cumo 0.1.0

data/ext/cumo/narray/gen/tmpl_bit/where2.c

@@ -0,0 +1,95 @@
+static void
+<%=c_iter%>(na_loop_t *const lp)
+{
+    size_t  i;
+    BIT_DIGIT *a;
+    size_t  p;
+    ssize_t s;
+    size_t *idx;
+    BIT_DIGIT x=0;
+    char   *idx0, *idx1;
+    size_t  count;
+    size_t  e;
+    where_opt_t *g;
+
+    // TODO(sonots): CUDA kernelize
+    SHOW_SYNCHRONIZE_FIXME_WARNING_ONCE("<%=name%>", "<%=type_name%>");
+    cumo_cuda_runtime_check_status(cudaDeviceSynchronize());
+
+    g = (where_opt_t*)(lp->opt_ptr);
+    count = g->count;
+    idx0  = g->idx0;
+    idx1  = g->idx1;
+    e     = g->elmsz;
+    INIT_COUNTER(lp, i);
+    INIT_PTR_BIT_IDX(lp, 0, a, p, s, idx);
+    if (idx) {
+        for (; i--;) {
+            LOAD_BIT(a, p+*idx, x);
+            idx++;
+            if (x==0) {
+                STORE_INT(idx0,e,count);
+                idx0 += e;
+            } else {
+                STORE_INT(idx1,e,count);
+                idx1 += e;
+            }
+            count++;
+        }
+    } else {
+        for (; i--;) {
+            LOAD_BIT(a, p, x);
+            p+=s;
+            if (x==0) {
+                STORE_INT(idx0,e,count);
+                idx0 += e;
+            } else {
+                STORE_INT(idx1,e,count);
+                idx1 += e;
+            }
+            count++;
+        }
+    }
+    g->count = count;
+    g->idx0  = idx0;
+    g->idx1  = idx1;
+}
+
+/*
+  Returns two index arrays.
+  The first array contains index where the bit is one (true).
+  The second array contains index where the bit is zero (false).
+  @overload <%=op_map%>
+  @return [Cumo::Int32,Cumo::Int64]*2
+*/
+static VALUE
+<%=c_func(0)%>(VALUE self)
+{
+    VALUE idx_1, idx_0;
+    size_t size, n_1, n_0;
+    where_opt_t *g;
+
+    ndfunc_arg_in_t ain[1] = {{cT,0}};
+    ndfunc_t ndf = { <%=c_iter%>, FULL_LOOP, 1, 0, ain, 0 };
+
+    size = RNARRAY_SIZE(self);
+    n_1 = NUM2SIZET(<%=find_tmpl("count_true_cpu").c_func%>(0, NULL, self));
+    n_0 = size - n_1;
+    g = ALLOCA_N(where_opt_t,1);
+    g->count = 0;
+    if (size>4294967295ul) {
+        idx_1 = nary_new(cumo_cInt64, 1, &n_1);
+        idx_0 = nary_new(cumo_cInt64, 1, &n_0);
+        g->elmsz = 8;
+    } else {
+        idx_1 = nary_new(cumo_cInt32, 1, &n_1);
+        idx_0 = nary_new(cumo_cInt32, 1, &n_0);
+        g->elmsz = 4;
+    }
+    g->idx1 = na_get_pointer_for_write(idx_1);
+    g->idx0 = na_get_pointer_for_write(idx_0);
+    na_ndloop3(&ndf, g, 1, self);
+    na_release_lock(idx_0);
+    na_release_lock(idx_1);
+    return rb_assoc_new(idx_1,idx_0);
+}

data/ext/cumo/narray/index.c

@@ -0,0 +1,880 @@
+#include <string.h>
+#include <ruby.h>
+#include "cumo/narray.h"
+#include "cumo/cuda/runtime.h"
+#include "cumo/template.h"
+
+#if   SIZEOF_VOIDP == 8
+#define cIndex cumo_cInt64
+#elif SIZEOF_VOIDP == 4
+#define cIndex cumo_cInt32
+#endif
+
+// from ruby/enumerator.c
+struct enumerator {
+    VALUE obj;
+    ID    meth;
+    VALUE args;
+    // use only above in this source
+    VALUE fib;
+    VALUE dst;
+    VALUE lookahead;
+    VALUE feedvalue;
+    VALUE stop_exc;
+    VALUE size;
+    // incompatible below depending on ruby version
+    //VALUE procs;                      // ruby 2.4
+    //rb_enumerator_size_func *size_fn; // ruby 2.1-2.4
+    //VALUE (*size_fn)(ANYARGS);        // ruby 2.0
+};
+
+// note: the memory refed by this pointer is not freed and causes memroy leak.
