I know, the question was often asked in the past and perhaps the information are given in previous Stack Overflow postings. But learning Forth is a very complicated task and repetition helps to understand the advantages of a concatenative programming language over alternative languages like C.
What I have learned from Forth tutorials is that Forth doesn't provide commands for creating a 2D array, but the user has to realize everything from scratch in the program. I've found two options in occupying memory in Forth. At first by creating a new word:
: namelist s” hello” s” world” ;
or secondly by the CREATE statement:
create temperature 10 allot
temperature 10 cells dump
So far so good; we have created an array of 10 cells in which integer variables can be stored. But what is, if I need to save float numbers? Do I have to convert float always to int or can they saved into the normal cells?
Another open problem is how to store string values in the array. As far as I know, a string contains a pointer plus a size. So in theory I can store only 5 strings in a 10 cell array and additionally I need memory somewhere else which holds the string itself. That doesn't make much sense.
Is there some kind of higher abstraction available to store values in arrays which can be used to write easy to read programs? I mean, if every programmer is using his own Forth method to store something in the memory, other programmers will find it hard to read the code.
create creates a word that returns address of a buffer in the dictionary (data space); it is zero length initially, so you have to reserve required space for it right away.
allot reserves space that is measured in address units (usually bytes), so you have to calculate the required size in bytes.
For example:
create a 10 cells allot
create b 10 floats allot
It is just buffers, and you still need to deal with pointers arithmetic to get or set an item, e.g.:
0.35e 2 floats b + f! \ store the float number into third item (0-based indexing)
Example of a word that creates an array of floats in the dictionary:
: create-floats-array ( length "name" -- ) create floats allot does> swap 1- floats + ;
10 create-floats-array c
0.35e 3 c f! \ store the float number into third item (1-based indexing)
3 c f# f. \ print float number form third item
If you need many arrays and many strings it is better to use appropriate libraries.
For example, see Cell array module and Dynamic text string module from Forth Foundation Library.
A generalised 2darray of elements. Takes the element size as a parameter
\ size is the per element multiplier
\ n size * is the per_row factor
\ m n size * * is the total space required
: 2darray \ create> m n size -- ; does> mix nix -- a
\ size is the number of bytes for one element
\
create 2dup * , \ multiplier to apply to m index
dup , \ multiplier to apply to n index
* * allot \ calculate total bytes and allot them
does> \ mix nix -- a
>r r# cell+ # * \ offset from n index
swap r# # * + \ offset with m index
r> + 2 cells+ \ 2 cells offset for the 'admin' cells
;
Examples
50 40 /float 2darray 50x40floats
50 40 2 cells 2darray 50x40stringpairs
even
20 constant /record
10 200 /record 2darray 2000records
You're confused about strings. The string just goes into memory and memory at that address is allocated for that string, and it's there forever (unless you change that).
So if you wanted to store 5 ( c-addr u) strings in an allocated block of memory (calling it an array is a bit of a stretch), you can just store the c-addr in cell n and the length u in cell n+1.
If you're worried about 10 cells being a lot of space (it's really nothing to worry about) and only want to use 5 cells, you can store your strings as counted strings, using words like C" - counted strings store the length in the first byte, every subsequent byte is a character.
Also, you can store things into the dictionary at the current dp using the word , (comma).
Related
I'm looking for an SSE Bitwise OR between components of same vector. (Editor's note: this is potentially an X-Y problem, see below for the real comparison logic.)
I am porting some SIMD logic from SPU intrinsics. It has an instruction
spu_orx(a)
Which according to the docs
spu_orx: OR word across d = spu_orx(a) The four word elements of
vector a are logically Ored. The result is returned in word element 0
of vector d. All other elements (1,2,3) of d are assigned a value of
zero.
How can I do that with SSE 2 - 4 involving minimum instruction? _mm_or_ps is what I got here.
UPDATE:
Here is the scenario from SPU based code:
qword res = spu_orx(spu_or(spu_fcgt(x, y), spu_fcgt(z, w)))
So it first ORs two 'greater' comparisons, then ORs its result.
Later couples of those results are ANDed to get final comparison value.
