gary/lib/ECP5_RAM.bsv

605 lines
24 KiB
Plaintext

////////////////////////////////////////////////////////////
package ECP5_RAM;
import DReg::*;
import Printf::*;
import ToString::*;
import StmtFSM::*;
import DelayLine::*;
export EBRWriteMode(..);
export EBRPortConfig(..);
export EBRPort(..);
export EBR(..);
export mkEBRCore;
export mkEBR;
////////////////////////////////////////////////////////////
// Configuration types
//
// The exported block RAMs in this package have one or more ports,
// where each port is independently configurable. Not all parameters
// are exposed, notably reset behavior is hardcoded to synchronous
// reset and release. This is purely because I don't yet understand
// Bluespec's reset semantics well enough to be confident in exposing
// async reset without messing it up.
//
// The exported EBRPortConfig type is internally expanded into an
// EBRPortConfig_Resolved. This expansion process resolves defaults,
// (e.g. assigning a default clock if none was provided), derives some
// additional values that implementations need (e.g. the widths of the
// data and address I/Os as regular integers), and checks the
// configuration for consistency errors (e.g. an address type larger
// than what the hardware can support).
// EBRWriteMode specifies an EBR port's output for a write operation,
// if any.
typedef enum {
// In Normal mode, write operations do not output a value.
Normal,
// In WriteThrough mode, write operations output the value that was
// written.
WriteThrough,
// In ReadBeforeWrite mode, write operations output the value that
// was overwritten. This mode is only available on 9-bit and 18-bit
// EBR configurations.
ReadBeforeWrite
} EBRWriteMode deriving (Bits, Eq);
// EBRPortConfig is the configuration of an EBR port.
typedef struct {
// clk, if specified, is the Clock to use for the port. If
// unspecified, uses the module's default clock.
Maybe#(Clock) clk;
// rstN, if specified, is the Reset to use for the port. If
// unspecified, uses the module's default reset.
Maybe#(Reset) rstN;
// Whether to register the output of the EBR port.
//
// EBR ports always register their inputs, to present predictable
// signals to the memory circuitry. Ports can optionally also
// enable an output register, which adds latency to operations but
// decouples the memory's internal latency from the logic connected
// to the output. This may allow designs to run at higher clock
// speeds, outweighing the added cycle overhead.
//
// With non-registered output, EBR operations have a latency of 1
// cycle. Registering the output increases that to 2 cycles. By
// default, the output is not registered.
Bool register_output;
// chip_select_addr is the port's chip select address. The port
// ignores put operations that don't provide a matching chip_select
// argument.
//
// This is intended to make it easier to construct larger memories
// out of multiple EBR ports: by configuring different chip
// addresses for each port, the inputs to the overall memory can be
// routed directly to all EBR ports, rather than having to provide
// your own address decoding and routing logic.
UInt#(3) chip_select_addr;
// write_mode specifies what the EBR port outputs for write
// operations. In the default Normal mode, write operations do not
// produce any output.
EBRWriteMode write_mode;
} EBRPortConfig deriving (Eq);
instance DefaultValue#(EBRPortConfig);
defaultValue = EBRPortConfig{
clk: defaultValue,
rstN: defaultValue,
register_output: False,
chip_select_addr: 0,
write_mode: Normal
};
endinstance
// EBRPortConfig_Resolved is an elaborated version of EBRPortConfig,
// with all defaults and overrides resolved to their concrete values,
// port widths made explicit and verified.
typedef struct {
// These fields are the same as in EBRPortConfig. If the port is
// not in use, they are tied to default values that avoid any logic
// or wires being generated outside of the EBR.
Clock clk;
Reset rstN;
Bool register_output;
UInt#(3) chip_select_addr;
EBRWriteMode write_mode;
// These are values derived by resolvePortCfg from an EBRPortConfig
// and other contextual information from a module
// instantiation. These are values that modules need to derive, so
// we derive them all once here instead of forcing each module to
// do so.