+//
+// @example
+//     a[1..3,1] generates two na_index_arg_t(s). First is for 1..3, and second is for 1.
+typedef struct {
+    size_t  n; // the number of elements of the dimesnion
+    size_t  beg; // the starting point in the dimension
+    ssize_t step; // the step size of the dimension
+    size_t *idx; // list of indices
+    int     reduce; // true if the dimension is reduced by addition
+    int     orig_dim; // the dimension of original array
+} na_index_arg_t;
+
+
+static void
+print_index_arg(na_index_arg_t *q, int n)
+{
+    int i;
+    printf("na_index_arg_t = 0x%"SZF"x {\n",(size_t)q);
+    for (i=0; i<n; i++) {
+        printf("  q[%d].n=%"SZF"d\n",i,q[i].n);
+        printf("  q[%d].beg=%"SZF"d\n",i,q[i].beg);
+        printf("  q[%d].step=%"SZF"d\n",i,q[i].step);
+        printf("  q[%d].idx=0x%"SZF"x\n",i,(size_t)q[i].idx);
+        printf("  q[%d].reduce=0x%x\n",i,q[i].reduce);
+        printf("  q[%d].orig_dim=%d\n",i,q[i].orig_dim);
+    }
+    printf("}\n");
+}
+
+static VALUE sym_ast;
+static VALUE sym_all;
+//static VALUE sym_reduce;
+static VALUE sym_minus;
+static VALUE sym_new;
+static VALUE sym_reverse;
+static VALUE sym_plus;
+static VALUE sym_sum;
+static VALUE sym_tilde;
+static VALUE sym_rest;
+static ID id_beg;
+static ID id_end;
+static ID id_exclude_end;
+static ID id_each;
+static ID id_step;
+static ID id_dup;
+static ID id_bracket;
+static ID id_shift_left;
+static ID id_mask;
+
+
+static void
+na_index_set_step(na_index_arg_t *q, int i, size_t n, size_t beg, ssize_t step)
+{
+    q->n    = n;
+    q->beg  = beg;
+    q->step = step;
+    q->idx  = NULL;
+    q->reduce = 0;
+    q->orig_dim = i;
+}
+
+static void
+na_index_set_scalar(na_index_arg_t *q, int i, ssize_t size, ssize_t x)
+{
+    if (x < -size || x >= size)
+        rb_raise(rb_eRangeError,
+                  "array index (%"SZF"d) is out of array size (%"SZF"d)",
+                  x, size);
+    if (x < 0)
+        x += size;
+    q->n    = 1;
+    q->beg  = x;
+    q->step = 0;
+    q->idx  = NULL;
+    q->reduce = 0;
+    q->orig_dim = i;
+}
+
+static inline ssize_t
+na_range_check(ssize_t pos, ssize_t size, int dim)
+{
+    ssize_t idx=pos;
+
+    if (idx < 0) idx += size;
+    if (idx < 0 || idx >= size) {
+        rb_raise(rb_eIndexError, "index=%"SZF"d out of shape[%d]=%"SZF"d",
+                 pos, dim, size);
+    }
+    return idx;
+}
+
+static void
+na_parse_array(VALUE ary, int orig_dim, ssize_t size, na_index_arg_t *q)
+{
+    int k;
+    int n = RARRAY_LEN(ary);
+    q->idx = ALLOC_N(size_t, n);
+    for (k=0; k<n; k++) {
+        q->idx[k] = na_range_check(NUM2SSIZET(RARRAY_AREF(ary,k)), size, orig_dim);
+    }
+    q->n    = n;
+    q->beg  = 0;
+    q->step = 1;
+    q->reduce = 0;
+    q->orig_dim = orig_dim;
+}
+
+static void
+na_parse_narray_index(VALUE a, int orig_dim, ssize_t size, na_index_arg_t *q)
+{
+    VALUE idx;
+    narray_t *na;
+    narray_data_t *nidx;
+    size_t k, n;
+    ssize_t *nidxp;
+
+    GetNArray(a,na);
+    if (NA_NDIM(na) != 1) {
+        rb_raise(rb_eIndexError, "should be 1-d NArray");
+    }
+    n = NA_SIZE(na);
+    idx = nary_new(cIndex,1,&n);
+    na_store(idx,a);
+
+    GetNArrayData(idx,nidx);
+    nidxp   = (ssize_t*)nidx->ptr;
+    q->idx  = ALLOC_N(size_t, n);
+