This is effectively doing (A||B||C||D||E||F||G||H) && (I||J||K||L||M||N||O||P) && ... where A..D are the 4x 32-bit elements of the fcgt(x,y) and so on.
Obviously vertical _mm_or_ps of _mm_cmp_ps results is a good way to reduce down to 1 vector, but then what? Shuffle + OR, or something else?
UPDATE 1
Regarding "but then what?"
I perform
qword res = spu_orx(spu_or(spu_fcgt(x, y), spu_fcgt(z, w)))
On SPU it goes like this:
qword aRes = si_and(res, res1);
qword aRes1 = si_and(aRes, res2);
qword aRes2 = si_and(aRes1 , res3);
return si_to_uint(aRes2 );
several times on different inputs,then AND those all into a single result,which is finally cast to integer 0 or 1 (false/true test)
SSE4.1 PTEST bool any_nonzero = !_mm_testz_si128(v,v);
That would be a good way to horizontal OR + booleanize a vector into a 0/1 integer. It will compile to multiple instructions, and ptest same,same is 2 uops on its own. But once you have the result as a scalar integer, scalar AND is even cheaper than any vector instruction, and you can branch on the result directly because it sets integer flags.
#include <immintrin.h>
bool any_nonzero_bit(__m128i v) {
return !_mm_testz_si128(v,v);
}
On Godbolt with gcc9.1 -O3 -march=nehalem:
any_nonzero(long long __vector(2)):
ptest xmm0, xmm0 # 2 uops
setne al # 1 uop with false dep on old value of RAX
ret
This is only 3 uops on Intel for a horizontal OR into a single bit in an integer register. AMD Ryzen ptest is only 1 uop so it's even better.
The only risk here is if gcc or clang creates false dependencies by not xor-zeroing eax before doing a setcc into AL. Usually gcc is pretty fanatical about spending extra uops to break false dependencies so I don't know why it doesn't here. (I did check with -march=skylake and -mtune=generic in case it was relying on Nehalem partial-register renaming for -march=nehalem. Even -march=znver1 didn't get it to xor-zero EAX before the ptest.)
It would be nice if we could avoid the _mm_or_ps and have PTEST do all the work. But even if we consider inverting the comparisons, the vertical-AND / horizontal-OR behaviour doesn't let us check something about all 8 elements of 2 vectors, or about any of those 8 elements.
e.g. Can PTEST be used to test if two registers are both zero or some other condition?
// NOT USEFUL
// 1 if all the vertical pairs AND to zero.
// but 0 if even one vertical AND result is non-zero
_mm_testz_si128( _mm_castps_si128(_mm_cmpngt_ps(x,y)),
_mm_castps_si128(_mm_cmpngt_ps(z,w)));
I mention this only to rule it out and save you the trouble of considering this optimization idea. (#chtz suggested it in comments. Inverting the comparison is a good idea that can be useful for other ways of doing things.)
Without SSE4.1 / delaying the horizontal OR
We might be able to delay horizontal ORing / booleanizing until after combining some results from multiple vectors. This makes combining more expensive (imul or something), but saves 2 uops in the vector -> integer stage vs. PTEST.
x86 has cheap vector mask->integer bitmap with _mm_movemask_ps. Especially if you ultimately want to branch on the result, this might be a good idea. (But x86 doesn't have a || instruction that booleanizes its inputs either so you can't just & the movemask results).
One thing you can do is integer multiply movemask results: x * y is non-zero iff both inputs are non-zero. Unlike x & y which can be false for 0b0101 &0b1010for example. (Our inputs are 4-bit movemask results andunsigned` is 32-bit so we have some room before we overflow). AMD Bulldozer family has an integer multiply that isn't fully pipelined so this could be a bottleneck on old AMD CPUs. Using just 32-bit integers is also good for some low-power CPUs with slow 64-bit multiply.
This might be good if throughput is more of a bottleneck than latency, although movmskps can only run on one port.
I'm not sure if there are any cheaper integer operations that let us recover the logical-AND result later. Adding doesn't work; the result is non-zero even if only one of the inputs was non-zero. Concatenating the bits together (shift+or) is also of course like an OR if we eventually just test for any non-zero bit. We can't just bitwise AND because 2 & 1 == 0, unlike 2 && 1.
