Notebook

Performance optimization overview¶

The purpose of this tutorial is twofold

Illustrate the performance optimizations applied to the code generated by an Operator.
Describe the options Devito provides to users to steer the optimization process.

As we shall see, most optimizations are automatically applied as they're known to systematically improve performance. Others, whose impact varies across different Operator's, are instead to be enabled through specific flags.

An Operator has several preset optimization levels; the fundamental ones are noop and advanced. With noop, no performance optimizations are introduced. With advanced, several flop-reducing and data locality optimization passes are applied. Examples of flop-reducing optimization passes are common sub-expressions elimination and factorization, while examples of data locality optimization passes are loop fusion and cache blocking. Optimization levels in Devito are conceptually akin to the -O2, -O3, ... flags in classic C/C++/Fortran compilers.

An optimization pass may provide knobs, or options, for fine-grained tuning. As explained in the next sections, some of these options are given at compile-time, others at run-time.

**** Remark ****

Parallelism -- both shared-memory (e.g., OpenMP) and distributed-memory (MPI) -- is by default disabled and is not controlled via the optimization level. In this tutorial we will also show how to enable OpenMP parallelism (you'll see it's trivial!). Another mini-guide about parallelism in Devito and related aspects is available here.

********

API¶

The optimization level may be changed in various ways:

globally, through the DEVITO_OPT environment variable. For example, to disable all optimizations on all Operator's, one could run with

DEVITO_OPT=noop python ...

programmatically, adding the following lines to a program

from devito import configuration
configuration['opt'] = 'noop'

locally, as an Operator argument

Operator(..., opt='noop')

Local takes precedence over programmatic, and programmatic takes precedence over global.

The optimization options, instead, may only be changed locally. The syntax to specify an option is

Operator(..., opt=('advanced', {<optimization options>})

A concrete example (you can ignore the meaning for now) is

Operator(..., opt=('advanced', {'blocklevels': 2})

That is, options are to be specified together with the optimization level (advanced).

Default values¶

By default, all Operator's are run with the optimization level set to advanced. So this

Operator(Eq(...))

is equivalent to

Operator(Eq(...), opt='advanced')

and obviously also to

Operator(Eq(...), opt=('advanced', {}))

In virtually all scenarios, regardless of application and underlying architecture, this ensures very good performance -- but not necessarily the very best.

Misc¶

The following functions will be used throughout the notebook for printing generated code.

In [1]:

from examples.performance.utils import unidiff_output, print_kernel

The following cell is only needed for Continuous Integration. But actually it's also an example of how "programmatic takes precedence over global" (see API section).

In [2]:

from devito import configuration
configuration['language'] = 'C'
configuration['platform'] = 'bdw'  # Optimize for an Intel Broadwell
configuration['opt-options']['par-collapse-ncores'] = 1  # Always use loop collapsing

Running example¶

Throughout the notebook we will generate Operator's for the following time-marching Eq.

In [3]:

from sympy import sin
from devito import Eq, Grid, Operator, Function, TimeFunction

grid = Grid(shape=(80, 80, 80))

f = Function(name='f', grid=grid)
u = TimeFunction(name='u', grid=grid, space_order=2)

eq = Eq(u.forward, f**2*sin(f)*u.dy.dy)

Despite its simplicity, this Eq is all we need to showcase the key components of the Devito optimization engine.

OpenMP parallelism¶

There are several ways to enable OpenMP parallelism. The one we use here consists of supplying an option to an Operator. The next cell illustrates the difference between two Operator's generated with the noop optimization level, but with OpenMP enabled on the latter one.

In [4]:

op0 = Operator(eq, opt=('noop'))
op0_omp = Operator(eq, opt=('noop', {'openmp': True}))

# print(op0)
# print(unidiff_output(str(op0), str(op0_omp)))  # Uncomment to print out the diff only
print_kernel(op0_omp)

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

The OpenMP-ized op0_omp Operator includes:

the header file "omp.h"
a #pragma omp parallel num_threads(nthreads) directive
a #pragma omp for collapse(...) schedule(dynamic,1) directive

More complex Operator's will have more directives, more types of directives, different iteration scheduling strategies based on heuristics and empirical tuning (e.g., static instead of dynamic), etc.

We note how the OpenMP pass introduces a new symbol, nthreads. This allows users to explicitly control the number of threads with which an Operator is run.

op0_omp.apply(time_M=0)  # Picks up `nthreads` from the standard environment variable OMP_NUM_THREADS
op0_omp.apply(time_M=0, nthreads=2)  # Runs with 2 threads per parallel loop

A few optimization options are available for this pass (but not on all platforms, see here), though in our experience the default values do a fine job:

par-collapse-ncores: use a collapse clause only if the number of available physical cores is greater than this value (default=4).
par-collapse-work: use a collapse clause only if the trip count of the collapsable loops is statically known to exceed this value (default=100).
par-chunk-nonaffine: a coefficient to adjust the chunk size in non-affine parallel loops. The larger the coefficient, the smaller the chunk size (default=3).
par-dynamic-work: use dynamic scheduling if the operation count per iteration exceeds this value. Otherwise, use static scheduling (default=10).
par-nested: use nested parallelism if the number of hyperthreads per core is greater than this value (default=2).

So, for instance, we could switch to static scheduling by constructing the following Operator

In [5]:

op0_b0_omp = Operator(eq, opt=('noop', {'openmp': True, 'par-dynamic-work': 100}))
print_kernel(op0_b0_omp)

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(static,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

The `advanced` mode¶

The default optimization level in Devito is advanced. This mode performs several compilation passes to optimize the Operator for computation (number of flops), working set size, and data locality. In the next paragraphs we'll dissect the advanced mode to analyze, one by one, some of its key passes.

Loop blocking¶

The next cell creates a new Operator that adds loop blocking to what we had in op0_omp.

In [6]:

op1_omp = Operator(eq, opt=('blocking', {'openmp': True}))
# print(op1_omp)  # Uncomment to see the *whole* generated code
print_kernel(op1_omp)

void bf0(struct dataobj *restrict f_vec, const float h_y, struct dataobj *restrict u_vec, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y0_blk0_size, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;

  if (x0_blk0_size == 0 || y0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(2) schedule(dynamic,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int y0_blk0 = y_m; y0_blk0 <= y_M; y0_blk0 += y0_blk0_size)
      {
        for (int x = x0_blk0; x <= x0_blk0 + x0_blk0_size - 1; x += 1)
        {
          for (int y = y0_blk0; y <= y0_blk0 + y0_blk0_size - 1; y += 1)
          {
            for (int z = z_m; z <= z_M; z += 1)
            {
              u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
            }
          }
        }
      }
    }
  }
}

**** Remark ****

'blocking' is not an optimization level -- it rather identifies a specific compilation pass. In other words, the advanced mode defines an ordered sequence of passes, and blocking is one such pass.

********

The blocking pass creates additional loops over blocks. In this simple Operator there's just one loop nest, so only a pair of additional loops are created. In more complex Operator's, several loop nests may individually be blocked, whereas others may be left unblocked -- this is decided by the Devito compiler according to certain heuristics. The size of a block is represented by the symbols x0_blk0_size and y0_blk0_size, which are runtime parameters akin to nthreads.

