starpu-max/examples/starpufft/starpufftx1d.c

/* StarPU --- Runtime system for heterogeneous multicore architectures.
 *
 * Copyright (C) 2009-2011  Université de Bordeaux 1
 * Copyright (C) 2010  Centre National de la Recherche Scientifique
 *
 * StarPU is free software; you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published by
 * the Free Software Foundation; either version 2.1 of the License, or (at
 * your option) any later version.
 *
 * StarPU is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 *
 * See the GNU Lesser General Public License in COPYING.LGPL for more details.
 */

#define PARALLEL
#ifdef PARALLEL

/* Dumb parallel version, disabled */

#define DIV_1D 64

  /*
   * Overall strategy for an fft of size n:
   * - perform n1 ffts of size n2
   * - twiddle
   * - perform n2 ffts of size n1
   *
   * - n1 defaults to DIV_1D, thus n2 defaults to n / DIV_1D.
   *
   * Precise tasks:
   *
   * - twist1: twist the whole n-element input (called "in") into n1 chunks of
   *           size n2, by using n1 tasks taking the whole n-element input as a
   *           R parameter and one n2 output as a W parameter. The result is
   *           called twisted1.
   * - fft1:   perform n1 (n2) ffts, by using n1 tasks doing one fft each. Also
   *           twiddle the result to prepare for the fft2. The result is called
   *           fft1.
   * - join:   depends on all the fft1s, to gather the n1 results of size n2 in
   *           the fft1 vector.
   * - twist2: twist the fft1 vector into n2 chunks of size n1, called twisted2.
   *           since n2 is typically very large, this step is divided in DIV_1D
   *           tasks, each of them performing n2/DIV_1D of them
   * - fft2:   perform n2 ffts of size n1. This is divided in DIV_1D tasks of
   *           n2/DIV_1D ffts, to be performed in batches. The result is called
   *           fft2.
   * - twist3: twist back the result of the fft2s above into the output buffer.
   *           Only implemented on CPUs for simplicity of the gathering.
   *
   * The tag space thus uses 3 dimensions:
   * - the number of the plan.
   * - the step (TWIST1, FFT1, JOIN, TWIST2, FFT2, TWIST3, END)
   * - an index i between 0 and DIV_1D-1.
   */

#define STEP_TAG_1D(plan, step, i) _STEP_TAG(plan, step, i)

#ifdef STARPU_USE_CUDA
/* twist1:
 *
 * Twist the full input vector (first parameter) into one chunk of size n2
 * (second parameter) */
static void
STARPUFFT(twist1_1d_kernel_gpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	int i = args->i;
	int n1 = plan->n1[0];
	int n2 = plan->n2[0];

	_cufftComplex * restrict in = (_cufftComplex *)STARPU_VECTOR_GET_PTR(descr[0]);
	_cufftComplex * restrict twisted1 = (_cufftComplex *)STARPU_VECTOR_GET_PTR(descr[1]);
	
	STARPUFFT(cuda_twist1_1d_host)(in, twisted1, i, n1, n2);

	cudaStreamSynchronize(starpu_cuda_get_local_stream());
}

/* fft1:
 *
 * Perform one fft of size n2 */
static void
STARPUFFT(fft1_1d_plan_gpu)(void *args)
{
	STARPUFFT(plan) plan = args;
	cufftResult cures;
	int n2 = plan->n2[0];
	int workerid = starpu_worker_get_id();

	cures = cufftPlan1d(&plan->plans[workerid].plan1_cuda, n2, _CUFFT_C2C, 1);
	STARPU_ASSERT(cures == CUFFT_SUCCESS);
	cufftSetStream(plan->plans[workerid].plan1_cuda, starpu_cuda_get_local_stream());
	STARPU_ASSERT(cures == CUFFT_SUCCESS);
}

static void
STARPUFFT(fft1_1d_kernel_gpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	int i = args->i;
	int n2 = plan->n2[0];
	cufftResult cures;