Keeping it in the vector domain
Horizontal OR of 4 elements takes multiple steps.
The obvious way is _mm_movehl_ps + OR, then another shuffle+OR. (See Fastest way to do horizontal float vector sum on x86 but replace _mm_add_ps with _mm_or_ps)
But since we don't actually need an exact bitwise-OR when our inputs are compare results, we just care if any element is non-zero. We can and should think of the vectors as integer, and look at integer instructions like 64-bit element ==. One 64-bit element covers/aliases two 32-bit elements.
__m128i cmp = _mm_castps_si128(cmpps_result); // reinterpret: zero instructions
// SSE4.1 pcmpeqq 64-bit integer elements
__m128i cmp64 = _mm_cmpeq_epi64(cmp, _mm_setzero_si128()); // -1 if both elements were zero, otherwise 0
__m128i swap = _mm_shuffle_epi32(cmp64, _MM_SHUFFLE(1,0, 3,2)); // copy and swap, no movdqa instruction needed even without AVX
__m128i bothzero = _mm_and_si128(cmp64, swap); // both halves have the full result
After this logical inversion, ORing together multiple bothzero results will give you the AND of multiple conditions you're looking for.
Alternatively, SSE4.1 _mm_minpos_epu16(cmp64) (phminposuw) will tell us in 1 uop (but 5 cycle latency) if either qword is zero. It will place either 0 or 0xFFFF in the lowest word (16 bits) of the result in this case.
If we inverted the original compares, we could use phminposuw on that (without pcmpeqq) to check if any are zero. So basically a horizontal AND across the whole vector. (Assuming that it's elements of 0 / -1). I think that's a useful result for inverted inputs. (And saves us from using _mm_xor_si128 to flip the bits).
An alternative to pcmpeqq (_mm_cmpeq_epi64) would be SSE2 psadbw against a zeroed vector to get 0 or non-zero results in the bottom of each 64-bit element. It won't be a mask, though, it's 0xFF * 8. Still, it's always that or 0 so you can still AND it. And it doesn't invert.
The structure is,
struct {
char a;
short b;
short c;
short d;
char e;
} s1;
size of short is given as 2 bytes
size of char is given as 1 bytes
It is a 32-bit LITTLE ENDIAN processor
According to me, the answer should be:
1000 a[0]
1001 offset
1002 b[0]
1003 b[1]
1004 c[0]
1005 c[1]
1006 d[0]
1007 d[1]
1008 e[0]
size of S1 = 9 bytes
but according to the solution, the size of S1 is supposed to be 10 bytes
The answer here is that it is that the layout of the structure is entirely up to the compiler.
10 is likely to be the most common size of this structure.
The reason for the padding is that, if there is an array, it will keep all the members properly aligned. If the size were 9, every other array element would have misaligned structure members.
Unaligned did accesses are not permitted on some systems. On most systems, they cause the processor to use extra cycles to access the data.
A compiler could allocate 4 bytes for each element in such a structure.
The C Standard says (sorry, not at my computer, so no quote): structs are aligned to the alignment of the largest (base type) member. Your largest member field is a short, 2 bytes, so the first element 'a' is aligned at an even address. 'a' takes up 1 byte. 'b' has to be aligned again at an even address, so one byte gets wasted. The last element of your struct 'e' is also one byte, and the byte following that is likely to be wasted, but that doesn't have to show up in the size of the struct. If put 'a' to the end, ie rearrange the members, you are likely to find the size of your struct to be 8 bytes..which is as good as it gets.
In section 4, Tables, in The Implementation of Lua 5.0 there is and example:
local t = {100, 200, 300, x = 9.3}
So we have t[4] == nil. If I write t[0] = 0, this will go to hash part.
If I write t[5] = 500 where it will go? Array part or hash part?
I would eager to hear answer for Lua 5.1, Lua 5.2 and LuaJIT 2 implementation if there is difference.
Contiguous integer keys starting from 1 always go in the array part.
Keys that are not positive integers always go in the hash part.
Other than that, it is unspecified, so you cannot predict where t[5] will be stored according to the spec (and it may or may not move between the two, for example if you create then delete t[4].)
LuaJIT 2 is slightly different - it will also store t[0] in the array part.
