTA_ACC_WIDTH) $(VTA_OUT_WIDTH) \
-		$(VTA_BATCH) $(VTA_IN_BLOCK) $(VTA_OUT_BLOCK) \
-		$(VTA_UOP_BUFF_SIZE) $(VTA_INP_BUFF_SIZE) $(VTA_WGT_BUFF_SIZE) \
-		$(VTA_ACC_BUFF_SIZE) $(VTA_OUT_BUFF_SIZE)
-ifeq ($(SLURM), true)
-	mkdir -p $(BUILD_DIR)/hls
-	mv $(IP_BUILD_PATH) $(BUILD_DIR)/hls/.
-endif
+		$(VIVADO_HLS) \
+		-f $(SCRIPT_DIR)/hls.tcl \
+		-tclargs \
+			$(VTA_DIR) \
+			${VTA_CONFIG}
 
 $(BIT_PATH): $(IP_PATH)
 	mkdir -p $(HW_BUILD_PATH)
 	cd $(HW_BUILD_PATH) && \
-		$(VIVADO) -mode tcl -source $(SCRIPT_DIR)/vivado.tcl \
-		-tclargs $(BUILD_DIR)/hls/$(CONF) $(VTA_HW_COMP_THREADS) $(VTA_CLOCK_FREQ) \
-		$(VTA_INP_WIDTH) $(VTA_WGT_WIDTH) $(VTA_OUT_WIDTH) \
-		$(VTA_BATCH) $(VTA_IN_BLOCK) $(VTA_OUT_BLOCK) \
-		$(VTA_INP_BUFF_SIZE) $(VTA_WGT_BUFF_SIZE) $(VTA_OUT_BUFF_SIZE)
-ifeq ($(SLURM), true)
-	mkdir -p $(BUILD_DIR)/vivado
-	mv $(HW_BUILD_PATH) $(BUILD_DIR)/vivado/.
-endif
-
-bsp: $(BIT_PATH)
-	cd $(HW_BUILD_PATH) && $(HSI) -mode tcl -source $(SCRIPT_DIR)/hsi.tcl -nojournal -nolog
-	cd $(HW_BUILD_PATH)/bsp && make
+		$(VIVADO) \
+		-mode tcl \
+		-source $(SCRIPT_DIR)/vivado.tcl \
+		-tclargs \
+			$(BUILD_DIR)/hls/$(CONF) \
+			${VTA_CONFIG}
 
 clean:
-	rm -rf *.out *.log *.sb figures
+	rm -rf *.out *.log
 
 cleanall: clean
 	rm -rf $(BUILD_DIR)

vta/hardware/xilinx/sim/vta_test.cc

@@ -35,17 +35,6 @@ int main(void) {
     printParameters();
 #endif
 
-    // Micro op bound
-    assert(VTA_UOP_GEM_2_1 < VTA_UOP_WIDTH);
-    assert(VTA_UOP_ALU_1_1 < VTA_UOP_WIDTH);
-    // Make sure there is no misaligment
-    assert(VTA_INSN_GEM_9_1 < VTA_INSN_GEM_A_0);
-    assert(VTA_INSN_MEM_7_1 < VTA_INSN_MEM_8_0);
-    // Instruction bounds
-    assert(VTA_INSN_MEM_E_1 < VTA_INS_WIDTH);
-    assert(VTA_INSN_GEM_F_1 < VTA_INS_WIDTH);
-    assert(VTA_INSN_ALU_G_1 < VTA_INS_WIDTH);
-
     int status = 0;
 
     // Run ALU test (vector-scalar operators)
@@ -65,15 +54,15 @@ int main(void) {
     status |= alu_test(VTA_ALU_OPCODE_MAX, false, VTA_BLOCK_OUT, 128, false);
     status |= alu_test(VTA_ALU_OPCODE_ADD, false, VTA_BLOCK_OUT, 128, true);
     status |= alu_test(VTA_ALU_OPCODE_ADD, false, VTA_BLOCK_OUT, 128, false);
+    status |= alu_test(VTA_ALU_OPCODE_SHR, false, VTA_BLOCK_OUT, 128, true);
+    status |= alu_test(VTA_ALU_OPCODE_SHR, false, VTA_BLOCK_OUT, 128, false);
 
     // Run blocked GEMM test
-    status |= blocked_gemm_test(256, 256, VTA_BLOCK_OUT*4, true, 2);
     status |= blocked_gemm_test(256, 256, VTA_BLOCK_OUT*4, false, 2);
-    status |= blocked_gemm_test(256, 256, VTA_BLOCK_OUT*4, true, 1);
     status |= blocked_gemm_test(256, 256, VTA_BLOCK_OUT*4, false, 1);
 
     // Simple GEMM unit test
-    status |= gemm_test(64, 64, 64, true);
+    status |= gemm_test(4 * VTA_BATCH, 4 * VTA_BLOCK_OUT, 4 * VTA_BLOCK_IN, false);
 
     return status;
 }

vta/hardware/xilinx/src/vta.h

@@ -18,7 +18,6 @@
  */
 
 /*!
- *  Copyright (c) 2018 by Contributors
  * \file vta.h
  * \brief Type definitions and prototype for VTA HLS design.
  */
@@ -32,6 +31,16 @@
 
