I want to use module parameter as a size parameter of Vector, which contains registers, and I try next code:
package Test;
import Vector :: *;
(* synthesize *)
module mkTest #(
parameter UInt#(32) qsize
) (Empty);
Vector#(qsize,Reg#(Bit#(8))) queue <- replicateM (mkReg (0));
endmodule
endpackage
But compiling this module with bsc I get next error message:
Verilog generation
bsc -verilog -remove-dollar Test.bsv
Error: "Test.bsv", line 9, column 11: (T0008)
Unbound type variable `qsize'
bsc version:
Bluespec Compiler (build e55aa23)
If I use not Registers as a type of Vector elements, everything is OK. Next code will produce no errors:
package Test;
import Vector :: *;
(* synthesize *)
module mkTest #(
parameter UInt#(32) qsize
) (Empty);
Vector#(qsize,Bit#(8)) queue = replicate(0);
endmodule
endpackage
And I can not understand, why qsize is Unbound as it is clearly declared as a parameter? If I did something wrong, could you please help me and explain, how to make parameterized size Vector of Regs correctly?
I have asked this question in one of the Bluespec repositories on github and Rishiyur S. Nikhil gave me a very full explanation. See https://github.com/BSVLang/Main/issues/4
In short: Vector as a first parameter needs a type, not UInt (or Int or something else). So the right way to do will be:
Make an interface for module and make it type-polymorphic
Use type from that interface as a Vector size parameter
package Test;
import Vector :: *;
interface Queue_IFC #(numeric type qsize_t);
method Bool done;
endinterface
module mkQueue ( Queue_IFC #(qsize_t) );
Vector #(qsize_t, Reg #(Bit #(8))) queue <- replicateM (mkReg (0));
endmodule
endpackage
Related
I've use a few intrinsics before with GNAT, but I get an error for __builtin_cpu_is when trying to pass in an Chars_Ptr:
error: parameter to builtin must be a string constant or literal
I also tried plugging the "amd" target parameter in directly, but that didn't work.
with Ada.Text_IO;
with Interfaces.C.Strings;
procedure Intrinsics is
procedure CPU_Init;
pragma Import (Intrinsic, CPU_Init, "__builtin_cpu_init");
function Is_CPU (CPU_Name : Interfaces.C.Strings.chars_ptr) return Interfaces.C.Int;
pragma Import (Intrinsic, Is_CPU, "__builtin_cpu_is");
Target : constant Interfaces.C.Strings.Chars_Ptr := Interfaces.C.Strings.New_String ("amd");
begin
CPU_Init;
-- ERROR from the below line, from Is_CPU
Ada.Text_IO.Put_Line (Interfaces.C.Int'Image (Is_CPU (Target)));
end Intrinsics;
References I've been looking at:
GCC Built-ins
Learn Ada, Interfacing w/ C
I think you hit a (current) limitation in importing GCC intrinsics in Ada programs (at least for version GCC/GNAT FSF 11.2). The best workaround is to wrap the builtin/intrinsic with string literal in a C function and then import that C wrapper function in the Ada program.
The error is thrown by the GCC back-end (see here). The built-in only accepts a string literal. This is not clear from the equivalent C signature of the built-in. The equivalent C signature suggests that any constant pointer-to-char is accepted:
int __builtin_cpu_is (const char *cpuname)
However, a simple test shows that this works:
#include <stdbool.h>
bool is_amd() {
return __builtin_cpu_is("amd") != 0;
}
But this doesn't:
#include <stdbool.h>
bool is_cpu(const char *cpuname) {
return __builtin_cpu_is(cpuname) != 0;
}
During compilation the abstract syntax tree is analyzed and the reference to the built-in is being matched along with the actual parameter that is passed in. This actual parameter must be a string literal (a specific tree node type). The string literal is then parsed/matched by GCC. Upon success, the call to the built-in in the syntax tree is (as a whole) replaced by a comparison (done here).
$ gcc -c is_amd.c --dump-tree-original && cat is_amd.c.005t.original
;; Function is_amd (null)
;; enabled by -tree-original
{
return __cpu_model.__cpu_vendor == 2 ? 1 : 0;
}
Now, it seems that the GNAT front-end is currently unable to generate the exact nodes (or node pattern) in the syntax tree that will match those expected by the built-in parser. This is likely because of the declared signature of the built-in and the fact that Ada makes a clear distinction between string values and string pointers.
