|
I used Fortran (77, 90) for years back in the day. It's claim to fame is very fast execution of mathematical problems, maybe the fastest. I was told it was the main compiler used for NOAA's super computers to predict/simulate weather systems. Today, of course, GPU's have come to the fore front so unless Fortran has been ported to these systems it may be eclipsed. Lot of Fortran code is still around.
"A little time, a little trouble, your better day"
Badfinger
|
|
|
|
|
If I am not mistaken, CUDA can be used with FORTRAN.
"They have a consciousness, they have a life, they have a soul! Damn you! Let the rabbits wear glasses! Save our brothers! Can I get an amen?"
|
|
|
|
|
I would not be surprised. Thanx, I didn't know that.
"A little time, a little trouble, your better day"
Badfinger
|
|
|
|
|
Another good thing about old style Fortran: You wouldn't experience any stack overflow, no out of heap space, no null pointer exception. Pre-90 Fortran didn't allow recursion, and had no dynamic allocation. All memory could be statically allocated. No run time load, no risk.
Lots of computing tasks can be solved without recursion, without new()/pointers. If you absolutely must do a recursion, you can manage your own stack by an array. Similar with linked lists.
In my student days, I was a TA for the Introduction to Programming course, taught to all Tech. University students. Some of the "classical" departments were still clinging to Fortran as the only viable language; half of the students were more modern, learning Pascal. The courses were identical, except for the language (even the textbook was identical but for the coding examples). 3 out of 4 hand-ins were identical. For the last one, the Pascal students were to build and manipulate a linked list, so the Fortran students had a completely different #4 hand-in.
One of the 'Fortran students' approached me, rather crossed: Why shall the others learn something that we don't get to learn? So I tried to explain to her how you could have a record field tell where you could find the next piece of data. I believe that I explained it referring to memory as a large array, the pointer being the index in that array. A few days later, this freshman girl approached me again, this time with Fortran code to solve the Pascal linked list problem, the 'heap' was a Fortran array, pointers were integers indexing the array, and the code certainly did solve the problem, giving the proper output.
If a freshman, non-computer girl (I think she was studying chemistry) can do it in Fortran, then a seasoned Fortran programmer with thirty years of experience should be able to!
|
|
|
|
|
yes, you correct. Code was all statically determined at compile time. Recursion was simulated with one's own stack arrays again already pre-allocated. If it was memory resident at execution time, it was bullet fast.
For many problems, the only thing not static was disk I/O. True the memory it used was static, but I/O times were not guaranteed fixed. Virtual memory was another variable not guaranteed fixed, but for most purposes, it was as good as fixed.
"A little time, a little trouble, your better day"
Badfinger
|
|
|
|
|
In my Fortran IV days, I wrote stacks / heaps / linked lists / trees using array and integer indices. That experience proved useful when I worked on a mainframe assembler. I was able to write recursive routines (e.g. an optimised quicksort) using the same techniques I had used in Fortran, emulating procedure calls as an in situ stack of return keys.
|
|
|
|
|
jsc42 wrote: In my Fortran IV days, I wrote stacks / heaps / linked lists / trees using array and integer indices. That experience proved useful when I worked on a mainframe assembler.
40 years ago I wrote a coroutine mechanism in assembler. This was on a VAX-like 32 bit supermini CISC, and I loved "having full control".
I recently picked up Aarch64 documentation to learn the ARM instruction set. I've never worked on a machine that register oriented, and I'm really itching to see both how bad code compilers generate  and to learn how much you can save through assembler coding. My experience from that supermini, heavily microcoded CISC 40 years ago was that even if the generated code looked inefficient, it was very difficult to beat the compiler by more that a single digit percentage: Pipelining, fancy prefetching and other optimization techniques flushed linear sequences so rapidly through that the execution time was almost proportional to the number of jumps taken, killing all gain from the prefetch (branch prediction wasn't common in those days). Interrupts were extremely expensive.