// enabled is whether the port is in use at all. Modules omit all
// glue logic and wiring for disabled ports, resulting in zero
// burden during synthesis (other than consuming an EBR primitive,
// but presumably you're using the other port still).
//
// Enabled is true if the memory's type for values is a non-zero
// number of bits. In particular, eanbled=False if the caller uses
// 'void' as the port's data type.
Bool enabled;
// addr_width is the bit width of addresses. resolvePortCfg ensures
// that it is less than or equal to the maximum address width that
// makes sense for data_width.
Integer addr_width;
// data_width is the bit width of input and output values. It is
// always one of the valid values for the EBR primitive: 1, 2, 4, 9
// or 18.
Integer data_width;
// write_outputs_data is whether write_mode is one of the modes
// where write operations output a value. Modules use this to
// generate the appropriate conditions for port reads.
Bool write_outputs_data;
// operation_latency is how many cycles elapse between put()
// executing to read() being ready. It is used to generate the
// appropriate conditions for port reads.
//
// Operation latency on enabled ports is 2 if the output is
// registered, or 1 for unregistered output. Disabled ports have 0
// latency, meaning no timing logic is needed.
Integer operation_latency;
// chip_select_addr_str is the string encoding of chip_select_addr
// that the EBR hardware primitive wants for its configuration
// parameter.
String chip_select_addr_str;
// write_mode_str is the string encoding of write_mode that hte EBR
// hardware primitive wants for its configuration parameter.
String write_mode_str;
// register_output_str is the string encoding of register_output
// that the EBR hardware primitive wants for its configuration
// parameter.
String register_output_str;
} EBRPortConfig_Resolved;
function EBRPortConfig_Resolved resolvePortCfg(String module_name, String port_name, addr a, data d, EBRPortConfig cfg, Clock defaultClk, Reset defaultRstN)
provisos (Bits#(addr, addr_sz),
Bits#(data, data_sz));
let addr_sz = valueOf(addr_sz);
let data_sz = valueOf(data_sz);
let addr_max = case (data_sz) matches
0: 0;
1: 14;
2: 13;
4: 12;
9: 11;
18: 10;
default: error(sprintf("invalid data width %d for %s port %s, must be one of 0,1,2,4,9,18", data_sz, module_name, port_name));
endcase;
let enabled = data_sz != 0;
let ret = ?;
if (enabled)
ret = EBRPortConfig_Resolved{
enabled: True,
clk: cfg.clk matches tagged Valid .clk ? clk : defaultClk,
rstN: cfg.rstN matches tagged Valid .rstN ? rstN : defaultRstN,
addr_width: addr_sz,
data_width: data_sz,
register_output: cfg.register_output,
chip_select_addr: cfg.chip_select_addr,
write_mode: cfg.write_mode,
write_outputs_data: cfg.write_mode != Normal,
operation_latency: cfg.register_output ? 2 : 1,
chip_select_addr_str: sprintf("0b%03b", cfg.chip_select_addr),
write_mode_str: case (cfg.write_mode) matches
Normal: "NORMAL";
WriteThrough: "WRITETHROUGH";
ReadBeforeWrite: "READBEFOREWRITE";
endcase,
register_output_str: cfg.register_output ? "OUTREG": "NOREG"
};
else
ret = EBRPortConfig_Resolved{
enabled: False,
clk: noClock,
rstN: noReset,
addr_width: 14,
data_width: 18,
register_output: False,
chip_select_addr: 0,
write_mode: Normal,
write_outputs_data: False,
operation_latency: 0,
chip_select_addr_str: "0b000",
write_mode_str: "NORMAL",
register_output_str: "NOREG"
};
if (addr_sz > addr_max) begin
addr dummy = ?;
ret = error(sprintf("The address type for port %s of %s is wider than the hardware can implement. "+
"Address type %s has %d bits, maximum is %d",
port_name, module_name,
printType(typeOf(dummy)),
addr_sz,
addr_max));
end
return ret;
endfunction
////////////////////////////////////////////////////////////
// Exported interfaces
//
// EBRPort is a port of an EBR memory.