The GNAT front-end compares the binding to __builtin_cpu_is with the signature declared internally by the GCC back-end and concludes that the cpuname argument must be a constant pointer-to-string. So, something like this:
function Is_CPU (CPU_Name : access constant String) return Integer;
pragma Import (Intrinsic, Is_CPU, "__builtin_cpu_is");
However, when using this signature, you cannot pass a string literal directly; you must use some indirection:
AMD : aliased constant String := "amd"
and then
Is_CPU (AMD'Access);
This indirection is (as far as I can see) preserved before the GNAT front-end hands over the syntax tree to the GCC back-end; GNAT will not "inline" the string literal (that is: will not remove the indirection; which I guess is actually a good thing as you do not want a constant string to be inlined into function calls in general: multiple functions might reference the string and if the string is very big, the effect of inlining might cause the program size to grow significantly).
On the other hand, if you want to pass a string literal directly in Ada, then you need a signature similar to
function Is_CPU (CPU_Name : String) return Integer;
pragma Import (Intrinsic, Is_CPU, "__builtin_cpu_is");
This signature, however, conflicts with the signature declared by the GCC back-end. Moreover, the GNAT front-end will complain that a string literal cannot be passed-by-copy (something that is likely required for the call to be accepted and recognized by the back-end).
So, I guess some additional logic for handling GCC built-ins with string arguments would have to be added to the GNAT front-end in order for this to work and allow something like this to compile:
function Is_AMD return Boolean is
function Is_CPU (CPU_Name : String) return Integer;
pragma Import (Intrinsic, Is_CPU, "__builtin_cpu_is");
begin
return Is_CPU ("amd") /= 0;
end Is_AMD;
Until then, wrapping the intrinsic with string literal in a separate C function (like the is_amd() example above) and then importing this C wrapper function in the Ada program will be the way to go.
Eric found a working solution:
with Ada.Unchecked_Conversion;
with Ada.Text_IO;
with Interfaces.C.Strings;
procedure Main is
procedure CPU_Init;
pragma Import (Intrinsic, CPU_Init, "__builtin_cpu_init");
function Is_CPU (CPU_Name : Interfaces.C.Strings.chars_ptr) return Integer;
pragma Import (Intrinsic, Is_CPU, "__builtin_cpu_is");
function To_Chars_Ptr is
new Ada.Unchecked_Conversion (String, Interfaces.C.Strings.chars_ptr);
begin
CPU_Init;
Ada.Text_IO.Put_Line (Integer'Image (Is_CPU (To_Chars_Ptr ("intel"))));
end;
How about trying Target as shown below
Target : constant Interfaces.C.Char_Ptr := Interfaces.C.To_C ("amd");
I am seeing that Julia explicitly does NOT do classes... and I should instead embrace mutable structs.. am I going down the correct path here?? I diffed my trivial example against an official flux library but cannot gather how do I reference self like a python object.. is the cleanest way to simply pass the type as a parameter in the function??
Python
# Dense Layer
class Layer_Dense
def __init__(self, n_inputs, n_neurons):
self.weights = 0.01 * np.random.randn(n_inputs, n_neurons)
self.biases = np.zeros((1, n_neurons))
def forward(self, inputs):
pass
My JuliaLang version so far
mutable struct LayerDense
num_inputs::Int64
num_neurons::Int64
weights
biases
end
function forward(layer::LayerDense, inputs)
layer.weights = 0.01 * randn(layer.num_inputs, layer.num_neurons)
layer.biases = zeros((1, layer.num_neurons))
end
The flux libraries version of a dense layer... which looks very different to me.. and I do not know what they're doing or why.. like where is the forward pass call, is it here in flux just named after the layer Dense???