By default, Devito applies 2D blocking and sets the default block shape to 8x8. There are two ways to set a different block shape:

passing an explicit value. For instance, below we run with a 24x8 block shape

op1_omp.apply(..., x0_blk0_size=24)

letting the autotuner pick up a better block shape for us. There are several autotuning modes. A short summary is available here

op1_omp.apply(..., autotune='aggressive')

Loop blocking also provides two optimization options:

blockinner={False, True} -- to enable 3D (or any nD, n>2) blocking
blocklevels={int} -- to enable hierarchical blocking, to exploit multiple levels of the cache hierarchy

In the example below, we construct an Operator with six-dimensional loop blocking: the first three loops represent outer blocks, whereas the second three loops represent inner blocks within an outer block.

In [7]:

op1_omp_6D = Operator(eq, opt=('blocking', {'blockinner': True, 'blocklevels': 2, 'openmp': True}))
# print(op1_omp_6D)  # Uncomment to see the *whole* generated code
print_kernel(op1_omp_6D)

void bf0(struct dataobj *restrict f_vec, const float h_y, struct dataobj *restrict u_vec, const int t0, const int t1, const int x0_blk0_size, const int x0_blk1_size, const int x_M, const int x_m, const int y0_blk0_size, const int y0_blk1_size, const int y_M, const int y_m, const int z0_blk0_size, const int z0_blk1_size, const int z_M, const int z_m, const int nthreads)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;

  if (x0_blk0_size == 0 || y0_blk0_size == 0 || z0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int y0_blk0 = y_m; y0_blk0 <= y_M; y0_blk0 += y0_blk0_size)
      {
        for (int z0_blk0 = z_m; z0_blk0 <= z_M; z0_blk0 += z0_blk0_size)
        {
          for (int x0_blk1 = x0_blk0; x0_blk1 <= x0_blk0 + x0_blk0_size - 1; x0_blk1 += x0_blk1_size)
          {
            for (int y0_blk1 = y0_blk0; y0_blk1 <= y0_blk0 + y0_blk0_size - 1; y0_blk1 += y0_blk1_size)
            {
              for (int z0_blk1 = z0_blk0; z0_blk1 <= z0_blk0 + z0_blk0_size - 1; z0_blk1 += z0_blk1_size)
              {
                for (int x = x0_blk1; x <= x0_blk1 + x0_blk1_size - 1; x += 1)
                {
                  for (int y = y0_blk1; y <= y0_blk1 + y0_blk1_size - 1; y += 1)
                  {
                    for (int z = z0_blk1; z <= z0_blk1 + z0_blk1_size - 1; z += 1)
                    {
                      u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
                    }
                  }
                }
              }
            }
          }
        }
      }
    }
  }
}

SIMD vectorization¶

Devito enforces SIMD vectorization through OpenMP pragmas.

In [8]:

op2_omp = Operator(eq, opt=('blocking', 'simd', {'openmp': True}))
# print(op2_omp)  # Uncomment to see the generated code
# print(unidiff_output(str(op1_omp), str(op2_omp)))  # Uncomment to print out the diff only

Code motion¶

The advanced mode has a code motion pass. In explicit PDE solvers, this is most commonly used to lift expensive time-invariant sub-expressions out of the inner loops. The pass is however quite general in that it is not restricted to the concept of time-invariance -- any sub-expression invariant with respect to a subset of Dimensions is a code motion candidate. In our running example, sin(f) gets hoisted out of the inner loops since it is determined to be an expensive invariant sub-expression. In other words, the compiler trades the redundant computation of sin(f) for additional storage (the r1[...] array).

In [9]:

op3_omp = Operator(eq, opt=('lift', {'openmp': True}))
print_kernel(op3_omp)

float (*r1)[y_size][z_size];
posix_memalign((void**)&r1, 64, sizeof(float[x_size][y_size][z_size]));

/* Begin section0 */
START_TIMER(section0)
#pragma omp parallel num_threads(nthreads)
{
  #pragma omp for collapse(3) schedule(static,1)
  for (int x = x_m; x <= x_M; x += 1)
  {
    for (int y = y_m; y <= y_M; y += 1)
    {
      for (int z = z_m; z <= z_M; z += 1)
      {
        r1[x][y][z] = sin(f[x + 1][y + 1][z + 1]);
      }
    }
  }
}
STOP_TIMER(section0,timers)
/* End section0 */
for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section1 */
  START_TIMER(section1)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*pow(f[x + 1][y + 1][z + 1], 2)*r1[x][y][z];
        }
      }
    }
  }
  STOP_TIMER(section1,timers)
  /* End section1 */
}

free(r1);

Basic flop-reducing transformations¶

Among the simpler flop-reducing transformations applied by the advanced mode we find:

"classic" common sub-expressions elimination (CSE),
factorization,
optimization of powers

The cell below shows how the computation of u changes by incrementally applying these three passes. First of all, we observe how the symbolic spacing h_y gets assigned to a temporary, r0, as it appears in several sub-expressions. This is the effect of CSE.

In [10]:

op4_omp = Operator(eq, opt=('cse', {'openmp': True}))
print(unidiff_output(str(op0_omp), str(op4_omp)))

--- 
+++ 
@@ -42,7 +42,8 @@
         {
           for (int z = z_m; z <= z_M; z += 1)
           {
-            u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
+            float r0 = 1.0/h_y;
+            u[t1][x + 2][y + 2][z + 2] = (-r0*(-r0*u[t0][x + 2][y + 2][z + 2] + r0*u[t0][x + 2][y + 3][z + 2]) + r0*(-r0*u[t0][x + 2][y + 3][z + 2] + r0*u[t0][x + 2][y + 4][z + 2]))*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
           }
         }
       }

The factorization pass makes sure r0 is collected to reduce the number of multiplications. In larger, more complex examples, the impact of factorization can be quite significant.

In [11]:

op5_omp = Operator(eq, opt=('cse', 'factorize', {'openmp': True}))
print(unidiff_output(str(op4_omp), str(op5_omp)))

--- 
+++ 
@@ -43,7 +43,7 @@
           for (int z = z_m; z <= z_M; z += 1)
           {
             float r0 = 1.0/h_y;
-            u[t1][x + 2][y + 2][z + 2] = (-r0*(-r0*u[t0][x + 2][y + 2][z + 2] + r0*u[t0][x + 2][y + 3][z + 2]) + r0*(-r0*u[t0][x + 2][y + 3][z + 2] + r0*u[t0][x + 2][y + 4][z + 2]))*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
+            u[t1][x + 2][y + 2][z + 2] = (r0*(-r0*(-u[t0][x + 2][y + 2][z + 2]) - r0*u[t0][x + 2][y + 3][z + 2]) + r0*(r0*(-u[t0][x + 2][y + 3][z + 2]) + r0*u[t0][x + 2][y + 4][z + 2]))*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
           }
         }
       }

Finally, opt-pows turns costly pow calls into multiplications.