 #include <vta/hw_spec.h>
 
+/*!
+* Define HLS stream depth 
+*/
+#define PRAGMA_SUB(x) _Pragma (#x)
+#define PRAGMA_HLS(x) PRAGMA_SUB(x)
+#define STREAM_IN_DEPTH 8
+
+/* \typedef bus_T memory bus datatype*/
+typedef ap_uint<VTA_BUS_WIDTH> bus_T;
+
 /* \typedef uop_T Micro-op datatype*/
 typedef ap_uint<VTA_UOP_WIDTH> uop_T;
 
@@ -53,18 +62,6 @@ typedef ap_int<VTA_WGT_WIDTH+VTA_INP_WIDTH+1> mul_T;
 /* \typedef sum_T GEMM accumulator datatype*/
 typedef ap_int<VTA_WGT_WIDTH+VTA_INP_WIDTH+VTA_LOG_BLOCK_IN+1> sum_T;
 
-/* \typedef inp_vec_T Input vector datatype*/
-typedef ap_uint<VTA_INP_WIDTH*VTA_BLOCK_IN> inp_vec_T;
-
-/* \typedef wgt_vec_T Weight vector datatype*/
-typedef ap_uint<VTA_WGT_WIDTH*VTA_BLOCK_IN> wgt_vec_T;
-
-/* \typedef acc_vec_T Accumulator vector datatype*/
-typedef ap_uint<VTA_ACC_WIDTH*VTA_BLOCK_OUT> acc_vec_T;
-
-/* \typedef out_vec_T Output vector datatype*/
-typedef ap_uint<VTA_OUT_WIDTH*VTA_BLOCK_OUT> out_vec_T;
-
 /* \typedef uop_idx_T Micro-op SRAM index datatype*/
 typedef ap_uint<VTA_LOG_UOP_BUFF_DEPTH+1> uop_idx_T;
 
@@ -107,18 +104,14 @@ typedef ap_uint<VTA_MEMOP_PAD_BIT_WIDTH> memop_pad_T;
 /* \typedef aluop_opcode_T ALU operation opcode datatype*/
 typedef ap_uint<VTA_ALU_OPCODE_BIT_WIDTH> aluop_opcode_T;
 
-/* \typedef aluop_opcode_T ALU operation immediate datatype*/
+/* \typedef aluop_imm_T ALU operation immediate datatype*/
 typedef ap_int<VTA_ALUOP_IMM_BIT_WIDTH> aluop_imm_T;
 
-/* \typedef aluop_opcode_T ALU operation shift immediate datatype*/
-typedef ap_int<VTA_LOG_ACC_WIDTH> aluop_sh_imm_T;
+/* \typedef aluop_shr_arg_T ALU operation shift right immediate datatype*/
+typedef ap_int<VTA_SHR_ARG_BIT_WIDTH> aluop_shr_arg_T;
 
-/*!
-* Define HLS stream depth 
-*/
-#define PRAGMA_SUB(x) _Pragma (#x)
-#define PRAGMA_HLS(x) PRAGMA_SUB(x)
-#define STREAM_IN_DEPTH 8
+/* \typedef aluop_mul_arg_T ALU operation multiply datatype*/
+typedef ap_int<VTA_MUL_ARG_BIT_WIDTH> aluop_mul_arg_T;
 
 /*!
 * \brief Fetch module.
@@ -153,13 +146,13 @@ void fetch(
 * \param wgt_mem Local weight SRAM buffer. Write only single port BRAM.
 */
 void load(
-  volatile inp_vec_T *inputs,
-  volatile wgt_vec_T *weights,
+  volatile bus_T *inputs,
+  volatile bus_T *weights,
   hls::stream<insn_T> &load_queue,
   hls::stream<bool> &g2l_dep_queue,
   hls::stream<bool> &l2g_dep_queue,
-  inp_vec_T inp_mem[VTA_INP_BUFF_DEPTH][VTA_BATCH],
-  wgt_vec_T wgt_mem[VTA_WGT_BUFF_DEPTH][VTA_BLOCK_OUT]);
+  bus_T inp_mem[VTA_INP_BUFF_DEPTH][INP_MAT_AXI_RATIO],
+  bus_T wgt_mem[VTA_WGT_BUFF_DEPTH][WGT_MAT_AXI_RATIO]);
 
 /*!
 * \brief Compute module.
@@ -187,15 +180,15 @@ void load(
 void compute(
   volatile uint32_t &done,
   volatile uop_T *uops,
-  volatile acc_vec_T *biases,
+  volatile bus_T *biases,
   hls::stream<insn_T> &gemm_queue,
   hls::stream<bool> &l2g_dep_queue,
   hls::stream<bool> &s2g_dep_queue,
   hls::stream<bool> &g2l_dep_queue,
   hls::stream<bool> &g2s_dep_queue,
-  out_vec_T inp_mem[VTA_INP_BUFF_DEPTH][VTA_BATCH],
-  wgt_vec_T wgt_mem[VTA_WGT_BUFF_DEPTH][VTA_BLOCK_OUT],
-  out_vec_T out_mem[VTA_ACC_BUFF_DEPTH][VTA_BATCH]);
+  bus_T inp_mem[VTA_INP_BUFF_DEPTH][INP_MAT_AXI_RATIO],
+  bus_T wgt_mem[VTA_WGT_BUFF_DEPTH][WGT_MAT_AXI_RATIO],
+  bus_T out_mem[VTA_ACC_BUFF_DEPTH][OUT_MAT_AXI_RATIO]);
 
