[VTA] Design refactoring and bug fixes (apache#20)

* updating driver and dll paths for latest pynqv2.3 image

* unifying tcl scripting for multipleFPGA backends, making bus width parameterizable

* fix coherent interface on Ultra96

* 2d padded load template

* simplifying tensor load/store

* streamlining GEMM and ALU pipeline

* refactor VTA for simpler instruction decoding

* fixed bug, mixed DSP/LUT GEMM support

* adding latency directive in DSPs for better pipelining and timing closure

* dual channel memory interface for Ultra96

* hardware bug fixes, working compilation on PYNQ and ULTRA96

* checking for correctness

* support for pynq v2.3

* report inference time in ms

* update CMAKE for Pynq v2.3

* testing hardware support for batch norm

* adding support for multiplication in ALU

* defaulting to coherent buffers in runtime

* increasing size of CMA buffer allocation limit

Author: tmoreau89

cmake/modules/VTA.cmake

@@ -38,16 +38,9 @@ elseif(PYTHON)
     set_target_properties(vta PROPERTIES LINK_FLAGS "-undefined dynamic_lookup")
   endif(APPLE)
 
-  # PYNQ rules for Pynq v2.3
-  if(${VTA_TARGET} STREQUAL "pynq")
-    find_library(__cma_lib NAMES cma PATH /usr/lib)
-    target_link_libraries(vta ${__cma_lib})
-  endif()
-  # Ultra96 rules
-  if(${VTA_TARGET} STREQUAL "ultra96")
-    find_library(__sds_lib NAMES sds_lib PATH /usr/lib)
-    target_link_libraries(vta ${__sds_lib})
-  endif()
+  # PYNQ rules for pynq v2.3
+  find_library(__cma_lib NAMES cma PATH /usr/lib)
+  target_link_libraries(vta ${__cma_lib})
 else()
   message(STATUS "Cannot found python in env, VTA build is skipped..")
 endif()

vta/config/pynq_sample.json

@@ -14,6 +14,7 @@
   "LOG_BATCH" : 0,
   "LOG_BLOCK_IN" : 4,
   "LOG_BLOCK_OUT" : 4,
+  "LOG_BUS_WIDTH" : 6,
   "LOG_UOP_BUFF_SIZE" : 15,
   "LOG_INP_BUFF_SIZE" : 15,
   "LOG_WGT_BUFF_SIZE" : 18,

vta/config/ultra96_sample.json

@@ -14,6 +14,7 @@
   "LOG_BATCH" : 0,
   "LOG_BLOCK_IN" : 4,
   "LOG_BLOCK_OUT" : 4,
+  "LOG_BUS_WIDTH" : 7,
   "LOG_UOP_BUFF_SIZE" : 15,
   "LOG_INP_BUFF_SIZE" : 15,
   "LOG_WGT_BUFF_SIZE" : 18,

vta/config/vta_config.py

@@ -60,6 +60,8 @@ def main():
                         help="returns log of tensor block in dimension")
     parser.add_argument("--get-blockout", action="store_true",
                         help="returns log of tensor block out dimension")
+    parser.add_argument("--get-buswidth", action="store_true",
+                        help="returns log of bus width in b")
     parser.add_argument("--get-uopbuffsize", action="store_true",
                         help="returns log of micro-op buffer size in B")
     parser.add_argument("--get-inpbuffsize", action="store_true",
@@ -182,6 +184,9 @@ def main():
     if args.get_blockout:
         print(cfg["LOG_BLOCK_OUT"])
 
+    if args.get_buswidth:
+        print(cfg["LOG_BUS_WIDTH"])
+
     if args.get_uopbuffsize:
         print(cfg["LOG_UOP_BUFF_SIZE"])
 

