diff --git a/finn-rtllib/swg/swg_template_default.sv b/finn-rtllib/swg/swg_template_default.sv
index 2d255a35edf97e28053545b89512a4b1415b6f57..0aa309f890140143be75123840708b6811d47a48 100644
--- a/finn-rtllib/swg/swg_template_default.sv
+++ b/finn-rtllib/swg/swg_template_default.sv
@@ -36,28 +36,19 @@ module $TOP_MODULE_NAME$_controller #(
logic signed [$clog2(LOOP_KW_ITERATIONS +2)+1-1:0] Counter_loop_kw = LOOP_KW_ITERATIONS-1;
logic signed [$clog2(LOOP_SIMD_ITERATIONS+2)+1-1:0] Counter_loop_simd = LOOP_SIMD_ITERATIONS-1;
- logic [INCR_BITWIDTH-1:0] tail_incr_reg = 'x;
assign addr_incr = ADDR_INCREMENT_MAP[State];
- assign tail_incr = tail_incr_reg;
// combinational logic for tail_incr generation
- uwire tail_incr_inner_condition;
- generate
- if (IS_DEPTHWISE)
- assign tail_incr_inner_condition = (Counter_loop_kh >= 0);
- else
- assign tail_incr_inner_condition = 0;
- endgenerate
-
- always @ (tail_incr_inner_condition, Counter_loop_w, Counter_loop_h) begin
+ uwire tail_incr_inner_condition = IS_DEPTHWISE? (Counter_loop_kh >= 0) : 0;
+ always_comb begin : blkTail
if (tail_incr_inner_condition)
- tail_incr_reg = 1;
+ tail_incr = 1;
else if (Counter_loop_w >= 0)
- tail_incr_reg = $TAIL_INCR_W$;
+ tail_incr = $TAIL_INCR_W$;
else if (Counter_loop_h >= 0)
- tail_incr_reg = $TAIL_INCR_H$;
+ tail_incr = $TAIL_INCR_H$;
else
- tail_incr_reg = $TAIL_INCR_LAST$;
+ tail_incr = $TAIL_INCR_LAST$;
end
// combinational next state logic
@@ -132,13 +123,8 @@ module $TOP_MODULE_NAME$_cyclic_buffer_addressable #(
$RAM_STYLE$ logic [WIDTH-1:0] Ram[DEPTH];
logic [WIDTH-1:0] Out = 'x;
always_ff @(posedge clk) begin
- if (!rst_n) begin
- Out <= 'x;
- end
- else begin
- if (read_enable) Out <= Ram[read_addr];
- if (write_enable) Ram[write_addr] <= data_in;
- end
+ if (read_enable) Out <= Ram[read_addr];
+ if (write_enable) Ram[write_addr] <= data_in;
end
assign data_out = Out;
@@ -213,7 +199,7 @@ module $TOP_MODULE_NAME$_impl #(
logic signed [$clog2(LAST_READ_ELEM+1)+1-1:0] Newest_buffered_elem = -1;
logic [$clog2(LAST_READ_ELEM+1)+1-1:0] Current_elem = 0;
logic [$clog2(LAST_READ_ELEM+1)+1-1:0] First_elem_next_window = 0;
- logic [$clog2(ELEM_PER_WINDOW) -1:0] K = 0;
+ logic [$clog2(ELEM_PER_WINDOW) -1:0] Position_in_window = 0;
logic [$clog2(BUF_ELEM_TOTAL)+1 -1:0] Window_buffer_read_addr_reg = 0;
logic [$clog2(BUF_ELEM_TOTAL)-1:0] Window_buffer_write_addr_reg = 0;
@@ -255,7 +241,7 @@ module $TOP_MODULE_NAME$_impl #(
Newest_buffered_elem <= -1;
Current_elem <= 0;
First_elem_next_window <= 0;
- K <= 0;
+ Position_in_window <= 0;
Window_buffer_read_addr_reg <= 0;
Window_buffer_write_addr_reg <= 0;
Fetching_done <= 0;
@@ -295,10 +281,10 @@ module $TOP_MODULE_NAME$_impl #(
Window_buffer_read_addr_reg <= ra + ra_correct;
//keep track where we are within a window
- K <= (K != ELEM_PER_WINDOW - 1)? K+1 : 0;
+ Position_in_window <= (Position_in_window != ELEM_PER_WINDOW - 1)? Position_in_window+1 : 0;
//update first element of next window to allow buffer overwrite up until that point
- if (K == 0)
+ if (Position_in_window == 0)
First_elem_next_window <= First_elem_next_window + tail_incr;
//check if this is the last write cycle (Writing_done will be true afterwards)
diff --git a/finn-rtllib/swg/swg_template_wrapper.v b/finn-rtllib/swg/swg_template_wrapper.v
index 1b470817d61af4140141d8478c4ae0d538678ad7..4411348beb7c338151955810021803f140c27b25 100644
--- a/finn-rtllib/swg/swg_template_wrapper.v
+++ b/finn-rtllib/swg/swg_template_wrapper.v
@@ -1,14 +1,16 @@
`timescale 1 ns / 1 ps
module $TOP_MODULE_NAME$ (
- ap_clk,
- ap_rst_n,
- in0_V_TDATA,
- in0_V_TVALID,
- in0_V_TREADY,
- out_V_TDATA,