 /*!
 * \brief Store module.
@@ -211,11 +204,11 @@ void compute(
 * \param out_mem Local output SRAM buffer. Read only single port BRAM.
 */
 void store(
-  volatile out_vec_T *outputs,
+  volatile bus_T *outputs,
   hls::stream<insn_T> &store_queue,
   hls::stream<bool> &g2s_dep_queue,
   hls::stream<bool> &s2g_dep_queue,
-  out_vec_T out_mem[VTA_ACC_BUFF_DEPTH][VTA_BATCH]);
+  bus_T out_mem[VTA_ACC_BUFF_DEPTH][OUT_MAT_AXI_RATIO]);
 
 /*!
 * \brief VTA wrapper for simulation purpose only.
@@ -232,9 +225,9 @@ void vta(
   uint32_t insn_count,
   volatile insn_T *insns,
   volatile uop_T *uops,
-  volatile inp_vec_T *inputs,
-  volatile wgt_vec_T *weights,
-  volatile acc_vec_T *biases,
-  volatile out_vec_T *outputs);
+  volatile bus_T *inputs,
+  volatile bus_T *weights,
+  volatile bus_T *biases,
+  volatile bus_T *outputs);
 
 #endif  // VTA_VTA_H_

vta/include/vta/driver.h

@@ -136,19 +136,23 @@ void VTAMemCopyToHost(void* dst, const void* src, size_t size);
 
 /*!
  * \brief Flushes the region of memory out of the CPU cache to DRAM.
- * \param buf Pointer to memory region allocated with VTAMemAlloc to be flushed.
- *            This need to be the physical address.
+ * \param vir_addr Pointer to memory region allocated with VTAMemAlloc to be flushed.
+ *                 This need to be the virtual address.
+ * \param phy_addr Pointer to memory region allocated with VTAMemAlloc to be flushed.
+ *                 This need to be the physical address.
  * \param size Size of the region to flush in Bytes.
  */
-void VTAFlushCache(vta_phy_addr_t buf, int size);
+void VTAFlushCache(void* vir_addr, vta_phy_addr_t phy_addr, int size);
 
 /*!
  * \brief Invalidates the region of memory that is cached.
- * \param buf Pointer to memory region allocated with VTAMemAlloc to be invalidated.
- *            This need to be the physical address.
+ * \param vir_addr Pointer to memory region allocated with VTAMemAlloc to be invalidated.
+ *                 This need to be the virtual address.
+ * \param phy_addr Pointer to memory region allocated with VTAMemAlloc to be invalidated.
+ *                 This need to be the physical address.
  * \param size Size of the region to invalidate in Bytes.
  */
-void VTAInvalidateCache(vta_phy_addr_t buf, int size);
+void VTAInvalidateCache(void* vir_addr, vta_phy_addr_t phy_addr, int size);
 
 #ifdef __cplusplus
 }

vta/python/vta/bitstream.py

@@ -45,10 +45,11 @@ def get_bitstream_path():
     # Derive destination path
     cache_dir = os.getenv("VTA_CACHE_PATH", os.path.join(os.getenv("HOME"), ".vta_cache/"))
     cache_dir = os.path.join(cache_dir, env.TARGET)
+    cache_dir = os.path.join(cache_dir, env.HW_VER.replace('.', '_'))
     # Create the directory if it didn't exist
     if not os.path.exists(cache_dir):
         os.makedirs(cache_dir)
-    bit_path = os.path.join(cache_dir, env.BITSTREAM)
+    bit_path = os.path.join(cache_dir, env.BITSTREAM) + ".bit"
 
     return bit_path
 
@@ -63,7 +64,7 @@ def download_bitstream():
     bit = get_bitstream_path()
     url = os.path.join(BITSTREAM_URL, env.TARGET)
     url = os.path.join(url, env.HW_VER)
-    url = os.path.join(url, env.BITSTREAM)
+    url = os.path.join(url, env.BITSTREAM + ".bit")
 
     try:
         download(url, bit)

vta/python/vta/environment.py

@@ -113,15 +113,9 @@ class Environment(object):
 
