<div dir="ltr">Thank you.<div><br></div><div>But how to resolve it, my work depends on it.</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Oct 24, 2017 at 6:30 PM, Michael Kruse <span dir="ltr"><<a href="mailto:llvmdev@meinersbur.de" target="_blank">llvmdev@meinersbur.de</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Your problem is due to GVN partial reduction elimination (PRE) which<br>
introduces a PHI node the current loop vectorizer cannot handle:<br>
<br>
opt -O3 stencil.ll -pass-remarks=loop-vectorize<br>
-pass-remarks-missed=loop-<wbr>vectorize<br>
-pass-remarks-analysis=loop-<wbr>vectorize<br>
<span class="">remark: <unknown>:0:0: loop not vectorized: value that could not be<br>
identified as reduction is used outside the loop<br>
remark: <unknown>:0:0: loop not vectorized<br>
<br>
</span>The message is not entirely accurate. The PHI is not used outside of<br>
the loop, but is in the loop header, therefore cannot be if-converted<br>
and is not a reduction. Polly also had difficulties with this<br>
construction as well.<br>
<br>
As presented at the dev meeting, VPlan will be able to insert a<br>
shuffle instruction when encountering this situation.<br>
<span class="HOEnZb"><font color="#888888"><br>
<br>
Michael<br>
</font></span><div class="HOEnZb"><div class="h5"><br>
<br>
<br>
<br>
<br>
2017-10-23 8:58 GMT+02:00 hameeza ahmed via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>>:<br>
> stencil.ll is attached here.<br>
><br>
> On Mon, Oct 23, 2017 at 11:37 AM, Serge Preis <<a href="mailto:spreis@yandex-team.ru">spreis@yandex-team.ru</a>> wrote:<br>
>><br>
>><br>
>><br>
>> Hello,<br>
>><br>
>> To me this is an issue in llvm loop vectorizer (if N is large enough to<br>
>> prevent complete unrolling of j-loop).<br>
>><br>
>> Woud you mind to share stencil.ll than I would say more definitely what<br>
>> the issue is.<br>
>><br>
>> Regards,<br>
>> Serge.<br>
>><br>
>><br>
>> 22.10.2017, 03:21, "hameeza ahmed" <<a href="mailto:hahmed2305@gmail.com">hahmed2305@gmail.com</a>>:<br>
>><br>
>> Hello,<br>
>><br>
>> The following stencil function that you wrote few months ago (for<br>
>> computing 2D-5 point Jacobi stencil) is vectorizable.<br>
>><br>
>> void stencil(int a[][N], int b[N])<br>
>> {<br>
>> int i, j, k;<br>
>> for (k = 0; k < N; k++) {<br>
>> for (i = 1; i <= N-2; i++) {<br>
>> for (j = 1; j <= N-2; j++)<br>
>> b[j] = 0.25 * (a[i][j] + a[i-1][j] + a[i+1][j] +<br>
>> a[i][j-1] + a[i][j+1]);<br>
>> for (j = 1; j <= N-2; j++)<br>
>> a[i][j] = b[j];<br>
>> }<br>
>> }<br>
>> }<br>
>><br>
>> but when i write the above code in main i am getting error<br>
>><br>
>> opt -S -O3 stencil.ll -o stencil_o3.ll<br>
>> remark: <unknown>:0:0: loop not vectorized: value that could not be<br>
>> identified as reduction is used outside the loop<br>
>><br>
>> Why is that so?<br>
>><br>
>> my code is follows:<br>
>> int a[N][N], b[N];<br>
>> int main(void)<br>
>> {<br>
>> int i, j, k;<br>
>> for (k = 0; k < N; k++) {<br>
>> for (i = 1; i <= N-2; i++) {<br>
>> for (j = 1; j <= N-2; j++)<br>
>> b[j] = 0.25 * (a[i][j] + a[i-1][j] + a[i+1][j] +<br>
>> a[i][j-1] + a[i][j+1]);<br>
>> for (j = 1; j <= N-2; j++)<br>
>> a[i][j] = b[j];<br>
>> }<br>
>> }<br>
>> }<br>
>><br>
>><br>
>> Please help.<br>
>><br>
>> Thank You.<br>
>><br>
>><br>
>> On Mon, Jul 3, 2017 at 10:19 AM, Serge Preis <<a href="mailto:spreis@yandex-team.ru">spreis@yandex-team.ru</a>><br>
>> wrote:<br>
>><br>
>> Hello,<br>
>><br>
>> This is not equivalent rewrite: your original code definitely shouldn't<br>