+    // ndixp is cuda memory (cuda narray)
+    SHOW_SYNCHRONIZE_WARNING_ONCE("na_parse_narray_index", "any");
+    cumo_cuda_runtime_check_status(cudaDeviceSynchronize());
+
+    for (k=0; k<n; k++) {
+        q->idx[k] = na_range_check(nidxp[k], size, orig_dim);
+    }
+    q->n    = n;
+    q->beg  = 0;
+    q->step = 1;
+    q->reduce = 0;
+    q->orig_dim = orig_dim;
+}
+
+static void
+na_parse_range(VALUE range, ssize_t step, int orig_dim, ssize_t size, na_index_arg_t *q)
+{
+    int n;
+    VALUE excl_end;
+    ssize_t beg, end, beg_orig, end_orig;
+    const char *dot = "..", *edot = "...";
+
+    beg = beg_orig = NUM2SSIZET(rb_funcall(range,id_beg,0));
+    if (beg < 0) {
+        beg += size;
+    }
+    end = end_orig = NUM2SSIZET(rb_funcall(range,id_end,0));
+    if (end < 0) {
+        end += size;
+    }
+    excl_end = rb_funcall(range,id_exclude_end,0);
+    if (RTEST(excl_end)) {
+        end--;
+        dot = edot;
+    }
+    if (beg < 0 || beg >= size || end < 0 || end >= size) {
+        rb_raise(rb_eRangeError,
+                 "%"SZF"d%s%"SZF"d is out of range for size=%"SZF"d",
+                 beg_orig, dot, end_orig, size);
+    }
+    n = (end-beg)/step+1;
+    if (n<0) n=0;
+    na_index_set_step(q,orig_dim,n,beg,step);
+
+}
+
+static void
+na_parse_enumerator(VALUE enum_obj, int orig_dim, ssize_t size, na_index_arg_t *q)
+{
+    int len;
+    ssize_t step;
+    struct enumerator *e;
+
+    if (!RB_TYPE_P(enum_obj, T_DATA)) {
+        rb_raise(rb_eTypeError,"wrong argument type (not T_DATA)");
+    }
+    e = (struct enumerator *)DATA_PTR(enum_obj);
+
+    if (rb_obj_is_kind_of(e->obj, rb_cRange)) {
+        if (e->meth == id_each) {
+            na_parse_range(e->obj, 1, orig_dim, size, q);
+        }
+        else if (e->meth == id_step) {
+            if (TYPE(e->args) != T_ARRAY) {
+                rb_raise(rb_eArgError,"no argument for step");
+            }
+            len = RARRAY_LEN(e->args);
+            if (len != 1) {
+                rb_raise(rb_eArgError,"invalid number of step argument (1 for %d)",len);
+            }
+            step = NUM2SSIZET(RARRAY_AREF(e->args,0));
+            na_parse_range(e->obj, step, orig_dim, size, q);
+        } else {
+            rb_raise(rb_eTypeError,"unknown Range method: %s",rb_id2name(e->meth));
+        }
+    } else {
+        rb_raise(rb_eTypeError,"not Range object");
+    }
+}
+
+// Analyze *a* which is *i*-th index object and store the information to q
+//
+// a: a ruby object of i-th index
+// size: size of i-th dimension of original NArray
+// i: parse i-th index
+// q: parsed information is stored to *q
+static void
+na_index_parse_each(volatile VALUE a, ssize_t size, int i, na_index_arg_t *q)
+{
+    switch(TYPE(a)) {
+
+    case T_FIXNUM:
+        na_index_set_scalar(q,i,size,FIX2LONG(a));
+        break;
+
+    case T_BIGNUM:
+        na_index_set_scalar(q,i,size,NUM2SSIZET(a));
+        break;
+
+    case T_FLOAT:
+        na_index_set_scalar(q,i,size,NUM2SSIZET(a));
+        break;
+
+    case T_NIL:
+    case T_TRUE:
+        na_index_set_step(q,i,size,0,1);
+        break;
+
+    case T_SYMBOL:
+        if (a==sym_all || a==sym_ast) {
+            na_index_set_step(q,i,size,0,1);
+        }
+        else if (a==sym_reverse) {