     # initialization function
     def __init__(self, cfg):
-        self.__dict__.update(cfg)
-        for key in PkgConfig.cfg_keys:
-            if key not in cfg:
-                raise ValueError("Expect key %s in cfg" % key)
-        # derive output buffer size
-        self.LOG_OUT_BUFF_SIZE = (
-            self.LOG_ACC_BUFF_SIZE +
-            self.LOG_OUT_WIDTH -
-            self.LOG_ACC_WIDTH)
+        # Produce the derived parameters and update dict
+        self.pkg = self.pkg_config(cfg)
+        self.__dict__.update(self.pkg.cfg_dict)
         # data type width
         self.INP_WIDTH = 1 << self.LOG_INP_WIDTH
         self.WGT_WIDTH = 1 << self.LOG_WGT_WIDTH
@@ -154,25 +148,15 @@ class Environment(object):
         self.WGT_ELEM_BYTES = self.WGT_ELEM_BITS // 8
         self.ACC_ELEM_BYTES = self.ACC_ELEM_BITS // 8
         self.OUT_ELEM_BYTES = self.OUT_ELEM_BITS // 8
-        # Configuration bitstream name
-        self.BITSTREAM = "{}x{}x{}_{}bx{}b_{}_{}_{}_{}_{}MHz_{}ns_v{}.bit".format(
-            (1 << cfg["LOG_BATCH"]),
-            (1 << cfg["LOG_BLOCK_IN"]),
-            (1 << cfg["LOG_BLOCK_OUT"]),
-            (1 << cfg["LOG_INP_WIDTH"]),
-            (1 << cfg["LOG_WGT_WIDTH"]),
-            cfg["LOG_UOP_BUFF_SIZE"],
-            cfg["LOG_INP_BUFF_SIZE"],
-            cfg["LOG_WGT_BUFF_SIZE"],
-            cfg["LOG_ACC_BUFF_SIZE"],
-            cfg["HW_FREQ"],
-            cfg["HW_CLK_TARGET"],
-            cfg["HW_VER"].replace('.', '_'))
         # dtypes
         self.acc_dtype = "int%d" % self.ACC_WIDTH
         self.inp_dtype = "int%d" % self.INP_WIDTH
         self.wgt_dtype = "int%d" % self.WGT_WIDTH
         self.out_dtype = "int%d" % self.OUT_WIDTH
+        # bistream name
+        self.BITSTREAM = self.pkg.bitstream
+        # model string
+        self.MODEL = self.TARGET + "_" + self.BITSTREAM
         # lazy cached members
         self.mock_mode = False
         self._mock_env = None
@@ -187,11 +171,15 @@ class Environment(object):
     def __exit__(self, ptype, value, trace):
         Environment.current = self._last_env
 
-    def pkg_config(self):
+    def pkg_config(self, cfg):
         """PkgConfig instance"""
         curr_path = os.path.dirname(os.path.abspath(os.path.expanduser(__file__)))
         proj_root = os.path.abspath(os.path.join(curr_path, "../../"))
-        return PkgConfig(self.__dict__, proj_root)
+        return PkgConfig(cfg, proj_root)
+
+    @property
+    def cfg_dict(self):
+        return self.pkg.cfg_dict
 
     @property
     def dev(self):
@@ -236,13 +224,15 @@ class Environment(object):
 
     @property
     def target(self):
-        return tvm.target.vta(model=self.TARGET)
+        return tvm.target.vta(model=self.MODEL)
 
     @property
     def target_host(self):
         """The target host"""
         if self.TARGET == "pynq":
             return "llvm -target=armv7-none-linux-gnueabihf"
+        if self.TARGET == "ultra96":
+            return "llvm -target=aarch64-linux-gnu"
         if self.TARGET == "sim" or self.TARGET == "tsim":
             return "llvm"
         raise ValueError("Unknown target %s" % self.TARGET)
@@ -316,21 +306,18 @@ def coproc_dep_pop(op):
 
 
 def _init_env():
-    """Iniitalize the default global env"""
+    """Initialize the default global env"""
     curr_path = os.path.dirname(
         os.path.abspath(os.path.expanduser(__file__)))
     proj_root = os.path.abspath(os.path.join(curr_path, "../../../"))
     path_list = [
-        os.path.join(curr_path, "vta_config.json"),
-        os.path.join(proj_root, "build", "vta_config.json"),
-        os.path.join(proj_root, "vta_config.json"),
         os.path.join(proj_root, "vta/config/vta_config.json")
     ]
     path_list = [p for p in path_list if os.path.exists(p)]
     if not path_list:
         raise RuntimeError(
-            "Error: {} not found.make sure you have config.json in your vta root"
-            .format(filename))
-    return Environment(json.load(open(path_list[0])))
+            "Error: vta_config.json not found.")
+    cfg = json.load(open(path_list[0]))
+    return Environment(cfg)
 
 Environment.current = _init_env()

vta/python/vta/program_bitstream.py

@@ -48,9 +48,12 @@ def pynq_bitstream_program(bitstream_path):
     bitstream.download()
 
 def bitstream_program(target, bitstream):
-    if target == 'pynq':
+    if target in ['pynq', 'ultra96']:
         pynq_bitstream_program(bitstream)
-    elif target != 'sim':
+    elif target in ['sim', 'tsim']:
+        # In simulation, bit stream programming is a no-op
+        return
+    else:
         raise RuntimeError("Unknown target {}".format(target))
 
 if __name__ == "__main__":

vta/python/vta/rpc_client.py

@@ -30,7 +30,7 @@ def reconfig_runtime(remote):
     """
     env = get_env()
     freconfig = remote.get_function("tvm.contrib.vta.reconfig_runtime")
-    freconfig(env.pkg_config().cfg_json)
+    freconfig(env.pkg.cfg_json)
 
 
 def program_fpga(remote, bitstream=None):

vta/python/vta/testing/util.py

@@ -33,7 +33,6 @@ def run(run_func):
     env = get_env()
 
     if env.TARGET in ["sim", "tsim"]:
-
         # Talk to local RPC if necessary to debug RPC server.
         # Compile vta on your host with make at the root.
         # Make sure TARGET is set to "sim" in the config.json file.
@@ -53,21 +52,20 @@ def run(run_func):
                 assert simulator.enabled()
             run_func(env, rpc.LocalSession())
 