>> vectorize because it has backward cross-iteration (loop-carried) dependency:<br>
>> your value on iteration j+1 depend on value from iteration j you've just<br>
>> written. In case of vectorization you need to do load-operation-store on<br>
>> multiple consecutive values at once and it is impossible in this case.<br>
>><br>
>> In your new code all values of a[] in right-hand side of the main loop are<br>
>> from previous k iteration (because on current k iteration you're writing to<br>
>> 'b'). So there is no way to vectorize loop in its original form, but new<br>
>> form is definitely vectorizable.<br>
>><br>
>> I am second to recommend you filing a bug over 'restrict' behavior. And<br>
>> you may in fact save some memory by making 'b' 1D array (this is not<br>
>> equivalent rewrite once again though)<br>
>><br>
>> // This function computes 2D-5 point Jacobi stencil<br>
>> void stencil(int a[][N], int b[N])<br>
>> {<br>
>> int i, j, k;<br>
>> for (k = 0; k < N; k++) {<br>
>> for (i = 1; i <= N-2; i++) {<br>
>> for (j = 1; j <= N-2; j++)<br>
>> b[j] = 0.25 * (a[i][j] + a[i-1][j] + a[i+1][j] +<br>
>> a[i][j-1] + a[i][j+1]);<br>
>> for (j = 1; j <= N-2; j++)<br>
>> a[i][j] = b[j];<br>
>> }<br>
>> }<br>
>> }<br>
>><br>
>> There is a way to vectorize your code even more efficient. This will look<br>
>> like more low-level, but probably closest to what you would expect from<br>
>> vectorization.<br>
>><br>
>> void stencil(int a[][N])<br>
>> {<br>
>> int i, j, k;<br>
>> for (k = 0; k < 100; k++) {<br>
>> for (i = 1; i <= N-2; i++) {<br>
>> for (j = 1; j <= N-2; j+=16) {<br>
>> int b[16]; // This should become a register on KNL<br>
>> for (v = 0; v < 16; v++) {<br>
>> b[v] = 0.25 * (a[i][j+v] + a[i-1][j+v] + a[i+1][j+v] +<br>
>> a[i][j-1+v] + a[i][j+1+v]);<br>
>> }<br>
>> for (v = 0; v < 16; v++) { // This will be a single store<br>
>> operation<br>
>> a[i][j+v] = b[v];<br>
>> }<br>
>> }<br>
>> // You should explicitly take care about the tail of j-loop<br>
>> #if !MASKED_SHORT_LOOP_<wbr>VECTORIZATION_SUPPORTED // This is not an actual<br>
>> name, just a designator<br>
>> for (;j <= N-2; j++) {<br>
>> a[i][j] = 0.25 * (a[i][j] + a[i-1][j] + a[i+1][j] +<br>
>> a[i][j-1] + a[i][j+1]);<br>
>> }<br>
>> #else<br>
>> for (v = 0; v < 16-(N-2-j); v++) { // This would become masked<br>
>> non-loop<br>
>> b[v] = 0.25 * (a[i][j+v] + a[i-1][j+v] + a[i+1][j+v] +<br>
>> a[i][j-1+v] + a[i][j+1+v]);<br>
>> }<br>
>> for (v = 0; v < 16-(N-2-j); v++) { // This will be a single<br>
>> masked store operation<br>
>> a[i][j+v] = b[v];<br>
>> }<br>
>> #endif<br>
>> }<br>
>> }<br>
>><br>
>> Unfortunalely compiler cannot do this for you: this is not equivalent<br>
>> transformation of original code. I am also not aware of any way to express<br>
>> this desired behavior less explicitly (e.g. OpenMP SIMD pragma won't work in<br>
>> this case).<br>
>><br>
>> Minor note: You're using 'int' for data, than multiply by 0.25 (divide by<br>
>> 4) and than write it back to 'int'. This will cost you 2 conversion to/from<br>
>> double while you may just place (...) / 4 which should be optimized to<br>
>> simple sequecnce with shifts (not to single shift due to signedness, but<br>
>> still better than conversions with changes of element size 4->8->4 and data<br>
>> size INT->FP->INT).<br>
>><br>
>> And by the way why do you divide by 4, not by 5 as number of points<br>
>> suggest?<br>
>><br>
>> Serge Preis<br>
>><br>
>><br>
>> 02.07.2017, 05:11, "hameeza ahmed via llvm-dev" <<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>>:<br>
>> > further i modified the code to the following;<br>