interface EBRPort#(type addr, type data);
method Action put(UInt#(3) chip_select, Bool write, addr address, data datain);
method data read();
endinterface
// EBR is an EBR memory.
interface EBR#(type portA_addr, type portA_data, type portB_addr, type portB_data);
interface EBRPort#(portA_addr, portA_data) portA;
interface EBRPort#(portB_addr, portB_data) portB;
endinterface
////////////////////////////////////////////////////////////
// Verilog import
//
// The raw primitive for EBR is called DP16KD. However, Lattice and
// Yosys both expose it with the I/O ports exploded out into
// individual bit signals, which is pretty horrible to plumb up here.
//
// Instead, ECP5_RAM.v defines a tiny Verilog wrapper, whose only
// purpose is to group those individual bit signals back into
// multi-bit ports that Bluespec can manipulate more elegantly.
//
// This wrapper exposes all the I/O ports with their maximum bit
// width, even though there is no configuration that can use all the
// bits. For example if you use all 14 address bits, you're only using
// 1 data bit (16384x1b configuration). If you're using all 18 bits of
// data, you're only using 10 address bits (1024x18b
// configuration). We do this because we want to drive unused signals
// to defined values, so we have to be able to see all of them.
//
// The exported wrapper modules defined further down translate these
// large raw ports into proper Bluespec types, and handle the
// necessary padding and truncation.
(* always_ready *)
interface V_EBRPort;
// Put starts an operation, if select's value matches the port's
// configured chip_select_addr.
method Action put(UInt#(3) select, Bool write, Bit#(14) address, Bit#(18) data);
// Read provides the EBR's output value. At this raw layer, read
// always returns a value, but that value is undefined unless a put
// which generates output happened N cycles prior, where N is the
// port's configured latency (see EBRPortConfig).
//
// It is the caller's responsibility to time reads correctly
// relative to puts.
method Bit#(18) read();
endinterface
interface V_EBR;
interface V_EBRPort portA;
interface V_EBRPort portB;
endinterface
// vEBRCoreInner instantiates a raw EBR primitive with the given
// configuration.
//
// The returned interface has maximally wide types on all I/O, and
// uses plain bit arrays. It also has no conditions on any methods,
// it's the caller's reponsibility to time method calls appropriately.
//
// Nothing should use this module directly, except for mkEBRCore
// below. mkEBRCore wraps the Verilog primitive in stronger types and
// handles configuration edge cases (detecting invalid configs, tying
// off unused ports), but otherwise presents the same "raw" primitive
// from a semantic perspective. Anything you can build using
// vMkEBRCore, you can build better with mkEBRCore.
import "BVI" ECP5_RAM =
module vMkEBRCore#(EBRPortConfig_Resolved cfgA,
EBRPortConfig_Resolved cfgB)
(V_EBR);
// EBRs are dual-port with independent clocks and resets on each
// port, so we need to be careful to map things correctly. Unset
// the default clock and reset entirely, so that the compiler
// complains loudly if we forget to explicitly specify the
// clocking/reset on a signal.
default_clock no_clock;
default_reset no_reset;
input_clock portA_clk(CLKA, (* unused *)CLKA_GATE) = cfgA.clk;
input_reset portA_rstN(RSTA) clocked_by(portA_clk) = cfgA.rstN;
input_clock portB_clk(CLKB, (* unused *)CLKB_GATE) = cfgB.clk;
input_reset portB_rstN(RSTB) clocked_by(portB_clk) = cfgB.rstN;
parameter DATA_WIDTH_A = cfgA.data_width;
parameter REGMODE_A = cfgA.register_output ? "OUTREG" : "NOREG";
parameter CSDECODE_A = cfgA.chip_select_addr_str;
parameter WRITEMODE_A = cfgA.write_mode_str;
parameter DATA_WIDTH_B = cfgB.data_width;
parameter REGMODE_B = cfgB.register_output ? "OUTREG" : "NOREG";
parameter CSDECODE_B = cfgB.chip_select_addr_str;
parameter WRITEMODE_B = cfgB.write_mode_str;
// The outputs of EBR ports also have an enable signal. It's
// unclear why you'd want to suppress the output of things you
// asked the memory to give you. Since I can't think of any use
// for them, leave them always enabled if the corresponding port
// is active.