	_cufftComplex * restrict in = (_cufftComplex *)STARPU_VECTOR_GET_PTR(descr[0]);
	_cufftComplex * restrict out = (_cufftComplex *)STARPU_VECTOR_GET_PTR(descr[1]);
	const _cufftComplex * restrict roots = (_cufftComplex *)STARPU_VECTOR_GET_PTR(descr[2]);

	int workerid = starpu_worker_get_id();

	task_per_worker[workerid]++;

	cures = _cufftExecC2C(plan->plans[workerid].plan1_cuda, in, out, plan->sign == -1 ? CUFFT_FORWARD : CUFFT_INVERSE);
	STARPU_ASSERT(cures == CUFFT_SUCCESS);

	STARPUFFT(cuda_twiddle_1d_host)(out, roots, n2, i);

	cudaStreamSynchronize(starpu_cuda_get_local_stream());
}

/* fft2:
 *
 * Perform n3 = n2/DIV_1D ffts of size n1 */
static void
STARPUFFT(fft2_1d_plan_gpu)(void *args)
{
	STARPUFFT(plan) plan = args;
	cufftResult cures;
	int n1 = plan->n1[0];
	int n2 = plan->n2[0];
	int n3 = n2/DIV_1D;
	int workerid = starpu_worker_get_id();

	cures = cufftPlan1d(&plan->plans[workerid].plan2_cuda, n1, _CUFFT_C2C, n3);
	STARPU_ASSERT(cures == CUFFT_SUCCESS);
	cufftSetStream(plan->plans[workerid].plan2_cuda, starpu_cuda_get_local_stream());
	STARPU_ASSERT(cures == CUFFT_SUCCESS);
}

static void
STARPUFFT(fft2_1d_kernel_gpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	cufftResult cures;

	_cufftComplex * restrict in = (_cufftComplex *)STARPU_VECTOR_GET_PTR(descr[0]);
	_cufftComplex * restrict out = (_cufftComplex *)STARPU_VECTOR_GET_PTR(descr[1]);

	int workerid = starpu_worker_get_id();

	task_per_worker[workerid]++;

	/* NOTE using batch support */
	cures = _cufftExecC2C(plan->plans[workerid].plan2_cuda, in, out, plan->sign == -1 ? CUFFT_FORWARD : CUFFT_INVERSE);
	STARPU_ASSERT(cures == CUFFT_SUCCESS);

	cudaStreamSynchronize(starpu_cuda_get_local_stream());
}
#endif

/* twist1:
 *
 * Twist the full input vector (first parameter) into one chunk of size n2
 * (second parameter) */
static void
STARPUFFT(twist1_1d_kernel_cpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	int i = args->i;
	int j;
	int n1 = plan->n1[0];
	int n2 = plan->n2[0];

	STARPUFFT(complex) * restrict in = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[0]);
	STARPUFFT(complex) * restrict twisted1 = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[1]);

	/* printf("twist1 %d %g\n", i, (double) cabs(plan->in[i])); */

	for (j = 0; j < n2; j++)
		twisted1[j] = in[i+j*n1];
}

#ifdef STARPU_HAVE_FFTW
/* fft1:
 *
 * Perform one fft of size n2 */
static void
STARPUFFT(fft1_1d_kernel_cpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	int i = args->i;
	int j;
	int n2 = plan->n2[0];
	int workerid = starpu_worker_get_id();

	task_per_worker[workerid]++;

	STARPUFFT(complex) * restrict twisted1 = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[0]);
	STARPUFFT(complex) * restrict fft1 = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[1]);

	/* printf("fft1 %d %g\n", i, (double) cabs(twisted1[0])); */

	_FFTW(execute_dft)(plan->plans[workerid].plan1_cpu, twisted1, fft1);

	/* twiddle fft1 buffer */
	for (j = 0; j < n2; j++)
		fft1[j] = fft1[j] * plan->roots[0][i*j];
}
#endif

/* twist2:
 *
 * Twist the full vector (results of the fft1s) into one package of n2/DIV_1D
 * chunks of size n1 */
static void
STARPUFFT(twist2_1d_kernel_cpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	int jj = args->jj;	/* between 0 and DIV_1D */
	int jjj;		/* beetween 0 and n3 */
	int i;
	int n1 = plan->n1[0];
	int n2 = plan->n2[0];
	int n3 = n2/DIV_1D;