>> ><br>
>> > #include <stdio.h><br>
>> > #define N 100351<br>
>> ><br>
>> > // This function computes 2D-5 point Jacobi stencil<br>
>> > void stencil(int a[restrict][N], int b[restrict][N])<br>
>> > {<br>
>> > int i, j, k;<br>
>> > for (k = 0; k < N; k++) {<br>
>> > for (i = 1; i <= N-2; i++)<br>
>> > for (j = 1; j <= N-2; j++)<br>
>> > b[i][j] = 0.25 * (a[i][j] + a[i-1][j] + a[i+1][j] + a[i][j-1] +<br>
>> > a[i][j+1]);<br>
>> ><br>
>> > for (i = 1; i <= N-2; i++)<br>
>> > for (j = 1; j <= N-2; j++)<br>
>> > a[i][j] = b[i][j];<br>
>> ><br>
>> > }<br>
>> > }<br>
>> ><br>
>> > but still no vectorization in IR. Also, when I set vector width<br>
>> > explicitly to 64, it gives the following error:<br>
>> ><br>
>> > remark: <unknown>:0:0: loop not vectorized: call instruction cannot be<br>
>> > vectorized<br>
>> > remark: <unknown>:0:0: loop not vectorized: value that could not be<br>
>> > identified as reduction is used outside the loop<br>
>> ><br>
>> > I need serious help on this. Please guide me.<br>
>> ><br>
>> > On Sun, Jul 2, 2017 at 1:54 AM, hameeza ahmed <<a href="mailto:hahmed2305@gmail.com">hahmed2305@gmail.com</a>><br>
>> > wrote:<br>
>> >> Does it happen due to loop carried dependence? if yes what is the<br>
>> >> solution to vectorize such codes?<br>
>> >><br>
>> >> please reply. i m waiting.<br>
>> >><br>
>> >> On Jul 1, 2017 12:30 PM, "hameeza ahmed" <<a href="mailto:hahmed2305@gmail.com">hahmed2305@gmail.com</a>> wrote:<br>
>> >>> I even tried polly but still my llvm IR does not contain vector<br>
>> >>> instructions. i used the following command;<br>
>> >>><br>
>> >>> clang -S -emit-llvm stencil.c -march=knl -O3 -mllvm -polly -mllvm<br>
>> >>> -polly-vectorizer=stripmine -o stencil_poly.ll<br>
>> >>><br>
>> >>> Please specify what is wrong with my code?<br>
>> >>><br>
>> >>> On Sat, Jul 1, 2017 at 4:08 PM, hameeza ahmed <<a href="mailto:hahmed2305@gmail.com">hahmed2305@gmail.com</a>><br>
>> >>> wrote:<br>
>> >>>> Hello,<br>
>> >>>><br>
>> >>>> I am trying to vectorize following stencil code;<br>
>> >>>><br>
>> >>>> #include <stdio.h><br>
>> >>>> #define N 100351<br>
>> >>>><br>
>> >>>> // This function computes 2D-5 point Jacobi stencil<br>
>> >>>> void stencil(int a[restrict][N])<br>
>> >>>> {<br>
>> >>>> int i, j, k;<br>
>> >>>> for (k = 0; k < 100; k++)<br>
>> >>>> { for (i = 1; i <= N-2; i++)<br>
>> >>>> { for (j = 1; j <= N-2; j++)<br>
>> >>>> { a[i][j] = 0.25 * (a[i][j] + a[i-1][j] + a[i+1][j] +<br>
>> >>>> a[i][j-1] + a[i][j+1]);<br>
>> >>>> }<br>
>> >>>> }<br>
>> >>>> }}<br>
>> >>>><br>
>> >>>> I have used the following commands<br>
>> >>>><br>
>> >>>> clang -S -emit-llvm stencil.c -march=knl -O3 -mllvm<br>
>> >>>> -disable-llvm-optzns -o stencil.ll<br>
>> >>>><br>
>> >>>> opt -S -O3 stencil.ll -o stencil_o3.ll<br>
>> >>>><br>
>> >>>> llc -x86-asm-syntax=intel stencil_o3.ll -o stencil.s<br>
>> >>>><br>
>> >>>> But the code is not vectorized. It still uses the scalar<br>
>> >>>> instructions;<br>
>> >>>><br>
>> >>>> Please correct me.<br>
>> >>>><br>
>> >>>> Thank You<br>
>> > ,<br>
>> ><br>
>> > ______________________________<wbr>_________________<br>
>> > LLVM Developers mailing list<br>
>> > <a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><br>
>> > <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br>
><br>
><br>
><br>
> ______________________________<wbr>_________________<br>
> LLVM Developers mailing list<br>
> <a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><br>
> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br>
><br>
</div></div></blockquote></div><br></div>