-    elif env.TARGET == "pynq":
-
+    elif env.TARGET in ["pynq", "ultra96"]:
         # The environment variables below should be set if we are using
         # a tracker to obtain a remote for a test device
-        tracket_host = os.environ.get("TVM_TRACKER_HOST", None)
-        tracket_port = os.environ.get("TVM_TRACKER_PORT", None)
+        tracker_host = os.environ.get("TVM_TRACKER_HOST", None)
+        tracker_port = os.environ.get("TVM_TRACKER_PORT", None)
         # Otherwise, we can set the variables below to directly
         # obtain a remote from a test device
         pynq_host = os.environ.get("VTA_PYNQ_RPC_HOST", None)
         pynq_port = os.environ.get("VTA_PYNQ_RPC_PORT", None)
         # Run device from fleet node if env variables are defined
-        if tracket_host and tracket_port:
+        if tracker_host and tracker_port:
             remote = autotvm.measure.request_remote(env.TARGET,
-                                                    tracket_host,
-                                                    int(tracket_port),
+                                                    tracker_host,
+                                                    int(tracker_port),
                                                     timeout=10000)
             run_func(env, remote)
         else:
@@ -78,3 +76,6 @@ def run(run_func):
             else:
                 raise RuntimeError(
                     "Please set the VTA_PYNQ_RPC_HOST and VTA_PYNQ_RPC_PORT environment variables")
+
+    else:
+        raise RuntimeError("Unknown target %s" % env.TARGET)

vta/src/pynq/pynq_driver.cc

@@ -15,12 +15,9 @@
  * KIND, either express or implied.  See the License for the
  * specific language governing permissions and limitations
  * under the License.
- */
-
-/*!
- *  Copyright (c) 2018 by Contributors
+ *
  * \file pynq_driver.c
- * \brief VTA driver for Pynq board.
+ * \brief VTA driver for Zynq SoC boards with Pynq support (see pynq.io).
  */
 
 #include <vta/driver.h>
@@ -53,19 +50,19 @@ void VTAMemCopyToHost(void* dst, const void* src, size_t size) {
   memcpy(dst, src, size);
 }
 
-void VTAFlushCache(vta_phy_addr_t buf, int size) {
-  // Call the xlnkFlushCache on the CMA buffer
+void VTAFlushCache(void* vir_addr, vta_phy_addr_t phy_addr, int size) {
+  // Call the cma_flush_cache on the CMA buffer
   // so that the FPGA can read the buffer data.
-  xlnkFlushCache(reinterpret_cast<void*>(buf), size);
+  cma_flush_cache(vir_addr, phy_addr, size);
 }
 
-void VTAInvalidateCache(vta_phy_addr_t buf, int size) {
-  // Call the xlnkInvalidateCache on the CMA buffer
+void VTAInvalidateCache(void* vir_addr, vta_phy_addr_t phy_addr, int size) {
+  // Call the cma_invalidate_cache on the CMA buffer
   // so that the host needs to read the buffer data.
-  xlnkInvalidateCache(reinterpret_cast<void*>(buf), size);
+  cma_invalidate_cache(vir_addr, phy_addr, size);
 }
 
-void *VTAMapRegister(uint32_t addr, size_t length) {
+void *VTAMapRegister(uint32_t addr) {
   // Align the base address with the pages
   uint32_t virt_base = addr & ~(getpagesize() - 1);
   // Calculate base address offset w.r.t the base address
@@ -73,16 +70,16 @@ void *VTAMapRegister(uint32_t addr, size_t length) {
   // Open file and mmap
   uint32_t mmap_file = open("/dev/mem", O_RDWR|O_SYNC);
   return mmap(NULL,
-              (length+virt_offset),
+              (VTA_IP_REG_MAP_RANGE + virt_offset),
               PROT_READ|PROT_WRITE,
               MAP_SHARED,
               mmap_file,
               virt_base);
 }
 
-void VTAUnmapRegister(void *vta, size_t length) {
+void VTAUnmapRegister(void *vta) {
   // Unmap memory
-  int status = munmap(vta, length);
+  int status = munmap(vta, VTA_IP_REG_MAP_RANGE);
   assert(status == 0);
 }
 
@@ -98,39 +95,30 @@ class VTADevice {
  public:
   VTADevice() {
     // VTA stage handles
-    vta_fetch_handle_ = VTAMapRegister(VTA_FETCH_ADDR, VTA_RANGE);
-    vta_load_handle_ = VTAMapRegister(VTA_LOAD_ADDR, VTA_RANGE);
-    vta_compute_handle_ = VTAMapRegister(VTA_COMPUTE_ADDR, VTA_RANGE);
-    vta_store_handle_ = VTAMapRegister(VTA_STORE_ADDR, VTA_RANGE);
+    vta_fetch_handle_ = VTAMapRegister(VTA_FETCH_ADDR);
+    vta_load_handle_ = VTAMapRegister(VTA_LOAD_ADDR);
+    vta_compute_handle_ = VTAMapRegister(VTA_COMPUTE_ADDR);
+    vta_store_handle_ = VTAMapRegister(VTA_STORE_ADDR);
   }
 
