localize_bounds: explicitly project on parameter domain

[ppcg.git] / cuda.c

/*
 * Copyright 2010-2011 INRIA Saclay
 *
 * Use of this software is governed by the GNU LGPLv2.1 license
 *
 * Written by Sven Verdoolaege, INRIA Saclay - Ile-de-France,
 * Parc Club Orsay Universite, ZAC des vignes, 4 rue Jacques Monod,
 * 91893 Orsay, France
 */

#include <assert.h>
#include <stdlib.h>

#include <isl/polynomial.h>
#include <isl/union_set.h>
#include <isl/aff.h>
#include <isl/ilp.h>
#include <isl/flow.h>
#include <isl/band.h>
#include <isl/schedule.h>
#include <isl/options.h>
#include <cloog/isl/cloog.h>

#include "cuda.h"
#include "cuda_common.h"
#include "gpucode.h"
#include "schedule.h"
#include "ppcg_options.h"

/* The fields stride, shift and shift_map only contain valid information
 * if shift != NULL.
 * If so, they express that current index is such that if you add shift,
 * then the result is always a multiple of stride.
 * shift_map contains the mapping
 *
 *      i -> (i + shift)/stride
 */
struct cuda_array_bound {
        isl_int size;
        isl_aff *lb;

        isl_int stride;
        isl_qpolynomial *shift;
        isl_basic_map *shift_map;
};

struct cuda_array_info;

/* A group of array references in a kernel that should be handled together.
 * If private_bound is not NULL, then it is mapped to registers.
 * Otherwise, if shared_bound is not NULL, it is mapped to shared memory.
 * Otherwise, it is accesses from global memory.
 */
struct cuda_array_ref_group {
        /* The references in this group access this array. */
        struct cuda_array_info *array;
        /* Position of this group in the list of reference groups of array. */
        int nr;

        /* The following fields are use during the construction of the groups.
         * access is the combined access relation relative to the shared
         * memory tiling.
         * write is set if any access in the group is a write.
         */
        isl_map *access;
        int write;

        /* For each index, size and offset of piece in shared memory. */
        struct cuda_array_bound *shared_bound;

        /* For each index, size and offset of piece in private memory. */
        struct cuda_array_bound *private_bound;

        /* References in this group; point to elements of a linked list. */
        int n_ref;
        struct cuda_stmt_access **refs;
};

struct cuda_array_info {
        isl_space *dim;
        /* Element type. */
        char *type;
        /* Name of the array. */
        char *name;
        /* Number of indices. */
        unsigned n_index;
        /* For each index, a bound on the array in that direction. */
        isl_pw_aff **bound;
        /* For each index, bound[i] specialized to the current kernel. */
        isl_pw_aff **local_bound;

        /* All references to this array; point to elements of a linked list. */
        int n_ref;
        struct cuda_stmt_access **refs;

        /* The reference groups associated to this array. */
        int n_group;
        struct cuda_array_ref_group **groups;

        /* Last shared memory tile dimension that affects tile of this array. */
        int last_shared;
        /* Dimension at which copying to/from shared memory is printed.
         * if >= 0, then the value is >= last_shared
         * if -1, then the copying is done at the leaf level.
         */
        int print_shared_level;
};

/* Print the name of the local copy of a given group of array references.
 */
static void print_array_name(FILE *out, struct cuda_array_ref_group *group)
{
        int global = 0;

        if (group->private_bound)
                fprintf(out, "private_");
        else if (group->shared_bound)
                fprintf(out, "shared_");
        else
                global = 1;
        fprintf(out, "%s", group->array->name);
        if (!global && group->array->n_group > 1)
                fprintf(out, "_%d", group->nr);
}

/* Collect all references to the given array and store pointers to them
 * in array->refs.
 */
static void collect_references(struct cuda_gen *gen,
        struct cuda_array_info *array)
{
        int i;
        int n;

        n = 0;
        for (i = 0; i < gen->n_stmts; ++i) {
                struct cuda_stmt *stmt = &gen->stmts[i];
                struct cuda_stmt_access *access;

                for (access = stmt->accesses; access; access = access->next) {
                        const char *name;
                        name = isl_map_get_tuple_name(access->access,
                                                      isl_dim_out);
                        if (name && !strcmp(array->name, name))
                                n++;
                }
        }

        array->n_ref = n;
        array->refs = isl_alloc_array(gen->ctx, struct cuda_stmt_access *, n);
        assert(array->refs);

        n = 0;
        for (i = 0; i < gen->n_stmts; ++i) {
                struct cuda_stmt *stmt = &gen->stmts[i];
                struct cuda_stmt_access *access;

                for (access = stmt->accesses; access; access = access->next) {
                        const char *name;
                        name = isl_map_get_tuple_name(access->access,
                                                      isl_dim_out);
                        if (!name || strcmp(array->name, name))
                                continue;

                        array->refs[n++] = access;
                }
        }
}

static struct cuda_array_bound *create_bound_list(isl_ctx *ctx, int n_index)
{
        int i;
        struct cuda_array_bound *bound;

        bound = isl_alloc_array(ctx, struct cuda_array_bound, n_index);
        assert(bound);

        for (i = 0; i < n_index; ++i) {
                isl_int_init(bound[i].size);
                bound[i].lb = NULL;
                isl_int_init(bound[i].stride);
                bound[i].shift = NULL;
                bound[i].shift_map = NULL;
        }

        return bound;
}

static void free_bound_list(struct cuda_array_bound *bound, int n_index)
{
        int j;

        if (!bound)
                return;

        for (j = 0; j < n_index; ++j) {
                isl_int_clear(bound[j].size);
                isl_int_clear(bound[j].stride);
                isl_aff_free(bound[j].lb);
                isl_qpolynomial_free(bound[j].shift);
                isl_basic_map_free(bound[j].shift_map);
        }
        free(bound);
}

static struct pet_array *find_array(struct pet_scop *scop,
        __isl_keep isl_set *accessed)
{
        int i;
        isl_id *id;

        id = isl_set_get_tuple_id(accessed);

        for (i = 0; i < scop->n_array; ++i) {
                isl_id *id_i;

                id_i = isl_set_get_tuple_id(scop->arrays[i]->extent);
                isl_id_free(id_i);
                if (id == id_i)
                        break;
        }
        isl_id_free(id);

        return i < scop->n_array ? scop->arrays[i] : NULL;
}

/* Compute bounds on the host arrays based on the accessed elements
 * and collect all references to the array.
 */
static int extract_array_info(__isl_take isl_set *array, void *user)
{
        int i;
        struct cuda_gen *gen = (struct cuda_gen *)user;
        const char *name;
        int n_index;
        isl_pw_aff **bounds;
        isl_pw_aff **local_bounds;
        struct pet_array *pa;

        n_index = isl_set_dim(array, isl_dim_set);
        name = isl_set_get_tuple_name(array);
        bounds = isl_alloc_array(isl_set_get_ctx(array),
                                 isl_pw_aff *, n_index);
        assert(bounds);
        local_bounds = isl_calloc_array(isl_set_get_ctx(array),
                                 isl_pw_aff *, n_index);
        assert(local_bounds);
        gen->array[gen->n_array].dim = isl_set_get_space(array);
        gen->array[gen->n_array].name = strdup(name);
        gen->array[gen->n_array].n_index = n_index;
        gen->array[gen->n_array].bound = bounds;
        gen->array[gen->n_array].local_bound = local_bounds;

        pa = find_array(gen->scop, array);
        assert(pa);

        gen->array[gen->n_array].type = strdup(pa->element_type);

        for (i = 0; i < n_index; ++i) {
                isl_set *dom;
                isl_local_space *ls;
                isl_aff *one;
                isl_pw_aff *bound;
                isl_set *size = i == 0 ? array : pa->extent;

                bound = isl_set_dim_max(isl_set_copy(size), i);
                assert(bound);
                dom = isl_pw_aff_domain(isl_pw_aff_copy(bound));
                ls = isl_local_space_from_space(isl_set_get_space(dom));
                one = isl_aff_zero(ls);
                one = isl_aff_add_constant_si(one, 1);
                bound = isl_pw_aff_add(bound, isl_pw_aff_alloc(dom, one));
                bound = isl_pw_aff_gist(bound, isl_set_copy(gen->context));

                bounds[i] = bound;
        }

        collect_references(gen, &gen->array[gen->n_array]);

        gen->n_array++;

        isl_set_free(array);
        return 0;
}

void collect_array_info(struct cuda_gen *gen)
{
        isl_union_set *arrays;

        arrays = isl_union_map_range(isl_union_map_copy(gen->read));
        arrays = isl_union_set_union(arrays,
                        isl_union_map_range(isl_union_map_copy(gen->write)));
        arrays = isl_union_set_coalesce(arrays);

        gen->n_array = isl_union_set_n_set(arrays);
        gen->array = isl_alloc_array(gen->ctx,
                                     struct cuda_array_info, gen->n_array);
        assert(gen->array);
        gen->n_array = 0;
        isl_union_set_foreach_set(arrays, &extract_array_info, gen);
        isl_union_set_free(arrays);
}

static void free_array_info(struct cuda_gen *gen)
{
        int i, j;

        for (i = 0; i < gen->n_array; ++i) {
                int n_index = gen->array[i].n_index;
                free(gen->array[i].type);
                free(gen->array[i].name);
                for (j = 0; j < n_index; ++j) {
                        isl_pw_aff_free(gen->array[i].bound[j]);
                        isl_pw_aff_free(gen->array[i].local_bound[j]);
                }
                isl_space_free(gen->array[i].dim);
                free(gen->array[i].bound);
                free(gen->array[i].local_bound);
                free(gen->array[i].refs);
        }
        free(gen->array);
}

static void declare_device_arrays(struct cuda_gen *gen)
{
        int i;

        for (i = 0; i < gen->n_array; ++i)
                fprintf(gen->cuda.host_c, "%s *dev_%s;\n",
                        gen->array[i].type, gen->array[i].name);
        fprintf(gen->cuda.host_c, "\n");
}

static void print_array_size(struct cuda_gen *gen, FILE *out,
        struct cuda_array_info *array)
{
        int i;
        isl_printer *prn;

        prn = isl_printer_to_file(gen->ctx, out);
        prn = isl_printer_set_output_format(prn, ISL_FORMAT_C);
        for (i = 0; i < array->n_index; ++i) {
                prn = isl_printer_print_str(prn, "(");
                prn = isl_printer_print_pw_aff(prn, array->bound[i]);
                prn = isl_printer_print_str(prn, ") * ");
        }
        prn = isl_printer_print_str(prn, "sizeof(");
        prn = isl_printer_print_str(prn, array->type);
        prn = isl_printer_print_str(prn, ")");
        isl_printer_free(prn);
}

static void allocate_device_arrays(struct cuda_gen *gen)
{
        int i;

        for (i = 0; i < gen->n_array; ++i) {
                fprintf(gen->cuda.host_c,
                        "cudaCheckReturn(cudaMalloc((void **) &dev_%s, ",
                        gen->array[i].name);
                print_array_size(gen, gen->cuda.host_c, &gen->array[i]);
                fprintf(gen->cuda.host_c, "));\n");
        }
        fprintf(gen->cuda.host_c, "\n");
}

static void free_device_arrays(struct cuda_gen *gen)
{
        int i;

        for (i = 0; i < gen->n_array; ++i)
                fprintf(gen->cuda.host_c, "cudaCheckReturn(cudaFree(dev_%s));\n",
                        gen->array[i].name);
}

/* Check if a cuda array is a scalar.  A scalar is a value that is not stored
 * as an array or through a pointer reference, but as single data element.  At
 * the moment, scalars are represented as zero dimensional arrays.
 */
static int cuda_array_is_scalar(struct cuda_array_info *array)
{
        return (array->n_index == 0);
}

static void copy_arrays_to_device(struct cuda_gen *gen)
{
        int i;

        for (i = 0; i < gen->n_array; ++i) {
                isl_space *dim;
                isl_set *read_i;
                int empty;

                dim = isl_space_copy(gen->array[i].dim);
                read_i = isl_union_set_extract_set(gen->copy_in, dim);
                empty = isl_set_fast_is_empty(read_i);
                isl_set_free(read_i);
                if (empty)
                        continue;

                fprintf(gen->cuda.host_c, "cudaCheckReturn(cudaMemcpy(dev_%s,",
                        gen->array[i].name);

                if (cuda_array_is_scalar(&(gen->array[i])))
                        fprintf(gen->cuda.host_c, " &%s, ",
                                gen->array[i].name);
                else
                        fprintf(gen->cuda.host_c, " %s, ", gen->array[i].name);

                print_array_size(gen, gen->cuda.host_c, &gen->array[i]);
                fprintf(gen->cuda.host_c, ", cudaMemcpyHostToDevice));\n");
        }
        fprintf(gen->cuda.host_c, "\n");
}

static void copy_arrays_from_device(struct cuda_gen *gen)
{
        int i;
        isl_union_set *write;
        write = isl_union_map_range(isl_union_map_copy(gen->write));

        for (i = 0; i < gen->n_array; ++i) {
                isl_space *dim;
                isl_set *write_i;
                int empty;

                dim = isl_space_copy(gen->array[i].dim);
                write_i = isl_union_set_extract_set(write, dim);
                empty = isl_set_fast_is_empty(write_i);
                isl_set_free(write_i);
                if (empty)
                        continue;

                fprintf(gen->cuda.host_c,
                        "cudaCheckReturn(cudaMemcpy(%s, dev_%s, ",
                        gen->array[i].name, gen->array[i].name);
                print_array_size(gen, gen->cuda.host_c, &gen->array[i]);
                fprintf(gen->cuda.host_c, ", cudaMemcpyDeviceToHost));\n");
        }

        isl_union_set_free(write);
        fprintf(gen->cuda.host_c, "\n");
}

static void read_sizes_from_file(struct cuda_gen *gen, const char *filename,
        int *sizes, int len)
{
        int i;
        FILE *file;

        file = fopen(filename, "r");
        if (!file)
                return;

        for (i = 0; i < len; ++i)
                if (fscanf(file, "%d", &sizes[i]) < 1)
                        break;

        fclose(file);
}

static void reverse_list(int *list, int len)
{
        int i;
        int t;

        for (i = 0; 2 * i < len; ++i) {
                t = list[i];
                list[i] = list[len - 1 - i];
                list[len - 1 - i] = t;
        }
}

/* Read user specified sizes from "tile.sizes", "block.sizes" and "grid.sizes"
 * after filling in some potentially useful defaults.
 */
static void read_sizes(struct cuda_gen *gen)
{
        int n;

        gen->tile_size = isl_alloc_array(gen->ctx, int, gen->tile_len);
        assert(gen->tile_size);
        for (n = 0; n < gen->tile_len; ++n)
                gen->tile_size[n] = gen->options->tile_size;
        read_sizes_from_file(gen, "tile.sizes", gen->tile_size, gen->tile_len);

        n = gen->n_parallel;
        gen->n_block = (n <= 3) ? n : 3;
        switch (gen->n_block) {
        case 1:
                gen->block_dim[0] = 512;
                break;
        case 2:
                gen->block_dim[0] = 32;
                gen->block_dim[1] = 16;
                break;
        default:
                gen->block_dim[0] = 32;
                gen->block_dim[1] = 4;
                gen->block_dim[2] = 4;
                break;
        }
        read_sizes_from_file(gen, "block.sizes", gen->block_dim, gen->n_block);
        reverse_list(gen->block_dim, gen->n_block);

        gen->n_grid = (n <= 2) ? n : 2;
        switch (gen->n_grid) {
        case 1:
                gen->grid_dim[0] = 32768;
                break;
        default:
                gen->grid_dim[0] = 256;
                gen->grid_dim[1] = 256;
                break;
        }
        read_sizes_from_file(gen, "grid.sizes", gen->grid_dim, gen->n_grid);
        reverse_list(gen->grid_dim, gen->n_grid);
}

static void free_stmts(struct cuda_stmt *stmts, int n)
{
        int i;

        for (i = 0; i < n; ++i) {
                struct cuda_stmt_access *access, *next;

                for (access = stmts[i].accesses; access; access = next) {
                        next = access->next;
                        isl_map_free(access->access);
                        free(access);
                }

                isl_set_free(stmts[i].domain);
        }
        free(stmts);
}

void clear_cuda_gen(struct cuda_gen *gen)
{
        free_stmts(gen->stmts, gen->n_stmts);
        free_array_info(gen);
        isl_set_free(gen->context);
        isl_union_set_free(gen->copy_in);
        isl_union_map_free(gen->sched);
        isl_union_map_free(gen->read);
        isl_union_map_free(gen->write);
}

static void print_reverse_list(FILE *out, int len, int *list)
{
        int i;

        for (i = 0; i < len; ++i) {
                if (i)
                        fprintf(out, ", ");
                fprintf(out, "%d", list[len - 1 - i]);
        }
}

static void print_kernel_launch(struct cuda_gen *gen,
        __isl_keep isl_union_set *arrays)
{
        int i;
        int first = 1;
        unsigned nparam;
        isl_space *dim;

        print_indent(gen->code.dst, gen->code.indent);
        fprintf(gen->code.dst, "kernel%d <<<k%d_dimGrid, k%d_dimBlock>>> (",
                gen->kernel_id, gen->kernel_id, gen->kernel_id);
        fprintf(gen->cuda.kernel_c, "__global__ void kernel%d(",
                gen->kernel_id);
        fprintf(gen->cuda.kernel_h, "__global__ void kernel%d(",
                gen->kernel_id);

        for (i = 0; i < gen->n_array; ++i) {
                isl_space *dim;
                isl_set *arr;
                int empty;