So I am curious to see if a RISC type CPU is much different. (After studying the Aarch64 "reduced" instruction set, my reaction is "Thanks heaven that the ARM doesn't have a complex instruction set ) So I am itching to get myself something like that "Windows Dev Kit 2023", aka. "Volterra". But I fear that when looking back, five years from now, we'll view it like an early prototype. We will view the ARM version of Windows that comes with Volterra as an early prototype. So I will try to hold back, hoping that within a year or two, there will be a nice crop of competitors to choose from.
|
|
|
|
|
|
The 'Arm A64 Instruction Set' (Armv8-A) index of instructions has 402 entries, if my count is correct. Quite a few of the entries cover several instructions, e.g. for different operand sizes, or because 2 or 3 instructions are always used together. On the other hand, a number of instructions (or rather assembler mnemonics) are really synonyms for special cases of other instructions. Other instructions have separate entries for different operand specifications (e.g. immediate constant or register identification). So counting distinct instructions without duplication requires some effort. You would probably end up with between 400 and 500.
Aside from that: A processor that supports a virtualization, three privilege levels, 256 interrupt priority levels, 3 cache levels, transactional memory, virtual memory, and what else? These are not typical RISC features when the concept was defined . Not that I will criticize or object to it. Some of the instructions are really useful, such as 'Count Leading Zeros'. Or: Compare memory location to a register and if they are equal, replace the memory location with the value in another register, as an atomic, uninterruptible instruction. Or: Reverse bit order, byte order or halfword order - nice for handling little/big/mixed-endian data, but hardly super-simple instructions.
Actually, I like the look of the ARM instruction set, interrupt system, memory management, and synchronization mechanisms. I just need a machine where I can try it out. I am unsure about Volterra being a good investment. (And it isn't yet available in Norway.)
|
|
|
|
|
Moreover, since it should not be able to allow (memory) aliasing (since it does not have pointers), a Fortran compiler eventually optimized for esoteric architectures with a high degree of parallelism, could easily be able to distribute computations that involving matrices/tensors (even huge ones) by taking full advantage of specialized computational units. It could do this almost transparently by devolving such computations directly to the hardware.
AFAIK, in C99 you can eliminate or limit (mem) aliasing with the restrict type qualifier.
|
|
|
|
|
giulicard wrote: Moreover, since it should not be able to allow (memory) aliasing (since it does not have pointers),
Being around 40 years since I coded Fortran, I have forgotten a lot , such as the parameter transfer mechanism. The standard mechanism was call by reference, wasn't it? If a subroutine assigned a value to a formal parameter, the value of the actual parameter would be changed.
A reference parameter is equivalent to a pointer. So if you pass the same actual parameter for two formal parameters, you have an aliasing issue. E.g. if the compiler reorders the assignments to the two formal parameters, it might affect the final value of the actual parameter. If some COMMON block variable is used as parameter, and the subroutine refers to the same COMMON variable, you have a similar aliasing issue.
In my study years, Pascal was quite new. It didn't become the language for the Introduction to Programming course until the year after I started my studies (and then only for half of the students), so pointers were definitely 'something new'. Yet, in the Compilers course, aliasing issues certainly had a prominent place. Aliasing was treated as exceptional situations; it wasn't something you would worry too much about in Fortran.
Even with Pascal and its strongly typed pointers, aliasing wasn't that essential - if the compiler didn't do much optimizing (which was often the case for Pascal compilers), the compiler could play back a 'You asked for it, you got it!' if a programmer referenced the same location along two different paths. Hell didn't break loose until C came with void*, pointer arithmetic, more or less arbitrary casting, and pointers to non-heap locations. Today, we have the processing power and memory capacity to do a more or less complete flow analysis of the compilation unit; in the K&R days, you couldn't expect that. I suspect that compilers were quite fast to conclude, when they encountered a void* or an arbitrary cast, "Oh well, we'll just have to drop all optimization of this one!"
|
|
|
|
|
Unfortunately, I too am old. 