In [12]:

op6_omp = Operator(eq, opt=('cse', 'factorize', 'opt-pows', {'openmp': True}))
print(unidiff_output(str(op5_omp), str(op6_omp)))

--- 
+++ 
@@ -43,7 +43,7 @@
           for (int z = z_m; z <= z_M; z += 1)
           {
             float r0 = 1.0/h_y;
-            u[t1][x + 2][y + 2][z + 2] = (r0*(-r0*(-u[t0][x + 2][y + 2][z + 2]) - r0*u[t0][x + 2][y + 3][z + 2]) + r0*(r0*(-u[t0][x + 2][y + 3][z + 2]) + r0*u[t0][x + 2][y + 4][z + 2]))*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
+            u[t1][x + 2][y + 2][z + 2] = (f[x + 1][y + 1][z + 1]*f[x + 1][y + 1][z + 1])*(r0*(-r0*(-u[t0][x + 2][y + 2][z + 2]) - r0*u[t0][x + 2][y + 3][z + 2]) + r0*(r0*(-u[t0][x + 2][y + 3][z + 2]) + r0*u[t0][x + 2][y + 4][z + 2]))*sin(f[x + 1][y + 1][z + 1]);
           }
         }
       }

Cross-iteration redundancy elimination (CIRE)¶

This is perhaps the most advanced among the optimization passes in Devito. CIRE [1] searches for redundancies across consecutive loop iterations. These are often induced by a mixture of nested, high-order, partial derivatives. The underlying idea is very simple. Consider:

r0 = a[i-1] + a[i] + a[i+1]

at i=1, we have

r0 = a[0] + a[1] + a[2]

at i=2, we have

r0 = a[1] + a[2] + a[3]

So the sub-expression a[1] + a[2] is computed twice, by two consecutive iterations. What makes CIRE complicated is the generalization to arbitrary expressions, the presence of multiple dimensions, the scheduling strategy due to the trade-off between redundant compute and working set, the co-existance with other optimizations (e.g., blocking, vectorization), and much more. All these aspects won't be treated here. What instead we will show is the effect of CIRE in our running example and the optimization options at our disposal to drive the detection and scheduling of the captured redundancies.

In our running example, some cross-iteration redundancies are induced by the nested first-order derivatives along y. As we see below, these redundancies are captured and assigned to the two-dimensional temporary r3.

Note: the name cire-sops means "Apply CIRE to sum-of-product expressions". A sum-of-product is what taking a derivative via finite differences produces.

In [13]:

op7_omp = Operator(eq, opt=('cire-sops', {'openmp': True}))
print_kernel(op7_omp)
# print(unidiff_output(str(op7_omp), str(op0_omp)))  # Uncomment to print out the diff only

float **r7;
posix_memalign((void**)&r7, 64, sizeof(float*)*nthreads);
#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  posix_memalign((void**)&r7[tid], 64, sizeof(float[y_size + 1][z_size]));
}

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r5)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r7[tid];

    #pragma omp for collapse(1) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m - 1; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          r5[y + 1][z] = -u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y;
        }
      }
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-r5[y][z]/h_y + r5[y + 1][z]/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  free(r7[tid]);
}
free(r7);

We note that since there are no redundancies along x, the compiler is smart to figure out that r5 and u can safely be computed within the same x loop. This is nice -- not only is the reuse distance decreased, but also a grid-sized temporary avoided.

The `min-storage` option¶

Let's now consider a variant of our running example

In [14]:

eq = Eq(u.forward, f**2*sin(f)*(u.dy.dy + u.dx.dx))
op7_b0_omp = Operator(eq, opt=('cire-sops', {'openmp': True}))
print_kernel(op7_b0_omp)

float (*r7)[y_size + 1][z_size];
posix_memalign((void**)&r7, 64, sizeof(float[x_size + 1][y_size + 1][z_size]));
float (*r8)[y_size + 1][z_size];
posix_memalign((void**)&r8, 64, sizeof(float[x_size + 1][y_size + 1][z_size]));

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x = x_m - 1; x <= x_M; x += 1)
    {
      for (int y = y_m - 1; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          r7[x + 1][y + 1][z] = -u[t0][x + 3][y + 2][z + 2]/h_x + u[t0][x + 4][y + 2][z + 2]/h_x;
          r8[x + 1][y + 1][z] = -u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y;
        }
      }
    }
  }
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-r8[x + 1][y][z]/h_y + r8[x + 1][y + 1][z]/h_y - r7[x][y + 1][z]/h_x + r7[x + 1][y + 1][z]/h_x)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

free(r7);
free(r8);

As expected, there are now two temporaries, one stemming from u.dy.dy and the other from u.dx.dx. A key difference with respect to op7_omp here is that both are grid-size temporaries. This might seem odd at first -- why should the u.dy.dy temporary, that is r8, now be a three-dimensional temporary when we know already it could be a two-dimensional temporary? This is merely a compiler heuristic: by adding an extra dimension to r8, both temporaries can be scheduled within the same loop nest, thus augmenting data reuse and potentially enabling further cross-expression optimizations. We can disable this heuristic through the min-storage option.

In [15]:

op7_b1_omp = Operator(eq, opt=('cire-sops', {'openmp': True, 'min-storage': True}))
print_kernel(op7_b1_omp)

float (*r8)[y_size][z_size];
posix_memalign((void**)&r8, 64, sizeof(float[x_size + 1][y_size][z_size]));
float **r10;
posix_memalign((void**)&r10, 64, sizeof(float*)*nthreads);
#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  posix_memalign((void**)&r10[tid], 64, sizeof(float[y_size + 1][z_size]));
}

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(static,1)
    for (int x = x_m - 1; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          r8[x + 1][y][z] = -u[t0][x + 3][y + 2][z + 2]/h_x + u[t0][x + 4][y + 2][z + 2]/h_x;
        }
      }
    }
  }
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r7)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r10[tid];

    #pragma omp for collapse(1) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m - 1; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          r7[y + 1][z] = -u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y;
        }
      }
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-r7[y][z]/h_y + r7[y + 1][z]/h_y - r8[x][y][z]/h_x + r8[x + 1][y][z]/h_x)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  free(r10[tid]);
}
free(r8);
free(r10);

The `cire-mincost-sops` option¶

So far so good -- we've seen that Devito can capture and schedule cross-iteration redundancies. But what if we actually do not want certain redundancies to be captured? There are a few reasons we may like that option, for example we're allocating too much extra memory for the tensor temporaries, and we rather prefer to avoid that. For this, we can tell Devito what the minimum cost of a sub-expression should be in order to be a CIRE candidate. The cost is an integer number based on the operation count and the type of operations:

A basic arithmetic operation such as + and * has a cost of 1.
A / has a cost of 25.
A power with non-integer exponent has a cost of 50.
A power with non-negative integer exponent n has a cost of n-1 (i.e., the number of * it will be converted into).
A trascendental function (sin, cos, etc.) has a cost of 50.

The cire-mincost-sops option can be used to control the minimum cost of CIRE candidates.

In [16]:

eq = Eq(u.forward, f**2*sin(f)*u.dy.dy)  # Back to original running example
op8_omp = Operator(eq, opt=('cire-sops', {'openmp': True, 'cire-mincost-sops': 107}))
print_kernel(op8_omp)

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

We observe how setting cire-min-cost=107 makes the tensor temporary produced by op7_omp disappear. 106 is indeed the cost of the targeted sub-expression

-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y

One + => 1
Two / by h_y => 50
One two-way * with operands (-1/h_y, u) => 1
One two-way * with operands (1/h_y, u) => 1
The sub-expression appears n=2 times with shifted indices (precisely what CIRE looks for), so total cost to be multiplied by n.

Note: the cost of integer arithmetic for array indexing is deliberately ignored.

Next, we try again with a smaller cire-mincost-sops.

In [17]:

op8_b1_omp = Operator(eq, opt=('cire-sops', {'openmp': True, 'cire-mincost-sops': 11}))
print_kernel(op8_b1_omp)

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

Hang on! This doesn't make any sense, does it? We just decreased cire-mincost-sops to 11 and we still get the same code as op8_omp. Well, yeah... we lied about one tiny yet important detail for ease of explanation. It's true the cost of division is generally 25, but there's one special case: when the divisor is made up of invariant values (i.e., read-only constants) the cost of division becomes 1. In this case we divide by h_y, that is the grid spacing along the y dimension, which the compiler knows to be read-only, so the actual cost of the expression is 10. This tweak reflects the fact that invariant sub-expressions are ultimately lifted out of inner loops.