                dim = isl_space_copy(gen->array[i].dim);
                arr = isl_union_set_extract_set(arrays, dim);
                empty = isl_set_fast_is_empty(arr);
                isl_set_free(arr);
                if (empty)
                        continue;

                if (!first) {
                        fprintf(gen->code.dst, ", ");
                        fprintf(gen->cuda.kernel_c, ", ");
                        fprintf(gen->cuda.kernel_h, ", ");
                }

                fprintf(gen->code.dst, "dev_%s", gen->array[i].name);
                fprintf(gen->cuda.kernel_c, "%s *%s",
                        gen->array[i].type, gen->array[i].name);
                fprintf(gen->cuda.kernel_h, "%s *%s",
                        gen->array[i].type, gen->array[i].name);

                first = 0;
        }

        dim = isl_union_set_get_space(arrays);
        nparam = isl_space_dim(dim, isl_dim_param);
        for (i = 0; i < nparam; ++i) {
                const char *name = isl_space_get_dim_name(dim, isl_dim_param, i);
                if (!first) {
                        fprintf(gen->code.dst, ", ");
                        fprintf(gen->cuda.kernel_c, ", ");
                        fprintf(gen->cuda.kernel_h, ", ");
                }
                fprintf(gen->code.dst, "%s", name);
                fprintf(gen->cuda.kernel_c, "int %s", name);
                fprintf(gen->cuda.kernel_h, "int %s", name);
                first = 0;
        }
        isl_space_free(dim);

        for (i = 0; i < gen->tile_first; ++i) {
                if (!first) {
                        fprintf(gen->code.dst, ", ");
                        fprintf(gen->cuda.kernel_c, ", ");
                        fprintf(gen->cuda.kernel_h, ", ");
                }
                fprintf(gen->code.dst, "h%d", i);
                fprintf(gen->cuda.kernel_c, "int h%d", i);
                fprintf(gen->cuda.kernel_h, "int h%d", i);
                first = 0;
        }

        fprintf(gen->code.dst, ");\n");
        fprintf(gen->cuda.kernel_c, ")\n");
        fprintf(gen->cuda.kernel_h, ");\n");

        fprintf(gen->code.dst, "cudaCheckKernel();\n");
}

/* Construct a map from a domain of dimensionality "len"
 * to a domain of dimensionality "len" + "tile_len" that tiles
 * the "tile_len" coordinates starting at "first".
 * In particular, [s_i] -> [s_i / tile_size[i], s_i % tile_size[i]].
 * "dim" prescribes the parameters.
 */
static __isl_give isl_map *tile(__isl_take isl_space *dim, int len,
        int first, int tile_len, int *tile_size)
{
        int i;
        isl_int v;
        isl_basic_map *bmap;
        isl_constraint *c;

        isl_int_init(v);

        dim = isl_space_add_dims(dim, isl_dim_in, len);
        dim = isl_space_add_dims(dim, isl_dim_out, len + tile_len);
        bmap = isl_basic_map_universe(isl_space_copy(dim));

        for (i = 0; i < len - tile_len; ++i) {
                int j = i < first ? i : i + tile_len;
                int k = i < first ? i : i + 2 * tile_len;

                c = isl_equality_alloc(isl_space_copy(dim));
                isl_int_set_si(v, -1);
                isl_constraint_set_coefficient(c, isl_dim_in, j, v);
                isl_int_set_si(v, 1);
                isl_constraint_set_coefficient(c, isl_dim_out, k, v);
                bmap = isl_basic_map_add_constraint(bmap, c);
        }

        for (i = 0; i < tile_len; ++i) {
                c = isl_equality_alloc(isl_space_copy(dim));
                isl_int_set_si(v, -1);
                isl_constraint_set_coefficient(c, isl_dim_in, first + i, v);
                isl_int_set_si(v, tile_size[i]);
                isl_constraint_set_coefficient(c, isl_dim_out, first + i, v);
                isl_int_set_si(v, 1);
                isl_constraint_set_coefficient(c, isl_dim_out,
                                                first + i + tile_len, v);
                bmap = isl_basic_map_add_constraint(bmap, c);

                c = isl_inequality_alloc(isl_space_copy(dim));
                isl_int_set_si(v, 1);
                isl_constraint_set_coefficient(c, isl_dim_out,
                                                first + i + tile_len, v);
                bmap = isl_basic_map_add_constraint(bmap, c);
        
                c = isl_inequality_alloc(isl_space_copy(dim));
                isl_int_set_si(v, -1);
                isl_constraint_set_coefficient(c, isl_dim_out,
                                                first + i + tile_len, v);
                isl_int_set_si(v, tile_size[i] - 1);
                isl_constraint_set_constant(c, v);
                bmap = isl_basic_map_add_constraint(bmap, c);
        }

        isl_space_free(dim);
        isl_int_clear(v);

        return isl_map_from_basic_map(bmap);
}

/* Construct a map from a domain of dimensionality "len"
 * to a domain of dimensionality "len" + "wrap_len" that "wraps"
 * the "wrap_len" coordinates starting at "first" according to "wrap_size".
 * In particular, [s_i] -> [s_i, s_i % wrap_size[i]].
 * To do so, we need extra variables corresponding to [s_i / wrap_size[i]],
 * that are projected out at the end.
 * "dim" prescribes the parameters.
 */
static __isl_give isl_map *wrap(__isl_take isl_space *dim, int len,
        int first, int wrap_len, int *wrap_size)
{
        int i;
        isl_basic_map *bmap;
        isl_constraint *c;

        dim = isl_space_add_dims(dim, isl_dim_in, len);
        dim = isl_space_add_dims(dim, isl_dim_out, len + 2 * wrap_len);
        bmap = isl_basic_map_universe(isl_space_copy(dim));

        for (i = 0; i < len; ++i) {
                int k = i < first + wrap_len ? i : i + 2 * wrap_len;

                c = isl_equality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_in, i, -1);
                isl_constraint_set_coefficient_si(c, isl_dim_out, k, 1);
                bmap = isl_basic_map_add_constraint(bmap, c);
        }

        for (i = 0; i < wrap_len; ++i) {
                c = isl_equality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_out,
                                                    first + i, -1);
                isl_constraint_set_coefficient_si(c, isl_dim_out,
                                                    first + wrap_len + i, 1);
                isl_constraint_set_coefficient_si(c, isl_dim_out,
                                    first + 2 * wrap_len + i, wrap_size[i]);
                bmap = isl_basic_map_add_constraint(bmap, c);

                c = isl_inequality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_out,
                                                    first + wrap_len + i, 1);
                bmap = isl_basic_map_add_constraint(bmap, c);
        
                c = isl_inequality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_out,
                                                    first + wrap_len + i, -1);
                isl_constraint_set_constant_si(c, wrap_size[i] - 1);
                bmap = isl_basic_map_add_constraint(bmap, c);
        }

        isl_space_free(dim);

        bmap = isl_basic_map_project_out(bmap, isl_dim_out,
                                first + 2 * wrap_len, wrap_len);

        return isl_map_from_basic_map(bmap);
}

/* Add "n" parameters named prefix%d.
 */
static __isl_give isl_set *add_params( __isl_take isl_set *set,
        int n, const char *prefix)
{
        int i;
        unsigned nparam;
        char name[20];

        nparam = isl_set_dim(set, isl_dim_param);
        set = isl_set_add_dims(set, isl_dim_param, n);

        for (i = 0; i < n; ++i) {
                snprintf(name, sizeof(name), "%s%d", prefix, i);
                set = isl_set_set_dim_name(set, isl_dim_param,
                                            nparam + i, name);
        }

        return set;
}

/* Equate the "n" dimensions of "set" starting at "first" to
 * freshly created parameters named prefix%d.
 */
static __isl_give isl_set *parametrize(__isl_take isl_set *set,
        int first, int n, const char *prefix)
{
        int i;
        unsigned nparam;
        isl_int v;
        isl_space *dim;
        isl_basic_set *bset;
        isl_constraint *c;

        nparam = isl_set_dim(set, isl_dim_param);

        set = add_params(set, n, prefix);

        dim = isl_set_get_space(set);
        bset = isl_basic_set_universe(isl_space_copy(dim));

        isl_int_init(v);

        for (i = 0; i < n; ++i) {
                c = isl_equality_alloc(isl_space_copy(dim));
                isl_int_set_si(v, -1);
                isl_constraint_set_coefficient(c, isl_dim_param, nparam + i, v);
                isl_int_set_si(v, 1);
                isl_constraint_set_coefficient(c, isl_dim_set, first + i, v);
                bset = isl_basic_set_add_constraint(bset, c);
        }

        isl_int_clear(v);
        isl_space_free(dim);

        return isl_set_intersect(set, isl_set_from_basic_set(bset));
}

static __isl_give isl_set *parametrization(__isl_take isl_space *dim,
        int len, int first, int n, const char *prefix)
{
        isl_set *set;

        dim = isl_space_add_dims(dim, isl_dim_set, len);
        set = isl_set_universe(dim);

        return parametrize(set, first, n, prefix);
}

/* Tile the B loops over the tile sizes and then tile/wrap
 * the T1 loops over the blocks.
 */
static __isl_give isl_union_map *tile_schedule(struct cuda_gen *gen,
        __isl_take isl_union_map *sched)
{
        isl_space *dim;
        isl_map *tiling, *block_tiling;

        dim = isl_union_map_get_space(sched);
        tiling = tile(isl_space_copy(dim), gen->untiled_len,
                      gen->tile_first, gen->tile_len, gen->tile_size);

        if (gen->options->wrap)
                block_tiling = wrap(dim, gen->untiled_len + gen->tile_len,
                                gen->tile_first, gen->n_grid, gen->grid_dim);
        else
                block_tiling = tile(dim, gen->untiled_len + gen->tile_len,
                                gen->tile_first, gen->n_grid, gen->grid_dim);

        gen->tiled_len = gen->untiled_len + gen->tile_len + gen->n_grid;

        tiling = isl_map_apply_range(tiling, block_tiling);

        sched = isl_union_map_apply_range(sched,
                                             isl_union_map_from_map(tiling));

        gen->shared_len = gen->tile_first + gen->tile_len + gen->n_grid;

        return sched;
}

static __isl_give isl_union_map *parametrize_tiled_schedule(
        struct cuda_gen *gen, __isl_take isl_union_map *sched)
{
        isl_space *dim;
        isl_set *par;

        dim = isl_union_map_get_space(sched);
        par = parametrization(dim, gen->tiled_len, 0, gen->tile_first, "h");
        sched = isl_union_map_intersect_range(sched,
                                                isl_union_set_from_set(par));

        dim = isl_union_map_get_space(sched);
        par = parametrization(dim, gen->tiled_len,
                gen->tile_first + gen->n_grid, gen->n_grid, "b");
        sched = isl_union_map_intersect_range(sched,
                                                isl_union_set_from_set(par));

        return sched;
}

/* Tile/wrap the P1 loops over the threads.
 */
static __isl_give isl_union_map *thread_tile_schedule(struct cuda_gen *gen,
        __isl_take isl_union_map *sched)
{
        isl_space *dim;
        isl_map *tiling;
        isl_set *par;

        dim = isl_union_map_get_space(sched);

        if (gen->options->wrap)
                tiling = wrap(isl_space_copy(dim), gen->tiled_len,
                                gen->shared_len, gen->n_block, gen->block_dim);
        else
                tiling = tile(isl_space_copy(dim), gen->tiled_len,
                                gen->shared_len, gen->n_block, gen->block_dim);
        gen->thread_tiled_len = gen->tiled_len + gen->n_block;

        sched = isl_union_map_apply_range(sched,
                                             isl_union_map_from_map(tiling));

        par = parametrization(dim, gen->thread_tiled_len,
                gen->tile_first + gen->tile_len + gen->n_grid + gen->n_block,
                gen->n_block, "t");
        sched = isl_union_map_intersect_range(sched,
                                                isl_union_set_from_set(par));

        gen->shared_len = gen->tile_first + gen->tile_len + gen->n_grid;

        return sched;
}

/* If the user asked for it, scale the shared memory tile loops
 * (T1P and T2) of "sched" by gen->tile_size[i].
 * If we are not performing "wrapping", then additionally scale the T1P
 * loops by gen->grid_dim[i].
 */
static __isl_give isl_union_map *scale_tile_loops(struct cuda_gen *gen,
        __isl_take isl_union_map *sched)
{
        int i;
        isl_space *dim;
        isl_basic_map *scale;
        isl_constraint *c;

        if (!gen->options->scale_tile_loops)
                return sched;

        dim = isl_union_map_get_space(sched);
        dim = isl_space_add_dims(dim, isl_dim_in, gen->tiled_len);
        dim = isl_space_add_dims(dim, isl_dim_out, gen->tiled_len);
        scale = isl_basic_map_universe(isl_space_copy(dim));

        for (i = 0; i < gen->tiled_len; ++i) {
                int f = 1;

                if (i >= gen->tile_first && i < gen->tile_first + gen->n_grid) {
                        f = gen->tile_size[i - gen->tile_first];
                        if (!gen->options->wrap)
                                f *= gen->grid_dim[i - gen->tile_first];
                } else if (i >= gen->tile_first + gen->n_grid &&
                           i < gen->tile_first + gen->n_grid + gen->tile_len) {
                        f = gen->tile_size[i - (gen->tile_first + gen->n_grid)];
                }

                c = isl_equality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_in, i, f);
                isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
                scale = isl_basic_map_add_constraint(scale, c);
        }

        isl_space_free(dim);

        sched = isl_union_map_apply_range(sched,
                isl_union_map_from_map(isl_map_from_basic_map(scale)));

        return sched;
}

/* If we are not performing "wrapping" and if the user asked for it,
 * scale the thread tile loops (P1T) of "sched" by gen->block_dim[i].
 */
static __isl_give isl_union_map *scale_thread_tile_loops(struct cuda_gen *gen,
        __isl_take isl_union_map *sched)
{
        int i;
        isl_space *dim;
        isl_basic_map *scale;
        isl_constraint *c;

        if (gen->options->wrap)
                return sched;
        if (!gen->options->scale_tile_loops)
                return sched;

        dim = isl_union_map_get_space(sched);
        dim = isl_space_add_dims(dim, isl_dim_in, gen->thread_tiled_len);
        dim = isl_space_add_dims(dim, isl_dim_out, gen->thread_tiled_len);
        scale = isl_basic_map_universe(isl_space_copy(dim));

        for (i = 0; i < gen->thread_tiled_len; ++i) {
                int f = 1;

                if (i >= gen->shared_len &&
                    i < gen->shared_len + gen->n_block)
                        f = gen->block_dim[i - gen->shared_len];

                c = isl_equality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_in, i, f);
                isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
                scale = isl_basic_map_add_constraint(scale, c);
        }

        isl_space_free(dim);

        sched = isl_union_map_apply_range(sched,
                isl_union_map_from_map(isl_map_from_basic_map(scale)));

        return sched;
}

/* If we are not performing "wrapping" and if the user asked for it,
 * scale the "n_tile" loops starting at "first" of "sched" by gen->block_dim[i].
 */
static __isl_give isl_union_map *scale_access_tile_loops(struct cuda_gen *gen,
        __isl_take isl_union_map *sched, int len, int first, int n_tile)
{
        int i;
        isl_space *dim;
        isl_basic_map *scale;
        isl_constraint *c;

        if (gen->options->wrap)
                return sched;
        if (!gen->options->scale_tile_loops)
                return sched;

        dim = isl_union_map_get_space(sched);
        dim = isl_space_add_dims(dim, isl_dim_in, len);
        dim = isl_space_add_dims(dim, isl_dim_out, len);
        scale = isl_basic_map_universe(isl_space_copy(dim));

        for (i = 0; i < len; ++i) {
                int f = 1;

                if (i >= first && i < first + n_tile)
                        f = gen->block_dim[i - first];

                c = isl_equality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_in, i, f);
                isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
                scale = isl_basic_map_add_constraint(scale, c);
        }

        isl_space_free(dim);

        sched = isl_union_map_apply_range(sched,
                isl_union_map_from_map(isl_map_from_basic_map(scale)));

        return sched;
}

/* If print_user_stmt is set, we want to print the statements ourselves,
 * instead of relying on the C preprocessor.  If so, we need to use
 * the stop option so that the domains will be saved on the statement
 * nodes.
 */
static void print_cloog_shared_body(struct cuda_gen *gen,
        __isl_keep isl_set *context, __isl_keep isl_union_map *sched, int len,
        void (*print_user_stmt)(struct gpucode_info *info,
                                struct clast_user_stmt *s),
        int first_unroll)
{
        int i;
        CloogOptions *options;
        CloogDomain *cloog_context;
        CloogUnionDomain *ud;
        CloogInput *input;
        struct clast_stmt *stmt;
        char name[20];

        sched = isl_union_map_copy(sched);
        sched = isl_union_map_align_params(sched, isl_set_get_space(context));

        options = cloog_options_malloc(gen->state);
        options->language = LANGUAGE_C;
        options->strides = 1;
        options->sh = 1;
        options->f = len;
        options->l = -1;
        options->override = 1;
        options->save_domains = 1;
        options->noscalars = 1;
        options->first_unroll = first_unroll;

        ud = cloog_union_domain_from_isl_union_map(sched);
        for (i = 0; i < len; ++i) {
                snprintf(name, sizeof(name), "c%d", i);
                ud = cloog_union_domain_set_name(ud, CLOOG_SCAT, i, name);
        }
        cloog_context = cloog_domain_from_isl_set(isl_set_copy(context));
        input = cloog_input_alloc(cloog_context, ud);

        stmt = cloog_clast_create_from_input(input, options);