- out_V_TVALID,
- out_V_TREADY
+(* X_INTERFACE_PARAMETER = "ASSOCIATED_BUSIF in0_V:out_V" *)
+input ap_clk,
+(* X_INTERFACE_PARAMETER = "ASSOCIATED_BUSIF in0_V:out_V" *)
+input ap_rst_n,
+input [BUF_IN_WIDTH-1:0] in0_V_TDATA,
+input in0_V_TVALID,
+output in0_V_TREADY,
+output [BUF_OUT_WIDTH-1:0] out_V_TDATA,
+output out_V_TVALID,
+input out_V_TREADY
);
// top-level parameters (set via code-generation)
@@ -21,17 +23,6 @@ parameter MMV_OUT = $MMV_OUT$;
parameter BUF_IN_WIDTH = BIT_WIDTH * SIMD * MMV_IN;
parameter BUF_OUT_WIDTH = BIT_WIDTH * SIMD * MMV_OUT;
-(* X_INTERFACE_PARAMETER = "ASSOCIATED_BUSIF in0_V:out_V" *)
-input ap_clk;
-(* X_INTERFACE_PARAMETER = "ASSOCIATED_BUSIF in0_V:out_V" *)
-input ap_rst_n;
-input [BUF_IN_WIDTH-1:0] in0_V_TDATA;
-input in0_V_TVALID;
-output in0_V_TREADY;
-output [BUF_OUT_WIDTH-1:0] out_V_TDATA;
-output out_V_TVALID;
-input out_V_TREADY;
-
$TOP_MODULE_NAME$_impl
#(
.BIT_WIDTH(BIT_WIDTH),
diff --git a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator_rtl.py b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator_rtl.py
index 98351942b9b4abd1568c9d465710a181e9cab86c..366dd396d199e02846046b00d4b829d0871060dc 100755
--- a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator_rtl.py
+++ b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator_rtl.py
@@ -1,4 +1,4 @@
-# Copyright (c) 2020, Xilinx
+# Copyright (c) 2022, Xilinx
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
@@ -72,7 +72,7 @@ class ConvolutionInputGenerator_rtl(HLSCustomOp):
"OFMDim": ("ints", True, []), # [H, W] = [Y, X]
"SIMD": ("i", True, 0),
"M": ("i", False, 1),
- "parallel_window": ("i", False, 0, {0, 1}),
+ "parallel_window": ("i", False, 0, {0}),
"Stride": ("ints", True, []), # [H, W] = [Y, X]
"Dilation": ("ints", True, []), # [H, W] = [Y, X]
# FINN DataTypes for inputs, weights, outputs
@@ -212,6 +212,49 @@ class ConvolutionInputGenerator_rtl(HLSCustomOp):
return (ifm_ch, ifm_dim, ofm_dim, k, stride, dilation)
+ def get_buffer_depth(self):
+ ifm_ch = self.get_nodeattr("IFMChannels")
+ k = self.get_nodeattr("ConvKernelDim")
+ ifm_dim = self.get_nodeattr("IFMDim")
+ stride = self.get_nodeattr("Stride")
+ dilation = self.get_nodeattr("Dilation")
+ simd = self.get_nodeattr("SIMD")
+
+ k_h, k_w = k
+ h, w = ifm_dim
+ stride_h, stride_w = stride
+ dilation_h, dilation_w = dilation
+ mmv_in = 1
+ mmv_out = 1
+ channel_factor = int(ifm_ch / simd)
+
+ impl_style = self.select_impl_style()
+ if impl_style == "default":
+ # compute minimal buffer length (assuming it holds 1 complete window)
+ buffer_min_size = (
+ (k_h - 1) * dilation_h * w + (k_w - 1) * dilation_w + 1
+ ) * channel_factor
+
+ # add additional buffer space in case of stride > 1
+ # this minimizes cycle count as it allows an earlier pre-load of inputs
+ buffer_depth = (
+ buffer_min_size
+ + max(
+ 0,
+ ((stride_w - 1) - (int(mmv_out * k_h * k_w / mmv_in)))
+ * channel_factor,
+ )
+ + max(
+ 0,
+ ((stride_h - 1) * w - (int(mmv_out * k_h * k_w / mmv_in)))
+ * channel_factor,
+ )
+ )
+ else:
+ buffer_depth = 0
+ raise Exception("Requested impl. style not implemented")
+ return buffer_depth
+
def get_exp_cycles(self):
simd = self.get_nodeattr("SIMD")
ifm_ch = self.get_nodeattr("IFMChannels")
@@ -268,17 +311,11 @@ class ConvolutionInputGenerator_rtl(HLSCustomOp):
def bram_estimation(self):
simd = self.get_nodeattr("SIMD")
ram_style = self.get_nodeattr("ram_style")
- impl_style = self.select_impl_style()
- # call codegen preparation to populate self.buffer_depth
- if impl_style == "default":
- self.prepare_codegen_default()
- else:
- raise Exception("Requested impl. style not implemented")