The `cire-repeats-sops` option¶

Higher order derivatives -- typically higher than order two -- may lead to proportionately larger sum-of-products. Consider the following variant of the running example

In [18]:

eq = Eq(u.forward, f**2*sin(f)*u.dy.dy.dy.dy.dy)
op9_omp = Operator(eq, opt=('cire-sops', {'openmp': True}))
print_kernel(op9_omp)

float **r5;
posix_memalign((void**)&r5, 64, sizeof(float*)*nthreads);
#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  posix_memalign((void**)&r5[tid], 64, sizeof(float[y_size + 1 + 1][z_size]));
}

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r3)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r5[tid];

    #pragma omp for collapse(1) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m - 1; y <= y_M + 1; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          r3[y + 1][z] = -(-(-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 4][z + 2]/h_y + u[t0][x + 2][y + 5][z + 2]/h_y)/h_y)/h_y + (-(-u[t0][x + 2][y + 4][z + 2]/h_y + u[t0][x + 2][y + 5][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 5][z + 2]/h_y + u[t0][x + 2][y + 6][z + 2]/h_y)/h_y)/h_y;
        }
      }
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-(-r3[y][z]/h_y + r3[y + 1][z]/h_y)/h_y + (-r3[y + 1][z]/h_y + r3[y + 2][z]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  free(r5[tid]);
}
free(r5);

At first glance, it's the same story as before, with a two-dimensional tensor temporary extracted and computed in a separate loop nest. By paying closer attention, however, we may note that there are now four occurrences of r3 in the expression computing u, while before, in op7_omp for instance, we had only two. Devito seeks for a compromise between compilation speed and run-time performance, so it runs by default searching for redundancies across sum-of-products of depth=7 or lower. depth=7 sum-of-products are typically the by-product of second-order derivatives, which is what we find in many PDEs. There are other reasons behind such default value, which for brevity aren't discussed here. To search for deeper sum-of-products (and, therefore, for redundancies induced by higher order derivatives), one shall use the cire-repeats-sops option.

In [19]:

op10_omp = Operator(eq, opt=('cire-sops', {'openmp': True, 'cire-repeats-sops': 9}))
print_kernel(op10_omp)

float **r4;
posix_memalign((void**)&r4, 64, sizeof(float*)*nthreads);
#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  posix_memalign((void**)&r4[tid], 64, sizeof(float[y_size + 1][z_size]));
}

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r2)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r4[tid];

    #pragma omp for collapse(1) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m - 1; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          r2[y + 1][z] = -(-(-(-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 4][z + 2]/h_y + u[t0][x + 2][y + 5][z + 2]/h_y)/h_y)/h_y + (-(-u[t0][x + 2][y + 4][z + 2]/h_y + u[t0][x + 2][y + 5][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 5][z + 2]/h_y + u[t0][x + 2][y + 6][z + 2]/h_y)/h_y)/h_y)/h_y + (-(-(-u[t0][x + 2][y + 4][z + 2]/h_y + u[t0][x + 2][y + 5][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 5][z + 2]/h_y + u[t0][x + 2][y + 6][z + 2]/h_y)/h_y)/h_y + (-(-u[t0][x + 2][y + 5][z + 2]/h_y + u[t0][x + 2][y + 6][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 6][z + 2]/h_y + u[t0][x + 2][y + 7][z + 2]/h_y)/h_y)/h_y)/h_y;
        }
      }
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-r2[y][z]/h_y + r2[y + 1][z]/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  free(r4[tid]);
}
free(r4);

The `cire-mincost-inv` and `cire-repeats-inv` options¶

Akin to sum-of-products, cross-iteration redundancies may be searched across dimension-invariant sub-expressions, typically time-invariants. So, analogously to what we've seen before:

In [20]:

eq = Eq(u.forward, f**2*sin(f)*u.dy.dy)  # Back to original running example
op11_omp = Operator(eq, opt=('lift', {'openmp': True}))
# print_kernel(op11_omp)  # Uncomment to see the generated code

For convenience, the lift pass triggers CIRE for dimension-invariant sub-expressions. As seen before, this leads to the r1 temporary. By setting a larger value for cire-mincost-inv, we avoid a grid-size temporary to be allocated, in exchange for a trascendental function, sin, to be computed at each iteration

In [21]:

op12_omp = Operator(eq, opt=('lift', {'openmp': True, 'cire-mincost-inv': 51}))
print_kernel(op12_omp)

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(dynamic,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-(-u[t0][x + 2][y + 2][z + 2]/h_y + u[t0][x + 2][y + 3][z + 2]/h_y)/h_y + (-u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y)/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

The `cire-maxpar` option¶

Sometimes it's possible to trade storage for parallelism (i.e., for more parallel dimensions). For this, Devito provides the cire-maxpar option which is by default set to:

False on CPU backends
True on GPU backends

Let's see what happens when we switch it on

In [22]:

op13_omp = Operator(eq, opt=('cire-sops', {'openmp': True, 'cire-maxpar': True}))
print_kernel(op13_omp)

float (*r5)[y_size + 1][z_size];
posix_memalign((void**)&r5, 64, sizeof(float[x_size][y_size + 1][z_size]));

for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
{
  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(static,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m - 1; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          r5[x][y + 1][z] = -u[t0][x + 2][y + 3][z + 2]/h_y + u[t0][x + 2][y + 4][z + 2]/h_y;
        }
      }
    }
  }
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(3) schedule(static,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        for (int z = z_m; z <= z_M; z += 1)
        {
          u[t1][x + 2][y + 2][z + 2] = (-r5[x][y][z]/h_y + r5[x][y + 1][z]/h_y)*sin(f[x + 1][y + 1][z + 1])*pow(f[x + 1][y + 1][z + 1], 2);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
}

free(r5);

The generated code uses a three-dimensional temporary that gets written and subsequently read in two separate x-y-z loop nests. Now, both loops can safely be openmp-collapsed, which is vital when running on GPUs.

Impact of CIRE in the `advanced` mode¶

The advanced mode triggers all of the passes we've seen so far... and in fact, many more! Some of them, however, aren't visible in our running example (e.g., all of the MPI-related optimizations). These will be treated in a future notebook.

Obviously, all of the optimization options (e.g., min-cost, cire-mincost-sops, blocklevels) are applicable and composable in any arbitrary way.