        gen->stmt_code.indent = gen->kernel_code.indent;
        gen->stmt_code.dst = gen->cuda.kernel_c;
        gen->stmt_code.print_user_stmt = print_user_stmt;
        gen->stmt_code.print_user_stmt_list = NULL;
        gen->stmt_code.print_for_head = NULL;
        gen->stmt_code.print_for_foot = NULL;
        gen->stmt_code.user = gen;
        gpu_print_host_stmt(&gen->stmt_code, stmt);

        cloog_clast_free(stmt);
        cloog_options_free(options);
}

/* Add "len" parameters p[i] called prefix%d,
 * with bounds to 0 <= p[i] < size[i].
 */
__isl_give isl_set *add_bounded_parameters(__isl_take isl_set *set,
        int len, int *size, const char *prefix)
{
        int i;
        unsigned nparam;
        isl_int v;
        isl_space *dim;
        isl_basic_set *bset;
        isl_constraint *c;
        char name[20];

        nparam = isl_set_dim(set, isl_dim_param);
        set = isl_set_add_dims(set, isl_dim_param, len);

        for (i = 0; i < len; ++i) {
                snprintf(name, sizeof(name), "%s%d", prefix, i);
                set = isl_set_set_dim_name(set, isl_dim_param,
                                            nparam + i, name);
        }

        dim = isl_set_get_space(set);
        bset = isl_basic_set_universe(isl_space_copy(dim));

        isl_int_init(v);

        for (i = 0; i < len; ++i) {
                c = isl_inequality_alloc(isl_space_copy(dim));
                isl_int_set_si(v, 1);
                isl_constraint_set_coefficient(c, isl_dim_param, nparam + i, v);
                bset = isl_basic_set_add_constraint(bset, c);
        
                c = isl_inequality_alloc(isl_space_copy(dim));
                isl_int_set_si(v, -1);
                isl_constraint_set_coefficient(c, isl_dim_param, nparam + i, v);
                isl_int_set_si(v, size[i] - 1);
                isl_constraint_set_constant(c, v);
                bset = isl_basic_set_add_constraint(bset, c);
        }

        isl_int_clear(v);
        isl_space_free(dim);

        return isl_set_intersect(set, isl_set_from_basic_set(bset));
}

static void print_shared_body(struct cuda_gen *gen,
        __isl_keep isl_set *shared_domain, __isl_keep isl_union_map *sched,
        int len, void (*print_user_stmt)(struct gpucode_info *info,
                                         struct clast_user_stmt *s),
        int first_unroll)
{
        isl_set *context;

        context = isl_set_copy(shared_domain);
        context = parametrize(context, 0, gen->shared_len, "g");
        context = isl_set_project_out(context, isl_dim_set, 0, gen->shared_len);
        context = add_bounded_parameters(context,
                                        gen->n_block, gen->block_dim, "t");

        print_cloog_shared_body(gen, context, sched, len, print_user_stmt,
                                first_unroll);

        isl_set_free(context);
}

/* Given a tile of an array, construct a map that maps each element
 * of the tile to a copy of the tile shifted to the origin
 * (based on the lower bounds in group->private_bound or group->shared_bound).
 * If any of the indices is strided, then {private,shared}_bound[i].shift_map
 * is applied to the index first.
 * The domain of the resulting map is "access",
 * while the range space is anonymous.
 */
static __isl_give isl_map *shift_access(__isl_take isl_set *access,
        struct cuda_array_ref_group *group)
{
        int i;
        isl_space *dim;
        isl_basic_set *bset;
        isl_basic_map *bmap;
        isl_aff *lb;
        isl_basic_set *offset;
        isl_basic_map *shift;
        isl_basic_map *pre_shift;
        isl_map *sched;
        const char *name;
        struct cuda_array_bound *bounds;
        int n_index = group->array->n_index;

        bounds = group->private_bound;
        if (!bounds)
                bounds = group->shared_bound;

        dim = isl_set_get_space(access);
        dim = isl_space_drop_dims(dim, isl_dim_set, 0, n_index);
        offset = isl_basic_set_universe(dim);
        for (i = 0; i < n_index; ++i) {
                lb = isl_aff_copy(bounds[i].lb);
                bmap = isl_basic_map_from_aff(lb);
                bset = isl_basic_map_range(bmap);
                offset = isl_basic_set_flat_product(offset, bset);
        }
        offset = isl_basic_set_neg(offset);

        dim = isl_space_map_from_set(isl_set_get_space(access));
        shift = isl_basic_map_identity(dim);
        shift = isl_basic_map_set_tuple_name(shift, isl_dim_out, NULL);

        bset = isl_basic_set_universe(isl_set_get_space(access));
        bmap = isl_basic_map_from_domain_and_range(bset, offset);

        shift = isl_basic_map_sum(shift, bmap);

        dim = isl_set_get_space(access);
        dim = isl_space_drop_dims(dim, isl_dim_set, 0, n_index);
        dim = isl_space_map_from_set(dim);
        pre_shift = isl_basic_map_universe(isl_space_copy(dim));
        dim = isl_space_add_dims(dim, isl_dim_in, 1);
        dim = isl_space_add_dims(dim, isl_dim_out, 1);
        for (i = 0; i < n_index; ++i) {
                if (!bounds[i].shift_map)
                        bmap = isl_basic_map_identity(isl_space_copy(dim));
                else
                        bmap = isl_basic_map_copy(bounds[i].shift_map);
                pre_shift = isl_basic_map_flat_product(pre_shift, bmap);
        }
        isl_space_free(dim);
        name = isl_basic_map_get_tuple_name(shift, isl_dim_in);
        pre_shift = isl_basic_map_set_tuple_name(pre_shift, isl_dim_in, name);
        pre_shift = isl_basic_map_set_tuple_name(pre_shift, isl_dim_out, name);
        shift = isl_basic_map_apply_range(pre_shift, shift);

        sched = isl_map_from_basic_map(shift);
        sched = isl_map_intersect_domain(sched, access);

        return sched;
}

/* Construct a schedule for iterating over all elements in the given
 * piece of an array.  The schedule iterates over a copy of the piece
 * that is shifted to the origin.
 * We subsequently also perform the tiling/wrapping over the threads.
 *
 * In particular, we tile the final iterators so that the final thread
 * dimension runs over the final array dimension.
 * However, if those final iterators have only a single iteration,
 * we try to tile earlier iterators instead.
 */
static __isl_give isl_union_map *access_schedule(struct cuda_gen *gen,
        __isl_take isl_set *access, struct cuda_array_ref_group *group)
{
        isl_space *dim;
        isl_map *sched;
        isl_union_map *usched;
        isl_map *tiling;
        isl_set *par;
        unsigned nvar = isl_set_dim(access, isl_dim_set);
        int n_tile;
        int first;

        sched = shift_access(access, group);

        n_tile = gen->n_block;
        if (n_tile > nvar) {
                int i;
                sched = isl_map_insert_dims(sched,
                                                isl_dim_out, 0, n_tile - nvar);
                for (i = 0; i < n_tile - nvar; ++i)
                        sched = isl_map_fix_si(sched, isl_dim_out, i, 0);
                nvar = n_tile;
        }

        first = nvar - n_tile;

        for (; first > 0; first --)
                if (!isl_map_plain_is_fixed(sched, isl_dim_out,
                                                first + n_tile - 1, NULL))
                        break;

        dim = isl_map_get_space(sched);
        dim = isl_space_params(dim);
        if (gen->options->wrap)
                tiling = wrap(isl_space_copy(dim), nvar, first,
                                n_tile, gen->block_dim);
        else
                tiling = tile(isl_space_copy(dim), nvar, first,
                                n_tile, gen->block_dim);
        sched = isl_map_apply_range(sched, tiling);

        par = parametrization(dim, nvar + n_tile, first + n_tile, n_tile, "t");
        usched = isl_union_map_from_map(sched);
        usched = isl_union_map_intersect_range(usched,
                                                isl_union_set_from_set(par));

        usched = scale_access_tile_loops(gen, usched, nvar + n_tile,
                                         first, n_tile);

        return usched;
}

static void print_shared_access(struct cuda_gen *gen,
        __isl_keep isl_set *shared_domain, __isl_take isl_set *access,
        const char *type, struct cuda_array_ref_group *group)
{
        const char *array_name;
        char *name;
        isl_ctx *ctx;
        isl_union_map *sched;
        unsigned nvar = isl_set_dim(access, isl_dim_set);
        int n_tile;

        ctx = isl_set_get_ctx(access);
        array_name = isl_set_get_tuple_name(access);
        name = isl_alloc_array(ctx, char,
                strlen(type) + sizeof("_shared_") + strlen(array_name) + 20);
        if (group->array->n_group > 1)
                sprintf(name, "%s_shared_%s_%d", type, array_name, group->nr);
        else
                sprintf(name, "%s_shared_%s", type, array_name);
        access = isl_set_set_tuple_name(access, name);
        free(name);

        sched = access_schedule(gen, access, group);

        n_tile = gen->n_block;
        if (n_tile > nvar)
                n_tile = nvar;

        print_shared_body(gen, shared_domain, sched, nvar + n_tile, NULL, -1);

        isl_union_map_free(sched);
}

/* Return the union of all read (read = 1) and/or write (write = 1)
 * access relations in the group.
 */
static __isl_give isl_union_map *group_access_relation(
        struct cuda_array_ref_group *group, int read, int write)
{
        int i;
        isl_union_map *access;

        access = isl_union_map_empty(isl_map_get_space(group->access));
        for (i = 0; i < group->n_ref; ++i) {
                isl_map *map_i;

                if (!((read && group->refs[i]->read) ||
                     (write && group->refs[i]->write)))
                        continue;
                map_i = isl_map_copy(group->refs[i]->access);
                access = isl_union_map_union(access,
                                            isl_union_map_from_map(map_i));
        }

        return access;
}

/* Check that none of the shared memory tiles involve any strides.
 */
static int no_strides(struct cuda_array_ref_group *group)
{
        int i;
        int n_index = group->array->n_index;

        for (i = 0; i < n_index; ++i)
                if (group->shared_bound[i].shift)
                        return 0;

        return 1;
}

/* Return a set containing the values of the given index i
 * of the elements in the array tile in global memory that corresponds
 * to the shared memory copy.
 * In particular, if a is the index, we return a set with constraints
 *
 *      tile_offset <= a <= tile_offset + tile_size - 1
 *
 * and
 *
 *      0 <= a <= array_size - 1
 *
 */
static __isl_give isl_set *group_tile_dim(struct cuda_array_ref_group *group,
        int i)
{
        isl_basic_set *tile;
        isl_aff *aff;
        isl_constraint *c;
        isl_local_space *ls;
        isl_pw_aff *bound;
        isl_set *dom;
        isl_set *tile_set;

        aff = isl_aff_copy(group->shared_bound[i].lb);
        aff = isl_aff_add_dims(aff, isl_dim_set, 1);
        ls = isl_aff_get_local_space(aff);
        aff = isl_aff_neg(aff);
        aff = isl_aff_add_coefficient_si(aff, isl_dim_set, 0, 1);
        c = isl_inequality_from_aff(isl_aff_copy(aff));
        tile = isl_basic_set_from_constraint(c);

        aff = isl_aff_neg(aff);
        aff = isl_aff_add_constant(aff, group->shared_bound[i].size);
        aff = isl_aff_add_constant_si(aff, -1);
        c = isl_inequality_from_aff(aff);
        tile = isl_basic_set_add_constraint(tile, c);

        aff = isl_aff_zero(ls);
        aff = isl_aff_add_coefficient_si(aff, isl_dim_set, 0, 1);
        c = isl_inequality_from_aff(aff);
        tile = isl_basic_set_add_constraint(tile, c);

        bound = isl_pw_aff_copy(group->array->bound[i]);
        bound = isl_pw_aff_add_dims(bound, isl_dim_set, 1);
        ls = isl_local_space_from_space(isl_pw_aff_get_space(bound));
        aff = isl_aff_zero(ls);
        aff = isl_aff_add_coefficient_si(aff, isl_dim_set, 0, 1);
        aff = isl_aff_add_constant_si(aff, 1);
        dom = isl_pw_aff_domain(isl_pw_aff_copy(bound));

        tile_set = isl_pw_aff_ge_set(bound, isl_pw_aff_alloc(dom, aff));
        tile_set = isl_set_align_params(tile_set, isl_basic_set_get_space(tile));
        tile_set = isl_set_intersect(tile_set, isl_set_from_basic_set(tile));

        return tile_set;
}

/* Return a set containing the elements in the array tile in
 * global memory that corresponds to the shared memory copy.
 */
static __isl_give isl_set *group_tile(struct cuda_array_ref_group *group)
{
        int i;
        int n_index = group->array->n_index;
        isl_set *tile;

        tile = group_tile_dim(group, 0);
        for (i = 1; i < n_index; ++i) {
                isl_set *tile_i;

                tile_i = group_tile_dim(group, i);
                tile = isl_set_flat_product(tile, tile_i);
        }

        tile = isl_set_set_tuple_name(tile, group->array->name);

        return tile;
}

/* Print code for reading into or writing from shared memory
 * the given array reference group.
 *
 * sched maps the original iteration domains to the shared memory tile loops.
 *
 * If we are performing a read from global memory to shared memory,
 * if the array involved is not a scalar and if the definition of the
 * shared memory tiles does not involve any strides, then we copy
 * the entire tile to shared memory.  This may result in some extra
 * elements getting copied, but it should lead to simpler code
 * (which means that fewer registers may be needed) and less divergence.
 *
 * Otherwise, we only copy the elements that will be read or have been written
 * in the kernel.
 *
 * Note that the absence of stride requirement can easily be lifted.
 * We would just need to add constraints of the form
 *
 *      shift + a = stride * alpha
 */
static int print_group_shared_accesses(struct cuda_gen *gen,
        struct cuda_array_ref_group *group, const char *type,
        __isl_keep isl_set *shared_domain, __isl_keep isl_union_map *sched)
{
        int read;
        isl_union_map *access;
        isl_union_set *uset;
        isl_set *access_set;

        if (group->private_bound)
                return 0;
        if (!group->shared_bound)
                return 0;

        read = !strcmp(type, "read");

        access = group_access_relation(group, read, !read);
        access = isl_union_map_apply_domain(access, isl_union_map_copy(sched));
        uset = isl_union_map_range(access);

        if (isl_union_set_is_empty(uset)) {
                isl_union_set_free(uset);
                return 0;
        }

        if (read && group->array->n_index > 0 && no_strides(group)) {
                isl_union_set_free(uset);
                access_set = group_tile(group);
                print_shared_access(gen, shared_domain, access_set,
                                    type, group);
                return 1;
        }

        access_set = isl_set_from_union_set(uset);
        access_set = isl_set_coalesce(access_set);

        print_shared_access(gen, shared_domain, access_set, type, group);

        return 1;
}

/* Print code for reading into or writing from shared memory at
 * the given level (-1 for innermost).
 *
 * If we are not printing at the innermost level, then the dimensionality
 * of shared_domain may be smaller than gen->shared_len.
 * As the rest of the code assumes that the domain of access has
 * gen->shared_len dimensions, we therefore may need to embed this domain
 * in a higher dimensional space after intersection with shared_domain.
 */
static void print_shared_accesses(struct cuda_gen *gen,
        __isl_keep isl_set *shared_domain, __isl_keep isl_union_map *access,
        const char *type, int level)
{
        int i, j;
        isl_space *dim;
        isl_map *proj;
        isl_set *par;
        int shared_len = isl_set_dim(shared_domain, isl_dim_set);
        int sync = 0;
        isl_union_map *sched;

        shared_domain = isl_set_copy(shared_domain);
        sched = isl_union_map_copy(gen->tiled_sched);
        dim = isl_union_map_get_space(sched);
        proj = projection(dim, gen->tiled_len, shared_len);
        sched = isl_union_map_apply_range(sched, isl_union_map_from_map(proj));
        sched = isl_union_map_intersect_range(sched,
                        isl_union_set_from_set(isl_set_copy(shared_domain)));
        if (shared_len != gen->shared_len) {
                dim = isl_union_map_get_space(sched);
                proj = projection(dim, gen->shared_len, shared_len);
                proj = isl_map_reverse(proj);
                shared_domain = isl_set_apply(shared_domain,
                                                isl_map_copy(proj));
                sched = isl_union_map_apply_range(sched,
                                isl_union_map_from_map(proj));
        }

        dim = isl_union_map_get_space(sched);
        par = parametrization(dim, gen->shared_len, 0, gen->shared_len, "g");
        sched = isl_union_map_intersect_range(sched,
                                                isl_union_set_from_set(par));

        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                if (gen->array[i].print_shared_level != level)
                        continue;

                for (j = 0; j < array->n_group; ++j) {
                        if (print_group_shared_accesses(gen, array->groups[j],
                                            type, shared_domain, sched))
                                sync = 1;
                }
        }

        isl_union_map_free(sched);
        isl_set_free(shared_domain);

        if (sync) {
                print_indent(gen->cuda.kernel_c, gen->kernel_code.indent);
                fprintf(gen->cuda.kernel_c, "__syncthreads();\n");
        }
}

