Files
b210-k7-fpga/lib/dsp/mult_add_clip.v
+18 6b67702ad7 Merge FPGA repository back into UHD repository
The FPGA codebase was removed from the UHD repository in 2014 to reduce
the size of the repository. However, over the last half-decade, the
split between the repositories has proven more burdensome than it has
been helpful. By merging the FPGA code back, it will be possible to
create atomic commits that touch both FPGA and UHD codebases. Continuous
integration testing is also simplified by merging the repositories,
because it was previously difficult to automatically derive the correct
UHD branch when testing a feature branch on the FPGA repository.

This commit also updates the license files and paths therein.

We are therefore merging the repositories again. Future development for
FPGA code will happen in the same repository as the UHD host code and
MPM code.

== Original Codebase and Rebasing ==

The original FPGA repository will be hosted for the foreseeable future
at its original local location: https://github.com/EttusResearch/fpga/

It can be used for bisecting, reference, and a more detailed history.

The final commit from said repository to be merged here is
05003794e2da61cabf64dd278c45685a7abad7ec. This commit is tagged as
v4.0.0.0-pre-uhd-merge.

If you have changes in the FPGA repository that you want to rebase onto
the UHD repository, simply run the following commands:

- Create a directory to store patches (this should be an empty
  directory):

    mkdir ~/patches

- Now make sure that your FPGA codebase is based on the same state as
  the code that was merged:

    cd src/fpga # Or wherever your FPGA code is stored
    git rebase v4.0.0.0-pre-uhd-merge

  Note: The rebase command may look slightly different depending on what
  exactly you're trying to rebase.

- Create a patch set for your changes versus v4.0.0.0-pre-uhd-merge:

    git format-patch v4.0.0.0-pre-uhd-merge -o ~/patches

  Note: Make sure that only patches are stored in your output directory.
  It should otherwise be empty. Make sure that you picked the correct
  range of commits, and only commits you wanted to rebase were exported
  as patch files.

- Go to the UHD repository and apply the patches:

    cd src/uhd # Or wherever your UHD repository is stored
    git am --directory fpga ~/patches/*
    rm -rf ~/patches # This is for cleanup

== Contributors ==

The following people have contributed mainly to these files (this list
is not complete):

Co-authored-by: Alex Williams <alex.williams@ni.com>
Co-authored-by: Andrej Rode <andrej.rode@ettus.com>
Co-authored-by: Ashish Chaudhari <ashish@ettus.com>
Co-authored-by: Ben Hilburn <ben.hilburn@ettus.com>
Co-authored-by: Ciro Nishiguchi <ciro.nishiguchi@ni.com>
Co-authored-by: Daniel Jepson <daniel.jepson@ni.com>
Co-authored-by: Derek Kozel <derek.kozel@ettus.com>
Co-authored-by: EJ Kreinar <ej@he360.com>
Co-authored-by: Humberto Jimenez <humberto.jimenez@ni.com>
Co-authored-by: Ian Buckley <ian.buckley@gmail.com>
Co-authored-by: Jörg Hofrichter <joerg.hofrichter@ni.com>
Co-authored-by: Jon Kiser <jon.kiser@ni.com>
Co-authored-by: Josh Blum <josh@joshknows.com>
Co-authored-by: Jonathon Pendlum <jonathan.pendlum@ettus.com>
Co-authored-by: Martin Braun <martin.braun@ettus.com>
Co-authored-by: Matt Ettus <matt@ettus.com>
Co-authored-by: Michael West <michael.west@ettus.com>
Co-authored-by: Moritz Fischer <moritz.fischer@ettus.com>
Co-authored-by: Nick Foster <nick@ettus.com>
Co-authored-by: Nicolas Cuervo <nicolas.cuervo@ettus.com>
Co-authored-by: Paul Butler <paul.butler@ni.com>
Co-authored-by: Paul David <paul.david@ettus.com>
Co-authored-by: Ryan Marlow <ryan.marlow@ettus.com>
Co-authored-by: Sugandha Gupta <sugandha.gupta@ettus.com>
Co-authored-by: Sylvain Munaut <tnt@246tNt.com>
Co-authored-by: Trung Tran <trung.tran@ettus.com>
Co-authored-by: Vidush Vishwanath <vidush.vishwanath@ettus.com>
Co-authored-by: Wade Fife <wade.fife@ettus.com>


Original-commit: bafa9d95453387814ef25e6b6256ba8db2df612f
2020-01-28 09:35:36 -08:00