In [23]:

op14_omp = Operator(eq, openmp=True)
print(op14_omp)
# op14_b0_omp = Operator(eq, opt=('advanced', {'min-storage': True}))

#define _POSIX_C_SOURCE 200809L
#define START_TIMER(S) struct timeval start_ ## S , end_ ## S ; gettimeofday(&start_ ## S , NULL);
#define STOP_TIMER(S,T) gettimeofday(&end_ ## S, NULL); T->S += (double)(end_ ## S .tv_sec-start_ ## S.tv_sec)+(double)(end_ ## S .tv_usec-start_ ## S .tv_usec)/1000000;

#include "stdlib.h"
#include "math.h"
#include "sys/time.h"
#include "xmmintrin.h"
#include "pmmintrin.h"
#include "omp.h"

struct dataobj
{
  void *restrict data;
  int * size;
  int * npsize;
  int * dsize;
  int * hsize;
  int * hofs;
  int * oofs;
} ;

struct profiler
{
  double section0;
  double section1;
} ;

void bf0(struct dataobj *restrict f_vec, const float h_y, float *restrict r1_vec, struct dataobj *restrict u_vec, const int x_size, const int y0_blk0_size, const int y_size, const int z_size, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, float * *restrict r9_vec);

int Kernel(struct dataobj *restrict f_vec, const float h_y, struct dataobj *restrict u_vec, const int x_size, const int y0_blk0_size, const int y_size, const int z_size, const int time_M, const int time_m, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, struct profiler * timers)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;

  float (*r1)[y_size][z_size];
  posix_memalign((void**)&r1, 64, sizeof(float[x_size][y_size][z_size]));
  float **r9;
  posix_memalign((void**)&r9, 64, sizeof(float*)*nthreads);
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    posix_memalign((void**)&r9[tid], 64, sizeof(float[y0_blk0_size + 1][z_size]));
  }

  /* Flush denormal numbers to zero in hardware */
  _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
  _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);

  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(2) schedule(static,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        #pragma omp simd aligned(f:32)
        for (int z = z_m; z <= z_M; z += 1)
        {
          r1[x][y][z] = sin(f[x + 1][y + 1][z + 1]);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
  for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
  {
    /* Begin section1 */
    START_TIMER(section1)
    bf0(f_vec,h_y,(float *)r1,u_vec,x_size,y0_blk0_size,y_size,z_size,t0,t1,x0_blk0_size,x_M - (x_M - x_m + 1)%(x0_blk0_size),x_m,y_M - (y_M - y_m + 1)%(y0_blk0_size),y_m,z_M,z_m,nthreads,(float * *)r9);
    bf0(f_vec,h_y,(float *)r1,u_vec,x_size,(y_M - y_m + 1)%(y0_blk0_size),y_size,z_size,t0,t1,x0_blk0_size,x_M - (x_M - x_m + 1)%(x0_blk0_size),x_m,y_M,y_M - (y_M - y_m + 1)%(y0_blk0_size) + 1,z_M,z_m,nthreads,(float * *)r9);
    bf0(f_vec,h_y,(float *)r1,u_vec,x_size,y0_blk0_size,y_size,z_size,t0,t1,(x_M - x_m + 1)%(x0_blk0_size),x_M,x_M - (x_M - x_m + 1)%(x0_blk0_size) + 1,y_M - (y_M - y_m + 1)%(y0_blk0_size),y_m,z_M,z_m,nthreads,(float * *)r9);
    bf0(f_vec,h_y,(float *)r1,u_vec,x_size,(y_M - y_m + 1)%(y0_blk0_size),y_size,z_size,t0,t1,(x_M - x_m + 1)%(x0_blk0_size),x_M,x_M - (x_M - x_m + 1)%(x0_blk0_size) + 1,y_M,y_M - (y_M - y_m + 1)%(y0_blk0_size) + 1,z_M,z_m,nthreads,(float * *)r9);
    STOP_TIMER(section1,timers)
    /* End section1 */
  }

  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    free(r9[tid]);
  }
  free(r1);
  free(r9);

  return 0;
}

void bf0(struct dataobj *restrict f_vec, const float h_y, float *restrict r1_vec, struct dataobj *restrict u_vec, const int x_size, const int y0_blk0_size, const int y_size, const int z_size, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, float * *restrict r9_vec)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict r1)[y_size][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size][z_size]) r1_vec;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;
  float **r9 = (float**) r9_vec;

  if (x0_blk0_size == 0 || y0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r7)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r9[tid];

    #pragma omp for collapse(2) schedule(static,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int y0_blk0 = y_m; y0_blk0 <= y_M; y0_blk0 += y0_blk0_size)
      {
        for (int x = x0_blk0; x <= x0_blk0 + x0_blk0_size - 1; x += 1)
        {
          for (int y = y0_blk0 - 1, ys = 0; y <= y0_blk0 + y0_blk0_size - 1; y += 1, ys += 1)
          {
            #pragma omp simd aligned(u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              r7[ys][z] = (-u[t0][x + 2][y + 3][z + 2] + u[t0][x + 2][y + 4][z + 2])/h_y;
            }
          }
          for (int y = y0_blk0, ys = 0; y <= y0_blk0 + y0_blk0_size - 1; y += 1, ys += 1)
          {
            #pragma omp simd aligned(f,u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              u[t1][x + 2][y + 2][z + 2] = (f[x + 1][y + 1][z + 1]*f[x + 1][y + 1][z + 1])*(-r7[ys][z] + r7[ys + 1][z])*r1[x][y][z]/h_y;
            }
          }
        }
      }
    }
  }
}

A crucial observation here is that CIRE is applied on top of loop blocking -- the r7 temporary is computed within the same block as u, which in turn requires an additional iteration at the block edge along y to be performed (the first y loop starts at y0_blk0 - 1, while the second one at y0_blk0). Further, the size of r7 is now a function of the block shape. Altogether, this implements what in the literature is often referred to as "overlapped tiling" (or "overlapped blocking"): data reuse across consecutive loop nests is obtained by cross-loop blocking, which in turn requires a certain degree of redundant computation at the block edge. Clearly, there's a tension between the block shape and the amount of redundant computation. For example, a small block shape guarantees a small(er) working set, and thus better data reuse, but also requires more redundant computation.

The `cire-rotate` option¶

So far we've seen two ways to compute the tensor temporaries:

The temporary dimensions span the whole grid;
The temporary dimensions span a block created by the loop blocking pass.

There are a few other ways, and in particular there's a third way supported in Devito, enabled through the cire-rotate option:

The temporary outermost-dimension is a function of the stencil radius; all other temporary dimensions are a function of the loop blocking shape.

Let's jump straight into an example

In [24]:

op15_omp = Operator(eq, opt=('advanced', {'openmp': True, 'cire-rotate': True}))
print_kernel(op15_omp)

void bf0(struct dataobj *restrict f_vec, const float h_y, float *restrict r1_vec, struct dataobj *restrict u_vec, const int x_size, const int y_size, const int z_size, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y0_blk0_size, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, float * *restrict r9_vec)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict r1)[y_size][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size][z_size]) r1_vec;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;
  float **r9 = (float**) r9_vec;

  if (x0_blk0_size == 0 || y0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r7)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r9[tid];

    #pragma omp for collapse(2) schedule(static,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int y0_blk0 = y_m; y0_blk0 <= y_M; y0_blk0 += y0_blk0_size)
      {
        for (int x = x0_blk0; x <= x0_blk0 + x0_blk0_size - 1; x += 1)
        {
          for (int y = y0_blk0, ys = 0, yr0 = (ys + 1)%(2), yr1 = (ys)%(2), yii = -1; y <= y0_blk0 + y0_blk0_size - 1; y += 1, ys += 1, yr0 = (ys + 1)%(2), yr1 = (ys)%(2), yii = 0)
          {
            for (int yy = yii, ysi = (yy + ys + 1)%(2); yy <= 0; yy += 1, ysi = (yy + ys + 1)%(2))
            {
              #pragma omp simd aligned(u:32)
              for (int z = z_m; z <= z_M; z += 1)
              {
                r7[ysi][z] = (-u[t0][x + 2][yy + y + 3][z + 2] + u[t0][x + 2][yy + y + 4][z + 2])/h_y;
              }
            }
            #pragma omp simd aligned(f,u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              u[t1][x + 2][y + 2][z + 2] = (f[x + 1][y + 1][z + 1]*f[x + 1][y + 1][z + 1])*(r7[yr0][z] - r7[yr1][z])*r1[x][y][z]/h_y;
            }
          }
        }
      }
    }
  }
}

There are two key things to notice here:

The r7 temporary is a pointer to a two-dimensional array of size [2][z_size]. It's obtained via casting of r9[tid], which in turn is defined as

In [25]:

print(op15_omp.body[1].header[4])

#pragma omp parallel num_threads(nthreads)
{
  const int tid = omp_get_thread_num();
  posix_memalign((void**)&r9[tid], 64, sizeof(float[2][z_size]));
}

Within the y loop there are several iteration variables, some of which (ys0, ys1) employ modulo increment to cyclically produce the indices 0 and 1.