/* Given an index expression into a tile of an array, adjust the expression
 * to a shift of the tile to the origin
 * (based on the lower bounds in array->shared_bound).
 * If the index is strided, then we first add
 * bound->shift and divide by bound->stride.
 */
static __isl_give isl_qpolynomial *shift_index(__isl_take isl_qpolynomial *qp,
        struct cuda_array_info *array,
        struct cuda_array_bound *bound, __isl_take isl_set *domain)
{
        isl_qpolynomial *lb;

        if (bound->shift) {
                isl_qpolynomial *shift, *t;
                isl_int one;
                isl_space *dim;
                shift = bound->shift;
                shift = isl_qpolynomial_copy(shift);
                shift = isl_qpolynomial_drop_dims(shift, isl_dim_set, 0,
                            isl_qpolynomial_dim(shift, isl_dim_set));
                shift = isl_qpolynomial_align_params(shift,
                                          isl_qpolynomial_get_space(qp));
                qp = isl_qpolynomial_add(qp, shift);
                dim = isl_qpolynomial_get_space(qp);
                isl_int_init(one);
                isl_int_set_si(one, 1);
                t = isl_qpolynomial_rat_cst(dim, one, bound->stride);
                isl_int_clear(one);
                qp = isl_qpolynomial_mul(qp, t);
        }

        lb = isl_qpolynomial_from_aff(isl_aff_copy(bound->lb));
        lb = isl_qpolynomial_drop_dims(lb, isl_dim_set, 0,
                        isl_qpolynomial_dim(lb, isl_dim_set));

        lb = isl_qpolynomial_align_params(lb, isl_qpolynomial_get_space(qp));

        qp = isl_qpolynomial_sub(qp, lb);
        qp = isl_qpolynomial_gist(qp, domain);

        return qp;
}

/* This function is called for each access to an array in some statement
 * in the original code.
 * Replace that access by an access to shared or (linearized) global memory.
 * Since the array in shared memory is just
 * a shifted copy of part of the original array, we simply need
 * to subtract the lower bound, which was computed
 * in can_tile_for_shared_memory.
 * If any of the indices is strided, then we first add
 * shared_bound[i].shift and divide by shared_bound[i].stride.
 *
 * If the given array is accessed directly from global memory,
 * we don't need to perform any shifting and simply simplify
 * expression in the context of the domain instead.
 *
 * If the array space (range of access) has no name, then we are
 * accessing an iterator in the original program.
 */
static void print_access(struct cuda_gen *gen, __isl_take isl_map *access,
        int group_nr)
{
        int i;
        const char *name;
        unsigned n_index;
        struct cuda_array_info *array = NULL;
        isl_printer *prn;
        isl_basic_set *aff;
        isl_set *data_set;
        isl_set *domain;
        struct cuda_array_bound *bounds = NULL;

        access = isl_map_align_params(access,
                                        isl_set_get_space(gen->stmt_domain));

        data_set = isl_set_apply(isl_set_copy(gen->stmt_domain), access);

        name = isl_set_get_tuple_name(data_set);

        if (!name)
                fprintf(gen->cuda.kernel_c, "(");
        else {
                struct cuda_array_ref_group *group;

                for (i = 0; i < gen->n_array; ++i) {
                        if (strcmp(name, gen->array[i].name))
                                continue;
                        array = &gen->array[i];
                }
                assert(array);
                group = array->groups[group_nr];
                bounds = group->private_bound;
                if (!bounds)
                        bounds = group->shared_bound;

                print_array_name(gen->cuda.kernel_c, group);

                if (cuda_array_is_scalar(array)) {
                        isl_set_free(data_set);
                        return;
                }

                fprintf(gen->cuda.kernel_c, "[");
        }


        n_index = isl_set_dim(data_set, isl_dim_set);
        aff = isl_set_affine_hull(data_set);

        prn = isl_printer_to_file(gen->ctx, gen->cuda.kernel_c);
        prn = isl_printer_set_output_format(prn, ISL_FORMAT_C);

        if (!bounds)
                for (i = 0; i + 1 < n_index; ++i)
                        prn = isl_printer_print_str(prn, "(");

        for (i = 0; i < n_index; ++i) {
                isl_constraint *c;
                isl_qpolynomial *qp;
                int ok;

                ok = isl_basic_set_has_defining_equality(aff,
                                                        isl_dim_out, i, &c);
                assert(ok);
                qp = isl_qpolynomial_from_constraint(c, isl_dim_out, i);
                qp = isl_qpolynomial_drop_dims(qp, isl_dim_set, 0,
                                isl_qpolynomial_dim(qp, isl_dim_set));

                if (!array) {
                        prn = isl_printer_print_qpolynomial(prn, qp);
                        isl_qpolynomial_free(qp);
                        continue;
                }

                domain = isl_set_copy(gen->stmt_domain);
                domain = isl_set_project_out(domain, isl_dim_set, 0,
                                            isl_set_dim(domain, isl_dim_set));
                if (!bounds)
                        qp = isl_qpolynomial_gist(qp, domain);
                else
                        qp = shift_index(qp, array, &bounds[i], domain);

                if (i) {
                        if (!bounds) {
                                prn = isl_printer_print_str(prn, ") * (");
                                prn = isl_printer_print_pw_aff(prn,
                                                        array->local_bound[i]);
                                prn = isl_printer_print_str(prn, ") + ");
                        } else
                                prn = isl_printer_print_str(prn, "][");
                }
                prn = isl_printer_print_qpolynomial(prn, qp);
                isl_qpolynomial_free(qp);
        }
        if (!name)
                prn = isl_printer_print_str(prn, ")");
        else
                prn = isl_printer_print_str(prn, "]");
        isl_printer_free(prn);

        isl_basic_set_free(aff);
}

static struct cuda_stmt_access *print_expr(struct cuda_gen *gen, FILE *out,
        struct pet_expr *expr, struct cuda_stmt_access *access, int outer)
{
        int i;

        switch (expr->type) {
        case pet_expr_double:
                fprintf(out, "%g", expr->d);
                break;
        case pet_expr_access:
                print_access(gen, isl_map_copy(access->access), access->group);
                access = access->next;
                break;
        case pet_expr_unary:
                if (!outer)
                        fprintf(out, "(");
                fprintf(out, " %s ", pet_op_str(expr->op));
                access = print_expr(gen, out, expr->args[pet_un_arg],
                                        access, 0);
                if (!outer)
                        fprintf(out, ")");
                break;
        case pet_expr_binary:
                if (!outer)
                        fprintf(out, "(");
                access = print_expr(gen, out, expr->args[pet_bin_lhs],
                                        access, 0);
                fprintf(out, " %s ", pet_op_str(expr->op));
                access = print_expr(gen, out, expr->args[pet_bin_rhs],
                                        access, 0);
                if (!outer)
                        fprintf(out, ")");
                break;
        case pet_expr_ternary:
                if (!outer)
                        fprintf(out, "(");
                access = print_expr(gen, out, expr->args[pet_ter_cond],
                                        access, 0);
                fprintf(out, " ? ");
                access = print_expr(gen, out, expr->args[pet_ter_true],
                                        access, 0);
                fprintf(out, " : ");
                access = print_expr(gen, out, expr->args[pet_ter_false],
                                        access, 0);
                if (!outer)
                        fprintf(out, ")");
                break;
        case pet_expr_call:
                fprintf(out, "%s(", expr->name);
                for (i = 0; i < expr->n_arg; ++i) {
                        if (i)
                                fprintf(out, ", ");
                        access = print_expr(gen, out, expr->args[i],
                                                access, 1);
                }
                fprintf(out, ")");
        }
        return access;
}

static void print_stmt_body(struct cuda_gen *gen,
        FILE *out, struct cuda_stmt *stmt)
{
        print_expr(gen, out, stmt->body, stmt->accesses, 1);
        fprintf(out, ";\n");
}

/* This function is called for each leaf in the innermost clast,
 * i.e., for each statemetn.
 * We print the statement body, simplifying the accesses based
 * on the schedule.
 */
static void print_statement(struct gpucode_info *code,
        struct clast_user_stmt *u)
{
        struct cuda_gen *gen = code->user;
        isl_space *dim;
        isl_set *par;
        isl_set *stmt_domain;
        isl_union_map *stmt_sched;
        isl_union_set *uset;
        int nr;
        struct cuda_stmt *stmt;

        nr = atoi(u->statement->name + 2);
        stmt = &gen->stmts[nr];

        stmt_domain = extract_host_domain(u);

        stmt_sched = isl_union_map_intersect_range(
                    isl_union_map_copy(gen->local_sched),
                    isl_union_set_from_set(extend(stmt_domain,
                                                  gen->thread_tiled_len)));
        dim = isl_union_map_get_space(stmt_sched);
        par = parametrization(dim, gen->thread_tiled_len, 0,
                                gen->thread_tiled_len, "c");
        stmt_sched = isl_union_map_intersect_range(stmt_sched,
                                                isl_union_set_from_set(par));

        uset = isl_union_map_domain(stmt_sched);
        dim = isl_union_set_get_space(uset);
        dim = isl_space_add_dims(dim, isl_dim_set,
                          isl_set_dim(stmt->domain, isl_dim_set));
        dim = isl_space_set_tuple_name(dim, isl_dim_set, u->statement->name);
        gen->stmt_domain = isl_union_set_extract_set(uset, dim);
        isl_union_set_free(uset);

        print_indent(code->dst, code->indent);
        print_stmt_body(gen, code->dst, stmt);

        isl_set_free(gen->stmt_domain);
}

/* Print an access to the element in the global memory copy of the
 * given array that corresponds to element [qp[0]][qp[1]]...
 * of the original array.
 * The copy in global memory has been linearized, so we need to take
 * the array size into account.
 */
static void print_private_global_index(isl_ctx *ctx, FILE *out,
        struct cuda_array_info *array, __isl_keep isl_qpolynomial **qp)
{
        int i;
        isl_printer *prn;

        fprintf(out, "%s[", array->name);
        prn = isl_printer_to_file(ctx, out);
        prn = isl_printer_set_output_format(prn, ISL_FORMAT_C);
        for (i = 0; i + 1 < array->n_index; ++i)
                prn = isl_printer_print_str(prn, "(");
        for (i = 0; i < array->n_index; ++i) {
                if (i) {
                        prn = isl_printer_print_str(prn, ") * (");
                        prn = isl_printer_print_pw_aff(prn,
                                                        array->local_bound[i]);
                        prn = isl_printer_print_str(prn, ") + ");
                }
                prn = isl_printer_print_qpolynomial(prn, qp[i]);
        }
        isl_printer_free(prn);
        fprintf(out, "]");
}

/* Print an access to the element in the shared memory copy of the
 * given array reference group that corresponds to element [qps[0]][qps[1]]...
 * of the original array.
 * Since the array in shared memory is just a shifted copy of part
 * of the original array, we simply need to subtract the lower bound,
 * which was computed in can_tile_for_shared_memory.
 * If any of the indices is strided, then we first add
 * shared_bound[i].shift and divide by shared_bound[i].stride.
 */
static void print_private_local_index(isl_ctx *ctx, FILE *out,
        struct cuda_array_ref_group *group,
        __isl_keep isl_qpolynomial **qps, __isl_keep isl_set *domain)
{
        int i;
        isl_printer *prn;
        struct cuda_array_info *array = group->array;
        struct cuda_array_bound *bounds = group->private_bound;

        print_array_name(out, group);
        for (i = 0; i < array->n_index; ++i) {
                isl_qpolynomial *qp = isl_qpolynomial_copy(qps[i]);

                qp = shift_index(qp, array, &bounds[i], isl_set_copy(domain));

                fprintf(out, "[");
                prn = isl_printer_to_file(ctx, out);
                prn = isl_printer_set_output_format(prn, ISL_FORMAT_C);
                prn = isl_printer_print_qpolynomial(prn, qp);
                isl_printer_free(prn);
                fprintf(out, "]");
                isl_qpolynomial_free(qp);
        }
}

/* This function is called for each leaf in the clast of the code
 * for copying to or from private memory.
 * The statement name is read_private_<array> or write_private_<array>.
 *
 * The schedule iterates over the array elements, so we can use
 * the domain of private_sched at the current scheduling position
 * as the index of the array.
 */
static void print_private_copy_statement(struct gpucode_info *code,
        struct clast_user_stmt *u)
{
        struct cuda_gen *gen = code->user;
        isl_set *domain;
        isl_map *sched;
        struct cuda_array_ref_group *group = gen->private_group;
        int i;
        unsigned n_in;
        unsigned n_out;
        isl_space *dim;
        isl_set *param;
        isl_set *index;
        isl_basic_set *aff;
        isl_ctx *ctx;
        isl_qpolynomial **qp;
        int read;

        read = !strncmp(u->statement->name, "read", 4);

        domain = extract_host_domain(u);
        assert(domain);

        sched = isl_map_copy(gen->private_sched);
        sched = isl_map_reverse(sched);
        sched = isl_map_intersect_domain(sched, domain);
        n_in = isl_map_dim(sched, isl_dim_in);
        n_out = isl_map_dim(sched, isl_dim_out);
        dim = isl_map_get_space(sched);
        dim = isl_space_drop_dims(dim, isl_dim_in, 0, n_in);
        dim = isl_space_drop_dims(dim, isl_dim_out, 0, n_out);
        param = parametrization(dim, n_in, 0, n_in, "c");
        sched = isl_map_align_params(sched, isl_set_get_space(param));
        sched = isl_map_intersect_domain(sched, param);
        index = isl_map_range(sched);
        domain = isl_set_copy(index);
        aff = isl_set_affine_hull(index);
        domain = isl_set_project_out(domain, isl_dim_set, 0, n_out);

        ctx = isl_basic_set_get_ctx(aff);
        qp = isl_alloc_array(ctx, isl_qpolynomial *, n_out);
        assert(qp);

        for (i = 0; i < n_out; ++i) {
                isl_constraint *c;
                int ok;

                ok = isl_basic_set_has_defining_equality(aff,
                                                        isl_dim_set, i, &c);
                assert(ok);
                qp[i] = isl_qpolynomial_from_constraint(c, isl_dim_set, i);
                qp[i] = isl_qpolynomial_drop_dims(qp[i], isl_dim_set, 0, n_out);
        }

        print_indent(code->dst, code->indent);
        if (read) {
                print_private_local_index(ctx, code->dst, group, qp, domain);
                fprintf(code->dst, " = ");
                print_private_global_index(ctx, code->dst, group->array, qp);
        } else {
                print_private_global_index(ctx, code->dst, group->array, qp);
                fprintf(code->dst, " = ");
                print_private_local_index(ctx, code->dst, group, qp, domain);
        }
        fprintf(code->dst, ";\n");

        for (i = 0; i < n_out; ++i)
                isl_qpolynomial_free(qp[i]);
        free(qp);

        isl_basic_set_free(aff);
        isl_set_free(domain);
}

static void print_private_access(struct cuda_gen *gen,
        __isl_keep isl_set *shared_domain, __isl_take isl_set *access,
        const char *type, struct cuda_array_ref_group *group)
{
        const char *array_name;
        char *name;
        isl_ctx *ctx;
        unsigned nvar = isl_set_dim(access, isl_dim_set);
        isl_union_map *usched;

        if (isl_set_fast_is_empty(access)) {
                isl_set_free(access);
                return;
        }

        ctx = isl_set_get_ctx(access);
        array_name = isl_set_get_tuple_name(access);
        name = isl_alloc_array(ctx, char,
                strlen(type) + sizeof("_private_") + strlen(array_name) + 20);
        if (group->array->n_group > 1)
                sprintf(name, "%s_private_%s_%d", type, array_name, group->nr);
        else
                sprintf(name, "%s_private_%s", type, array_name);
        access = isl_set_set_tuple_name(access, name);
        free(name);

        gen->private_sched = shift_access(access, group);
        gen->private_group = group;

        usched = isl_union_map_from_map(isl_map_copy(gen->private_sched));
        print_shared_body(gen, shared_domain, usched, nvar,
                                &print_private_copy_statement, 1);
        isl_union_map_free(usched);

        isl_map_free(gen->private_sched);
}

/* Print code for reading into or writing from private memory
 * the given array reference group.
 *
 * sched maps the original iteration domains to the shared memory tile loops.
 */
static void print_group_private_accesses(struct cuda_gen *gen,
        struct cuda_array_ref_group *group,
         const char *type, __isl_keep isl_set *shared_domain,
        unsigned first_shared, int shared_len, __isl_keep isl_union_map *sched)
{
        int read;
        isl_union_map *access;
        isl_union_set *uset;
        isl_set *access_set;

        if (!group->private_bound)
                return;

        read = !strcmp(type, "read");

        access = group_access_relation(group, read, !read);
        access = isl_union_map_apply_domain(access, isl_union_map_copy(sched));
        access = isl_union_map_intersect(access,
                                    isl_union_map_copy(gen->private_access));
        uset = isl_union_map_range(access);

        if (isl_union_set_is_empty(uset)) {
                isl_union_set_free(uset);
                return;
        }

        access_set = isl_set_from_union_set(uset);
        access_set = isl_set_coalesce(access_set);
        access_set = isl_set_eliminate(access_set, isl_dim_param,
                        first_shared + shared_len,
                        gen->shared_len - shared_len);

        print_private_access(gen, shared_domain, access_set, type, group);
}

/* Print code for reading into or writing from private memory at
 * the given level (-1 for innermost).
 *
 * If we are not printing at the innermost level, then the dimensionality
 * of shared_domain may be smaller than gen->shared_len.
 * As the rest of the code assumes that the domain of access has
 * gen->shared_len dimensions, we therefore may need to embed this domain
 * in a higher dimensional space after intersection with shared_domain.
 *
 * This code is very similar to print_shared_accesses.
 * The main difference is that we to take into account gen->private_access.
 */
static void print_private_accesses(struct cuda_gen *gen,
        __isl_keep isl_set *shared_domain, __isl_keep isl_union_map *access,
        const char *type, int level)
{
        int i, j;
        isl_space *dim;
        isl_map *proj;
        int shared_len = isl_set_dim(shared_domain, isl_dim_set);
        unsigned first_shared;
        isl_union_map *sched;

        shared_domain = isl_set_copy(shared_domain);
        sched = isl_union_map_copy(gen->tiled_sched);
        dim = isl_union_map_get_space(sched);
        first_shared = isl_space_dim(dim, isl_dim_param);
        proj = projection(dim, gen->tiled_len, shared_len);
        sched = isl_union_map_apply_range(sched, isl_union_map_from_map(proj));
        sched = isl_union_map_intersect_range(sched,
                        isl_union_set_from_set(isl_set_copy(shared_domain)));
        if (shared_len != gen->shared_len) {
                dim = isl_union_map_get_space(sched);
                proj = projection(dim, gen->shared_len, shared_len);
                proj = isl_map_reverse(proj);
                shared_domain = isl_set_apply(shared_domain,
                                                isl_map_copy(proj));
                sched = isl_union_map_apply_range(sched,
                                isl_union_map_from_map(proj));
        }