   ~VTADevice() {
     // Close VTA stage handle
-    VTAUnmapRegister(vta_fetch_handle_, VTA_RANGE);
-    VTAUnmapRegister(vta_load_handle_, VTA_RANGE);
-    VTAUnmapRegister(vta_compute_handle_, VTA_RANGE);
-    VTAUnmapRegister(vta_store_handle_, VTA_RANGE);
+    VTAUnmapRegister(vta_fetch_handle_);
+    VTAUnmapRegister(vta_load_handle_);
+    VTAUnmapRegister(vta_compute_handle_);
+    VTAUnmapRegister(vta_store_handle_);
   }
 
   int Run(vta_phy_addr_t insn_phy_addr,
           uint32_t insn_count,
           uint32_t wait_cycles) {
-    // NOTE: Register address map is derived from the auto-generated
-    // driver files available under hardware/build/vivado/<design>/export/driver
-    // FETCH @ 0x10 : Data signal of insn_count_V
-    VTAWriteMappedReg(vta_fetch_handle_, 0x10, insn_count);
-    // FETCH @ 0x18 : Data signal of insns_V
-    VTAWriteMappedReg(vta_fetch_handle_, 0x18, insn_phy_addr);
-    // LOAD @ 0x10 : Data signal of inputs_V
-    VTAWriteMappedReg(vta_load_handle_, 0x10, 0);
-    // LOAD @ 0x18 : Data signal of weight_V
-    VTAWriteMappedReg(vta_load_handle_, 0x18, 0);
-    // COMPUTE @ 0x20 : Data signal of uops_V
-    VTAWriteMappedReg(vta_compute_handle_, 0x20, 0);
-    // COMPUTE @ 0x28 : Data signal of biases_V
-    VTAWriteMappedReg(vta_compute_handle_, 0x28, 0);
-    // STORE @ 0x10 : Data signal of outputs_V
-    VTAWriteMappedReg(vta_store_handle_, 0x10, 0);
+    VTAWriteMappedReg(vta_fetch_handle_, VTA_FETCH_INSN_COUNT_OFFSET, insn_count);
+    VTAWriteMappedReg(vta_fetch_handle_, VTA_FETCH_INSN_ADDR_OFFSET, insn_phy_addr);
+    VTAWriteMappedReg(vta_load_handle_, VTA_LOAD_INP_ADDR_OFFSET, 0);
+    VTAWriteMappedReg(vta_load_handle_, VTA_LOAD_WGT_ADDR_OFFSET, 0);
+    VTAWriteMappedReg(vta_compute_handle_, VTA_COMPUTE_UOP_ADDR_OFFSET, 0);
+    VTAWriteMappedReg(vta_compute_handle_, VTA_COMPUTE_BIAS_ADDR_OFFSET, 0);
+    VTAWriteMappedReg(vta_store_handle_, VTA_STORE_OUT_ADDR_OFFSET, 0);
 
     // VTA start
     VTAWriteMappedReg(vta_fetch_handle_, 0x0, VTA_START);
@@ -141,7 +129,7 @@ class VTADevice {
     // Loop until the VTA is done
     unsigned t, flag = 0;
     for (t = 0; t < wait_cycles; ++t) {
-      flag = VTAReadMappedReg(vta_compute_handle_, 0x18);
+      flag = VTAReadMappedReg(vta_compute_handle_, VTA_COMPUTE_DONE_RD_OFFSET);
       if (flag == VTA_DONE) break;
       std::this_thread::yield();
     }

vta/src/pynq/pynq_driver.h

@@ -6,21 +6,18 @@
  * to you under the Apache License, Version 2.0 (the
  * "License"); you may not use this file except in compliance
  * with the License.  You may obtain a copy of the License at
- * 
+ *
  *   http://www.apache.org/licenses/LICENSE-2.0
- * 
+ *
  * Unless required by applicable law or agreed to in writing,
  * software distributed under the License is distributed on an
  * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
  * KIND, either express or implied.  See the License for the
  * specific language governing permissions and limitations
  * under the License.
- */
-
-/*!
- *  Copyright (c) 2018 by Contributors
- * \file vta_pynq_driver.h
- * \brief VTA driver for Pynq board.
+ *
+ * \file pynq_driver.h
+ * \brief VTA driver for Zynq SoC boards with Pynq support (see pynq.io).
  */
 
 #ifndef VTA_PYNQ_PYNQ_DRIVER_H_
@@ -41,23 +38,21 @@ extern "C" {
 #include <time.h>
 #include <unistd.h>
 