In essence, with cire-rotate, instead of computing an entire slice of y values, at each y iteration we only keep track of the values that are strictly necessary to evaluate u at y -- only two values in this case. This results in a working set reduction, at the price of turning one parallel loop (y) into a sequential one.

The `cire-maxalias` option¶

Let's consider the following variation of our running example, in which the outer y derivative now comprises several terms, other than just u.dy.

In [26]:

eq = Eq(u.forward, (2*f*f*u.dy).dy)
op16_omp = Operator(eq, opt=('advanced', {'openmp': True}))
print_kernel(op16_omp)

void bf0(struct dataobj *restrict f_vec, const float h_y, struct dataobj *restrict u_vec, const int y0_blk0_size, const int z_size, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, float * *restrict r4_vec)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;
  float **r4 = (float**) r4_vec;

  if (x0_blk0_size == 0 || y0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r2)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r4[tid];

    #pragma omp for collapse(2) schedule(dynamic,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int y0_blk0 = y_m; y0_blk0 <= y_M; y0_blk0 += y0_blk0_size)
      {
        for (int x = x0_blk0; x <= x0_blk0 + x0_blk0_size - 1; x += 1)
        {
          for (int y = y0_blk0 - 1, ys = 0; y <= y0_blk0 + y0_blk0_size - 1; y += 1, ys += 1)
          {
            #pragma omp simd aligned(u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              r2[ys][z] = (-u[t0][x + 2][y + 3][z + 2] + u[t0][x + 2][y + 4][z + 2])/h_y;
            }
          }
          for (int y = y0_blk0, ys = 0; y <= y0_blk0 + y0_blk0_size - 1; y += 1, ys += 1)
          {
            #pragma omp simd aligned(f,u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              float r3 = 1.0/h_y;
              u[t1][x + 2][y + 2][z + 2] = r3*(2*(f[x + 1][y + 2][z + 1]*f[x + 1][y + 2][z + 1])*r2[ys + 1][z]) + (-2*r3*r2[ys][z])*(f[x + 1][y + 1][z + 1]*f[x + 1][y + 1][z + 1]);
            }
          }
        }
      }
    }
  }
}

By paying close attention to the generated code, we see that the r2 temporary only stores the u.dy component, while the 2*f*f term is left intact in the loop nest computing u. This is due to a heuristic applied by the Devito compiler: in a derivative, only the sub-expressions representing additions -- and therefore inner derivatives -- are used as CIRE candidates. The logic behind this heuristic is that of minimizing the number of required temporaries at the price of potentially leaving some redundancies on the table. One can disable this heuristic via the cire-maxalias option.

In [27]:

op16_b0_omp = Operator(eq, opt=('advanced', {'openmp': True, 'cire-maxalias': True}))
print_kernel(op16_b0_omp)

void bf0(struct dataobj *restrict f_vec, const float h_y, struct dataobj *restrict u_vec, const int y0_blk0_size, const int z_size, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, float * *restrict r3_vec)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;
  float **r3 = (float**) r3_vec;

  if (x0_blk0_size == 0 || y0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r2)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r3[tid];

    #pragma omp for collapse(2) schedule(static,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int y0_blk0 = y_m; y0_blk0 <= y_M; y0_blk0 += y0_blk0_size)
      {
        for (int x = x0_blk0; x <= x0_blk0 + x0_blk0_size - 1; x += 1)
        {
          for (int y = y0_blk0 - 1, ys = 0; y <= y0_blk0 + y0_blk0_size - 1; y += 1, ys += 1)
          {
            #pragma omp simd aligned(f,u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              r2[ys][z] = (f[x + 1][y + 2][z + 1]*f[x + 1][y + 2][z + 1])*(-u[t0][x + 2][y + 3][z + 2] + u[t0][x + 2][y + 4][z + 2])/((h_y*h_y));
            }
          }
          for (int y = y0_blk0, ys = 0; y <= y0_blk0 + y0_blk0_size - 1; y += 1, ys += 1)
          {
            #pragma omp simd aligned(u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              u[t1][x + 2][y + 2][z + 2] = 2*(-r2[ys][z] + r2[ys + 1][z]);
            }
          }
        }
      }
    }
  }
}

Now we have a "fatter" temporary, and that's good -- but if, instead, we had had an Operator such as the one below, the gain from a reduced operation count might be outweighed by the presence of more temporaries, which means larger working set and increased memory traffic.

In [28]:

eq = Eq(u.forward, (2*f*f*u.dy).dy + (3*f*u.dy).dy)
op16_b1_omp = Operator(eq, opt=('advanced', {'openmp': True}))
op16_b2_omp = Operator(eq, opt=('advanced', {'openmp': True, 'cire-maxalias': True}))
# print_kernel(op16_b1_omp)  # Uncomment to see generated code with one temporary but more flops
# print_kernel(op16_b2_omp)  # Uncomment to see generated code with two temporaries but fewer flops

The `advanced-fsg` mode¶

The alternative advanced-fsg optimization level applies the same passes as advanced, but in a different order. The key difference is that -fsg does not generate overlapped blocking code across CIRE-generated loop nests.

In [29]:

eq = Eq(u.forward, f**2*sin(f)*u.dy.dy)  # Back to original running example
op17_omp = Operator(eq, opt=('advanced-fsg', {'openmp': True}))
print(op17_omp)

#define _POSIX_C_SOURCE 200809L
#define START_TIMER(S) struct timeval start_ ## S , end_ ## S ; gettimeofday(&start_ ## S , NULL);
#define STOP_TIMER(S,T) gettimeofday(&end_ ## S, NULL); T->S += (double)(end_ ## S .tv_sec-start_ ## S.tv_sec)+(double)(end_ ## S .tv_usec-start_ ## S .tv_usec)/1000000;

#include "stdlib.h"
#include "math.h"
#include "sys/time.h"
#include "xmmintrin.h"
#include "pmmintrin.h"
#include "omp.h"

struct dataobj
{
  void *restrict data;
  int * size;
  int * npsize;
  int * dsize;
  int * hsize;
  int * hofs;
  int * oofs;
} ;

struct profiler
{
  double section0;
  double section1;
} ;

void bf0(struct dataobj *restrict f_vec, const float h_y, float *restrict r1_vec, struct dataobj *restrict u_vec, const int x_size, const int y_size, const int z_size, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, float * *restrict r9_vec);

int Kernel(struct dataobj *restrict f_vec, const float h_y, struct dataobj *restrict u_vec, const int x_size, const int y_size, const int z_size, const int time_M, const int time_m, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, struct profiler * timers)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;

  float (*r1)[y_size][z_size];
  posix_memalign((void**)&r1, 64, sizeof(float[x_size][y_size][z_size]));
  float **r9;
  posix_memalign((void**)&r9, 64, sizeof(float*)*nthreads);
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    posix_memalign((void**)&r9[tid], 64, sizeof(float[y_size + 1][z_size]));
  }