        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                if (gen->array[i].print_shared_level != level)
                        continue;

                for (j = 0; j < array->n_group; ++j)
                        print_group_private_accesses(gen, array->groups[j],
                                            type, shared_domain,
                                            first_shared, shared_len, sched);
        }

        isl_union_map_free(sched);
        isl_set_free(shared_domain);
}

/* Set unroll[j] if the input dimension j is involved in
 * the index expression represented by bmap.
 */
static int check_unroll(__isl_take isl_basic_map *bmap, void *user)
{
        int i, j;
        int n_in = isl_basic_map_dim(bmap, isl_dim_in);
        int n_out = isl_basic_map_dim(bmap, isl_dim_out);
        int *unroll = user;

        for (i = 0; i < n_out; ++i) {
                isl_constraint *c;
                int ok;

                ok = isl_basic_map_has_defining_equality(bmap,
                                                        isl_dim_out, i, &c);
                assert(ok);
                for (j = 0; j < n_in; ++j)
                        if (isl_constraint_involves_dims(c, isl_dim_in, j, 1))
                                unroll[j] = 1;
                isl_constraint_free(c);
        }

        isl_basic_map_free(bmap);
        return 0;
}

/* Given an array pos mapping input dimensions to the corresponding
 * output dimension, construct the corresponding map.
 */
static __isl_give isl_map *permutation(__isl_take isl_space *dim,
        int *pos, int len)
{
        int i;
        isl_constraint *c;
        isl_basic_map *bmap;

        dim = isl_space_add_dims(dim, isl_dim_in, len);
        dim = isl_space_add_dims(dim, isl_dim_out, len);
        bmap = isl_basic_map_universe(isl_space_copy(dim));

        for (i = 0; i < len; ++i) {
                c = isl_equality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_in, i, -1);
                isl_constraint_set_coefficient_si(c, isl_dim_out, pos[i], 1);
                bmap = isl_basic_map_add_constraint(bmap, c);
        }
        isl_space_free(dim);

        return isl_map_from_basic_map(bmap);
}

/* Find all loops involved in any of the index expressions for any of
 * the private accesses, move them innermost and then mark them as
 * requiring unrolling by setting gen->first_unroll.
 * The loops involved should all be parallel because of the checks
 * we performed in check_private_group_access.  Moving them innermost
 * is therefore a valid transformation.
 */
static __isl_give isl_union_map *interchange_for_unroll(struct cuda_gen *gen,
        __isl_take isl_union_map *sched)
{
        int i, j;
        int unroll[gen->thread_tiled_len];
        int perm[gen->thread_tiled_len];
        isl_space *dim;
        isl_map *permute;
        int len = gen->shared_len + gen->n_parallel + gen->n_block;

        gen->first_unroll = -1;

        for (i = 0; i < gen->thread_tiled_len; ++i)
                unroll[i] = 0;
        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                for (j = 0; j < array->n_group; ++j) {
                        isl_union_map *access;
                        isl_map *acc;

                        if (!array->groups[j]->private_bound)
                                continue;

                        access = group_access_relation(array->groups[j], 1, 1);
                        access = isl_union_map_apply_domain(access,
                                                isl_union_map_copy(sched));

                        acc = isl_map_from_union_map(access);
                        isl_map_foreach_basic_map(acc, &check_unroll, unroll);

                        isl_map_free(acc);
                }
        }

        for (i = 0; i < gen->shared_len; ++i)
                if (unroll[i])
                        return sched;

        for (i = gen->shared_len; i < len; ++i)
                if (unroll[i])
                        break;

        if (i >= len)
                return sched;

        for (i = len; i < gen->thread_tiled_len; ++i)
                if (unroll[i])
                        return sched;

        j = 0;
        for (i = 0; i < gen->thread_tiled_len; ++i)
                if (!unroll[i])
                        perm[i] = j++;
        gen->first_unroll = 1 + j;
        for (i = 0; i < len; ++i)
                if (unroll[i])
                        perm[i] = j++;

        dim = isl_union_map_get_space(sched);
        permute = permutation(dim, perm, gen->thread_tiled_len);
        sched = isl_union_map_apply_range(sched,
                                          isl_union_map_from_map(permute));

        return sched;
}

/* This function is called for each leaf in the clast of the kernel code.
 * We first specialize the schedule to the site of the leaf and
 * print code for reading into shared memory, performing the actual
 * computations and writing from shared memory, with the required
 * synchronizations.
 */
static void print_kernel_user(struct gpucode_info *code,
        struct clast_user_stmt *u)
{
        struct cuda_gen *gen = code->user;
        isl_set *shared_domain;

        shared_domain = extract_entire_host_domain(u);

        print_shared_accesses(gen, shared_domain, gen->read, "read", -1);

        print_private_accesses(gen, shared_domain, gen->read, "read", -1);

        print_shared_body(gen, shared_domain, gen->local_sched,
                            gen->thread_tiled_len, &print_statement,
                            gen->first_unroll);

        print_private_accesses(gen, shared_domain, gen->write, "write", -1);

        print_indent(gen->cuda.kernel_c, gen->kernel_code.indent);
        fprintf(gen->cuda.kernel_c, "__syncthreads();\n");

        print_shared_accesses(gen, shared_domain, gen->write, "write", -1);

        isl_set_free(shared_domain);
}

/* Check if we need to perform any copying to shared memory at this level
 * and if so, print the copying instructions.
 * Any array for which we are allowed to print copying instructions at
 * this level, but haven't done so already, is printed.
 */
static void print_kernel_for_head(struct gpucode_info *code,
        struct clast_for *f)
{
        int i;
        struct cuda_gen *gen = code->user;
        isl_set *domain;
        int level;
        int print = 0;

        domain = isl_set_from_cloog_domain(cloog_domain_copy(f->domain));
        level = isl_set_dim(domain, isl_dim_set) - 1;

        for (i = 0; i < gen->n_array; ++i) {
                if (gen->array[i].print_shared_level >= 0)
                        continue;
                if (gen->array[i].last_shared > level)
                        continue;
                gen->array[i].print_shared_level = level;
                print = 1;
        }

        if (print) {
                print_shared_accesses(gen, domain, gen->read, "read", level);
                print_private_accesses(gen, domain, gen->read, "read", level);
        }

        isl_set_free(domain);
}

/* Print instructions for copying from shared memory for each array
 * for which print_kernel_for_head has added copying instructions
 * to shared memory.
 */
static void print_kernel_for_foot(struct gpucode_info *code,
        struct clast_for *f)
{
        int i;
        struct cuda_gen *gen = code->user;
        isl_set *domain;
        int level;
        int print = 0;

        domain = isl_set_from_cloog_domain(cloog_domain_copy(f->domain));
        level = isl_set_dim(domain, isl_dim_set) - 1;

        for (i = 0; i < gen->n_array; ++i) {
                if (gen->array[i].print_shared_level != level)
                        continue;
                print = 1;
                break;
        }

        if (print) {
                print_private_accesses(gen, domain, gen->write, "write", level);
                print_shared_accesses(gen, domain, gen->write, "write", level);
        }

        isl_set_free(domain);
}

/* Use CLooG to generate code for the outer gen->shared_first loops
 * of the local schedule "sched".
 * The pretty printing of this code is handled by gpu_print_host_stmt,
 * which calls print_kernel_user for each iteration of the shared tile loops.
 */
static void print_cloog_kernel_body(struct cuda_gen *gen,
        __isl_keep isl_set *context, __isl_keep isl_union_map *sched)
{
        int i;
        CloogOptions *options;
        CloogDomain *cloog_context;
        CloogUnionDomain *ud;
        CloogInput *input;
        struct clast_stmt *stmt;
        char name[20];

        sched = isl_union_map_copy(sched);
        sched = isl_union_map_align_params(sched, isl_set_get_space(context));

        options = cloog_options_malloc(gen->state);
        options->language = LANGUAGE_C;
        options->strides = 1;
        options->sh = 1;
        options->stop = gen->shared_len;
        options->f = gen->tiled_len;
        options->l = gen->tiled_len;
        options->save_domains = 1;
        options->noscalars = 1;

        ud = cloog_union_domain_from_isl_union_map(sched);
        for (i = 0; i < gen->shared_len; ++i) {
                snprintf(name, sizeof(name), "g%d", i);
                ud = cloog_union_domain_set_name(ud, CLOOG_SCAT, i, name);
        }
        cloog_context = cloog_domain_from_isl_set(isl_set_copy(context));
        input = cloog_input_alloc(cloog_context, ud);

        stmt = cloog_clast_create_from_input(input, options);

        gen->kernel_code.indent = 4;
        gen->kernel_code.dst = gen->cuda.kernel_c;
        gen->kernel_code.print_user_stmt = NULL;
        gen->kernel_code.print_user_stmt_list = &print_kernel_user;
        gen->kernel_code.print_for_head = &print_kernel_for_head;
        gen->kernel_code.print_for_foot = &print_kernel_for_foot;
        gen->kernel_code.user = gen;
        gpu_print_host_stmt(&gen->kernel_code, stmt);

        cloog_clast_free(stmt);
        cloog_options_free(options);
}

static void print_kernel_iterators(struct cuda_gen *gen)
{
        int i;
        const char *block_dims[] = { "blockIdx.x", "blockIdx.y" };
        const char *thread_dims[] = { "threadIdx.x", "threadIdx.y",
                                        "threadIdx.z" };

        if (gen->n_grid > 0) {
                print_indent(gen->cuda.kernel_c, 4);
                fprintf(gen->cuda.kernel_c, "int ");
                for (i = 0; i < gen->n_grid; ++i) {
                        if (i)
                                fprintf(gen->cuda.kernel_c, ", ");
                        fprintf(gen->cuda.kernel_c, "b%d = %s",
                                i, block_dims[gen->n_grid - 1 - i]);
                }
                fprintf(gen->cuda.kernel_c, ";\n");
        }

        if (gen->n_block > 0) {
                print_indent(gen->cuda.kernel_c, 4);
                fprintf(gen->cuda.kernel_c, "int ");
                for (i = 0; i < gen->n_block; ++i) {
                        if (i)
                                fprintf(gen->cuda.kernel_c, ", ");
                        fprintf(gen->cuda.kernel_c, "t%d = %s",
                                i, thread_dims[gen->n_block - 1 - i]);
                }
                fprintf(gen->cuda.kernel_c, ";\n");
        }
}

static void print_group_shared_array(struct cuda_gen *gen,
        struct cuda_array_ref_group *group)
{
        int j;
        struct cuda_array_bound *bounds;

        bounds = group->private_bound;
        if (!bounds)
                bounds = group->shared_bound;
        if (!bounds)
                return;

        print_indent(gen->cuda.kernel_c, 4);
        fprintf(gen->cuda.kernel_c, "%s%s ",
                group->private_bound ? "" : "__shared__ ", group->array->type);
        print_array_name(gen->cuda.kernel_c, group);
        for (j = 0; j < group->array->n_index; ++j) {
                fprintf(gen->cuda.kernel_c, "[");
                isl_int_print(gen->cuda.kernel_c, bounds[j].size, 0);
                fprintf(gen->cuda.kernel_c, "]");
        }
        fprintf(gen->cuda.kernel_c, ";\n");
}

static void print_shared_arrays(struct cuda_gen *gen)
{
        int i, j;

        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                for (j = 0; j < array->n_group; ++j)
                        print_group_shared_array(gen, array->groups[j]);
        }
}

static void print_kernel_body(struct cuda_gen *gen,
        __isl_keep isl_set *host_domain, __isl_keep isl_union_map *sched)
{
        isl_set *context;

        context = isl_set_copy(host_domain);
        context = parametrize(context, 0, gen->tile_first, "h");
        context = isl_set_project_out(context, isl_dim_set, 0, gen->tile_first);
        context = add_bounded_parameters(context,
                                        gen->n_grid, gen->grid_dim, "b");

        print_kernel_iterators(gen);
        print_shared_arrays(gen);

        fprintf(gen->cuda.kernel_c, "\n");

        print_cloog_kernel_body(gen, context, sched);

        isl_set_free(context);
}

/* Given a constraint
 *
 *              a(p,i) + j = g f(e)
 *
 * or -a(p,i) - j = g f(e) if sign < 0,
 * store a(p,i) in bound->shift and g (stride) in bound->stride.
 * a(p,i) is assumed to be an expression in only the parameters.
 */
static void extract_stride(__isl_keep isl_constraint *c,
        struct cuda_array_bound *bound, isl_int stride, int sign)
{
        int i;
        isl_int v;
        isl_int one;
        isl_space *dim;
        unsigned nparam;
        isl_qpolynomial *qp;

        isl_int_set(bound->stride, stride);

        dim = isl_constraint_get_space(c);
        dim = isl_space_params(dim);

        nparam = isl_space_dim(dim, isl_dim_param);

        isl_int_init(v);
        isl_int_init(one);
        isl_int_set_si(one, 1);

        isl_constraint_get_constant(c, &v);
        if (sign < 0)
                isl_int_neg(v, v);
        qp = isl_qpolynomial_rat_cst(isl_space_copy(dim), v, one);

        for (i = 0; i < nparam; ++i) {
                isl_qpolynomial *t, *p;

                isl_constraint_get_coefficient(c, isl_dim_param, i, &v);
                if (isl_int_is_zero(v))
                        continue;
                if (sign < 0)
                        isl_int_neg(v, v);
                t = isl_qpolynomial_rat_cst(isl_space_copy(dim), v, one);
                p = isl_qpolynomial_var(isl_space_copy(dim), isl_dim_param, i);
                t = isl_qpolynomial_mul(t, p);
                qp = isl_qpolynomial_add(qp, t);
        }

        isl_space_free(dim);
        isl_int_clear(one);
        isl_int_clear(v);

        bound->shift = qp;
}

/* Given an equality constraint of a map with a single output dimension j,
 * check if the constraint is of the form
 *
 *              a(p,i) + j = g f(e)
 *
 * with a(p,i) an expression in the parameters and input dimensions
 * and f(e) an expression in the existentially quantified variables.
 * If so, and if g is larger than any such g from a previously considered
 * constraint, then call extract_stride. to record the stride information
 * in bound.
 */
static int check_stride_constraint(__isl_take isl_constraint *c, void *user)
{
        int i;
        isl_int v, stride;
        unsigned n_div;
        struct cuda_array_bound *bound = user;

        isl_int_init(v);
        isl_int_init(stride);

        n_div = isl_constraint_dim(c, isl_dim_div);
        isl_constraint_get_coefficient(c, isl_dim_out, 0, &v);

        if (n_div && (isl_int_is_one(v) || isl_int_is_negone(v))) {
                int s = isl_int_sgn(v);
                isl_int_set_si(stride, 0);
                for (i = 0; i < n_div; ++i) {
                        isl_constraint_get_coefficient(c, isl_dim_div, i, &v);
                        isl_int_gcd(stride, stride, v);
                }
                if (!isl_int_is_zero(stride) &&
                    isl_int_gt(stride, bound->stride))
                        extract_stride(c, bound, stride, s);
        }

        isl_int_clear(stride);
        isl_int_clear(v);

        isl_constraint_free(c);
        return 0;
}

/* Given contraints on an array index i, check if we can find
 * a shift a(p) and a stride g such that
 *
 *      a(p) + i = 0 mod g
 *
 * If so, record the information in bound and apply the mapping
 * i -> (i + a(p))/g to the array index in bounds and return
 * the new constraints.
 * If not, simply return the original constraints.
 */
static __isl_give isl_basic_map *check_stride(struct cuda_gen *gen,
        struct cuda_array_bound *bound, __isl_take isl_basic_map *bounds)
{
        isl_space *dim;
        isl_basic_map *aff;
        isl_basic_map *shift;
        isl_qpolynomial *qp, *t;
        isl_int one;

        isl_int_set_si(bound->stride, -1);

        aff = isl_basic_map_affine_hull(isl_basic_map_copy(bounds));

        isl_basic_map_foreach_constraint(aff, &check_stride_constraint, bound);

        isl_basic_map_free(aff);

        if (isl_int_is_neg(bound->stride))
                return bounds;

        qp = isl_qpolynomial_copy(bound->shift);
        qp = isl_qpolynomial_add_dims(qp, isl_dim_set, 1);
        dim = isl_qpolynomial_get_space(qp);
        t = isl_qpolynomial_var(isl_space_copy(dim), isl_dim_set, 0);
        qp = isl_qpolynomial_add(qp, t);
        isl_int_init(one);
        isl_int_set_si(one, 1);
        t = isl_qpolynomial_rat_cst(dim, one, bound->stride);
        isl_int_clear(one);
        qp = isl_qpolynomial_mul(qp, t);
        shift = isl_basic_map_from_qpolynomial(qp);

        bound->shift_map = isl_basic_map_copy(shift);
        bounds = isl_basic_map_apply_range(bounds, shift);

        return bounds;
}

struct cuda_size_info {
        isl_basic_set *bset;
        struct cuda_array_bound *bound;
        int pos;
};