source : https://github.com/FluxML/Flux.jl/blob/b78a27b01c9629099adb059a98657b995760b617/src/layers/basic.jl#L71-L111
struct Dense{F, M<:AbstractMatrix, B}
weight::M
bias::B
σ::F
function Dense(W::M, bias = true, σ::F = identity) where {M<:AbstractMatrix, F}
b = create_bias(W, bias, size(W,1))
new{F,M,typeof(b)}(W, b, σ)
end
end
function Dense(in::Integer, out::Integer, σ = identity;
initW = nothing, initb = nothing,
init = glorot_uniform, bias=true)
W = if initW !== nothing
Base.depwarn("keyword initW is deprecated, please use init (which similarly accepts a funtion like randn)", :Dense)
initW(out, in)
else
init(out, in)
end
b = if bias === true && initb !== nothing
Base.depwarn("keyword initb is deprecated, please simply supply the bias vector, bias=initb(out)", :Dense)
initb(out)
else
bias
end
return Dense(W, b, σ)
end
This is an equivalent of your Python code in Julia:
mutable struct Layer_Dense
weights::Matrix{Float64}
biases::Matrix{Float64}
Layer_Dense(n_inputs::Integer, n_neurons::Integer) =
new(0.01 * randn(n_inputs, n_neurons),
zeros((1, n_neurons)))
end
forward(ld::Layer_Dense, inputs) = nothing
What is important here:
here I create an inner constructor only, as outer constructor is not needed; as opposed in the Flux.jl code you have linked the Dense type defines both inner and outer constructors
in python forward function does not do anything, so I copied it in Julia (your Julia code worked a bit differently); note that instead of self one should pass an instance of the object to the function as the first argument (and add ::Layer_Dense type signature so that Julia knows how to correctly dispatch it)
similarly in Python you store only weights and biases in the class, I have reflected this in the Julia code; note, however, that for performance reasons it is better to provide an explicit type of these two fields of Layer_Dense struct
like where is the forward pass call
In the code you have shared only constructors of Dense object are defined. However, in the lines below here and here the Dense type is defined to be a functor.
Functors are explained here (in general) and in here (more specifically for your use case)
I have the following short fortran code.
!==============================================
MODULE PREC
INTEGER, PARAMETER :: q=8
END MODULE PREC
!==============================================
MODULE MOD_FIT
USE prec ! q
TYPE spec
INTEGER HL,HR
COMPLEX(q), POINTER :: HMAT(:,:) ! (HL,HR)
END TYPE
END MODULE MOD_FIT
!==============================================
PROGRAM MAIN
USE prec
USE MOD_FIT ! spec
IMPLICIT NONE
!
TYPE(spec) SMP
write(*,*)'check associated:',associated(SMP%HMAT)
END
I compiled it with the newest version gfortran, and ran it. The following is what I got
check associated: T
Should it be F as I hadn't initialize it at all?
No, the status of your pointer is undefined. You are not allowed to inquire it using associated() because it can result in anything.
What you should always do is to use default initialization of all pointer components and initialize them to null().
TYPE spec
COMPLEX(q), POINTER :: HMAT(:,:) => null()
END TYPE
After that you are guaranteed to get the expected result false.
I'm attempting to pass a structure from (x86) assembler to Ada on the stack. I've been able to successfully use this pattern in C to accept to wrap a large number of arguments passed from assembly inside a struct and I'm wondering if this will work in a similar way in Ada.
Here is a (contrived, minimal) example:
When I do this, debugging the callee shows that the passed record contains uninitialised data. It appears that Ada is interpreting the C calling convention differently despite the export directive.
The RM contains information about passing structs from Ada to C, saying that it will automatically pass a record as a pointer type, but the inverse does not appear to be true. If you accept a single access type it will simply be filled with the first value on the stack, as one would expect from cdecl.
( Please excuse any minor errors, this isn't my actual code. )
#####################################################################
# Caller
#
# This pushes the values onto the stack and calls the Ada function
#####################################################################
.global __example_function
.type __example_function, #function
__example_function:
push $1
push $2
push $3
push $4
call accepts_struct
ret
----------------------------------------------------------------------------
-- Accepts_Struct
--
-- Purpose:
-- Attempts to accept arguments pass on the stack as a struct.
----------------------------------------------------------------------------
procedure Accepts_Struct (
Struct : Struct_Passed_On_Stack
)
with Export,
Convention => C,
External_Name => "accepts_struct";
----------------------------------------------------------------------------
-- Ideally the four variables passed on the stack would be accepted as
-- the values of this struct.
