Re: How to make parallelizing loops and vectorization work at the same time?

[Richard Biener via Gcc]

On Fri, Sep 15, 2023 at 3:09 PM Hanke Zhang <hkzhang...@gmail.com> wrote:
>
> Richard Biener <richard.guent...@gmail.com> 于2023年9月15日周五 19:59写道：
>
> >
> > On Fri, Sep 15, 2023 at 1:21 PM Hanke Zhang via Gcc <gcc@gcc.gnu.org> wrote:
> > >
> > > Hi I'm trying to accelerate my program with -ftree-vectorize and
> > > -ftree-parallelize-loops.
> > >
> > > Here are my test results using the different options (based on
> > > gcc10.3.0 on i9-12900KF):
> > > gcc-10 test.c -O3 -flto
> > > > time: 29000 ms
> > > gcc-10 test.c -O3 -flto -mavx2 -ftree-vectorize
> > > > time: 17000 ms
> > > gcc-10 test.c -O3 -flto -ftree-parallelize-loops=24
> > > > time: 5000 ms
> > > gcc-10 test.c -O3 -flto -ftree-parallelize-loops=24 -mavx2 
> > > -ftree-vectorize
> > > > time: 5000 ms
> > >
> >
> > First of all -O3 already enables -ftree-vectorize, adding -mavx2 is what 
> > brings
> > the first gain.  So adding -ftree-vectorize to the last command-line is not
> > expected to change anything.  Instead you can use -fno-tree-vectorize on
> > the second last one.  Doing that I get 111s vs 41s thus doing both helps.
> >
> > Note parallelization hasn't seen any development in the last years.
> >
> > Richard.
>
> Hi Richard:
>
> Thank you for your sincere reply.
>
> I get what you mean above. But I still see the following after I add
> `-fipo-info-vec`:
>
> gcc-10 test.c -O3 -flto -mavx2 -fopt-info-vec
> > test.c:29:5: optimized: loop vectorized using 32 byte vectors
> gcc-10 test.c -O3 -flto -mavx2 -fopt-info-vec -ftree-parallelize-loops=24
> > nothing happened
>
> That means the vectorization does nothing help actually.
>
> At the same time, I added `-fno-tree-vectorize` to the second last one
> command. It did not bring about a performance change on my computer.
>
> So I still think only parallel loops work.

I checked GCC 13 and do see vectorized loops when parallelizing.

Richard.

> Hanke Zhang
>
> >
> > > I found that these two options do not work at the same time, that is,
> > > if I use the `-ftree-vectorize` option alone, it can bring a big
> > > efficiency gain compared to doing nothing; At the same time, if I use
> > > the option of `-ftree-parallelize-loops` alone, it will also bring a
> > > big efficiency gain. But if I use both options, vectorization fails,
> > > that is, I can't get the benefits of vectorization, I can only get the
> > > benefits of parallelizing loops.
> > >
> > > I know that the reason may be that after parallelizing the loop,
> > > vectorization cannot be performed, but is there any way I can reap the
> > > benefits of both optimizations?
> > >
> > > Here is my example program, adapted from the 462.libquantum in 
> > > speccpu2006:
> > >
> > > ```
> > > #include <stdio.h>
> > > #include <stdlib.h>
> > > #include <time.h>
> > >
> > > #define MAX_UNSIGNED unsigned long long
> > >
> > > struct quantum_reg_node_struct {
> > >     float _Complex *amplitude; /* alpha_j */
> > >     MAX_UNSIGNED *state;       /* j */
> > > };
> > >
> > > typedef struct quantum_reg_node_struct quantum_reg_node;
> > >
> > > struct quantum_reg_struct {
> > >     int width; /* number of qubits in the qureg */
> > >     int size;  /* number of non-zero vectors */
> > >     int hashw; /* width of the hash array */
> > >     quantum_reg_node *node;
> > >     int *hash;
> > > };
> > >
> > > typedef struct quantum_reg_struct quantum_reg;
> > >
> > > void quantum_toffoli(int control1, int control2, int target, quantum_reg 
> > > *reg) {
> > >     for (int i = 0; i < reg->size; i++) {
> > >         if (reg->node->state[i] & ((MAX_UNSIGNED)1 << control1)) {
> > >             if (reg->node->state[i] & ((MAX_UNSIGNED)1 << control2))  {
> > >                 reg->node->state[i] ^= ((MAX_UNSIGNED)1 << target);
> > >             }
> > >         }
> > >     }
> > > }
> > >
> > > int get_random() {
> > >     return rand() % 64;
> > > }
> > >
> > > void init(quantum_reg *reg) {
> > >     reg->size = 2097152;
> > >     for (int i = 0; i < reg->size; i++)  {
> > >         reg->node = (quantum_reg_node *)malloc(sizeof(quantum_reg_node));
> > >         reg->node->state = (MAX_UNSIGNED *)malloc(sizeof(MAX_UNSIGNED)
> > > * reg->size);
> > >         reg->node->amplitude = (float _Complex *)malloc(sizeof(float
> > > _Complex) * reg->size);
> > >         if (i >= 1) break;
> > >     }
> > >     for (int i = 0; i < reg->size; i++)  {
> > >         reg->node->amplitude[i] = 0;
> > >         reg->node->state[i] = 0;
> > >     }
> > > }
> > >
> > > int main() {
> > >     quantum_reg reg;
> > >     init(&reg);
> > >     for (int i = 0; i < 65000; i++) {
> > >         quantum_toffoli(get_random(), get_random(), get_random(), &reg);
> > >     }
> > > }
> > > ```
> > >
> > > Thanks so much.