port OCEA = cfgA.enabled;
port OCEB = cfgB.enabled;
interface V_EBRPort portA;
method put((*reg*)CSA, (*reg*)WEA, (*reg*)ADA, (*reg*)DIA) enable(CEA) clocked_by(portA_clk) reset_by(portA_rstN);
method DOA read() clocked_by(portA_clk) reset_by(portA_rstN);
endinterface
interface V_EBRPort portB;
method put((*reg*)CSB, (*reg*)WEB, (*reg*)ADB, (*reg*)DIB) enable(CEB) clocked_by(portB_clk) reset_by(portB_rstN);
method DOB read() clocked_by(portB_clk) reset_by(portB_rstN);
endinterface
// A quick crash course on Bluespec's scheduling instructions.
//
// Bluespec's fundamental property is that rule execution is
// serializable: all designs behave as if they execute a single
// rule at a time, in some order. In the actual hardware
// typically many rules execute in parallel on every cycle, but
// that's just an optimization: the observed behavior of the
// system must always be explainable by executing rules one at a
// time, where each rule sees the effects of all previously
// executed rules.
//
// When pulling Verilog modules into a Bluespec universe, the
// compiler must be told explicitly what orders of execution are
// valid, given the hardware's behavior. The canonical example
// is a read of a register's value and a write to the same
// register. Those two actions produce different system states
// depending on which one executes first: if read-before-write,
// the read sees the register's old value. In write-before-read,
// the read sees the updated value.
//
// That's why, if you go digging into the low level Bluespec
// definition of what a register is, you'll find a scheduling
// annotation which says that if a read and a write both want to
// happen (both methods are "enabled" in a clock cycle), the
// read must execute before the write. When translated into
// hardware, this matches familiar synchronous logic: on a given
// cycle, your logic sees the previous cycle's value, and all
// writes to registers happen at the "end" of the cycle.
//
// And so we come to the scheduling rules. Our annotations tell
// the compiler how the memory's methods can be called, if
// several of them are able to execute. Each scheduling
// annotation is written as:
//
// schedule <method(s) A> ORDERING <method(s) B>
//
// This means: assuming that method(s) A and method(s) B both
// want both execute, can both be executed without issues? And
// if yes, do they need to execute in a specific order?
//
// The orderings you can specify are:
//
// - C : "conflict". The scheduler must pick a single one of A
// or B to execute.
// - CF : "conflict-free". A and B can both execute, and the
// outcome is the same regardless of which executes first.
// - SB : "schedule before". A and B can both execute, but A
// must execute first to get correct results.
// - SBR: "schedule before (restricted)". Same as SB, but A
// and B must also execute from different rules.
//
// With that, here are the scheduling annotations for
// vMkEBRCore.
// TODO: why is portA.read CF portA.put? Shouldn't that be SB to
// match register semantics?
schedule (portA.read) CF (portA.read);
schedule (portA.read) SB (portA.put);
schedule (portA.put) C (portA.put);
schedule (portB.read) CF (portB.read);
schedule (portB.read) SB (portB.put);
schedule (portB.put) C (portB.put);
endmodule : vMkEBRCore
////////////////////////////////////////////////////////////
// Exported modules
// mkEBRCore instantiates one EBR memory block with the given
// configuration.