-#ifdef __arm__
+#if defined(__arm__) || defined(__aarch64__)
 #include <libxlnk_cma.h>
 #else
 void* cma_alloc(size_t size, int cached);
 void cma_free(void* buf);
 uint32_t cma_get_phy_addr(void* buf);
+void cma_flush_cache(void* buf, unsigned int phys_addr, int size);
+void cma_invalidate_cache(void* buf, unsigned int phys_addr, int size);
 #endif
-void xlnkFlushCache(void* buf, int size);
-void xlnkInvalidateCache(void* buf, int size);
 
-void *VTAMapRegister(uint32_t addr, size_t length);
-void VTAUnmapRegister(void *vta, size_t length);
+void *VTAMapRegister(uint32_t addr);
+void VTAUnmapRegister(void *vta);
 void VTAWriteMappedReg(void* base_addr, uint32_t offset, uint32_t val);
 uint32_t VTAReadMappedReg(void* base_addr, uint32_t offset);
 
-/*! \brief VTA configuration register address range */
-#define VTA_RANGE 0x100
 /*! \brief VTA configuration register start value */
 #define VTA_START 0x1
 /*! \brief VTA configuration register auto-restart value */
@@ -65,27 +60,6 @@ uint32_t VTAReadMappedReg(void* base_addr, uint32_t offset);
 /*! \brief VTA configuration register done value */
 #define VTA_DONE 0x1
 
-/*! \brief VTA fetch stage configuration register address
-*   from auto-generated XPAR_FETCH_0_S_AXI_CONTROL_BUS_BASEADDR define
-*   in xparameters.h (under build/vivado/<design name>/export/bsp/ps7_cortexa9_0/include)
-*/
-#define VTA_FETCH_ADDR    0x43C00000
-/*! \brief VTA compute stage configuration register address
-*   from auto-generated XPAR_COMPUTE_0_S_AXI_CONTROL_BUS_BASEADDR define
-*   in xparameters.h (under build/vivado/<design name>/export/bsp/ps7_cortexa9_0/include)
-*/
-#define VTA_COMPUTE_ADDR  0x43C10000
-/*! \brief VTA compute stage configuration register address
-*   from auto-generated XPAR_LOAD_0_S_AXI_CONTROL_BUS_BASEADDR define
-*   in xparameters.h (under build/vivado/<design name>/export/bsp/ps7_cortexa9_0/include)
-*/
-#define VTA_LOAD_ADDR     0x43C20000
-/*! \brief VTA store stage configuration register address
-*   from auto-generated XPAR_STORE_0_S_AXI_CONTROL_BUS_BASEADDR define
-*   in xparameters.h (under build/vivado/<design name>/export/bsp/ps7_cortexa9_0/include)
-*/
-#define VTA_STORE_ADDR    0x43C30000
-
 #ifdef __cplusplus
 }
 #endif

vta/src/runtime.cc

@@ -44,8 +44,10 @@ namespace vta {
 static_assert(VTA_UOP_WIDTH == sizeof(VTAUop) * 8,
               "VTA_UOP_WIDTH do not match VTAUop size");
 
-/*! \brief Enable coherent access between VTA and CPU (used on shared mem systems). */
-static const bool kBufferCoherent = true;
+/*! \brief Enable coherent access of data buffers between VTA and CPU */
+static const bool kBufferCoherent = VTA_COHERENT_ACCESSES;
+/*! \brief Always cache buffers (otherwise, write back to DRAM from CPU) */
+static const bool kAlwaysCache = true;
 
 /*!
  * \brief Data buffer represents data on CMA.
@@ -65,8 +67,10 @@ struct DataBuffer {
    * \param size The size of the data.
    */
   void InvalidateCache(size_t offset, size_t size) {
-    if (!kBufferCoherent) {
-      VTAInvalidateCache(phy_addr_ + offset, size);
+    if (!kBufferCoherent && kAlwaysCache) {
+      VTAInvalidateCache(reinterpret_cast<char *>(data_) + offset,
+                         phy_addr_ + offset,
+                         size);
     }
   }
   /*!
@@ -75,8 +79,10 @@ struct DataBuffer {
    * \param size The size of the data.
    */
   void FlushCache(size_t offset, size_t size) {
-    if (!kBufferCoherent) {
-      VTAFlushCache(phy_addr_ + offset, size);
+    if (!kBufferCoherent && kAlwaysCache) {
+      VTAFlushCache(reinterpret_cast<char *>(data_) + offset,
+                    phy_addr_ + offset,
+                    size);
     }
   }
   /*!
@@ -102,7 +108,7 @@ struct DataBuffer {
    * \param size The size of the buffer.
    */
   static DataBuffer* Alloc(size_t size) {
-    void* data = VTAMemAlloc(size, 1);
+    void* data = VTAMemAlloc(size, kAlwaysCache);
     CHECK(data != nullptr);
     DataBuffer* buffer = new DataBuffer();
     buffer->data_ = data;
@@ -469,7 +475,9 @@ class UopQueue : public BaseQueue<VTAUop> {
     // Flush if we're using a shared memory system
     // and if interface is non-coherent
     if (!coherent_ && always_cache_) {
-      VTAFlushCache(fpga_buff_phy_, offset);
+      VTAFlushCache(fpga_buff_,
+                    fpga_buff_phy_,
+                    offset);
     }
   }
 
