Я строю код Verilog с множителем со знаком на основе Архитектура сумматора строк (двоичное дерево) и модифицированного алгоритма Боу-Вули .
Однако яКогда я добавляю частичные продукты в последующий слой двоичного дерева, возникает проблема с циклом генерации.
У вас есть какие-нибудь идеи, как избежать этой ошибки?
Онлайн-код edaplayground
Использует цикл создания единственно возможным способом (учитывая большую длину умножения и множителя) для добавления частичных произведений в слои двоичного файлаtree?
module multiply(clk, reset, in_valid, out_valid, in_A, in_B, out_C); // C=A*B
parameter A_WIDTH = 16;
parameter B_WIDTH = 16;
input clk, reset;
input in_valid; // to signify that in_A, in_B are valid
input signed [(A_WIDTH-1):0] in_A;
input signed [(B_WIDTH-1):0] in_B;
output reg signed [(A_WIDTH+B_WIDTH-1):0] out_C;
output reg out_valid; // to signify that out_C is valid
/*
This multiplier code architecture requires an area of O(N*M*logN) and time O(logN)
with M being the length or bitwidth of the multiplicand
see https://i.imgur.com/NaqjC6G.png or
Row Adder Tree Multipliers in http://www.andraka.com/multipli.php or
https://pdfs.semanticscholar.org/415c/d98dafb5c9cb358c94189927e1f3216b7494.pdf#page=10
regarding the mechanisms within all layers
In the case of an adder tree, the adders making up the levels closer to the input
take up real estate (remember the structure of row adder tree). As the size of
the input multiplicand bitwidth grows, it becomes more and more difficult to find a
placement that does not use long routes involving multiple switch nodes. The result
is the maximum clocking speed degrades quickly as the size of the bitwidth grows.
For signed multiplication, see also modified baugh-wooley algorithm for trick in
skipping sign extension, thus smaller final routed silicon area.
/10753315/ponimanie-modifitsirovannogo-algoritma-umnozheniya-bougli
All layers are pipelined, so throughput = one result for each clock cycle
but each multiplication result still have latency = NUM_OF_INTERMEDIATE_LAYERS
*/
// The multiplication of two numbers is equivalent to adding as many copies of one
// of them, the multiplicand, as the value of the other one, the multiplier.
localparam SMALLER_WIDTH = (A_WIDTH <= B_WIDTH) ? A_WIDTH : B_WIDTH;
localparam LARGER_WIDTH = (A_WIDTH > B_WIDTH) ? A_WIDTH : B_WIDTH;
wire [(LARGER_WIDTH-1):0] MULTIPLICAND = (A_WIDTH > B_WIDTH) ? in_A : in_B ;
wire [(SMALLER_WIDTH-1):0] MULTIPLIPLIER = (A_WIDTH <= B_WIDTH) ? in_A : in_B ;
localparam NUM_OF_INTERMEDIATE_LAYERS = $clog2(SMALLER_WIDTH);
/*Stage 1: Binary multiplications to generate partial products rows*/
// first layer has "SMALLER_WIDTH" entries of data of width "LARGER_WIDTH"
// This resulted in a binary tree with faster vertical addition processes as we have
// lesser (NUM_OF_INTERMEDIATE_LAYERS) rows to add
reg [(LARGER_WIDTH-1):0] partial_products [0:(SMALLER_WIDTH-1)];
generate
genvar first_layer_index; // all partial products rows are in first layer
for(first_layer_index=0; first_layer_index<SMALLER_WIDTH; first_layer_index=first_layer_index+1) begin: first_layer
always @(posedge clk, posedge reset)
begin
if(reset) partial_products[first_layer_index] <= 0;
else begin
partial_products[first_layer_index] <= (MULTIPLICAND & MULTIPLIPLIER[first_layer_index]); // generation of partial products rows
end
end
end
endgenerate
/*Stage 2 : Intermediate partial products additions*/
// intermediate partial product rows
// Imagine a rhombus of height of "NUM_OF_INTERMEDIATE_LAYERS"
// and width of "LARGER_WIDTH" being re-arranged into binary row adder tree
// such that additions can be done in O(logN) time
generate
genvar layer;
for(layer=1; layer<NUM_OF_INTERMEDIATE_LAYERS; layer=layer+1) begin: middle_layers
// number of leafs (or children) in each layer within the binary tree
localparam NUM_OF_PP_ADDITION = (SMALLER_WIDTH >> layer);
reg [(LARGER_WIDTH+layer-1):0] middle_rows[0:(NUM_OF_PP_ADDITION-1)];
integer pp_index; // leaf index within each layer of the tree
always @(posedge clk, posedge reset)
begin
if(reset)
begin
for(pp_index=0; pp_index<NUM_OF_PP_ADDITION ; pp_index=pp_index+1)
middle_rows[pp_index] <= 0;
end
else begin
for(pp_index=0; pp_index<NUM_OF_PP_ADDITION ; pp_index=pp_index+1)
middle_rows[pp_index] <=
middle_layers[layer-1].middle_rows[1<<pp_index] +
(middle_layers[layer-1].middle_rows[(1<<pp_index) + 1]) << 1;
end
end
end
endgenerate
/*Stage 3 : Adding the final two partial products*/
wire sign_bit = in_A[A_WIDTH-1] ^ in_B[B_WIDTH-1];
always @(posedge clk, posedge reset)
begin
if(reset)
begin
out_C <= 0;
out_valid <= 0;
end
else out_C <= 0;// {sign_bit, };
end
endmodule
iverilog '-Wall' '-g2012' design.sv testbench.sv && unbuffer vvp a.out
design.sv: 107: ошибка: Невозможнодля связывания wire / reg / memory 'middle_layers [(layer) - (' sd1)]. middle_rows [('sd1) << (pp_index)]' in 'test.mul.middle_layers [1]' </p>
design.sv: 108: ошибка: Невозможно связать wire / reg / memory 'middle_layers [(layer) - (' sd1)]. middle_rows [(('sd1) << (pp_index)) + (' sd1)] 'in' test.mul.middle_layers [1] '</p>
2 ошибка (s) во время разработки.