  /* Flush denormal numbers to zero in hardware */
  _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
  _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);

  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(2) schedule(static,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        #pragma omp simd aligned(f:32)
        for (int z = z_m; z <= z_M; z += 1)
        {
          r1[x][y][z] = sin(f[x + 1][y + 1][z + 1]);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
  for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
  {
    /* Begin section1 */
    START_TIMER(section1)
    bf0(f_vec,h_y,(float *)r1,u_vec,x_size,y_size,z_size,t0,t1,x0_blk0_size,x_M - (x_M - x_m + 1)%(x0_blk0_size),x_m,y_M,y_m,z_M,z_m,nthreads,(float * *)r9);
    bf0(f_vec,h_y,(float *)r1,u_vec,x_size,y_size,z_size,t0,t1,(x_M - x_m + 1)%(x0_blk0_size),x_M,x_M - (x_M - x_m + 1)%(x0_blk0_size) + 1,y_M,y_m,z_M,z_m,nthreads,(float * *)r9);
    STOP_TIMER(section1,timers)
    /* End section1 */
  }

  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    free(r9[tid]);
  }
  free(r1);
  free(r9);

  return 0;
}

void bf0(struct dataobj *restrict f_vec, const float h_y, float *restrict r1_vec, struct dataobj *restrict u_vec, const int x_size, const int y_size, const int z_size, const int t0, const int t1, const int x0_blk0_size, const int x_M, const int x_m, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, float * *restrict r9_vec)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict r1)[y_size][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size][z_size]) r1_vec;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;
  float **r9 = (float**) r9_vec;

  if (x0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    const int tid = omp_get_thread_num();
    float (*restrict r7)[z_size] __attribute__ ((aligned (64))) = (float (*)[z_size]) r9[tid];

    #pragma omp for collapse(1) schedule(static,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int x = x0_blk0; x <= x0_blk0 + x0_blk0_size - 1; x += 1)
      {
        for (int y = y_m - 1; y <= y_M; y += 1)
        {
          #pragma omp simd aligned(u:32)
          for (int z = z_m; z <= z_M; z += 1)
          {
            r7[y + 1][z] = (-u[t0][x + 2][y + 3][z + 2] + u[t0][x + 2][y + 4][z + 2])/h_y;
          }
        }
        for (int y = y_m; y <= y_M; y += 1)
        {
          #pragma omp simd aligned(f,u:32)
          for (int z = z_m; z <= z_M; z += 1)
          {
            u[t1][x + 2][y + 2][z + 2] = (f[x + 1][y + 1][z + 1]*f[x + 1][y + 1][z + 1])*(-r7[y][z] + r7[y + 1][z])*r1[x][y][z]/h_y;
          }
        }
      }
    }
  }
}

The x loop here is still shared by the two loop nests, but the y one isn't. Analogously, if we consider the alternative eq already used in op7_b0_omp, we get two completely separate, and therefore individually blocked, loop nests.

In [30]:

eq = Eq(u.forward, f**2*sin(f)*(u.dy.dy + u.dx.dx))
op17_b0 = Operator(eq, opt=('advanced-fsg', {'openmp': True}))
print(op17_b0)

#define _POSIX_C_SOURCE 200809L
#define START_TIMER(S) struct timeval start_ ## S , end_ ## S ; gettimeofday(&start_ ## S , NULL);
#define STOP_TIMER(S,T) gettimeofday(&end_ ## S, NULL); T->S += (double)(end_ ## S .tv_sec-start_ ## S.tv_sec)+(double)(end_ ## S .tv_usec-start_ ## S .tv_usec)/1000000;

#include "stdlib.h"
#include "math.h"
#include "sys/time.h"
#include "xmmintrin.h"
#include "pmmintrin.h"
#include "omp.h"

struct dataobj
{
  void *restrict data;
  int * size;
  int * npsize;
  int * dsize;
  int * hsize;
  int * hofs;
  int * oofs;
} ;

struct profiler
{
  double section0;
  double section1;
} ;

void bf0(const float h_x, const float h_y, float *restrict r10_vec, float *restrict r9_vec, struct dataobj *restrict u_vec, const int t0, const int x0_blk0_size, const int x_M, const int x_m, const int x_size, const int y0_blk0_size, const int y_M, const int y_m, const int y_size, const int z_M, const int z_m, const int z_size, const int nthreads);
void bf1(struct dataobj *restrict f_vec, const float h_x, const float h_y, float *restrict r1_vec, float *restrict r10_vec, float *restrict r9_vec, struct dataobj *restrict u_vec, const int x_size, const int y_size, const int z_size, const int t1, const int x1_blk0_size, const int x_M, const int x_m, const int y1_blk0_size, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads);

int Kernel(struct dataobj *restrict f_vec, const float h_x, const float h_y, struct dataobj *restrict u_vec, const int x_size, const int y_size, const int z_size, const int time_M, const int time_m, const int x0_blk0_size, const int x1_blk0_size, const int x_M, const int x_m, const int y0_blk0_size, const int y1_blk0_size, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads, struct profiler * timers)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;

  float (*r1)[y_size][z_size];
  posix_memalign((void**)&r1, 64, sizeof(float[x_size][y_size][z_size]));
  float (*r10)[y_size + 1][z_size];
  posix_memalign((void**)&r10, 64, sizeof(float[x_size + 1][y_size + 1][z_size]));
  float (*r9)[y_size + 1][z_size];
  posix_memalign((void**)&r9, 64, sizeof(float[x_size + 1][y_size + 1][z_size]));

  /* Flush denormal numbers to zero in hardware */
  _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
  _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);