//
// The returned ports have no implicit conditions. The caller is
// responsible for upholding the block's timing and synchronization
// requirements, following Lattice TN 02204.
//
// read() yields valid data 1 cycle after put() for ports configured
// with unregistered output, or 2 cycles for registered outputs. At
// all other times, the returned value is undefined.
//
// portA and portB must not concurrently write the same bits, or read
// bits while the other is writing them. The stored value in a
// write-write race is undefined, as is the read value in a write-read
// race.
module mkEBRCore#(EBRPortConfig cfgA,
EBRPortConfig cfgB)
(EBR#(addr_a, data_a, addr_b, data_b))
provisos (Bits#(addr_a, addr_sz_a),
Bits#(data_a, data_sz_a),
Bits#(addr_b, addr_sz_b),
Bits#(data_b, data_sz_b),
Add#(addr_a_pad, addr_sz_a, 14),
Add#(data_a_pad, data_sz_a, 18),
Add#(addr_b_pad, addr_sz_b, 14),
Add#(data_b_pad, data_sz_b, 18));
let defaultClk <- exposeCurrentClock;
let defaultRstN <- exposeCurrentReset;
let rcfgA = resolvePortCfg("mkEBRCore", "A", addr_a ' (?), data_a ' (?), cfgA, defaultClk, defaultRstN);
let rcfgB = resolvePortCfg("mkEBRCore", "B", addr_b ' (?), data_b ' (?), cfgB, defaultClk, defaultRstN);
let vEBR <- vMkEBRCore(rcfgA, rcfgB);
interface EBRPort portA;
method Action put(UInt#(3) chip_select, Bool write, addr_a address, data_a datain);
if (!rcfgA.enabled)
noAction;
else
vEBR.portA.put(chip_select, write, zeroExtend(pack(address)), zeroExtend(pack(datain)));
endmethod
method data_a read();
if (!rcfgA.enabled)
return ?;
else
return unpack(truncate(vEBR.portA.read()));
endmethod
endinterface
interface EBRPort portB;
method Action put(UInt#(3) chip_select, Bool write, addr_b address, data_b datain);
if (!rcfgB.enabled)
noAction;
else
vEBR.portB.put(chip_select, write, zeroExtend(pack(address)), zeroExtend(pack(datain)));
endmethod
method data_b read();
if (!rcfgB.enabled)
return ?;
else
return unpack(truncate(vEBR.portB.read()));
endmethod
endinterface
endmodule
// mkEBRCore instantiates one EBR memory block with the given
// configuration.
//
// This module includes flow control for reads, but unlike the
// standard library BRAM servers there is no flow control on puts. Put
// is always_ready, and read behaves like a Wire: the result of each
// put is available for a single cycle, and is lost if not read at
// that time.
module mkEBR#(EBRPortConfig cfgA,
EBRPortConfig cfgB)
(EBR#(addr_a, data_a, addr_b, data_b))
provisos (Bits#(addr_a, addr_sz_a),
Bits#(data_a, data_sz_a),
Bits#(addr_b, addr_sz_b),
Bits#(data_b, data_sz_b),
Add#(addr_a_pad, addr_sz_a, 14),
Add#(data_a_pad, data_sz_a, 18),
Add#(addr_b_pad, addr_sz_b, 14),
Add#(data_b_pad, data_sz_b, 18));
let defaultClk <- exposeCurrentClock;
let defaultRstN <- exposeCurrentReset;
let rcfgA = resolvePortCfg("mkEBR", "A", addr_a ' (?), data_a ' (?), cfgA, defaultClk, defaultRstN);
let rcfgB = resolvePortCfg("mkEBR", "B", addr_b ' (?), data_b ' (?), cfgB, defaultClk, defaultRstN);
let mem <- mkEBRCore(cfgA, cfgB);
DelayLine#(void) latencyA <- mkDelayLine(rcfgA.operation_latency, clocked_by(rcfgA.clk), reset_by(rcfgA.rstN));
DelayLine#(void) latencyB <- mkDelayLine(rcfgB.operation_latency, clocked_by(rcfgB.clk), reset_by(rcfgB.rstN));
// WriteOnly#(Bool) portA_start_op = ?;
// ReadOnly#(Bool) portA_op_complete = ?;
// WriteOnly#(Bool) portB_start_op = ?;
// ReadOnly#(Bool) portB_op_complete = ?;
// // TODO: this variable-depth register chain should be pulled into a
// // separate "delay line" module.