If you need it to be predictable (which is probably a design smell), stick to pure-array tables (contiguous integer keys starting from 1 - if you want to leave gap use a value of false instead of nil) or pure hash tables (avoid non-negative integer keys.)
Quoting from Implementation of Lua 5.0
The array part tries to store the values corresponding to integer keys from 1 to some limit n.Values corresponding to non-integer keys or to integer keys outside the array range are
stored in the hash part.
The index of the array part starts from 1, that's why t[0] = 0 will go to hash part.
The computed size of the array part is the largest nsuch that at least half the slots between 1 and n are in use (to avoid wasting space with sparse arrays) and there is at least one used slot between n/2+1 and n(to avoid a size n when n/2 would do).
According from this rule, in the example table:
local t = {100, 200, 300, x = 9.3}
The array part which holds 3 elements, may have a size of 3, 4 or 5. (EDIT: the size should be 4, see #dualed's comment.)
Assume that the array has a size of 4, when writing t[5] = 500, the array part can no longer hold the element t[5], what if the array part resize to 8? With a size of 8, the array part holds 4 elements, which is equal to (so, not less that) half of the array size. And the index from between n/2+1 and n, which in this case, is 5 to 8, has one element:t[5]. So an array size of 8 can accomplish the requirement. In this case, t[5] will go to the array part.
I'm trying to solve some huffman coding problems, but I always get different values for the codewords (values not lengths).
for example, if the codeword of character 'c' was 100, in my solution it is 101.
Here is an example:
Character Frequency codeword my solution
A 22 00 10
B 12 100 010
C 24 01 11
D 6 1010 0110
E 27 11 00
F 9 1011 0111
Both solutions have the same length for codewords, and there is no codeword that is prefix of another codeword.
Does this make my solution valid ? or it has to be only 2 solutions, the optimal one and flipping the bits of the optimal one ?
There are 96 possible ways to assign the 0's and 1's to that set of lengths, and all would be perfectly valid, optimal, prefix codes. You have shown two of them.
There exist conventions to define "canonical" Huffman codes which resolve the ambiguity. The value of defining canonical codes is in the transmission of the code from the compressor to the decompressor. As long as both sides know and agree on how to unambiguously assign the 0's and 1's, then only the code length for each symbol needs to be transmitted -- not the codes themselves.
The deflate format starts with zero for the shortest code, and increments up. Within each code length, the codes are ordered by the symbol values, i.e. sorting by symbol. So for your code that canonical Huffman code would be:
A - 00
C - 01
E - 10
B - 110
D - 1110
F - 1111
So there the two bit codes are assigned in the symbol order A, C, E, and similarly, the four bit codes are assigned in the order D, F. Shorter codes are assigned before longer codes.
There is a different and interesting ambiguity that arises in finding the code lengths. Depending on the order of combination of equal frequency nodes, i.e. when you have a choice of more than two lowest frequency nodes, you can actually end up with different sets of code lengths that are exactly equally optimal. Even though the code lengths are different, when you multiply the lengths by the frequencies and add them up, you get exactly the same number of bits for the two different codes.
There again, the different codes are all optimal and equally valid. There are ways to resolve that ambiguity as well at the time the nodes to combine are chosen, where the benefit can be minimizing the depth of the tree. That can reduce the table size for table-driven Huffman decoding.
For example, consider the frequencies A: 2, B: 2, C: 1, D: 1. You first combine C and D to get 2. Then you have A, B, and C+D all with frequency 2. Now you can choose to combine either A and B, or C+D with A or B. This gives two different sets of bit lengths. If you combine A and B, you get lengths: A-2, B-2, C-2, and D-2. If you combine C+D with B, you get A-1, B-2, C-3, D-3. Both are optimal codes, since 2x2 + 2x2 + 1x2 + 1x2 = 2x1 + 2x2 + 1x3 + 1x3 = 12, so both codes use 12 bits to represent those symbols that many times.
The problem is, that there is no problem.
You huffman tree is valid, it also gives the exactly same results after encoding and decoding. Just think if you would build a huffman tree by hand, there are always more ways to combine items with equal (or least difference) value. E.g. if you have A B C (everyone frequency 1), you can at first combine A and B, and the result with C, or at first B and C, and the result with a.