I started writing software as a professional back in 1984 (Alpha Microsystem with MC68000, AMOS/L, AlphaBasic, then AlphaC... Assembly, etc, etc), and years ago the information came from magazine articles. I remember reading things about the Fortran compiler of a Cray system or other supercomputer manufacturer, saying that their Fortran compiler pushed entire matrices into the memory of hardware dedicated to matrix computation or something like that. The level of hardware parallelism, according to them, was very high for that kind of computation. But it's been almost 40 years (or maybe more) and I don't remember any of the details anymore. But this information, that Fortran facilitated the use of scientific algorithms on dedicated hardware because of the fact that it freed the optimizers of the time from looking for aliasing, has always stuck in my memory. For that matter, expecting a pointer in main memory to be able to point to an address in a block of RAM within a number-crunching subsystem, I see that as a very difficult thing to achieve. It may be because for 20 years I have also been designing digital hardware, from discrete systems to using the first Quicklogic FPGAs that appeared in the early 1990s, so I can humbly say that a modicum of the idea of how things work inside hardware I may well have.
I guess if we are referring to a mainstream compiler things are straightforward as you state. But on architectures of yesteryear devoted to hard numerical computation, I think the principles of a current compiler are not easily applicable. Others architectures, others optimizers.
That said, I have only touched on Fortran very occasionally. The last time, well over a decade ago, I helped a colleague import a library for crunching huge matrices. It was written in Fortran. These algorithms worked a particular type of not regular matrices. The purpose was of calculating price indices for inflation forecasting for a government agency, but I ignore completely the mathematical principles at the base.
Sorry for the bad elglish.
Regards
|
|
|
|
|
You are correct about NOAA. Most, if not all, the main climate models used to predict weather are written in Fortran.
When my daughter first started college she was looking at meteorology. We talked to a professor about it and I noticed several Fortran books on his shelf. I asked about them and he told us about the models which were written in the 50s and 60s in Fortran.
They are large and complex so rewriting them would be difficult and error prone. And, Fortran has been likened to NASCAR - go fast and turn left. In other words, flexibility/readability is not the main thing, speed is. The professor said the the models have evolved a lot over the years and aren't completely understood. Which would make rewrites to a "modern" language even more problematic.
|
|
|
|
|
Try GNU Fortran. Eclipse IDE used to have Fortran hooks.
|
|
|
|
|
I recommend the IBM 2540 card reader punch. Just needs some cleaning and oil and remember to empty the chip box.
|
|
|
|
|
Linux supports Fortran most of the way back to Fortran 66.
You can install the compiler(s) in WSL or on a Linux box if you have access to one
|
|
|
|
|
I think I might prefer Fortran, which I once knew, to translating from perl, which I've never properly learned. Still, it's fun in small doses.
|
|
|
|
|
I had the similar problem this past week. I had to troubleshoot old code written in FORTRAN. I downloaded and installed INTEL FORTRAN compiler. It integrated directly with latest Visual Studio, no issues. I was able to build and debug code directly from VS as if I worked with C/C++/C# (and found and fixed the problem) .
|
|
|
|
|
Check out the GNU Fortran compiler:
en.wikipedia.org/wiki/GNU_Fortran
gcc.gnu.org/fortran/
|
|
|
|
|
If you alrady know visual studio, then just get the intel fortan compiler - its FREE to download and use and works well with VS and GITHUB for source control.
There are a lot of things fortran can do now as a "modern" version of it, while still being able to leverage existing code.
|
|
|
|
|
thanks for the great info! I definitely try this configuration!
diligent hands rule....
|
|
|
|
|
The founder of a paint company was found dead on top of a cold snow covered mountain. The rescue team said he could have have used another coat.
|
|
|
|
|
... repaint and thin no more.
|
|
|
|
|
Well, they right.
If all you see is snow, build an iglo?
Bastard Programmer from Hell 
"If you just follow the bacon Eddy, wherever it leads you, then you won't have to think about politics." -- Some Bell.
|
|
|
|
|
He should have brushed up on winter wear recommendations!
|
|
|
|
|
General 
News 
Suggestion 
Question 
Bug 
Answer 
Joke 
Praise 
Rant 
Admin 
Submit an article or tip