	STARPUFFT(complex) * restrict twisted2 = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[0]);

	/* printf("twist2 %d %g\n", jj, (double) cabs(plan->fft1[jj])); */

	for (jjj = 0; jjj < n3; jjj++) {
		int j = jj * n3 + jjj;
		for (i = 0; i < n1; i++)
			twisted2[jjj*n1+i] = plan->fft1[i*n2+j];
	}
}

#ifdef STARPU_HAVE_FFTW
/* fft2:
 *
 * Perform n3 = n2/DIV_1D ffts of size n1 */
static void
STARPUFFT(fft2_1d_kernel_cpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	/* int jj = args->jj; */
	int workerid = starpu_worker_get_id();

	task_per_worker[workerid]++;

	STARPUFFT(complex) * restrict twisted2 = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[0]);
	STARPUFFT(complex) * restrict fft2 = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[1]);

	/* printf("fft2 %d %g\n", jj, (double) cabs(twisted2[plan->totsize4-1])); */

	_FFTW(execute_dft)(plan->plans[workerid].plan2_cpu, twisted2, fft2);
}
#endif

/* twist3:
 *
 * Spread the package of n2/DIV_1D chunks of size n1 into the output vector */
static void
STARPUFFT(twist3_1d_kernel_cpu)(void *descr[], void *_args)
{
	struct STARPUFFT(args) *args = _args;
	STARPUFFT(plan) plan = args->plan;
	int jj = args->jj;	/* between 0 and DIV_1D */
	int jjj;		/* beetween 0 and n3 */
	int i;
	int n1 = plan->n1[0];
	int n2 = plan->n2[0];
	int n3 = n2/DIV_1D;

	const STARPUFFT(complex) * restrict fft2 = (STARPUFFT(complex) *)STARPU_VECTOR_GET_PTR(descr[0]);

	/* printf("twist3 %d %g\n", jj, (double) cabs(fft2[0])); */

	for (jjj = 0; jjj < n3; jjj++) {
		int j = jj * n3 + jjj;
		for (i = 0; i < n1; i++)
			plan->out[i*n2+j] = fft2[jjj*n1+i];
	}
}

/* Performance models for the 5 kinds of tasks */
static struct starpu_perfmodel STARPUFFT(twist1_1d_model) = {
	.type = STARPU_HISTORY_BASED,
	.symbol = TYPE"twist1_1d"
};

static struct starpu_perfmodel STARPUFFT(fft1_1d_model) = {
	.type = STARPU_HISTORY_BASED,
	.symbol = TYPE"fft1_1d"
};

static struct starpu_perfmodel STARPUFFT(twist2_1d_model) = {
	.type = STARPU_HISTORY_BASED,
	.symbol = TYPE"twist2_1d"
};

static struct starpu_perfmodel STARPUFFT(fft2_1d_model) = {
	.type = STARPU_HISTORY_BASED,
	.symbol = TYPE"fft2_1d"
};

static struct starpu_perfmodel STARPUFFT(twist3_1d_model) = {
	.type = STARPU_HISTORY_BASED,
	.symbol = TYPE"twist3_1d"
};

/* codelet pointers for the 5 kinds of tasks */
static struct starpu_codelet STARPUFFT(twist1_1d_codelet) = {
	.where =
#ifdef STARPU_USE_CUDA
		STARPU_CUDA|
#endif
		STARPU_CPU,
#ifdef STARPU_USE_CUDA
	.cuda_func = STARPUFFT(twist1_1d_kernel_gpu),
#endif
	.cpu_func = STARPUFFT(twist1_1d_kernel_cpu),
	.model = &STARPUFFT(twist1_1d_model),
	.nbuffers = 2
};