/* Given a constraint from the basic set describing the bounds on
 * an array index, check if it is a lower bound, say m i >= b(x), and,
 * if so, check whether the expression "i - ceil(b(x)/m) + 1" has a constant
 * upper bound.  If so, and if this bound is smaller than any bound
 * derived from earlier constraints, set the size to this bound on
 * the expression and the lower bound to ceil(b(x)/m).
 */
static int compute_size_in_direction(__isl_take isl_constraint *c, void *user)
{
        struct cuda_size_info *size = user;
        unsigned nparam;
        unsigned n_div;
        isl_int v;

        nparam = isl_basic_set_dim(size->bset, isl_dim_param);
        n_div = isl_constraint_dim(c, isl_dim_div);

        if (isl_constraint_involves_dims(c, isl_dim_div, 0, n_div)) {
                isl_constraint_free(c);
                return 0;
        }

        isl_int_init(v);

        isl_constraint_get_coefficient(c, isl_dim_set, size->pos, &v);

        if (isl_int_is_pos(v)) {
                isl_aff *aff;
                isl_aff *lb;
                enum isl_lp_result res;

                aff = isl_constraint_get_bound(c, isl_dim_set, size->pos);
                aff = isl_aff_ceil(aff);

                lb = isl_aff_copy(aff);

                aff = isl_aff_neg(aff);
                aff = isl_aff_add_coefficient_si(aff, isl_dim_set, size->pos, 1);

                res = isl_basic_set_max(size->bset, aff, &v);
                isl_aff_free(aff);

                if (res == isl_lp_ok) {
                        isl_int_add_ui(v, v, 1);
                        if (isl_int_is_neg(size->bound->size) ||
                            isl_int_lt(v, size->bound->size)) {
                                isl_int_set(size->bound->size, v);
                                lb = isl_aff_drop_dims(lb, isl_dim_set,
                                                        0, size->pos + 1);
                                isl_aff_free(size->bound->lb);
                                size->bound->lb = isl_aff_copy(lb);
                        }
                }
                isl_aff_free(lb);
        }

        isl_int_clear(v);
        isl_constraint_free(c);

        return 0;
}

/* Given a basic map "bounds" that maps parameters and input dimensions
 * to a single output dimension, look for an expression in the parameters
 * and input dimensions such that the range of the output dimension shifted
 * by this expression is a constant.
 *
 * In particular, we currently only consider lower bounds on the output
 * dimension as candidate expressions.
 */
static int compute_array_dim_size(struct cuda_gen *gen,
        struct cuda_array_bound *bound, __isl_take isl_basic_map *bounds)
{
        struct cuda_size_info size;

        bounds = check_stride(gen, bound, bounds);

        isl_int_set_si(bound->size, -1);
        bound->lb = NULL;

        size.bound = bound;
        size.pos = isl_basic_map_dim(bounds, isl_dim_in);
        size.bset = isl_basic_map_wrap(bounds);
        size.bset = isl_basic_set_flatten(size.bset);
        isl_basic_set_foreach_constraint(size.bset, &compute_size_in_direction,
                                        &size);
        isl_basic_set_free(size.bset);

        return isl_int_is_nonneg(bound->size) ? 0 : -1;
}

/* Check if we can find a shared memory tile for the given array
 * based on the given accesses, and if so, put the results
 * in array->shared_bound.
 *
 * We project the accesses on each index in turn and look for a parametric
 * offset such that the size is constant.
 */
static int can_tile_for_shared_memory(struct cuda_gen *gen,
        struct cuda_array_info *array, __isl_keep isl_map *access,
        struct cuda_array_bound *bounds)
{
        int i;

        for (i = 0; i < array->n_index; ++i) {
                isl_map *access_i;
                isl_basic_map *hull;

                access_i = isl_map_copy(access);
                access_i = isl_map_project_out(access_i, isl_dim_out, 0, i);
                access_i = isl_map_project_out(access_i, isl_dim_out,
                                            1, array->n_index - (i + 1));
                access_i = isl_map_compute_divs(access_i);
                hull = isl_map_simple_hull(access_i);
                if (compute_array_dim_size(gen, &bounds[i], hull) < 0)
                        return 0;
        }

        return 1;
}

/* Construct a map with input the shared tile loops and the loops that
 * will be wrapped around the threads that relates these later loops
 * to the thread indices and the projects them out.
 */
static __isl_give isl_map *compute_privatization(struct cuda_gen *gen)
{
        isl_map *priv;
        isl_map *tiling;
        isl_map *proj;
        isl_set *par;
        isl_space *dim;

        dim = isl_union_map_get_space(gen->shared_sched);

        if (gen->options->wrap)
                tiling = wrap(isl_space_copy(dim), gen->shared_len + gen->n_block,
                                gen->shared_len, gen->n_block, gen->block_dim);
        else
                tiling = tile(isl_space_copy(dim), gen->shared_len + gen->n_block,
                                gen->shared_len, gen->n_block, gen->block_dim);

        priv = tiling;

        par = parametrization(dim, gen->shared_len + 2 * gen->n_block,
                gen->tile_first + gen->tile_len + gen->n_grid + gen->n_block,
                gen->n_block, "t");

        priv = isl_map_align_params(priv, isl_set_get_space(par));
        priv = isl_map_intersect_range(priv, par);

        dim = isl_map_get_space(priv);
        dim = isl_space_drop_dims(dim, isl_dim_in, 0, isl_space_dim(dim, isl_dim_in));
        dim = isl_space_drop_dims(dim, isl_dim_out, 0, isl_space_dim(dim, isl_dim_out));
        proj = projection(dim, gen->shared_len + 2 * gen->n_block,
                          gen->shared_len);

        priv = isl_map_apply_range(priv, proj);

        return priv;
}

/* Construct a map from domain_dim to domain_dim that increments
 * the dimension at position "pos" and leaves all other dimensions
 * constant.
 */
static __isl_give isl_map *next(__isl_take isl_space *domain_dim, int pos)
{
        int i;
        int len = isl_space_dim(domain_dim, isl_dim_set);
        isl_space *dim;
        isl_basic_map *next;

        dim = isl_space_map_from_set(domain_dim);
        next = isl_basic_map_universe(isl_space_copy(dim));

        for (i = 0; i < len; ++i) {
                isl_constraint *c;

                c = isl_equality_alloc(isl_space_copy(dim));
                isl_constraint_set_coefficient_si(c, isl_dim_in, i, 1);
                isl_constraint_set_coefficient_si(c, isl_dim_out, i, -1);
                if (i == pos)
                        isl_constraint_set_constant_si(c, 1);
                next = isl_basic_map_add_constraint(next, c);
        }

        isl_space_free(dim);

        return isl_map_from_basic_map(next);
}

/* Check if the given access is coalesced.
 * That is, check whether incrementing the dimension that will get
 * wrapped over the last thread index results in incrementing
 * the last array index.
 *
 * This function is only called for access relations without reuse.
 */
static int access_is_coalesced(struct cuda_gen *gen,
        __isl_keep isl_union_map *access)
{
        isl_space *dim;
        isl_map *access_map;
        isl_map *next_thread_x;
        isl_map *next_element;
        isl_map *map;
        int coalesced;

        access = isl_union_map_copy(access);
        access = isl_union_map_apply_domain(access,
                                isl_union_map_copy(gen->tiled_sched));
        access_map = isl_map_from_union_map(access);

        dim = isl_map_get_space(access_map);
        dim = isl_space_domain(dim);
        next_thread_x = next(dim, gen->shared_len + gen->n_block - 1);

        dim = isl_map_get_space(access_map);
        dim = isl_space_range(dim);
        next_element = next(dim, isl_space_dim(dim, isl_dim_set) - 1);

        map = isl_map_apply_domain(next_thread_x, isl_map_copy(access_map));
        map = isl_map_apply_range(map, access_map);

        coalesced = isl_map_is_subset(map, next_element);

        isl_map_free(next_element);
        isl_map_free(map);

        return coalesced;
}

/* For the given array reference group, check whether the access is private
 * to the thread.  That is, check that any given array element
 * is only accessed by a single thread.
 * We compute an access relation that maps the shared tile loop iterators
 * and the shared point loop iterators that will be wrapped over the
 * threads to the array elements.
 * We actually check that those iterators that will be wrapped
 * partition the array space.  This check is stricter than necessary
 * since several iterations may be mapped onto the same thread
 * and then they could be allowed to access the same memory elements,
 * but our check does not allow this situation.
 *
 * We also check that the index expression only depends on parallel
 * loops.  That way, we can move those loops innermost and unroll them.
 * Again, we use a test that is stricter than necessary.
 * We actually check whether the index expression only depends
 * on the iterators that are wrapped over the threads.
 * These are necessarily parallel, but there may be more parallel loops.
 *
 * Combining the injectivity of the first test with the single-valuedness
 * of the second test, we simply test for bijectivity.
 *
 * If it turns out we can use registers, we compute the private memory
 * tile size using can_tile_for_shared_memory, after introducing a dependence
 * on the thread indices.
 *
 * Before performing any of the above computations, we first check
 * if there is any reuse on the reference group.  If not, we simply
 * return.  If, moreover, the access is coalesced then we also remove
 * the shared memory tiling since we should just use global memory instead.
 */
static void check_private_group_access(struct cuda_gen *gen,
        struct cuda_array_ref_group *group)
{
        isl_map *acc;
        isl_union_map *access;
        int n_index = group->array->n_index;

        access = group_access_relation(group, 1, 1);
        if (isl_union_map_is_injective(access)) {
                if (group->shared_bound && access_is_coalesced(gen, access)) {
                        free_bound_list(group->shared_bound, n_index);
                        group->shared_bound = NULL;
                }
                isl_union_map_free(access);
                return;
        }
        access = isl_union_map_apply_domain(access,
                                        isl_union_map_copy(gen->shared_sched));

        acc = isl_map_from_union_map(access);

        if (!isl_map_is_bijective(acc)) {
                isl_map_free(acc);
                return;
        }

        group->private_bound = create_bound_list(gen->ctx, n_index);
        acc = isl_map_align_params(acc, isl_map_get_space(gen->privatization));
        acc = isl_map_apply_domain(acc, isl_map_copy(gen->privatization));
        if (!can_tile_for_shared_memory(gen, group->array, acc,
                                        group->private_bound)) {
                free_bound_list(group->private_bound, n_index);
                group->private_bound = NULL;
        }

        isl_map_free(acc);
}

/* Look for the last shared tile loop that affects the offset of the
 * shared or private tile and store the result in array->last_shared.
 */
static void set_last_shared(struct cuda_gen *gen,
        struct cuda_array_ref_group *group)
{
        int i, j;
        struct cuda_array_bound *bounds;
        unsigned first_shared = gen->first_shared;
        int n_index = group->array->n_index;

        bounds = group->private_bound;
        if (!bounds)
                bounds = group->shared_bound;
        if (!bounds)
                return;

        for (j = gen->shared_len - 1; j >= 0; --j) {
                for (i = 0; i < n_index; ++i) {
                        isl_aff *lb;
                        isl_qpolynomial *shift;

                        lb = bounds[i].lb;
                        if (isl_aff_involves_dims(lb, isl_dim_param,
                                                        first_shared + j, 1))
                                break;

                        shift = bounds[i].shift;
                        if (!shift)
                                continue;
                        if (isl_qpolynomial_involves_dims(shift, isl_dim_param,
                                                        first_shared + j, 1))
                                break;
                }
                if (i < n_index)
                        break;
        }
        group->array->last_shared = j;
}

/* Compute the sizes of all private arrays for the current kernel,
 * as well as the offsets of the private pieces in the original arrays.
 * If we cannot or don't want to privatize a given array group,
 * we use the shared memory tile sizes computed in
 * compute_group_shared_bound instead.
 *
 * If a given Array only has a single reference group and if we have
 * been able to find a privated or shared tile,
 * we also look for the last shared tile loop that affects the offset
 * (and therefore the array tile) and store the result in array->last_shared.
 *
 * A privatized copy of all access relations from reference groups that
 * are mapped to private memory is stored in gen->privatization.
 */
static void compute_private_size(struct cuda_gen *gen)
{
        int i, j;
        isl_union_map *private;

        if (!gen->options->use_private_memory)
                return;

        private = isl_union_map_empty(isl_union_map_get_space(gen->shared_sched));

        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                for (j = 0; j < array->n_group; ++j) {
                        check_private_group_access(gen, array->groups[j]);

                        if (!array->groups[j]->private_bound)
                                continue;

                        private = isl_union_map_union(private,
                                group_access_relation(array->groups[j], 1, 1));
                }

                array->last_shared = gen->shared_len - 1;
                array->print_shared_level = -1;

                if (array->n_group != 1)
                        continue;
                set_last_shared(gen, array->groups[0]);
        }

        if (isl_union_map_is_empty(private))
                isl_union_map_free(private);
        else {
                isl_union_map *priv;

                private = isl_union_map_apply_domain(private,
                                        isl_union_map_copy(gen->shared_sched));
                priv = isl_union_map_from_map(isl_map_copy(gen->privatization));
                private = isl_union_map_apply_domain(private, priv);
                gen->private_access = private;
        }
}

/* Fill up the groups array with singleton groups, i.e., one group
 * per reference, initializing the array, access, write and refs fields.
 * In particular the access field is initialized to the scheduled
 * access relation of the array reference.
 *
 * Return the number of elements initialized, i.e., the number of
 * active references in the current kernel.
 */
static int populate_array_references(struct cuda_gen *gen,
        struct cuda_array_info *array, __isl_keep isl_union_map *sched,
        struct cuda_array_ref_group **groups)
{
        int i;
        int n;
        isl_ctx *ctx = isl_union_map_get_ctx(sched);

        n = 0;
        for (i = 0; i < array->n_ref; ++i) {
                isl_union_map *umap;
                isl_map *map;
                struct cuda_array_ref_group *group;
                struct cuda_stmt_access *access = array->refs[i];

                map = isl_map_copy(access->access);
                umap = isl_union_map_from_map(map);
                umap = isl_union_map_apply_domain(umap,
                                isl_union_map_copy(sched));

                if (isl_union_map_is_empty(umap)) {
                        isl_union_map_free(umap);
                        continue;
                }

                map = isl_map_from_union_map(umap);

                group = isl_calloc_type(ctx, struct cuda_array_ref_group);
                assert(group);
                group->array = array;
                group->access = map;
                group->write = access->write;
                group->refs = &array->refs[i];

                groups[n++] = group;
        }

        return n;
}

static void free_array_ref_group(struct cuda_array_ref_group *group,
        int n_index)
{
        if (!group)
                return;
        free_bound_list(group->shared_bound, n_index);
        free_bound_list(group->private_bound, n_index);
        isl_map_free(group->access);
        free(group->refs);
        free(group);
}

/* If two groups have overlapping access relations and if one of them
 * involves a write, then merge the two groups into one.
 *
 * We keep track of the grouping in "leader".  leader[j] points to
 * an earlier group array element that belongs to the same group,
 * or the array element j itself if this element is the first in the group.
 *
 * Return the number of group leaders.
 */
static int group_overlapping_writes(int n,
        struct cuda_array_ref_group **groups, int *leader)
{
        int i, j;
        int n_group = n;

        for (i = 0; i < n; ++i) {
                int l = i;
                groups[l]->n_ref = 1;
                for (j = i - 1; j >= 0; --j) {
                        isl_map *map;
                        int empty;

                        if (leader[j] != j)
                                continue;
                        if (!groups[l]->write && !groups[j]->write)
                                continue;

                        map = isl_map_intersect(isl_map_copy(groups[l]->access),
                                            isl_map_copy(groups[j]->access));
                        empty = isl_map_is_empty(map);
                        isl_map_free(map);

                        if (empty)
                                continue;
                        
                        groups[j]->access = isl_map_union(groups[j]->access,
                                                        groups[l]->access);
                        groups[j]->write = 1;
                        groups[l]->access = NULL;
                        groups[j]->n_ref += groups[l]->n_ref;
                        l = leader[l] = j;
                        n_group--;
                }
                leader[i] = l;
        }

        return n_group;
}

/* Compute the size of the shared array corresponding to the given array
 * array refrence group, based on the accesses from the current kernel,
 * as well as the offset of the shared piece in the original array.
 */
static void compute_group_shared_bound(struct cuda_gen *gen,
        struct cuda_array_info *array, struct cuda_array_ref_group *group)
{
        isl_ctx *ctx = isl_space_get_ctx(array->dim);

        if (!gen->options->use_shared_memory)
                return;

        group->shared_bound = create_bound_list(ctx, array->n_index);
        if (!can_tile_for_shared_memory(gen, array, group->access,
                                        group->shared_bound)) {
                free_bound_list(group->shared_bound, array->n_index);
                group->shared_bound = NULL;
        }
}

/* Given an initial grouping of array references and shared memory tiles
 * for each group that allows for a shared memory tile, merge two groups
 * if both have a shared memory tile and if the merged group also has
 * a shared memory tile.
 *
 * Return the number of group leaders after merging.
 */
static int group_common_shared_memory_tile(struct cuda_gen *gen,
        struct cuda_array_info *array, int n,
        struct cuda_array_ref_group **groups, int *leader, int n_group)
{
        int i, j;
        isl_ctx *ctx = isl_space_get_ctx(array->dim);

        for (i = 0; n_group > 1 && i < n; ++i) {
                int l = i;
                if (leader[i] != i)
                        continue;
                if (!groups[i]->shared_bound)
                        continue;
                for (j = i - 1; j >= 0; --j) {
                        isl_map *map;
                        int empty;
                        struct cuda_array_bound *shared_bound;

                        if (leader[j] != j)
                                continue;
                        if (!groups[j]->shared_bound)
                                continue;

                        map = isl_map_intersect(isl_map_copy(groups[l]->access),
                                            isl_map_copy(groups[j]->access));
                        empty = isl_map_is_empty(map);
                        isl_map_free(map);

                        if (empty)
                                continue;