You see, there are more correct ways.
Edit: Even with only one possible way to combine the items by frequency, you can get different results because you can assign 1 for the left or for the right branch, so you would get different (correct) results.
I'm puzzling over how to map a set of sequences to consecutive integers.
All the sequences follow this rule:
A_0 = 1
A_n >= 1
A_n <= max(A_0 .. A_n-1) + 1
I'm looking for a solution that will be able to, given such a sequence, compute a integer for doing a lookup into a table and given an index into the table, generate the sequence.
Example: for length 3, there are 5 the valid sequences. A fast function for doing the following map (preferably in both direction) would be a good solution
1,1,1 0
1,1,2 1
1,2,1 2
1,2,2 3
1,2,3 4
The point of the exercise is to get a packed table with a 1-1 mapping between valid sequences and cells.
The size of the set in bounded only by the number of unique sequences possible.
I don't know now what the length of the sequence will be but it will be a small, <12, constant known in advance.
I'll get to this sooner or later, but though I'd throw it out for the community to have "fun" with in the meantime.
these are different valid sequences
1,1,2,3,2,1,4
1,1,2,3,1,2,4
1,2,3,4,5,6,7
1,1,1,1,2,3,2
these are not
1,2,2,4
2,
1,1,2,3,5
Related to this
There is a natural sequence indexing, but no so easy to calculate.
Let look for A_n for n>0, since A_0 = 1.
Indexing is done in 2 steps.
Part 1:
Group sequences by places where A_n = max(A_0 .. A_n-1) + 1. Call these places steps.
On steps are consecutive numbers (2,3,4,5,...).
On non-step places we can put numbers from 1 to number of steps with index less than k.
Each group can be represent as binary string where 1 is step and 0 non-step. E.g. 001001010 means group with 112aa3b4c, a<=2, b<=3, c<=4. Because, groups are indexed with binary number there is natural indexing of groups. From 0 to 2^length - 1. Lets call value of group binary representation group order.
Part 2:
Index sequences inside a group. Since groups define step positions, only numbers on non-step positions are variable, and they are variable in defined ranges. With that it is easy to index sequence of given group inside that group, with lexicographical order of variable places.
It is easy to calculate number of sequences in one group. It is number of form 1^i_1 * 2^i_2 * 3^i_3 * ....
Combining:
This gives a 2 part key: <Steps, Group> this then needs to be mapped to the integers. To do that we have to find how many sequences are in groups that have order less than some value. For that, lets first find how many sequences are in groups of given length. That can be computed passing through all groups and summing number of sequences or similar with recurrence. Let T(l, n) be number of sequences of length l (A_0 is omitted ) where maximal value of first element can be n+1. Than holds:
T(l,n) = n*T(l-1,n) + T(l-1,n+1)
T(1,n) = n
Because l + n <= sequence length + 1 there are ~sequence_length^2/2 T(l,n) values, which can be easily calculated.
Next is to calculate number of sequences in groups of order less or equal than given value. That can be done with summing of T(l,n) values. E.g. number of sequences in groups with order <= 1001010 binary, is equal to
T(7,1) + # for 1000000
2^2 * T(4,2) + # for 001000
2^2 * 3 * T(2,3) # for 010
Optimizations:
This will give a mapping but the direct implementation for combining the key parts is >O(1) at best. On the other hand, the Steps portion of the key is small and by computing the range of Groups for each Steps value, a lookup table can reduce this to O(1).
I'm not 100% sure about upper formula, but it should be something like it.
With these remarks and recurrence it is possible to make functions sequence -> index and index -> sequence. But not so trivial :-)
I think hash with out sorting should be the thing.
As A0 always start with 0, may be I think we can think of the sequence as an number with base 12 and use its base 10 as the key for look up. ( Still not sure about this).
This is a python function which can do the job for you assuming you got these values stored in a file and you pass the lines to the function
def valid_lines(lines):
for line in lines:
line = line.split(",")
if line[0] == 1 and line[-1] and line[-1] <= max(line)+1:
yield line
lines = (line for line in open('/tmp/numbers.txt'))
for valid_line in valid_lines(lines):
print valid_line
Given the sequence, I would sort it, then use the hash of the sorted sequence as the index of the table.