vta/hardware/xilinx/Makefile

@@ -14,7 +14,7 @@ VIVADO = vivado
 HSI = hsi
 
 # HLS mode
-MODE = skip_sim
+MODE = all
 # Debug flag
 DEBUG = False
 # SLURM
@@ -35,6 +35,7 @@ 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_BUS_WIDTH := $(shell ${VTA_CONFIG} --get-buswidth)
 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)
@@ -90,7 +91,7 @@ $(IP_PATH): $(SRC_DIR)/*
 		$(MODE) $(DEBUG) $(VTA_ALU_EN) $(VTA_MUL_EN) \
 		$(VTA_TARGET_PER) $(VTA_GEMM_II) $(VTA_TALU_II) \
 		$(VTA_INP_WIDTH) $(VTA_WGT_WIDTH) $(VTA_ACC_WIDTH) $(VTA_OUT_WIDTH) \
-		$(VTA_BATCH) $(VTA_IN_BLOCK) $(VTA_OUT_BLOCK) \
+		$(VTA_BATCH) $(VTA_IN_BLOCK) $(VTA_OUT_BLOCK) $(VTA_BUS_WIDTH) \
 		$(VTA_UOP_BUFF_SIZE) $(VTA_INP_BUFF_SIZE) $(VTA_WGT_BUFF_SIZE) \
 		$(VTA_ACC_BUFF_SIZE) $(VTA_OUT_BUFF_SIZE)
 ifeq ($(SLURM), True)
@@ -101,7 +102,7 @@ endif
 $(BIT_PATH): $(IP_PATH)
 	mkdir -p $(HW_BUILD_PATH)
 	cd $(HW_BUILD_PATH) && \
-		$(VIVADO) -mode tcl -source $(SCRIPT_DIR)/ultra96.tcl \
+		$(VIVADO) -mode tcl -source $(SCRIPT_DIR)/vivado.tcl \
 		-tclargs $(VTA_TARGET) $(BUILD_DIR)/hls/$(CONF) $(VTA_HW_COMP_THREADS) \
 		$(VTA_CLOCK_FREQ) $(VTA_GEMM_II) \
 		$(VTA_INP_WIDTH) $(VTA_WGT_WIDTH) $(VTA_OUT_WIDTH) \

vta/hardware/xilinx/scripts/hls.tcl

@@ -24,13 +24,14 @@
 # Arg 17: batch size (log)
 # Arg 18: in block size (log)
 # Arg 19: out block size (log)
-# Arg 20: uop buffer size in B (log)
-# Arg 21: inp buffer size in B (log)
-# Arg 22: wgt buffer size in B (log)
-# Arg 23: acc buffer size in B (log)
-# Arg 24: out buffer size in B (log)
-
-if { [llength $argv] eq 26 } {
+# Arg 20: bus width in b (log)
+# Arg 21: uop buffer size in B (log)
+# Arg 22: inp buffer size in B (log)
+# Arg 23: wgt buffer size in B (log)
+# Arg 24: acc buffer size in B (log)
+# Arg 25: out buffer size in B (log)
+
+if { [llength $argv] eq 27 } {
 	set target [lindex $argv 2]
 	set src_dir [lindex $argv 3]
 	set sim_dir [lindex $argv 4]
@@ -50,43 +51,24 @@ if { [llength $argv] eq 26 } {
 	set batch [lindex $argv 18]
 	set block_in [lindex $argv 19]
 	set block_out [lindex $argv 20]
-	set uop_buff_size [lindex $argv 21]
-	set inp_buff_size [lindex $argv 22]
-	set wgt_buff_size [lindex $argv 23]
-	set acc_buff_size [lindex $argv 24]
-	set out_buff_size [lindex $argv 25]
+	set bus_width [lindex $argv 21]
+	set uop_buff_size [lindex $argv 22]
+	set inp_buff_size [lindex $argv 23]
+	set wgt_buff_size [lindex $argv 24]
+	set acc_buff_size [lindex $argv 25]
+	set out_buff_size [lindex $argv 26]
 } else {
-	set target "pynq"
-	set src_dir "../src"
-	set sim_dir "../sim"
-	set test_dir "../../src/test"
-	set include_dir "../../include"
-	set mode "all"
-	set debug "False"
-	set alu_ena "True"
-	set mul_ena "True"
-	set target_period 8
-	set target_gemm_ii 10
-	set target_alu_ii 16
-	set inp_width 3
-	set wgt_width 3
-	set acc_width 5
-	set out_width 3
-	set batch 1
-	set block_in 4
-	set block_out 4
-	set uop_buff_size 15
-	set inp_buff_size 15
-	set wgt_buff_size 15
-	set acc_buff_size 17
-	set out_buff_size 15
+	puts "Not enough arguments provided!"
 	exit
 }
 
+puts "about to start doing some stuff"
+
+
 # Initializes the HLS design and sets HLS pragmas for memory partitioning.
 # This is necessary because of a Vivado restriction that doesn't allow for
 # buses wider than 1024 bits.
-proc init_design {target per g_ii a_ii inp_width wgt_width out_width acc_width batch block_in block_out alu_ena} {
+proc init_design {target per g_ii a_ii bus_width inp_width wgt_width out_width acc_width batch block_in block_out alu_ena} {
 
 	# Set device number
 	if {$target=="pynq"} {
@@ -95,28 +77,25 @@ proc init_design {target per g_ii a_ii inp_width wgt_width out_width acc_width b
 		set_part {xczu3eg-sbva484-1-e}
 	} elseif {$target=="zcu102"} {
 		set_part {xczu9eg-ffvb1156-2-e}
+	} elseif {$target=="f1"} {
+		set_part {xcvu9p-flgb2104-2-i}
+		# config_interface -m_axi_addr64
 	}
 