  /* Begin section0 */
  START_TIMER(section0)
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(2) schedule(static,1)
    for (int x = x_m; x <= x_M; x += 1)
    {
      for (int y = y_m; y <= y_M; y += 1)
      {
        #pragma omp simd aligned(f:32)
        for (int z = z_m; z <= z_M; z += 1)
        {
          r1[x][y][z] = sin(f[x + 1][y + 1][z + 1]);
        }
      }
    }
  }
  STOP_TIMER(section0,timers)
  /* End section0 */
  for (int time = time_m, t0 = (time)%(2), t1 = (time + 1)%(2); time <= time_M; time += 1, t0 = (time)%(2), t1 = (time + 1)%(2))
  {
    /* Begin section1 */
    START_TIMER(section1)
    bf0(h_x,h_y,(float *)r10,(float *)r9,u_vec,t0,x0_blk0_size,x_M - (x_M - x_m + 2)%(x0_blk0_size),x_m - 1,x_size,y0_blk0_size,y_M - (y_M - y_m + 2)%(y0_blk0_size),y_m - 1,y_size,z_M,z_m,z_size,nthreads);
    bf0(h_x,h_y,(float *)r10,(float *)r9,u_vec,t0,x0_blk0_size,x_M - (x_M - x_m + 2)%(x0_blk0_size),x_m - 1,x_size,(y_M - y_m + 2)%(y0_blk0_size),y_M,y_M - (y_M - y_m + 2)%(y0_blk0_size) + 1,y_size,z_M,z_m,z_size,nthreads);
    bf0(h_x,h_y,(float *)r10,(float *)r9,u_vec,t0,(x_M - x_m + 2)%(x0_blk0_size),x_M,x_M - (x_M - x_m + 2)%(x0_blk0_size) + 1,x_size,y0_blk0_size,y_M - (y_M - y_m + 2)%(y0_blk0_size),y_m - 1,y_size,z_M,z_m,z_size,nthreads);
    bf0(h_x,h_y,(float *)r10,(float *)r9,u_vec,t0,(x_M - x_m + 2)%(x0_blk0_size),x_M,x_M - (x_M - x_m + 2)%(x0_blk0_size) + 1,x_size,(y_M - y_m + 2)%(y0_blk0_size),y_M,y_M - (y_M - y_m + 2)%(y0_blk0_size) + 1,y_size,z_M,z_m,z_size,nthreads);
    bf1(f_vec,h_x,h_y,(float *)r1,(float *)r10,(float *)r9,u_vec,x_size,y_size,z_size,t1,x1_blk0_size,x_M - (x_M - x_m + 1)%(x1_blk0_size),x_m,y1_blk0_size,y_M - (y_M - y_m + 1)%(y1_blk0_size),y_m,z_M,z_m,nthreads);
    bf1(f_vec,h_x,h_y,(float *)r1,(float *)r10,(float *)r9,u_vec,x_size,y_size,z_size,t1,x1_blk0_size,x_M - (x_M - x_m + 1)%(x1_blk0_size),x_m,(y_M - y_m + 1)%(y1_blk0_size),y_M,y_M - (y_M - y_m + 1)%(y1_blk0_size) + 1,z_M,z_m,nthreads);
    bf1(f_vec,h_x,h_y,(float *)r1,(float *)r10,(float *)r9,u_vec,x_size,y_size,z_size,t1,(x_M - x_m + 1)%(x1_blk0_size),x_M,x_M - (x_M - x_m + 1)%(x1_blk0_size) + 1,y1_blk0_size,y_M - (y_M - y_m + 1)%(y1_blk0_size),y_m,z_M,z_m,nthreads);
    bf1(f_vec,h_x,h_y,(float *)r1,(float *)r10,(float *)r9,u_vec,x_size,y_size,z_size,t1,(x_M - x_m + 1)%(x1_blk0_size),x_M,x_M - (x_M - x_m + 1)%(x1_blk0_size) + 1,(y_M - y_m + 1)%(y1_blk0_size),y_M,y_M - (y_M - y_m + 1)%(y1_blk0_size) + 1,z_M,z_m,nthreads);
    STOP_TIMER(section1,timers)
    /* End section1 */
  }

  free(r1);
  free(r10);
  free(r9);

  return 0;
}

void bf0(const float h_x, const float h_y, float *restrict r10_vec, float *restrict r9_vec, struct dataobj *restrict u_vec, const int t0, const int x0_blk0_size, const int x_M, const int x_m, const int x_size, const int y0_blk0_size, const int y_M, const int y_m, const int y_size, const int z_M, const int z_m, const int z_size, const int nthreads)
{
  float (*restrict r10)[y_size + 1][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size + 1][z_size]) r10_vec;
  float (*restrict r9)[y_size + 1][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size + 1][z_size]) r9_vec;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;

  if (x0_blk0_size == 0 || y0_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(2) schedule(static,1)
    for (int x0_blk0 = x_m; x0_blk0 <= x_M; x0_blk0 += x0_blk0_size)
    {
      for (int y0_blk0 = y_m; y0_blk0 <= y_M; y0_blk0 += y0_blk0_size)
      {
        for (int x = x0_blk0; x <= x0_blk0 + x0_blk0_size - 1; x += 1)
        {
          for (int y = y0_blk0; y <= y0_blk0 + y0_blk0_size - 1; y += 1)
          {
            #pragma omp simd aligned(u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              r9[x + 1][y + 1][z] = (-u[t0][x + 3][y + 2][z + 2] + u[t0][x + 4][y + 2][z + 2])/h_x;
              r10[x + 1][y + 1][z] = (-u[t0][x + 2][y + 3][z + 2] + u[t0][x + 2][y + 4][z + 2])/h_y;
            }
          }
        }
      }
    }
  }
}

void bf1(struct dataobj *restrict f_vec, const float h_x, const float h_y, float *restrict r1_vec, float *restrict r10_vec, float *restrict r9_vec, struct dataobj *restrict u_vec, const int x_size, const int y_size, const int z_size, const int t1, const int x1_blk0_size, const int x_M, const int x_m, const int y1_blk0_size, const int y_M, const int y_m, const int z_M, const int z_m, const int nthreads)
{
  float (*restrict f)[f_vec->size[1]][f_vec->size[2]] __attribute__ ((aligned (64))) = (float (*)[f_vec->size[1]][f_vec->size[2]]) f_vec->data;
  float (*restrict r1)[y_size][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size][z_size]) r1_vec;
  float (*restrict r10)[y_size + 1][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size + 1][z_size]) r10_vec;
  float (*restrict r9)[y_size + 1][z_size] __attribute__ ((aligned (64))) = (float (*)[y_size + 1][z_size]) r9_vec;
  float (*restrict u)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]] __attribute__ ((aligned (64))) = (float (*)[u_vec->size[1]][u_vec->size[2]][u_vec->size[3]]) u_vec->data;

  if (x1_blk0_size == 0 || y1_blk0_size == 0)
  {
    return;
  }
  #pragma omp parallel num_threads(nthreads)
  {
    #pragma omp for collapse(2) schedule(dynamic,1)
    for (int x1_blk0 = x_m; x1_blk0 <= x_M; x1_blk0 += x1_blk0_size)
    {
      for (int y1_blk0 = y_m; y1_blk0 <= y_M; y1_blk0 += y1_blk0_size)
      {
        for (int x = x1_blk0; x <= x1_blk0 + x1_blk0_size - 1; x += 1)
        {
          for (int y = y1_blk0; y <= y1_blk0 + y1_blk0_size - 1; y += 1)
          {
            #pragma omp simd aligned(f,u:32)
            for (int z = z_m; z <= z_M; z += 1)
            {
              u[t1][x + 2][y + 2][z + 2] = (f[x + 1][y + 1][z + 1]*f[x + 1][y + 1][z + 1])*((-r10[x + 1][y][z] + r10[x + 1][y + 1][z])/h_y + (-r9[x][y + 1][z] + r9[x + 1][y + 1][z])/h_x)*r1[x][y][z];
            }
          }
        }
      }
    }
  }
}

References¶

[1] Fabio Luporini, Mathias Louboutin, Michael Lange, Navjot Kukreja, Philipp Witte, Jan Hückelheim, Charles Yount, Paul H. J. Kelly, Felix J. Herrmann, and Gerard J. Gorman. 2020. Architecture and Performance of Devito, a System for Automated Stencil Computation. ACM Trans. Math. Softw. 46, 1, Article 6 (April 2020), 28 pages. DOI:https://doi.org/10.1145/3374916

Performance optimization overview¶

Outline¶

API¶

Default values¶

Misc¶

Running example¶

OpenMP parallelism¶

The advanced mode¶

Loop blocking¶

SIMD vectorization¶

Code motion¶

Basic flop-reducing transformations¶

Cross-iteration redundancy elimination (CIRE)¶

The min-storage option¶

The cire-mincost-sops option¶

The cire-repeats-sops option¶

The cire-mincost-inv and cire-repeats-inv options¶

The cire-maxpar option¶

Impact of CIRE in the advanced mode¶

The cire-rotate option¶

The cire-maxalias option¶

The advanced-fsg mode¶