// if (!rcfgA.enabled) begin
// portA_start_op = discardingWriteOnly;
// portA_op_complete = constToReadOnly(False);
// end
// else if (rcfgA.register_output) begin
// let syncA1 <- mkDReg(False, clocked_by(rcfgA.clk), reset_by(rcfgA.rstN));
// let syncA2 <- mkReg(False, clocked_by(rcfgA.clk), reset_by(rcfgA.rstN));
// portA_start_op = regToWriteOnly(syncA1);
// portA_op_complete = regToReadOnly(syncA2);
// (* no_implicit_conditions, fire_when_enabled *)
// rule syncA1_to_syncA2;
// syncA2 <= syncA1;
// endrule
// end
// else begin
// let syncA <- mkDReg(False, clocked_by(rcfgA.clk), reset_by(rcfgA.rstN));
// portA_start_op = regToWriteOnly(syncA);
// portA_op_complete = regToReadOnly(syncA);
// end
// if (!rcfgB.enabled) begin
// portB_start_op = discardingWriteOnly;
// portB_op_complete = constToReadOnly(False);
// end
// else if (rcfgB.register_output) begin
// let syncB1 <- mkDReg(False, clocked_by(rcfgB.clk), reset_by(rcfgB.rstN));
// let syncB2 <- mkReg(False, clocked_by(rcfgB.clk), reset_by(rcfgB.rstN));
// portB_start_op = regToWriteOnly(syncB1);
// portB_op_complete = regToReadOnly(syncB2);
// (* no_implicit_conditions, fire_when_enabled *)
// rule syncB1_to_syncB2;
// syncB2 <= syncB1;
// endrule
// end
// else begin
// let syncB1 <- mkDReg(False, clocked_by(rcfgB.clk), reset_by(rcfgB.rstN));
// portB_start_op = regToWriteOnly(syncB1);
// portB_op_complete = regToReadOnly(syncB1);
// end
interface EBRPort portA;
method Action put(UInt#(3) chip_select, Bool write, addr_a address, data_a datain);
mem.portA.put(chip_select, write, address, datain);
if (rcfgA.write_outputs_data || !write)
latencyA <= ?;
endmethod
method data_a read() if (rcfgA.enabled && latencyA.ready);
return mem.portA.read();
endmethod
endinterface
interface EBRPort portB;
method Action put(UInt#(3) chip_select, Bool write, addr_b address, data_b datain);
mem.portB.put(chip_select, write, address, datain);
if (rcfgB.write_outputs_data || !write)
latencyB <= ?;
endmethod
method data_b read() if (rcfgB.enabled && latencyB.ready);
return mem.portB.read();
endmethod
endinterface
endmodule : mkEBR
////////////////////////////////////////////////////////////
// Utilities
//
// These are little helpers that I expected to find in the stdlib, but
// aren't there. Thankfully, they are easy to write by following the
// examples of similar helpers.
function WriteOnly#(a) discardingWriteOnly();
return (interface WriteOnly
method Action _write(a x);
noAction;
endmethod
endinterface);
endfunction
function WriteOnly#(a) regToWriteOnly(Reg#(a) r);
return (interface WriteOnly
method _write = r._write;
endinterface);
endfunction
function ReadOnly#(a) constToReadOnly(a x);
return (interface ReadOnly
method _read;
return x;
endmethod
endinterface);
endfunction
endpackage