static struct starpu_codelet STARPUFFT(fft1_1d_codelet) = {
	.where =
#ifdef STARPU_USE_CUDA
		STARPU_CUDA|
#endif
#ifdef STARPU_HAVE_FFTW
		STARPU_CPU|
#endif
		0,
#ifdef STARPU_USE_CUDA
	.cuda_func = STARPUFFT(fft1_1d_kernel_gpu),
#endif
#ifdef STARPU_HAVE_FFTW
	.cpu_func = STARPUFFT(fft1_1d_kernel_cpu),
#endif
	.model = &STARPUFFT(fft1_1d_model),
	.nbuffers = 3
};

static struct starpu_codelet STARPUFFT(twist2_1d_codelet) = {
	.where = STARPU_CPU,
	.cpu_func = STARPUFFT(twist2_1d_kernel_cpu),
	.model = &STARPUFFT(twist2_1d_model),
	.nbuffers = 1
};

static struct starpu_codelet STARPUFFT(fft2_1d_codelet) = {
	.where =
#ifdef STARPU_USE_CUDA
		STARPU_CUDA|
#endif
#ifdef STARPU_HAVE_FFTW
		STARPU_CPU|
#endif
		0,
#ifdef STARPU_USE_CUDA
	.cuda_func = STARPUFFT(fft2_1d_kernel_gpu),
#endif
#ifdef STARPU_HAVE_FFTW
	.cpu_func = STARPUFFT(fft2_1d_kernel_cpu),
#endif
	.model = &STARPUFFT(fft2_1d_model),
	.nbuffers = 2
};

static struct starpu_codelet STARPUFFT(twist3_1d_codelet) = {
	.where = STARPU_CPU,
	.cpu_func = STARPUFFT(twist3_1d_kernel_cpu),
	.model = &STARPUFFT(twist3_1d_model),
	.nbuffers = 1
};

#endif /* PARALLEL */

/* Planning:
 *
 * - For each CPU worker, we need to plan the two fftw stages.
 * - For GPU workers, we need to do the planning in the CUDA context, so we do
 *   this lazily through the initialised1 and initialised2 flags ; TODO: use
 *   starpu_execute_on_each_worker instead (done in the omp branch).
 * - We allocate all the temporary buffers and register them to starpu.
 * - We create all the tasks, but do not submit them yet. It will be possible
 *   to reuse them at will to perform several ffts with the same planning.
 */
STARPUFFT(plan)
STARPUFFT(plan_dft_1d)(int n, int sign, unsigned flags)
{
	int workerid;
	int n1 = DIV_1D;
	int n2 = n / n1;
	int n3;
	int z;
	struct starpu_task *task;

#ifdef STARPU_USE_CUDA
	/* cufft 1D limited to 8M elements */
	while (n2 > 8 << 20) {
		n1 *= 2;
		n2 /= 2;
	}
#endif
	STARPU_ASSERT(n == n1*n2);
	STARPU_ASSERT(n1 < (1ULL << I_BITS));

	/* distribute the n2 second ffts into DIV_1D packages */
	n3 = n2 / DIV_1D;
	STARPU_ASSERT(n2 == n3*DIV_1D);

	/* TODO: flags? Automatically set FFTW_MEASURE on calibration? */
	STARPU_ASSERT(flags == 0);

	STARPUFFT(plan) plan = malloc(sizeof(*plan));
	memset(plan, 0, sizeof(*plan));

	plan->number = STARPU_ATOMIC_ADD(&starpufft_last_plan_number, 1) - 1;

	/* The plan number has a limited size */
	STARPU_ASSERT(plan->number < (1ULL << NUMBER_BITS));

	/* Just one dimension */
	plan->dim = 1;
	plan->n = malloc(plan->dim * sizeof(*plan->n));
	plan->n[0] = n;

	check_dims(plan);

	plan->n1 = malloc(plan->dim * sizeof(*plan->n1));
	plan->n1[0] = n1;
	plan->n2 = malloc(plan->dim * sizeof(*plan->n2));
	plan->n2[0] = n2;

	/* Note: this is for coherency with the 2D case */
	plan->totsize = n;
	plan->totsize1 = n1;
	plan->totsize2 = n2;
	plan->totsize3 = DIV_1D;
	plan->totsize4 = plan->totsize / plan->totsize3;
	plan->type = C2C;
	plan->sign = sign;

	/* Compute the w^k just once. */
	compute_roots(plan);