 	# Max bus width (supported by Vivado)
 	set max_width 1024
 
-	# Set axi width (TODO derive from top level config)
-	if {$target=="pynq"} {
-		set axi_width 64
-	} elseif {$target=="ultra96"} {
-		set axi_width 128
-	} elseif {$target=="zcu102"} {
-		set axi_width 128
-	}
+	# Set axi width
+	set axi_width [expr {1 << $bus_width}]
 
 	# Set the clock frequency
 	create_clock -period $per -name default
 
 	# Set pipeline directive
-	set_directive_pipeline -II $g_ii "compute/READ_GEMM_UOP"
+	set_directive_pipeline -II $g_ii "gemm/READ_GEMM_UOP"
 
 	if {$alu_ena=="True"} {
-		set_directive_pipeline -II $a_ii "compute/READ_ALU_UOP"
+		set_directive_pipeline -II $a_ii "alu/READ_ALU_UOP"
 	}
 
 	# Set input partition factor to (INP_VECTOR_WIDTH*BATCH/(1024*g_ii)
@@ -174,7 +153,8 @@ set cflags "-I $include_dir -I $src_dir -I $test_dir \
 	-DVTA_LOG_BATCH=$batch -DVTA_LOG_BLOCK_OUT=$block_out -DVTA_LOG_BLOCK_IN=$block_in \
 	-DVTA_LOG_UOP_BUFF_SIZE=$uop_buff_size -DVTA_LOG_INP_BUFF_SIZE=$inp_buff_size \
 	-DVTA_LOG_WGT_BUFF_SIZE=$wgt_buff_size -DVTA_LOG_ACC_BUFF_SIZE=$acc_buff_size \
-	-DVTA_LOG_OUT_BUFF_SIZE=$out_buff_size"
+	-DVTA_LOG_OUT_BUFF_SIZE=$out_buff_size -DVTA_LOG_BUS_WIDTH=$bus_width \
+	-DVTA_GEMM_II=$target_gemm_ii"
 if {$debug=="True"} {
 	append cflags " -DVTA_DEBUG=1"
 }
@@ -185,6 +165,7 @@ if {$mul_ena=="True"} {
 	append cflags " -DMUL_EN"
 }
 
+
 # HLS behavioral sim
 if {$mode=="all" || $mode=="sim"} {
 	open_project vta_sim
@@ -193,7 +174,7 @@ if {$mode=="all" || $mode=="sim"} {
 	add_files -tb $sim_dir/vta_test.cc -cflags $cflags
 	add_files -tb $test_dir/test_lib.cc -cflags $cflags
 	open_solution "solution0"
-	init_design $target $target_period $target_gemm_ii $target_alu_ii $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
+	init_design $target $target_period $target_gemm_ii $target_alu_ii $bus_width $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
 	csim_design -clean
 	close_project
 }
@@ -204,7 +185,7 @@ if {$mode=="all" || $mode=="skip_sim" || $mode=="fetch"} {
 	set_top fetch
 	add_files $src_dir/vta.cc -cflags $cflags
 	open_solution "solution0"
-	init_design $target $target_period $target_gemm_ii $target_alu_ii $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
+	init_design $target $target_period $target_gemm_ii $target_alu_ii $bus_width $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
 	csynth_design
 	if {$mode=="all" || $mode=="skip_sim"} {
 		export_design -format ip_catalog
@@ -218,7 +199,7 @@ if {$mode=="all" || $mode=="skip_sim" || $mode=="load"} {
 	set_top load
 	add_files $src_dir/vta.cc -cflags $cflags
 	open_solution "solution0"
-	init_design $target $target_period $target_gemm_ii $target_alu_ii $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
+	init_design $target $target_period $target_gemm_ii $target_alu_ii $bus_width $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
 	csynth_design
 	if {$mode=="all" || $mode=="skip_sim"} {
 		export_design -format ip_catalog
@@ -232,7 +213,7 @@ if {$mode=="all" || $mode=="skip_sim" || $mode=="compute"} {
 	set_top compute
 	add_files $src_dir/vta.cc -cflags $cflags
 	open_solution "solution0"
-	init_design $target $target_period $target_gemm_ii $target_alu_ii $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
+	init_design $target $target_period $target_gemm_ii $target_alu_ii $bus_width $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
 	csynth_design
 	if {$mode=="all" || $mode=="skip_sim"} {
 		export_design -format ip_catalog
@@ -246,7 +227,7 @@ if {$mode=="all" || $mode=="skip_sim" || $mode=="store"} {
 	set_top store
 	add_files $src_dir/vta.cc -cflags $cflags
 	open_solution "solution0"
-	init_design $target $target_period $target_gemm_ii $target_alu_ii $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
+	init_design $target $target_period $target_gemm_ii $target_alu_ii $bus_width $inp_width $wgt_width $out_width $acc_width $batch $block_in $block_out $alu_ena
 	csynth_design
 	if {$mode=="all" || $mode=="skip_sim"} {
 		export_design -format ip_catalog