----------------------------------------------------------------------------
type Struct_Passed_On_Stack is
record
A : Unsigned_32;
B : Unsigned_32;
C : Unsigned_32;
D : Unsigned_32;
end record
with Convention => C;
On the other hand, this works just fine:
procedure Accepts_Struct (
A : Unsigned_32;
B : Unsigned_32;
C : Unsigned_32;
D : Unsigned_32
)
with Export,
Convention => C,
External_Name => "accepts_struct";
That's not a big deal in this minimal case, but if I'm passing 16 or more variables it gets a bit onerous. If you're wondering why I'm doing this, it's an exception handler where the processor automatically passes variables onto the stack to show register states.
Any help here would be greatly appreciated.
The record version does not work because a record is not stored on the stack. Instead 4 Unsigned_32 elements are stored on the stack. If you really want to work with a record instead of four separate unsigned integer values you can assign the four values to the members of your record within the call to "accepts_struct".
Ada expects the first entry in the stack to be a record, not an unsigned_32.
The Ada LRM, section 6.4.1 states:
For the evaluation of a parameter_association: The actual parameter is
first evaluated. For an access parameter, the access_definition is
elaborated, which creates the anonymous access type. For a parameter
(of any mode) that is passed by reference (see 6.2), a view conversion
of the actual parameter to the nominal subtype of the formal parameter
is evaluated, and the formal parameter denotes that conversion. For an
in or in out parameter that is passed by copy (see 6.2), the formal
parameter object is created, and the value of the actual parameter is
converted to the nominal subtype of the formal parameter and assigned
to the formal.
Furthermore, the passing mode for parameters is described in section 6.2:
6.2 Formal Parameter Modes
A parameter_specification declares a formal parameter of mode in, in
out, or out. Static Semantics
A parameter is passed either by copy or by reference. When a parameter
is passed by copy, the formal parameter denotes a separate object from
the actual parameter, and any information transfer between the two
occurs only before and after executing the subprogram. When a
parameter is passed by reference, the formal parameter denotes (a view
of) the object denoted by the actual parameter; reads and updates of
the formal parameter directly reference the actual parameter object.
A type is a by-copy type if it is an elementary type, or if it is a
descendant of a private type whose full type is a by-copy type. A
parameter of a by-copy type is passed by copy, unless the formal
parameter is explicitly aliased.
A type is a by-reference type if it is a descendant of one of the
following:
a tagged type;
a task or protected type;
an explicitly limited record type;
a composite type with a subcomponent of a by-reference type;
a private type whose full type is a by-reference type.
A parameter of a by-reference type is passed by reference, as is an
explicitly aliased parameter of any type. Each value of a by-reference
type has an associated object. For a parenthesized expression,
qualified_expression, or type_conversion, this object is the one
associated with the operand. For a conditional_expression, this object
is the one associated with the evaluated dependent_expression.
For other parameters, it is unspecified whether the parameter is
passed by copy or by reference.
It appears that your compiler is trying to pass the struct by reference rather than by copy. In C all parameters are passed by copy.
Maybe you already solved the problem, but if not, then you might also want to have at look at the interrupt function attribute provided by GCC (see here). I've translated a test of the GCC testsuite which pushes values to the stack (as described in section 6.12 of the Intel SDM) and reads them back in an ISR. The translated Ada version seems to work well. See here for the original C version. See the GCC ChangeLog for some additional info.
main.adb
with PR68037_1;
procedure Main is
begin
PR68037_1.Run;
end Main;
pr68037_1.ads
package PR68037_1 is
procedure Run;
end PR68037_1;
pr68037_1.adb
with System.Machine_Code;
with Ada.Assertions;
with Interfaces.C;
with GNAT.OS_Lib;
package body PR68037_1 is
-- Ada-like re-implementation of
-- gcc/testsuite/gcc.dg/guality/pr68037-1.c
subtype uword_t is Interfaces.C.unsigned_long; -- for x86-64
ERROR : constant uword_t := 16#1234567_0#;
IP : constant uword_t := 16#1234567_1#;
CS : constant uword_t := 16#1234567_2#;
FLAGS : constant uword_t := 16#1234567_3#;
SP : constant uword_t := 16#1234567_4#;
SS : constant uword_t := 16#1234567_5#;
type interrupt_frame is
record
ip : uword_t;
cs : uword_t;
flags : uword_t;
sp : uword_t;
ss : uword_t;
end record
with Convention => C;
procedure fn (frame : interrupt_frame; error : uword_t)
with Export, Convention => C, Link_Name => "__fn";
pragma Machine_Attribute (fn, "interrupt");
--------
-- fn --
--------
procedure fn (frame : interrupt_frame; error : uword_t) is
use Ada.Assertions;
use type uword_t;
begin
-- Using the assertion function here. In general, be careful when
-- calling subprograms from an ISR. For now it's OK as we will not
-- return from the ISR and not continue the execution of an interrupted
-- program.