	/* Initialize per-worker working set */
	for (workerid = 0; workerid < starpu_worker_get_count(); workerid++) {
		switch (starpu_worker_get_type(workerid)) {
		case STARPU_CPU_WORKER:
#ifdef STARPU_HAVE_FFTW
			/* first fft plan: one fft of size n2.
			 * FFTW imposes that buffer pointers are known at
			 * planning time. */
			plan->plans[workerid].plan1_cpu = _FFTW(plan_dft_1d)(n2, NULL, NULL, sign, _FFTW_FLAGS);
			STARPU_ASSERT(plan->plans[workerid].plan1_cpu);

			/* second fft plan: n3 ffts of size n1 */
			plan->plans[workerid].plan2_cpu = _FFTW(plan_many_dft)(plan->dim,
					plan->n1, n3,
					NULL, NULL, 1, plan->totsize1,
					NULL, NULL, 1, plan->totsize1,
					sign, _FFTW_FLAGS);
			STARPU_ASSERT(plan->plans[workerid].plan2_cpu);
#else
#warning libstarpufft can not work correctly if libfftw3 is not installed
#endif
			break;
		case STARPU_CUDA_WORKER:
			break;
		default:
			STARPU_ABORT();
			break;
		}
	}
#ifdef STARPU_USE_CUDA
	starpu_execute_on_each_worker(STARPUFFT(fft1_1d_plan_gpu), plan, STARPU_CUDA);
	starpu_execute_on_each_worker(STARPUFFT(fft2_1d_plan_gpu), plan, STARPU_CUDA);
#endif

	/* Allocate buffers. */
	plan->twisted1 = STARPUFFT(malloc)(plan->totsize * sizeof(*plan->twisted1));
	memset(plan->twisted1, 0, plan->totsize * sizeof(*plan->twisted1));
	plan->fft1 = STARPUFFT(malloc)(plan->totsize * sizeof(*plan->fft1));
	memset(plan->fft1, 0, plan->totsize * sizeof(*plan->fft1));
	plan->twisted2 = STARPUFFT(malloc)(plan->totsize * sizeof(*plan->twisted2));
	memset(plan->twisted2, 0, plan->totsize * sizeof(*plan->twisted2));
	plan->fft2 = STARPUFFT(malloc)(plan->totsize * sizeof(*plan->fft2));
	memset(plan->fft2, 0, plan->totsize * sizeof(*plan->fft2));

	/* Allocate handle arrays */
	plan->twisted1_handle = malloc(plan->totsize1 * sizeof(*plan->twisted1_handle));
	plan->fft1_handle = malloc(plan->totsize1 * sizeof(*plan->fft1_handle));
	plan->twisted2_handle = malloc(plan->totsize3 * sizeof(*plan->twisted2_handle));
	plan->fft2_handle = malloc(plan->totsize3 * sizeof(*plan->fft2_handle));

	/* Allocate task arrays */
	plan->twist1_tasks = malloc(plan->totsize1 * sizeof(*plan->twist1_tasks));
	plan->fft1_tasks = malloc(plan->totsize1 * sizeof(*plan->fft1_tasks));
	plan->twist2_tasks = malloc(plan->totsize3 * sizeof(*plan->twist2_tasks));
	plan->fft2_tasks = malloc(plan->totsize3 * sizeof(*plan->fft2_tasks));
	plan->twist3_tasks = malloc(plan->totsize3 * sizeof(*plan->twist3_tasks));

	/* Allocate codelet argument arrays */
	plan->fft1_args = malloc(plan->totsize1 * sizeof(*plan->fft1_args));
	plan->fft2_args = malloc(plan->totsize3 * sizeof(*plan->fft2_args));

	/* Create first-round tasks: DIV_1D tasks of type twist1 and fft1 */
	for (z = 0; z < plan->totsize1; z++) {
		int i = z;
#define STEP_TAG(step)	STEP_TAG_1D(plan, step, i)

		plan->fft1_args[z].plan = plan;
		plan->fft1_args[z].i = i;