261 lines
7.0 KiB
Verilog

// Copyright 2018 Ettus Research, a National Instruments Company
//
// SPDX-License-Identifier: LGPL-3.0-or-later
// Write xilinx DSP48E1 primitive for mult-add-clip (signed)
`default_nettype none
module mult_add_clip #(
parameter WIDTH_A=25, // Max 25
parameter BIN_PT_A=24,
parameter WIDTH_B=18, // Max 18
parameter BIN_PT_B=17,
parameter WIDTH_C=43, // Max 43
// Min (47-WIDTH_C-1)+BIN_PT_A+BIN_PT_B,
// Max WIDTH_C-1+BIN_PT_A+BIN_PT_B
parameter BIN_PT_C=42,
parameter WIDTH_O=43, // Max 43-(BIN_PT_A+BIN_PT_B-BIN_PT_O)
parameter BIN_PT_O=42,
parameter LATENCY=2 // Maximum is 4
) (
input wire clk,
input wire reset,
input wire CE, // Ordinarily set to 1'b1
input wire [WIDTH_A-1:0] A,
input wire [WIDTH_B-1:0] B,
input wire [WIDTH_C-1:0] C,
output reg [WIDTH_O-1:0] O
);
// DSP operations:
// O = clip(A * B + C)
//
// Mux settings:
// X,Y (01,01) = M
// Z (011) = C
localparam MREG_IN = (LATENCY >= 1) ? 1 : 0;
localparam CREG_IN = MREG_IN;
localparam PREG_IN = (LATENCY >= 2) ? 1 : 0;
localparam A2REG_IN = (LATENCY >= 3) ? 1 : 0;
localparam A1REG_IN = (LATENCY == 4) ? 1 : 0;
localparam AREG_IN = A1REG_IN + A2REG_IN;
// See OPMODE Control Bits Settings, Table 2-7,2-8,2-9
localparam ZMUX_C = 3'b011;
localparam YMUX_M = 2'b01;
localparam XMUX_M = 2'b01;
localparam [6:0] OPMODE = {ZMUX_C, YMUX_M, XMUX_M};
// A_IN is 25 bits; B_IN is 18 bits. Product M's binary point shifts:
localparam BIN_PT_M = BIN_PT_A+(25-WIDTH_A) + BIN_PT_B+(18-WIDTH_B);
// Calculate shift for C to align binary point to A*B product (M)
// Determine top and bottom indices of C (in C_IN), normalized to M
// Divide by 2**BIN_PT_C then multiply up by 2**BIN_PT_M
localparam C_TOP = WIDTH_C-1 - BIN_PT_C + BIN_PT_M;
localparam C_BOT = 0 - BIN_PT_C + BIN_PT_M;
// Determine number of sign-extended bits above C_TOP
localparam C_EXT = 47 - C_TOP;
// P is a 43-bit fixed point number with bin pt BIN_PT_M
// O is extracted from those bits
// Sign extend if more bits to left of bin pt
localparam O_EXT = ((WIDTH_O-BIN_PT_O) > (43-BIN_PT_M)) ?
(WIDTH_O-BIN_PT_O) - (43-BIN_PT_M) : 0;
// If extending, use highest bit of P, else extract bits based on bin pt
localparam P_TOP = (O_EXT > 0) ? 42 :
(42 + (WIDTH_O-BIN_PT_O) - (43-BIN_PT_M));
// Pad bottom of O if remaining P not enough bits
localparam O_PAD = (WIDTH_O > P_TOP+1) ? (WIDTH_O-P_TOP-1) : 0;
// If padding O, grab lowest bit of P, else determine based on O's width
localparam P_BOT = (O_PAD > 0) ? 0 : (P_TOP+1-WIDTH_O);
//------------------------------------------------
// Normalize C input to A*B product's binary point
//------------------------------------------------
function automatic [47:0] align_c;
input [WIDTH_C-1:0] c;
begin
// Do sign extension
if (C_EXT > 0) begin
align_c[47 -: C_EXT] = {C_EXT{c[WIDTH_C-1]}};
end
if (C_BOT < 0) begin
// Chop off lower bits of C
align_c[C_TOP:0] = c[WIDTH_C-1:(-C_BOT)];
end else begin
// Place C and zero pad if necessary
align_c[C_TOP:C_BOT] = c;
if (C_BOT > 0) begin
align_c[C_BOT-1:0] = {C_BOT{1'b0}};
end
end
end
endfunction
wire [24:0] A_IN = (WIDTH_A < 25) ? { A, {(25-(WIDTH_A)){1'b0}}} : A;