Assert (frame.ip = IP , "Mismatch IP");
Assert (frame.cs = CS , "Mismatch CS");
Assert (frame.flags = FLAGS, "Mismatch FLAGS");
Assert (frame.sp = SP , "Mismatch SP");
Assert (frame.ss = SS , "Mismatch SS");
-- At the end of this function IRET will be executed. This will
-- result in a segmentation fault as the value for EIP is nonsense.
-- Hence, abort the program before IRET is executed.
GNAT.OS_Lib.OS_Exit (0);
end fn;
---------
-- Run --
---------
procedure Run is
use System.Machine_Code;
use ASCII;
begin
-- Mimic the processor behavior when an ISR is invoked. See also:
--
-- Intel (R) 64 and IA-32 Architectures / Software Developer's Manual
-- Volume 3 (3A, 3B, 3C & 3D) : System Programming Guide
-- Section 6.12: Exception and Interrupt Handling
--
-- Push the data to the stack and jump unconditionally to the
-- interrupt service routine.
Asm
(Template =>
"push %0" & LF &
"push %1" & LF &
"push %2" & LF &
"push %3" & LF &
"push %4" & LF &
"push %5" & LF &
"jmp __fn",
Inputs =>
(uword_t'Asm_Input ("l", SS),
uword_t'Asm_Input ("l", SP),
uword_t'Asm_Input ("l", FLAGS),
uword_t'Asm_Input ("l", CS),
uword_t'Asm_Input ("l", IP),
uword_t'Asm_Input ("l", ERROR)),
Volatile => True);
end Run;
end PR68037_1;
I compiled the program in GNAT CE 2019 with compiler options -g -mgeneral-regs-only (copied from the GCC test). Note that the parameter interrupt_frame will be passed by reference (see RM B.3 69/2).
I'm using the dgeev algorithm from the LAPACK implementation in the Accelerate framework to calculate eigenvectors and eigenvalues of a matrix. Sadly the LAPACK functions are not described in the Apple Documentation with a mere link to http://netlib.org/lapack/faq.html included.
If you look it up, you will find that the first two arguments in dgeev are characters signifying whether to calculate eigenvectors or not. In Swift, it is asking for UnsafeMutablePointer<Int8>. When I simply use "N", I get an error. The dgeev function and the error are described in the following screenshot
What should I do to solve this?
The "problem" is that the first two parameters are declared as char *
and not as const char *, even if the strings are not modified by the function:
int dgeev_(char *__jobvl, char *__jobvr, ...);
is mapped to Swift as
func dgeev_(__jobvl: UnsafeMutablePointer<Int8>, __jobvr: UnsafeMutablePointer<Int8>, ...) -> Int32;
A possible workaround is
let result = "N".withCString {
dgeev_(UnsafeMutablePointer($0), UnsafeMutablePointer($0), &N, ...)
}
Inside the block, $0 is a pointer to a NUL-terminated array of char with the
UTF-8 representation of the string.
Remark: dgeev_() does not modify the strings pointed to by the first two arguments,
so it "should be" declared as
int dgeev_(const char *__jobvl, const char *__jobvr, ...);
which would be mapped to Swift as
func dgeev_(__jobvl: UnsafePointer<Int8>, __jobvr: UnsafePointer<Int8>, ...) -> Int32;
and in that case you could simply call it as
let result = dgeev_("N", "N", &N, ...)
because Swift strings are converted to UnsafePointer<Int8>) automatically,
as explained in String value to UnsafePointer<UInt8> function parameter behavior.
It is ugly, but you can use:
let unsafePointerOfN = ("N" as NSString).UTF8String
var unsafeMutablePointerOfN: UnsafeMutablePointer<Int8> = UnsafeMutablePointer(unsafePointerOfN)
and use unsafeMutablePointerOfN as a parameter instead of "N".
With Swift 4.2 and 5 you can use this similar approach
let str = "string"
let unsafePointer = UnsafeMutablePointer<Int8>(mutating: (str as NSString).utf8String)
You can get the result from unsafePointer.