		/* Register the twisted1 buffer of size n2. */
		starpu_vector_data_register(&plan->twisted1_handle[z], 0, (uintptr_t) &plan->twisted1[z*plan->totsize2], plan->totsize2, sizeof(*plan->twisted1));
		/* Register the fft1 buffer of size n2. */
		starpu_vector_data_register(&plan->fft1_handle[z], 0, (uintptr_t) &plan->fft1[z*plan->totsize2], plan->totsize2, sizeof(*plan->fft1));

		/* We'll need the result of fft1 on the CPU for the second
		 * twist anyway, so tell starpu to not keep the fft1 buffer in
		 * the GPU. */
		starpu_data_set_wt_mask(plan->fft1_handle[z], 1<<0);

		/* Create twist1 task */
		plan->twist1_tasks[z] = task = starpu_task_create();
		task->cl = &STARPUFFT(twist1_1d_codelet);
		/* task->buffers[0].handle = to be filled at execution to point
		   to the application input. */
		task->buffers[0].mode = STARPU_R;
		task->buffers[1].handle = plan->twisted1_handle[z];
		task->buffers[1].mode = STARPU_W;
		task->cl_arg = &plan->fft1_args[z];
		task->tag_id = STEP_TAG(TWIST1);
		task->use_tag = 1;
		task->detach = 1;
		task->destroy = 0;

		/* Tell that fft1 depends on twisted1 */
		starpu_tag_declare_deps(STEP_TAG(FFT1),
				1, STEP_TAG(TWIST1));

		/* Create FFT1 task */
		plan->fft1_tasks[z] = task = starpu_task_create();
		task->cl = &STARPUFFT(fft1_1d_codelet);
		task->buffers[0].handle = plan->twisted1_handle[z];
		task->buffers[0].mode = STARPU_R;
		task->buffers[1].handle = plan->fft1_handle[z];
		task->buffers[1].mode = STARPU_W;
		task->buffers[2].handle = plan->roots_handle[0];
		task->buffers[2].mode = STARPU_R;
		task->cl_arg = &plan->fft1_args[z];
		task->tag_id = STEP_TAG(FFT1);
		task->use_tag = 1;
		task->detach = 1;
		task->destroy = 0;

		/* Tell that the join task will depend on the fft1 task. */
		starpu_tag_declare_deps(STEP_TAG_1D(plan, JOIN, 0),
				1, STEP_TAG(FFT1));
#undef STEP_TAG
	}

	/* Create the join task, only serving as a dependency point between
	 * fft1 and twist2 tasks */
	plan->join_task = task = starpu_task_create();
	task->cl = NULL;
	task->tag_id = STEP_TAG_1D(plan, JOIN, 0);
	task->use_tag = 1;
	task->detach = 1;
	task->destroy = 0;

	/* Create second-round tasks: DIV_1D batches of n2/DIV_1D twist2, fft2,
	 * and twist3 */
	for (z = 0; z < plan->totsize3; z++) {
		int jj = z;
#define STEP_TAG(step)	STEP_TAG_1D(plan, step, jj)

		plan->fft2_args[z].plan = plan;
		plan->fft2_args[z].jj = jj;

		/* Register n3 twisted2 buffers of size n1 */
		starpu_vector_data_register(&plan->twisted2_handle[z], 0, (uintptr_t) &plan->twisted2[z*plan->totsize4], plan->totsize4, sizeof(*plan->twisted2));
		starpu_vector_data_register(&plan->fft2_handle[z], 0, (uintptr_t) &plan->fft2[z*plan->totsize4], plan->totsize4, sizeof(*plan->fft2));

		/* We'll need the result of fft2 on the CPU for the third
		 * twist anyway, so tell starpu to not keep the fft2 buffer in
		 * the GPU. */
		starpu_data_set_wt_mask(plan->fft2_handle[z], 1<<0);

		/* Tell that twisted2 depends on the join task */
		starpu_tag_declare_deps(STEP_TAG(TWIST2),
				1, STEP_TAG_1D(plan, JOIN, 0));