                        map = isl_map_union(isl_map_copy(groups[l]->access),
                                            isl_map_copy(groups[j]->access));
                        shared_bound = create_bound_list(ctx, array->n_index);
                        if (!can_tile_for_shared_memory(gen, array, map,
                                                        shared_bound)) {
                                isl_map_free(map);
                                free_bound_list(shared_bound, array->n_index);
                                continue;
                        }

                        free_bound_list(groups[j]->shared_bound,
                                        array->n_index);
                        groups[j]->shared_bound = shared_bound;
                        isl_map_free(groups[j]->access);
                        groups[j]->access = map;
                        groups[j]->n_ref += groups[l]->n_ref;
                        l = leader[l] = j;
                        n_group--;
                }
        }

        return n_group;
}

/* Extract an array of array reference groups from the array of references
 * and the grouping information in "leader".
 *
 * Store the results in array->n_group and array->groups.
 */
static void extract_array_groups(isl_ctx *ctx, struct cuda_array_info *array,
        int n, struct cuda_array_ref_group **groups, int *leader, int n_group)
{
        int i, j;

        for (i = 2; i < n; ++i)
                leader[i] = leader[leader[i]];

        array->n_group = n_group;
        array->groups = isl_alloc_array(ctx, struct cuda_array_ref_group *,
                                        n_group);
        assert(array->groups);

        j = 0;
        for (i = 0; i < n; ++i) {
                int k, l;
                struct cuda_stmt_access **refs;

                if (leader[i] != i) {
                        groups[i]->refs = NULL;
                        free_array_ref_group(groups[i], array->n_index);
                        continue;
                }

                refs = isl_alloc_array(ctx, struct cuda_stmt_access *,
                                        groups[i]->n_ref);
                assert(refs);
                l = 0;
                for (k = i; k < n; ++k)
                        if (leader[k] == i) {
                                refs[l++] = *groups[k]->refs;
                                (*groups[k]->refs)->group = j;
                        }

                groups[i]->refs = refs;
                groups[i]->nr = j;
                array->groups[j++] = groups[i];
        }
}

/* Group array references that should be considered together when
 * deciding whether to access them from private, shared or global memory.
 *
 * In particular, if two array references overlap and if one of them
 * is a write, then the two references are grouped together.
 * Furthermore, if two groups admit a shared memory tile and if the
 * combination of the two also admits a shared memory tile, we merge
 * the two groups.
 *
 * During the construction the group->refs field points to a single
 * array reference inside the array of array references, while
 * group->n_ref contains the number of element in leader that
 * (directly or indirectly) point to this group, provided the group
 * is a leader.
 */
static void group_array_references(struct cuda_gen *gen,
        struct cuda_array_info *array, __isl_keep isl_union_map *sched)
{
        int i;
        int n, n_group;
        isl_ctx *ctx = isl_union_map_get_ctx(sched);
        struct cuda_array_ref_group **groups;
        int *leader;

        groups = isl_calloc_array(ctx, struct cuda_array_ref_group *,
                                        array->n_ref);
        assert(groups);

        n = populate_array_references(gen, array, sched, groups);

        leader = isl_alloc_array(ctx, int, n);
        assert(leader);

        n_group = group_overlapping_writes(n, groups, leader);

        for (i = 0; i < n; ++i)
                if (leader[i] == i)
                        compute_group_shared_bound(gen, array, groups[i]);

        n_group = group_common_shared_memory_tile(gen, array, n, groups,
                                                  leader, n_group);

        extract_array_groups(ctx, array, n, groups, leader, n_group);

        free(leader);
        free(groups);
}

/* Take tiled_sched, project it onto the shared tile loops and
 * the loops that will be wrapped over the threads,
 * parametrize the shared tile loops and store the result in gen->shared_sched.
 * The position of the first of these parameters is stored in gen->first_shared.
 * Also compute a projection that projects out the loops that will be
 * wrapped over the threads and store this projection in gen->shared_proj.
 */
static void compute_shared_sched(struct cuda_gen *gen)
{
        isl_space *dim;
        isl_map *proj;
        isl_set *par;
        isl_union_map *sched;

        sched = isl_union_map_copy(gen->tiled_sched);

        dim = isl_union_map_get_space(sched);
        gen->first_shared = isl_space_dim(dim, isl_dim_param);
        proj = projection(dim, gen->tiled_len, gen->shared_len + gen->n_block);
        sched = isl_union_map_apply_range(sched, isl_union_map_from_map(proj));

        dim = isl_union_map_get_space(sched);
        par = parametrization(dim, gen->shared_len + gen->n_block,
                                0, gen->shared_len, "g");
        sched = isl_union_map_intersect_range(sched,
                                                isl_union_set_from_set(par));

        dim = isl_union_map_get_space(sched);
        proj = projection(dim, gen->shared_len + gen->n_block, gen->shared_len);

        gen->shared_sched = sched;
        gen->shared_proj = isl_union_map_from_map(proj);
}

/* Group references of all arrays in the program.
 */
static void group_references(struct cuda_gen *gen)
{
        int i;
        isl_union_map *sched;

        sched = isl_union_map_apply_range(isl_union_map_copy(gen->shared_sched),
                                          isl_union_map_copy(gen->shared_proj));

        for (i = 0; i < gen->n_array; ++i)
                group_array_references(gen, &gen->array[i], sched);

        isl_union_map_free(sched);
}

/* Free all array information that is local to the current kernel.
 */
static void free_local_array_info(struct cuda_gen *gen)
{
        int i, j;

        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                for (j = 0; j < array->n_group; ++j)
                        free_array_ref_group(array->groups[j], array->n_index);
                free(array->groups);

                if (array->n_group == 0)
                        continue;
                for (j = 0; j < gen->array[i].n_index; ++j) {
                        isl_pw_aff_free(gen->array[i].local_bound[j]);
                        gen->array[i].local_bound[j] = NULL;
                }
        }
}

static void print_iterator_list(FILE *out, int len, const char *prefix,
        int parens)
{
        int i;

        fprintf(out, "(");
        for (i = 0; i < len; ++i) {
                if (i)
                        fprintf(out, ", ");
                if (parens)
                        fprintf(out, "(%s%d)", prefix, i);
                else
                        fprintf(out, "%s%d", prefix, i);
        }
        fprintf(out, ")");
}

/* Print an access to the element in the global memory copy of the
 * given array that corresponds to element [a0][a1]... of the original array.
 * The copy in global memory has been linearized, so we need to take
 * the array size into account.
 */
static void print_global_index(isl_ctx *ctx, FILE *out,
        struct cuda_array_info *array)
{
        int i;
        isl_printer *prn;

        if (cuda_array_is_scalar(array)) {
                fprintf(out, "*%s", array->name);
                return;
        }

        fprintf(out, "%s[", array->name);
        for (i = 0; i + 1 < array->n_index; ++i)
                fprintf(out, "(");
        for (i = 0; i < array->n_index; ++i) {
                if (i) {
                        prn = isl_printer_to_file(ctx, out);
                        prn = isl_printer_set_output_format(prn, ISL_FORMAT_C);
                        prn = isl_printer_print_str(prn, ") * (");
                        prn = isl_printer_print_pw_aff(prn,
                                                        array->local_bound[i]);
                        prn = isl_printer_print_str(prn, ") + ");
                        isl_printer_free(prn);
                }
                fprintf(out, "a%d", i);
        }
        fprintf(out, "]");
}

/* Print an access to the element in the shared memory copy of the
 * given array that corresponds to element [a0][a1]... of the original array.
 * Since the array in shared memory is just a shifted copy of part
 * of the original array, we simply need to subtract the lower bound,
 * which was computed in can_tile_for_shared_memory.
 * If any of the indices is strided, then we first add
 * shared_bound[i].shift and divide by shared_bound[i].stride.
 */
static void print_local_index(FILE *out, struct cuda_array_ref_group *group)
{
        int i;
        isl_ctx *ctx;
        isl_printer *prn;
        struct cuda_array_bound *bounds = group->shared_bound;

        ctx = isl_space_get_ctx(group->array->dim);
        print_array_name(out, group);
        for (i = 0; i < group->array->n_index; ++i) {
                fprintf(out, "[(a%d", i);
                if (bounds[i].shift) {
                        fprintf(out, " + (");
                        prn = isl_printer_to_file(ctx, out);
                        prn = isl_printer_set_output_format(prn, ISL_FORMAT_C);
                        prn = isl_printer_print_qpolynomial(prn,
                                                bounds[i].shift);
                        prn = isl_printer_print_str(prn, "))/");
                        prn = isl_printer_print_isl_int(prn,
                                                bounds[i].stride);
                        isl_printer_free(prn);
                } else
                        fprintf(out, ")");
                fprintf(out, " - (");
                prn = isl_printer_to_file(ctx, out);
                prn = isl_printer_set_output_format(prn, ISL_FORMAT_C);
                prn = isl_printer_print_aff(prn, bounds[i].lb);
                isl_printer_free(prn);
                fprintf(out, ")]");
        }
}

/* Print '#define's for copying data from global memory to shared
 * memory and back for the given array.
 */
static void print_array_copy_defines(struct cuda_gen *gen,
        struct cuda_array_ref_group *group)
{
        int i;
        const char *type[] = { "read", "write" };
        struct cuda_array_info *array = group->array;
        int n_index = array->n_index;

        for (i = 0; i < 2; ++i) {
                fprintf(gen->cuda.kernel_c, "#define %s_", type[i]);
                print_array_name(gen->cuda.kernel_c, group);
                print_iterator_list(gen->cuda.kernel_c, n_index, "a", 0);
                fprintf(gen->cuda.kernel_c, " %s_", type[i]);
                print_array_name(gen->cuda.kernel_c, group);
                fprintf(gen->cuda.kernel_c, "_");
                print_iterator_list(gen->cuda.kernel_c, n_index, "a", 1);
                fprintf(gen->cuda.kernel_c, "\n");

                fprintf(gen->cuda.kernel_c, "#define %s_", type[i]);
                print_array_name(gen->cuda.kernel_c, group);
                fprintf(gen->cuda.kernel_c, "_");
                print_iterator_list(gen->cuda.kernel_c, n_index, "a", 0);
                if (i) {
                        fprintf(gen->cuda.kernel_c, " ");
                        print_global_index(gen->ctx, gen->cuda.kernel_c, array);
                        fprintf(gen->cuda.kernel_c, " = ");
                        print_local_index(gen->cuda.kernel_c, group);
                } else {
                        fprintf(gen->cuda.kernel_c, " ");
                        print_local_index(gen->cuda.kernel_c, group);
                        fprintf(gen->cuda.kernel_c, " = ");
                        print_global_index(gen->ctx, gen->cuda.kernel_c, array);
                }
                fprintf(gen->cuda.kernel_c, "\n");
        }
}

static void print_copy_defines(struct cuda_gen *gen)
{
        int i, j;

        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                for (j = 0; j < array->n_group; ++j) {
                        if (array->groups[j]->private_bound)
                                continue;
                        if (!array->groups[j]->shared_bound)
                                continue;
                        print_array_copy_defines(gen, array->groups[j]);
                }
        }
}

/* The sizes of the arrays on the host that have been computed by
 * extract_array_info may depend on the parameters.  Use the extra
 * constraints on the parameters that are valid at "host_domain"
 * to simplify these expressions.
 */
static void localize_bounds(struct cuda_gen *gen,
        __isl_keep isl_set *host_domain)
{
        int i, j;
        isl_set *context;

        context = isl_set_copy(host_domain);
        context = isl_set_params(host_domain);

        for (i = 0; i < gen->n_array; ++i) {
                struct cuda_array_info *array = &gen->array[i];

                if (array->n_group == 0)
                        continue;

                for (j = 0; j < array->n_index; ++j) {
                        isl_pw_aff *pwaff;

                        pwaff = isl_pw_aff_copy(array->bound[j]);
                        pwaff = isl_pw_aff_gist(pwaff, isl_set_copy(context));
                        array->local_bound[j] = pwaff;
                }
        }
        isl_set_free(context);
}

/* Set gen->tile_len and gen->n_parallel to those of the first statement
 * in the statement list u.
 * Because of the way the schedule is constructed, the other statements
 * in the list, if any, should have the same values for these properties.
 */
static void set_tile_len(struct cuda_gen *gen, struct clast_user_stmt *u)
{
        int nr;
        struct cuda_stmt *stmt;

        nr = atoi(u->statement->name + 2);
        stmt = &gen->stmts[nr];

        gen->tile_len = stmt->tile_len;
        gen->n_parallel = stmt->n_parallel;
}

/* This function is called for each leaf in the clast of the host code.
 * We first specialize the schedule to the site of the leaf, compute
 * the size of shared memory and then print the body of host code
 * and the associated kernel (through a call to print_kernel_body).
 */
static void print_host_user(struct gpucode_info *code,
        struct clast_user_stmt *u)
{
        struct cuda_gen *gen = code->user;
        isl_space *dim;
        isl_set *par;
        isl_set *host_domain;
        isl_union_map *access;
        isl_union_map *local_sched;
        isl_union_set *arrays;

        set_tile_len(gen, u);
        read_sizes(gen);

        host_domain = extract_entire_host_domain(u);

        local_sched = isl_union_map_intersect_range(
                    isl_union_map_copy(gen->sched),
                    isl_union_set_from_set(extend(isl_set_copy(host_domain),
                                                  gen->untiled_len)));
        access = isl_union_map_union(isl_union_map_copy(gen->read),
                                     isl_union_map_copy(gen->write));
        access = isl_union_map_apply_domain(access,
                                            isl_union_map_copy(local_sched));
        arrays = isl_union_map_range(access);

        print_indent(code->dst, code->indent);
        fprintf(code->dst, "dim3 k%d_dimBlock(", gen->kernel_id);
        print_reverse_list(code->dst, gen->n_block, gen->block_dim);
        fprintf(code->dst, ");\n");

        print_indent(code->dst, code->indent);
        fprintf(code->dst, "dim3 k%d_dimGrid(", gen->kernel_id);
        print_reverse_list(code->dst, gen->n_grid, gen->grid_dim);
        fprintf(code->dst, ");\n");

        gen->tiled_sched = tile_schedule(gen, local_sched);
        gen->tiled_sched = parametrize_tiled_schedule(gen, gen->tiled_sched);
        gen->tiled_sched = scale_tile_loops(gen, gen->tiled_sched);

        gen->local_sched = isl_union_map_copy(gen->tiled_sched);

        dim = isl_union_map_get_space(gen->local_sched);
        par = parametrization(dim, gen->tiled_len, 0, gen->shared_len, "g");
        gen->local_sched = isl_union_map_intersect_range(gen->local_sched,
                                                isl_union_set_from_set(par));

        gen->local_sched = thread_tile_schedule(gen, gen->local_sched);
        gen->local_sched = scale_thread_tile_loops(gen, gen->local_sched);

        gen->private_access = NULL;
        compute_shared_sched(gen);
        gen->privatization = compute_privatization(gen);
        group_references(gen);
        compute_private_size(gen);
        localize_bounds(gen, host_domain);

        gen->local_sched = interchange_for_unroll(gen, gen->local_sched);

        print_copy_defines(gen);
        print_kernel_launch(gen, arrays);

        fprintf(gen->cuda.kernel_c, "{\n");

        print_kernel_body(gen, host_domain, gen->tiled_sched);

        fprintf(gen->cuda.kernel_c, "}\n");

        free_local_array_info(gen);
        isl_map_free(gen->privatization);
        isl_union_map_free(gen->private_access);
        isl_union_map_free(gen->local_sched);
        isl_union_map_free(gen->tiled_sched);
        isl_union_map_free(gen->shared_sched);
        isl_union_map_free(gen->shared_proj);
        isl_union_set_free(arrays);
        isl_set_free(host_domain);

        free(gen->tile_size);
        gen->kernel_id++;
}

/* Use CLooG to generate code for the outer gen->tile_first loops
 * of the global schedule in gen->sched.
 * The pretty printing of this code is handled by gpu_print_host_stmt,
 * which calls print_host_user for each kernel invocation location.
 */
static void print_cloog_host_code(struct cuda_gen *gen)
{
        int i;
        isl_set *context;
        isl_union_map *sched;
        CloogOptions *options;
        CloogDomain *cloog_context;
        CloogUnionDomain *ud;
        CloogInput *input;
        struct clast_stmt *stmt;
        char name[20];

        options = cloog_options_malloc(gen->state);
        options->language = LANGUAGE_C;
        options->otl = 0;
        options->strides = 1;
        options->stop = gen->tile_first;
        options->f = gen->untiled_len;
        options->l = gen->untiled_len;
        options->save_domains = 1;
        options->noscalars = 1;

        sched = isl_union_map_copy(gen->sched);
        ud = cloog_union_domain_from_isl_union_map(sched);
        for (i = 0; i < options->stop; ++i) {
                snprintf(name, sizeof(name), "h%d", i);
                ud = cloog_union_domain_set_name(ud, CLOOG_SCAT, i, name);
        }
        context = isl_set_copy(gen->context);
        cloog_context = cloog_domain_from_isl_set(context);
        input = cloog_input_alloc(cloog_context, ud);

        stmt = cloog_clast_create_from_input(input, options);

        gen->code.indent = 0;
        gen->code.dst = gen->cuda.host_c;
        gen->code.print_user_stmt = NULL;
        gen->code.print_user_stmt_list = &print_host_user;
        gen->code.print_for_head = NULL;
        gen->code.print_for_foot = NULL;
        gen->code.user = gen;
        gpu_print_host_stmt(&gen->code, stmt);

        cloog_clast_free(stmt);
        cloog_options_free(options);
        fprintf(gen->cuda.host_c, "\n");
}

void print_cuda_macros(struct cuda_gen *gen)
{
        const char *macros =
                "#define cudaCheckReturn(ret) assert((ret) == cudaSuccess)\n"
                "#define cudaCheckKernel()"
                " assert(cudaGetLastError() == cudaSuccess)\n\n";
        fputs(macros, gen->cuda.host_c);
}

void print_host_code(struct cuda_gen *gen)
{
        fprintf(gen->cuda.host_c, "{\n");
        print_cloog_macros(gen->cuda.host_c);
        print_cloog_macros(gen->cuda.kernel_c);

        print_cuda_macros(gen);

        declare_device_arrays(gen);

        allocate_device_arrays(gen);
        copy_arrays_to_device(gen);

        gen->kernel_id = 0;
        print_cloog_host_code(gen);

        copy_arrays_from_device(gen);
        free_device_arrays(gen);

        fprintf(gen->cuda.host_c, "}\n");
}

__isl_give isl_set *add_context_from_str(__isl_take isl_set *set,
        const char *str)
{
        isl_ctx *ctx;
        isl_set *context;

        if (!str)
                return set;

        ctx = isl_set_get_ctx(set);
        context = isl_set_read_from_str(ctx, str, -1);
        context = isl_set_align_params(context, isl_set_get_space(set));
        set = isl_set_intersect(set, context);

        return set;
}

/* Return the union of all iteration domains of the gen->stmts[i].
 */
static __isl_give isl_union_set *extract_domain(struct cuda_gen *gen)
{
        int i;
        isl_union_set *domain;

        domain = isl_union_set_empty(isl_set_get_space(gen->context));
        for (i = 0; i < gen->n_stmts; ++i) {
                isl_set *domain_i;

                domain_i = isl_set_copy(gen->stmts[i].domain);
                domain = isl_union_set_union(domain,
                                             isl_union_set_from_set(domain_i));
        }

        return domain;
}

/* Information about the outermost tilable bands in the forest of bands.
 *
 * tile_len and n_parallel are only sets on band_info structures
 * that correspond to outermost bands.  For other bands (in particular,
 * ancestors of the outermost bands), n_parallal is set to 0.
 *
 * prefix is the (padded) schedule leading up to the outermost tilable bands.
 *
 * tile_first is the number of schedule dimensions in prefix.
 *
 * suffix is the schedule of the outermost tilable bands and their descendants.
 */
struct band_info {
        struct cuda_gen *gen;
        int tile_first;
        int tile_len;
        int n_parallel;
        isl_union_map *prefix;
        isl_union_map *suffix;
};