wire [17:0] B_IN = (WIDTH_B < 18) ? { B, {(18-(WIDTH_B)){1'b0}}} : B;
wire [47:0] C_IN;
wire [47:0] P_OUT;
//--------------------------------------------------
// C needs more pipeline registers at higher latency
//--------------------------------------------------
generate if (AREG_IN > 0) begin
reg [AREG_IN*WIDTH_C-1:0] c_r;
if (AREG_IN > 1) begin
always @ (posedge clk)
begin
if (CE) begin
c_r <= {c_r[0 +: (AREG_IN-1)*WIDTH_C], C};
end
end
end else begin
always @ (posedge clk)
begin
if (CE) begin
c_r <= C;
end
end
end
wire [WIDTH_C-1:0] c_pre = c_r[AREG_IN*WIDTH_C-1 -: WIDTH_C];
assign C_IN = align_c(c_pre);
end else begin
assign C_IN = align_c(C);
end endgenerate
//----------------------------------------------
// Track signs for overflow/underflow processing
//----------------------------------------------
reg [LATENCY-1:0] mult_sign;
reg [LATENCY-1:0] c_sign;
wire bin_pt_overflow;
wire adder_overflow;
wire [WIDTH_O-1:0] p_extract;
generate if (LATENCY > 1) begin
always @ (posedge clk)
begin
if (CE) begin
mult_sign <= {mult_sign[0 +: LATENCY-1], A[WIDTH_A-1] ^ B[WIDTH_B-1]};
c_sign <= {c_sign[0 +: LATENCY-1], C[WIDTH_C-1]};
end
end
end else begin
always @ (posedge clk)
begin
if (CE) begin
mult_sign <= A[WIDTH_A-1] ^ B[WIDTH_B-1];
c_sign <= C[WIDTH_C-1];
end
end
end endgenerate
assign adder_overflow = (mult_sign[LATENCY-1] == c_sign[LATENCY-1]) &&
(P_OUT[42] != c_sign[LATENCY-1]);
//----------------------------------------------
// Extract renormalized bits from P_OUT
//----------------------------------------------
generate
if (P_TOP < 42) begin
assign bin_pt_overflow = (|P_OUT[42:P_TOP]) != (&P_OUT[42:P_TOP]);
end else begin
assign bin_pt_overflow = 1'b0;
end
if (O_EXT > 0) begin
assign p_extract[WIDTH_O-1 -: O_EXT] = {O_EXT{P_OUT[42]}};
end
if (O_PAD > 0) begin
assign p_extract[O_PAD-1:0] = {O_PAD{1'b0}};
end
endgenerate
assign p_extract[WIDTH_O-1-O_EXT:O_PAD] = P_OUT[P_TOP:P_BOT];
//----------------------------------
// Clip if underflowed or overflowed
//----------------------------------
always @ (*)
begin
if (bin_pt_overflow || adder_overflow) begin
O <= {c_sign[LATENCY-1], {WIDTH_O-1{!c_sign[LATENCY-1]}}};
end else begin
O <= p_extract;
end
end
DSP48E1 #(
.ACASCREG(AREG_IN),
.AREG(AREG_IN),
.ADREG(0),
.DREG(0),
.BCASCREG(AREG_IN),
.BREG(AREG_IN),
.MREG(MREG_IN),
.CREG(CREG_IN),
.PREG(PREG_IN)
) DSP48_inst (
// Outputs
.ACOUT(),
.BCOUT(),
.CARRYCASCOUT(),
.CARRYOUT(),
.MULTSIGNOUT(),
.OVERFLOW(),
.P(P_OUT),
.PATTERNBDETECT(),
.PATTERNDETECT(),
.PCOUT(),
.UNDERFLOW(),
// Inputs
.A({5'b0,A_IN}),
.ACIN(30'b0),
.ALUMODE(4'b0000),
.B(B_IN),
.BCIN(18'b0),
.C(C_IN),
.CARRYCASCIN(1'b0),
.CARRYIN(1'b0),
.CARRYINSEL(3'b0),
.CEA1(CE),
.CEA2(CE),
.CEAD(1'b0),
.CEALUMODE(1'b1),
.CEB1(CE),
.CEB2(CE),
.CEC(CE),
.CECARRYIN(CE),
.CECTRL(CE),
.CED(1'b0),
.CEINMODE(CE),
.CEM(CE),
.CEP(CE),
.CLK(clk),
.D({25{1'b1}}),
.INMODE(5'b0),
.MULTSIGNIN(1'b0),
.OPMODE(OPMODE),
.PCIN(48'b0),
.RSTA(reset),
.RSTALLCARRYIN(reset),
.RSTALUMODE(reset),
.RSTB(reset),
.RSTC(reset),
.RSTD(reset),
.RSTCTRL(reset),
.RSTINMODE(reset),
.RSTM(reset),
.RSTP(reset)
);
endmodule // mult_add_clip
`default_nettype wire