		/* Create twist2 task */
		plan->twist2_tasks[z] = task = starpu_task_create();
		task->cl = &STARPUFFT(twist2_1d_codelet);
		task->buffers[0].handle = plan->twisted2_handle[z];
		task->buffers[0].mode = STARPU_W;
		task->cl_arg = &plan->fft2_args[z];
		task->tag_id = STEP_TAG(TWIST2);
		task->use_tag = 1;
		task->detach = 1;
		task->destroy = 0;

		/* Tell that fft2 depends on twisted2 */
		starpu_tag_declare_deps(STEP_TAG(FFT2),
				1, STEP_TAG(TWIST2));

		/* Create FFT2 task */
		plan->fft2_tasks[z] = task = starpu_task_create();
		task->cl = &STARPUFFT(fft2_1d_codelet);
		task->buffers[0].handle = plan->twisted2_handle[z];
		task->buffers[0].mode = STARPU_R;
		task->buffers[1].handle = plan->fft2_handle[z];
		task->buffers[1].mode = STARPU_W;
		task->cl_arg = &plan->fft2_args[z];
		task->tag_id = STEP_TAG(FFT2);
		task->use_tag = 1;
		task->detach = 1;
		task->destroy = 0;

		/* Tell that twist3 depends on fft2 */
		starpu_tag_declare_deps(STEP_TAG(TWIST3),
				1, STEP_TAG(FFT2));

		/* Create twist3 tasks */
		/* These run only on CPUs and thus write directly into the
		 * application output buffer. */
		plan->twist3_tasks[z] = task = starpu_task_create();
		task->cl = &STARPUFFT(twist3_1d_codelet);
		task->buffers[0].handle = plan->fft2_handle[z];
		task->buffers[0].mode = STARPU_R;
		task->cl_arg = &plan->fft2_args[z];
		task->tag_id = STEP_TAG(TWIST3);
		task->use_tag = 1;
		task->detach = 1;
		task->destroy = 0;

		/* Tell that to be completely finished we need to have finished
		 * this twisted3 */
		starpu_tag_declare_deps(STEP_TAG_1D(plan, END, 0),
				1, STEP_TAG(TWIST3));
#undef STEP_TAG
	}

	/* Create end task, only serving as a join point. */
	plan->end_task = task = starpu_task_create();
	task->cl = NULL;
	task->tag_id = STEP_TAG_1D(plan, END, 0);
	task->use_tag = 1;
	task->detach = 1;
	task->destroy = 0;

	return plan;
}

/* Actually submit all the tasks. */
static starpu_tag_t
STARPUFFT(start1dC2C)(STARPUFFT(plan) plan)
{
	STARPU_ASSERT(plan->type == C2C);
	int z;

	for (z=0; z < plan->totsize1; z++) {
		starpu_task_submit(plan->twist1_tasks[z]);
		starpu_task_submit(plan->fft1_tasks[z]);
	}

	starpu_task_submit(plan->join_task);

	for (z=0; z < plan->totsize3; z++) {
		starpu_task_submit(plan->twist2_tasks[z]);
		starpu_task_submit(plan->fft2_tasks[z]);
		starpu_task_submit(plan->twist3_tasks[z]);
	}

	starpu_task_submit(plan->end_task);

	return STEP_TAG_1D(plan, END, 0);
}

/* Free all the tags. The generic code handles freeing the buffers. */
static void
STARPUFFT(free_1d_tags)(STARPUFFT(plan) plan)
{
	unsigned i;
	int n1 = plan->n1[0];

	for (i = 0; i < n1; i++) {
		starpu_tag_remove(STEP_TAG_1D(plan, TWIST1, i));
		starpu_tag_remove(STEP_TAG_1D(plan, FFT1, i));
	}

	starpu_tag_remove(STEP_TAG_1D(plan, JOIN, 0));

	for (i = 0; i < DIV_1D; i++) {
		starpu_tag_remove(STEP_TAG_1D(plan, TWIST2, i));
		starpu_tag_remove(STEP_TAG_1D(plan, FFT2, i));
		starpu_tag_remove(STEP_TAG_1D(plan, TWIST3, i));
	}

	starpu_tag_remove(STEP_TAG_1D(plan, END, 0));
}