/* Set tile_len and n_parallel of the statement to that of
 * their outermost band, recorded in the band_info.
 */
static int set_stmt_tile_len(__isl_take isl_map *map, void *user)
{
        struct band_info *info = user;
        int nr;
        struct cuda_stmt *stmt;

        nr = atoi(isl_map_get_tuple_name(map, isl_dim_in) + 2);
        stmt = &info->gen->stmts[nr];

        stmt->tile_len = info->tile_len;
        stmt->n_parallel = info->n_parallel;

        isl_map_free(map);

        return 0;
}

static void list_select_outer_band(struct cuda_gen *gen,
        __isl_take isl_band_list *list, int pos, struct band_info *list_info);

/* Check if this band has any parallel loops.  If so, take it as
 * the outermost tilable band.  If not, continue looking for the
 * outermost tilable band in the children of the current band.
 */
static void band_select_outer_band(struct cuda_gen *gen,
        __isl_take isl_band *band, int pos, struct band_info *info)
{
        int n = isl_band_n_member(band);
        int n_parallel;

        for (n_parallel = 0; n_parallel < n; ++n_parallel)
                if (!isl_band_member_is_zero_distance(band, n_parallel))
                        break;

        info->n_parallel = n_parallel;
        if (n_parallel) {
                info->gen = gen;
                info->tile_first = pos;
                info->tile_len = n;
                info->prefix = isl_band_get_prefix_schedule(band);
                info->suffix = isl_union_map_flat_range_product(
                                isl_band_get_partial_schedule(band),
                                isl_band_get_suffix_schedule(band));
                isl_union_map_foreach_map(info->prefix,
                                            &set_stmt_tile_len, info);
        } else {
                isl_band_list *children;
                if (!isl_band_has_children(band))
                        isl_die(isl_band_get_ctx(band), isl_error_unknown,
                                "unable to detect any parallelism", abort());
                children = isl_band_get_children(band);
                list_select_outer_band(gen, children, pos + n, info);
        }

        isl_band_free(band);
}

/* Comparison function that returns a non-zero value for band_infos
 * with different tile_len fields or different n_parallel fields.
 */
static int cmp_band(const void *p1, const void *p2)
{
        const struct band_info *info1 = p1;
        const struct band_info *info2 = p2;

        if (info1->tile_len != info2->tile_len)
                return info1->tile_len - info2->tile_len;

        return info1->n_parallel - info2->n_parallel;
}

/* Extend "umap" with coordinates with fixed value "val"
 * to a total length of "dst_len", assuming the original dimension is "src_len".
 */
static __isl_give isl_union_map *extend_range(__isl_take isl_union_map *umap,
        int src_len, int dst_len, int val)
{
        isl_space *dim;
        isl_map *map;
        int i;

        dim = isl_union_map_get_space(umap);
        map = isl_map_reverse(projection(dim, dst_len, src_len));
        for (i = src_len; i < dst_len; ++i)
                map = isl_map_fix_si(map, isl_dim_out, i, val);

        umap = isl_union_map_apply_range(umap, isl_union_map_from_map(map));

        return umap;
}

/* Group bands with the same values for tile_len and n_parallel.
 * The prefix schedule is then extended with a fixed coordinate that
 * is different for each such group.
 * Note that the actual values for this coordinate are not important.
 * The bands have already been effectively separated at a higher level
 * or they are independent and may be executed in parallel.
 * The list of band_info has been sorted before this functions is called.
 */
static void separate_bands(struct band_info *info, int n)
{
        int i;
        int j = 0;

        for (i = 0; i < n; ++i) {
                int l = info[i].tile_first;

                if (i &&
                    (info[i].tile_len != info[i - 1].tile_len ||
                     info[i].n_parallel != info[i - 1].n_parallel))
                        j++;

                info[i].prefix = extend_range(info[i].prefix,
                                                l, l + 1, j);
                info[i].tile_first = l + 1;
        }
}

/* Select the outermost bands in the elements of the list, align
 * their prefix schedules, separate bands with different values
 * for tile_len and/or n_parallel and then combine the resulting
 * prefix and suffix schedules into a single pair of prefix and
 * suffix schedules for the entire list.
 */
static void list_select_outer_band(struct cuda_gen *gen,
        __isl_take isl_band_list *list, int pos, struct band_info *list_info)
{
        isl_band *band;
        int i;
        int n = isl_band_list_n_band(list);
        isl_ctx *ctx = isl_band_list_get_ctx(list);
        struct band_info *info;
        int max_tile_first;
        isl_union_map *prefix;
        isl_union_map *suffix;

        assert(n >= 1);
        info = isl_calloc_array(ctx, struct band_info, n);
        assert(info);

        max_tile_first = 0;
        for (i = 0; i < n; ++i) {
                band = isl_band_list_get_band(list, i);
                band_select_outer_band(gen, band, pos, &info[i]);
                if (info[i].tile_first > max_tile_first)
                        max_tile_first = info[i].tile_first;
        }

        for (i = 0; i < n; ++i) {
                if (info[i].tile_first == max_tile_first)
                        continue;
                info[i].prefix = extend_range(info[i].prefix,
                                        info[i].tile_first, max_tile_first, 0);
        }

        qsort(info, n, sizeof(struct band_info), &cmp_band);

        for (i = 0; i < n - 1; ++i)
                if (info[i].tile_len != info[i + 1].tile_len ||
                    info[i].n_parallel != info[i + 1].n_parallel)
                        break;

        if (i < n -1)
                separate_bands(info, n);

        prefix = info[0].prefix;
        suffix = info[0].suffix;

        for (i = 1; i < n; ++i) {
                prefix = isl_union_map_union(prefix, info[i].prefix);
                suffix = isl_union_map_union(suffix, info[i].suffix);
        }

        list_info->tile_first = info[0].tile_first;
        list_info->tile_len = -1;
        list_info->prefix = prefix;
        list_info->suffix = suffix;

        isl_band_list_free(list);
        free(info);
}

/* Set max_out to the maximal number of output dimensions over
 * all maps.
 */
static int update_max_out(__isl_take isl_map *map, void *user)
{
        int *max_out = user;
        int n_out = isl_map_dim(map, isl_dim_out);

        if (n_out > *max_out)
                *max_out = n_out;

        isl_map_free(map);
        return 0;
}

struct align_range_data {
        int max_out;
        isl_union_map *res;
};

/* Extend the dimension of the range of the given map to data->max_out and
 * then add the result to data->res.
 */
static int map_align_range(__isl_take isl_map *map, void *user)
{
        struct align_range_data *data = user;
        int i;
        isl_space *dim;
        isl_map *proj;
        int n_out = isl_map_dim(map, isl_dim_out);

        dim = isl_union_map_get_space(data->res);
        proj = isl_map_reverse(projection(dim, data->max_out, n_out));
        for (i = n_out; i < data->max_out; ++i)
                proj = isl_map_fix_si(proj, isl_dim_out, i, 0);

        map = isl_map_apply_range(map, proj);

        data->res = isl_union_map_add_map(data->res, map);

        return 0;
}

/* Extend the ranges of the maps in the union map such they all have
 * the same dimension.
 */
static __isl_give isl_union_map *align_range(__isl_take isl_union_map *umap)
{
        struct align_range_data data;

        data.max_out = 0;
        isl_union_map_foreach_map(umap, &update_max_out, &data.max_out);

        data.res = isl_union_map_empty(isl_union_map_get_space(umap));
        isl_union_map_foreach_map(umap, &map_align_range, &data);

        isl_union_map_free(umap);
        return data.res;
}

/* Select the outermost tilable band that (by construction)
 * has at least one parallel loop.
 * The starting position of the aligned band is stored in the pair
 * gen->tile_first.
 * The sizes and number of parallel loops may be different in different
 * parts of the band forest and are therefore stored in the cuda_stmts.
 *
 * Return the complete schedule, with the tilable bands aligned
 * at gen->tile_first and padded with zero, if needed.
 */
static __isl_give isl_union_map *select_outer_tilable_band(struct cuda_gen *gen,
        __isl_keep isl_schedule *schedule)
{
        isl_band_list *list;
        struct band_info info;

        gen->n_parallel = 0;
        gen->tile_len = -1;

        list = isl_schedule_get_band_forest(schedule);

        list_select_outer_band(gen, list, 0, &info);

        gen->tile_first = info.tile_first;
        info.suffix = align_range(info.suffix);

        return isl_union_map_flat_range_product(info.prefix, info.suffix);
}

/* Set gen->untiled_len to the number of scheduling dimensions
 * for the schedule of the first domain.
 * We assume here that this number is the same for all domains.
 */
static int set_untiled_len(__isl_take isl_map *map, void *user)
{
        unsigned *untiled_len = user;

        *untiled_len = isl_map_dim(map, isl_dim_out);

        isl_map_free(map);
        return -1;
}

/* Compute an appropriate schedule based on the accesses in
 * gen->read and gen->write.
 *
 * We first compute dependences and then use those to compute
 * a schedule that has a parallel loop in each tilable band.
 * Finally, we select the outermost tilable band.
 */
static void compute_schedule(struct cuda_gen *gen,
        __isl_take isl_union_map *sched)
{
        isl_ctx *ctx = isl_union_map_get_ctx(sched);
        isl_union_set *domain;
        isl_union_map *empty;
        isl_union_map *dep_raw, *dep2, *dep3, *dep;
        isl_union_map *uninitialized;
        isl_schedule *schedule;
        struct isl_options *options;

        empty = isl_union_map_empty(isl_union_map_get_space(sched));

        isl_union_map_compute_flow(isl_union_map_copy(gen->read),
                            isl_union_map_copy(gen->write), empty,
                            isl_union_map_copy(sched),
                            &dep_raw, NULL, &uninitialized, NULL);
        isl_union_map_compute_flow(isl_union_map_copy(gen->write),
                            isl_union_map_copy(gen->write),
                            isl_union_map_copy(gen->read),
                            isl_union_map_copy(sched),
                            &dep2, &dep3, NULL, NULL);
        isl_union_map_free(sched);

        gen->copy_in = isl_union_map_range(uninitialized);

        dep = isl_union_map_union(dep2, dep3);
        dep = isl_union_map_union(dep, dep_raw);
        dep = isl_union_map_coalesce(dep);

        domain = extract_domain(gen);
        options = isl_ctx_peek_options(ctx, isl_options_arg);
        options->schedule_outer_zero_distance = 1;
        schedule = isl_union_set_compute_schedule(isl_union_set_copy(domain),
                                isl_union_map_copy(dep), dep);

        sched = select_outer_tilable_band(gen, schedule);

        isl_union_map_foreach_map(sched, &set_untiled_len, &gen->untiled_len);
        sched = isl_union_map_intersect_domain(sched, domain);
        gen->sched = sched;

        isl_schedule_free(schedule);
}

static struct cuda_stmt_access **expr_extract_access(struct pet_expr *expr,
        struct cuda_stmt_access **next_access)
{
        struct cuda_stmt_access *access;
        isl_ctx *ctx = isl_map_get_ctx(expr->acc.access);

        access = isl_alloc_type(ctx, struct cuda_stmt_access);
        assert(access);
        access->next = NULL;
        access->read = expr->acc.read;
        access->write = expr->acc.write;
        access->access = isl_map_copy(expr->acc.access);

        *next_access = access;
        next_access = &(*next_access)->next;
        return next_access;
}

static struct cuda_stmt_access **expr_extract_accesses(struct pet_expr *expr,
        struct cuda_stmt_access **next_access)
{
        int i;

        for (i = 0; i < expr->n_arg; ++i)
                next_access = expr_extract_accesses(expr->args[i],
                                                        next_access);

        if (expr->type == pet_expr_access)
                next_access = expr_extract_access(expr, next_access);

        return next_access;
}

static void pet_stmt_extract_accesses(struct cuda_stmt *stmt)
{
        struct cuda_stmt_access **next_access = &stmt->accesses;

        stmt->accesses = NULL;
        expr_extract_accesses(stmt->body, next_access);
}

/* Return an array of cuda_stmt representing the statements in "scop".
 */
static struct cuda_stmt *extract_stmts(isl_ctx *ctx, struct pet_scop *scop,
        __isl_keep isl_set *context)
{
        int i;
        struct cuda_stmt *stmts;

        stmts = isl_calloc_array(ctx, struct cuda_stmt, scop->n_stmt);
        assert(stmts);

        for (i = 0; i < scop->n_stmt; ++i) {
                struct cuda_stmt *s = &stmts[i];

                s->domain = isl_set_copy(scop->stmts[i]->domain);
                s->domain = isl_set_intersect(s->domain, isl_set_copy(context));
                s->body = scop->stmts[i]->body;
                pet_stmt_extract_accesses(s);
        }

        return stmts;
}

/* Replace the scop in the "input" file by equivalent code
 * that uses the GPU.  "scop" is assumed to correspond to this scop.
 *
 * We first compute a schedule that respects the dependences
 * of the original program and select the outermost band
 * of tilable dimensions that has at least one parallel loop.
 * We then have three blocks of dimensions
 *
 *      H               B                       G
 *
 * The tilable band "B" is first tiled according to "tile.sizes", resulting
 * in
 *
 *      H       T               P               G
 *
 * For each iteration of the T loop and for each array, we compute
 * the array elements accessed by that iteration, construct a rectangular
 * box around it and shift it to the origin.  The result is used
 * as shared memory for the array.
 *
 * We then split off at most 2 parallel loops from the T loops and
 * at most 3 parallel loops from the P loops
 *
 *      H       T1      T2      P1      P2      G
 *
 * The T1/P1 loops are then tiled or "wrapped" over the blocks/threads,
 * according to "grid.sizes"/"block.sizes".
 *
 *      H       T1T T1P T2      P1T P1P P2      G
 *
 * Finally, the T1P and P1P iterators are equated to the block and
 * thread dimensions respectively and so are effectively removed.
 * The H loops are run on the host.  The T1T, T2, P1T, P2 and G loops
 * are run on the GPU.
 *
 * Code is generated in three stages.  We first generate code for the
 * host (the H loops), with iterators h%d.  Then, for each leaf node
 * of the resulting AST, we generate code for the shared loops (up to
 * and including T2), with iterators g%d and after equating the H loops
 * to h%d parameters and the T1P loops to the block dimensions.
 * Finally, we generate code for the remaining loops in a similar fashion.
 */
int cuda_pet(isl_ctx *ctx, struct pet_scop *scop, struct ppcg_options *options,
        const char *input)
{
        isl_union_map *sched;
        struct cuda_gen gen;

        if (!scop)
                return -1;

        scop = pet_scop_align_params(scop);

        gen.ctx = ctx;
        gen.context = isl_set_copy(scop->context);
        gen.context = add_context_from_str(gen.context, options->ctx);
        gen.n_stmts = scop->n_stmt;
        gen.stmts = extract_stmts(ctx, scop, gen.context);
        gen.read = pet_scop_collect_reads(scop);
        gen.write = pet_scop_collect_writes(scop);
        gen.options = options;
        gen.state = cloog_isl_state_malloc(gen.ctx);
        gen.scop = scop;

        cuda_open_files(&gen.cuda, input);

        collect_array_info(&gen);

        sched = pet_scop_collect_schedule(scop);

        compute_schedule(&gen, sched);

        print_host_code(&gen);

        cloog_state_free(gen.state);
        clear_cuda_gen(&gen);

        cuda_close_files(&gen.cuda);

        return 0;
}