# NOTE: Actual BRAM usage might be lower in some cases.
# This does not account for the exact Vivado behavior yet.
buffer_width = simd * self.get_input_datatype().bitwidth()
- buffer_depth = self.buffer_depth
+ buffer_depth = self.get_buffer_depth()
if ram_style == "block" or ram_style == "auto":
if buffer_depth <= 512:
ram_width = 36
@@ -321,15 +358,8 @@ class ConvolutionInputGenerator_rtl(HLSCustomOp):
def lut_estimation(self):
simd = self.get_nodeattr("SIMD")
ram_style = self.get_nodeattr("ram_style")
- impl_style = self.select_impl_style()
- # call codegen preparation to populate self.buffer_depth
- if impl_style == "default":
- self.prepare_codegen_default()
- else:
- raise Exception("Requested impl. style not implemented")
-
buffer_width = simd * self.get_input_datatype().bitwidth()
- buffer_depth = self.buffer_depth
+ buffer_depth = self.get_buffer_depth()
if ram_style == "distributed":
ram_luts = int(buffer_width * math.ceil(buffer_depth / 38))
else:
@@ -339,15 +369,8 @@ class ConvolutionInputGenerator_rtl(HLSCustomOp):
def uram_estimation(self):
simd = self.get_nodeattr("SIMD")
ram_style = self.get_nodeattr("ram_style")
- impl_style = self.select_impl_style()
- # call codegen preparation to populate self.buffer_depth
- if impl_style == "default":
- self.prepare_codegen_default()
- else:
- raise Exception("Requested impl. style not implemented")
-
buffer_width = simd * self.get_input_datatype().bitwidth()
- buffer_depth = self.buffer_depth
+ buffer_depth = self.get_buffer_depth()
if ram_style == "ultra":
ram_depth = 4096
@@ -460,21 +483,7 @@ class ConvolutionInputGenerator_rtl(HLSCustomOp):
(k_h - 1) * dilation_h * w + (k_w - 1) * dilation_w + 1
) * channel_factor
- # add additional buffer space in case of stride > 1
- # this minimizes cycle count as it allows an earlier pre-load of input elements
- buffer_actual_size = (
- buffer_min_size
- + max(
- 0,
- ((stride_w - 1) - (int(mmv_out * k_h * k_w / mmv_in))) * channel_factor,
- )
- + max(
- 0,
- ((stride_h - 1) * w - (int(mmv_out * k_h * k_w / mmv_in)))
- * channel_factor,
- )
- )
- self.buffer_depth = buffer_actual_size # for resource estimation
+ buffer_actual_size = self.get_buffer_depth()
code_gen_dict["$BUF_ELEM_TOTAL$"] = [str(buffer_actual_size)]
# compute some intermediate values, e.g., kernel "width" = k_w incl. dilation
@@ -643,9 +652,6 @@ class ConvolutionInputGenerator_rtl(HLSCustomOp):
and stride_w <= ifm_dim_w
), "Illegal conv configuration: kernel or stride > FM dimension"
- if k_h == 1 and k_w == 1:
- assert simd == ifm_ch, "1x1 Kernel only supported in parallel mode (SIMD=C)"
-
# init folding config
if self.get_nodeattr("parallel_window"):
# mmv_in = M * 1
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index 850bcf6616aebc251ebd12f9d766b89c43ee4208..540c217cbca8c47243a080ac493f19bd1c72abc8 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -132,7 +132,26 @@ class InferConvInpGen(Transformation):
)
graph.node.insert(node_ind, padding_node)
- if self.use_rtl_variant:
+ is_kernel_pointwise = k_h == 1 and k_w == 1
+ is_square_image = ConvInpGen_idim_h == ConvInpGen_idim_w
+ is_square_kernel = k_h == k_w
+ is_equal_stride = stride_h == stride_w
+ is_1d_convolution = (k_h == 1 and k_w > 1 and ifm_dim_h == 1) or (
+ k_h > 1 and k_w == 1 and ifm_dim_w == 1
+ )
+
+ # Ensure that RTL variant is not inserted for unsupported configuration
+ is_rtl_variant_compatible = True
+ if is_kernel_pointwise:
+ is_rtl_variant_compatible = False
+ if self.use_rtl_variant:
+ warnings.warn(
+ """%s : RTL ConvInpGen requested for unsupported
+ configuration. Falling back to HLS implementation."""