@@ -860,7 +868,9 @@ class InsnQueue : public BaseQueue<VTAGenericInsn> {
     // Flush if we're using a shared memory system
     // and if interface is non-coherent
     if (!coherent_ && always_cache_) {
-      VTAFlushCache(fpga_buff_phy_, buff_size);
+      VTAFlushCache(fpga_buff_,
+                    fpga_buff_phy_,
+                    buff_size);
     }
   }
 
@@ -1302,9 +1312,9 @@ class CommandQueue {
   // The kernel we are currently recording
   UopKernel* record_kernel_{nullptr};
   // Micro op queue
-  UopQueue<VTA_MAX_XFER, true, true> uop_queue_;
+  UopQueue<VTA_MAX_XFER, kBufferCoherent, kAlwaysCache> uop_queue_;
   // instruction queue
-  InsnQueue<VTA_MAX_XFER, true, true> insn_queue_;
+  InsnQueue<VTA_MAX_XFER, kBufferCoherent, kAlwaysCache> insn_queue_;
   // Device handle
   VTADeviceHandle device_{nullptr};
 #ifdef USE_TSIM

vta/src/sim/sim_driver.cc

@@ -615,10 +615,10 @@ void VTAMemCopyToHost(void* dst, const void* src, size_t size) {
   memcpy(dst, src, size);
 }
 
-void VTAFlushCache(vta_phy_addr_t buf, int size) {
+void VTAFlushCache(void* vir_addr, vta_phy_addr_t phy_addr, int size) {
 }
 
-void VTAInvalidateCache(vta_phy_addr_t buf, int size) {
+void VTAInvalidateCache(void* vir_addr, vta_phy_addr_t phy_addr, int size) {
 }
 
 VTADeviceHandle VTADeviceAlloc() {

vta/src/tsim/tsim_driver.cc

@@ -228,10 +228,10 @@ void VTAMemCopyToHost(void* dst, const void* src, size_t size) {
   memcpy(dst, src, size);
 }
 
-void VTAFlushCache(vta_phy_addr_t buf, int size) {
+void VTAFlushCache(void* vir_addr, vta_phy_addr_t phy_addr, int size) {
 }
 
-void VTAInvalidateCache(vta_phy_addr_t buf, int size) {
+void VTAInvalidateCache(void* vir_addr, vta_phy_addr_t phy_addr, int size) {
 }
 
 VTADeviceHandle VTADeviceAlloc() {

vta/tests/hardware/common/test_lib.h

@@ -18,7 +18,6 @@
  */
 
 /*!
- *  Copyright (c) 2018 by Contributors
  * \file test_lib.cpp
  * \brief Test library for the VTA design simulation and driver tests.
  */
@@ -40,7 +39,6 @@
 #include "../../../src/pynq/pynq_driver.h"
 #endif  // VTA_TARGET_PYNQ
 
-typedef uint64_t axi_T;
 typedef uint32_t uop_T;
 typedef int8_t wgt_T;
 typedef int8_t inp_T;
@@ -95,15 +93,25 @@ template <typename T, int T_WIDTH>
 void unpackBuffer(T **dst, T *src, int y_size, int x_size, int y_block, int x_block);
 
 /*!
-* \brief Allocates and initializes a 2D array in the heap.
+* \brief Allocates and randomly initializes a 2D array in the heap.
 * \param rows Number of rows.
 * \param cols Number of columns.
 * \return Pointer to the 2D array.
 */
-template <typename T, int T_WIDTH>
+template <typename T>
 T ** allocInit2dArray(int rows, int cols);
 
 /*!
+* \brief Allocates and initializes a 2D array to a set value in the heap.
+* \param rows Number of rows.
+* \param cols Number of columns.
+* \param val Value to set the whole array to.
+* \return Pointer to the 2D array.
+*/
+template <typename T>
+T ** allocSet2dArray(int rows, int cols, int val);
+
+/*!
 * \brief Allocates a 2D array in the heap.
 * \param rows Number of rows.
 * \param cols Number of columns.

vta/tests/python/unittest/test_environment.py

@@ -24,7 +24,7 @@ def test_env():
 
 def test_env_scope():
     env = vta.get_env()
-    cfg = env.pkg_config().cfg_dict
+    cfg = env.cfg_dict
     cfg["TARGET"] = "xyz"
     with vta.Environment(cfg):
         assert vta.get_env().TARGET == "xyz"

vta/tutorials/frontend/deploy_resnet_on_vta.py

@@ -100,9 +100,9 @@ if env.TARGET not in ["sim", "tsim"]:
     # the host, make sure you've set the variables below to the IP of
     # your board.
     device_host = os.environ.get("VTA_PYNQ_RPC_HOST", "192.168.2.99")
-    device_port = int(os.environ.get("VTA_PYNQ_RPC_PORT", "9091"))
+    device_port = os.environ.get("VTA_PYNQ_RPC_PORT", "9091")
     if not tracker_host or not tracker_port:
-        remote = rpc.connect(device_host, device_port)
+        remote = rpc.connect(device_host, int(device_port))
     else:
         remote = autotvm.measure.request_remote(env.TARGET, tracker_host, int(tracker_port), timeout=10000)