+ % n.name
+ )
+
+ if self.use_rtl_variant and is_rtl_variant_compatible:
ConvInpGen_node = helper.make_node(
"ConvolutionInputGenerator_rtl",
[ConvInpGen_input],
@@ -151,19 +170,11 @@ class InferConvInpGen(Transformation):
inputDataType=dt.name,
outputDataType=dt.name,
depthwise=depthwise,
- name="ConvolutionInputGenerator_rtl" + n.name,
+ name="ConvolutionInputGenerator_rtl_" + n.name,
)
graph.node.insert(ConvInpGen_node_idx, ConvInpGen_node)
else:
# Ensure that only supported HLS nodes are inserted
- is_square_image = ConvInpGen_idim_h == ConvInpGen_idim_w
- is_square_kernel = k_h == k_w
- is_kernel_pointwise = k_h == 1 and k_w == 1
- is_equal_stride = stride_h == stride_w
- is_1d_convolution = (k_h == 1 and k_w > 1 and ifm_dim_h == 1) or (
- k_h > 1 and k_w == 1 and ifm_dim_w == 1
- )
-
if (stride_h > 1 or stride_w > 1) and is_kernel_pointwise:
assert is_square_image, (
"""%s : DownSampler currently only supports square
diff --git a/src/finn/transformation/fpgadataflow/set_folding.py b/src/finn/transformation/fpgadataflow/set_folding.py
index 5c94272bad52fd6265a9bb0054fae87a3d77b93b..e24e24f1f8ebb2873c81617884cd333311d8aea9 100644
--- a/src/finn/transformation/fpgadataflow/set_folding.py
+++ b/src/finn/transformation/fpgadataflow/set_folding.py
@@ -172,11 +172,7 @@ class SetFolding(Transformation):
"Expected SWU on DW op input, found " + swu_node.op_type
)
elif op_type in simd_ops:
- if op_type in [
- "ConvolutionInputGenerator",
- "ConvolutionInputGenerator1D",
- "ConvolutionInputGenerator_rtl",
- ]:
+ if op_type.startswith("ConvolutionInputGenerator"):
depthwise = node_inst.get_nodeattr("depthwise")
if depthwise == 0:
max_simd = node_inst.get_nodeattr("IFMChannels")
diff --git a/tests/fpgadataflow/test_convert_to_hls_conv_layer.py b/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
index 56438ac6b6c5ac835ca35d9e66073042e467224f..8c9f110c315089ec03354863bf2213963197217a 100644
--- a/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
+++ b/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
@@ -164,14 +164,10 @@ def test_convert_to_hls_conv_layer(conv_config, depthwise, use_rtl_swg, exec_mod
inp_dict = {model.graph.input[0].name: x}
assert oxe.compare_execution(model, new_model, inp_dict)
- if use_rtl_swg:
- downsampler_op_type = "ConvolutionInputGenerator_rtl"
- else:
- downsampler_op_type = "DownSampler"
if kernel_size == 1 and stride > 1 and pad == 0:
- assert new_model.graph.node[1].op_type == downsampler_op_type
+ assert new_model.graph.node[1].op_type == "DownSampler"
if exec_mode == "rtlsim":
- node = new_model.get_nodes_by_op_type(downsampler_op_type)[0]
+ node = new_model.get_nodes_by_op_type("DownSampler")[0]
inst = getCustomOp(node)
cycles_rtlsim = inst.get_nodeattr("cycles_rtlsim")
exp_cycles_dict = new_model.analysis(exp_cycles_per_layer)
diff --git a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator_rtl.py b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator_rtl.py
index eeeb09329448f546f4a668fde3d32ffaa36f5aaf..5da1fa6eb1f63251769b9e88e06087cf51e863a1 100755
--- a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator_rtl.py
+++ b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator_rtl.py
@@ -142,7 +142,7 @@ def prepare_inputs(input_tensor):
# kernel size
@pytest.mark.parametrize("k", [[2, 2], [3, 3], [1, 3]])
# input dimension
-@pytest.mark.parametrize("ifm_dim", [[24, 24], [13, 13], [1, 14]])
+@pytest.mark.parametrize("ifm_dim", [[24, 24], [15, 6], [13, 13], [1, 14]])
# input channels
@pytest.mark.parametrize("ifm_ch", [6])
# Stride