+            na_index_set_step(q,i,size,size-1,-1);
+        }
+        else if (a==sym_new) {
+            na_index_set_step(q,i,1,0,1);
+        }
+        else if (a==sym_reduce || a==sym_sum || a==sym_plus) {
+            na_index_set_step(q,i,size,0,1);
+            q->reduce = 1;
+        } else {
+            rb_raise(rb_eIndexError, "invalid symbol for index");
+        }
+        break;
+
+    case T_ARRAY:
+        na_parse_array(a, i, size, q);
+        break;
+
+    default:
+        if (rb_obj_is_kind_of(a, rb_cRange)) {
+            na_parse_range(a, 1, i, size, q);
+        }
+        else if (rb_obj_is_kind_of(a, rb_cEnumerator)) {
+            na_parse_enumerator(a, i, size, q);
+        }
+        else if (rb_obj_is_kind_of(a, na_cStep)) {
+            ssize_t beg, step, n;
+            nary_step_array_index(a, size, (size_t*)(&n), &beg, &step);
+            na_index_set_step(q,i,n,beg,step);
+        }
+        // NArray index
+        else if (NA_IsNArray(a)) {
+            na_parse_narray_index(a, i, size, q);
+        }
+        else {
+            rb_raise(rb_eIndexError, "not allowed type");
+        }
+    }
+}
+
+
+static size_t
+na_index_parse_args(VALUE args, narray_t *na, na_index_arg_t *q, int ndim)
+{
+    int i, j, k, l, nidx;
+    size_t total=1;
+    VALUE v;
+
+    if (ndim == 0) {
+        return /*total*/1;
+    }
+
+    nidx = RARRAY_LEN(args);
+
+    for (i=j=k=0; i<nidx; i++) {
+        v = RARRAY_AREF(args,i);
+        // rest (ellipsis) dimension
+        if (v==Qfalse) {
+            for (l = ndim - (nidx-1); l>0; l--) {
+                //printf("i=%d j=%d k=%d l=%d ndim=%d nidx=%d\n",i,j,k,l,ndim,nidx);
+                na_index_parse_each(Qtrue, na->shape[k], k, &q[j]);
+                if (q[j].n > 1) {
+                    total *= q[j].n;
+                }
+                j++;
+                k++;
+            }
+        }
+        // new dimension
+        else if (v==sym_new) {
+            na_index_parse_each(v, 1, k, &q[j]);
+            j++;
+        }
+        // other dimention
+        else {
+            na_index_parse_each(v, na->shape[k], k, &q[j]);
+            if (q[j].n > 1) {
+                total *= q[j].n;
+            }
+            j++;
+            k++;
+        }
+    }
+    return total;
+}
+
+
+static void
+na_get_strides_nadata(const narray_data_t *na, ssize_t *strides, ssize_t elmsz)
+{
+    int i = na->base.ndim - 1;
+    strides[i] = elmsz;
+    for (; i>0; i--) {
+        strides[i-1] = strides[i] * na->base.shape[i];
+    }
+}
+
+static void
+na_index_aref_nadata(narray_data_t *na1, narray_view_t *na2,
+                     na_index_arg_t *q, ssize_t elmsz, int ndim, int keep_dim)
+{
+    int i, j;
+    ssize_t size, k, total=1;
+    ssize_t stride1;
+    ssize_t *strides_na1;
+    size_t  *index;
+    ssize_t beg, step;
+    VALUE m;
+
+    strides_na1 = ALLOCA_N(ssize_t, na1->base.ndim);
+    na_get_strides_nadata(na1, strides_na1, elmsz);
+
+    for (i=j=0; i<ndim; i++) {
+        stride1 = strides_na1[q[i].orig_dim];
+
+        // numeric index -- trim dimension
+        if (!keep_dim && q[i].n==1 && q[i].step==0) {
+            beg  = q[i].beg;
+            na2->offset += stride1 * beg;
+            continue;
+        }
+
+        na2->base.shape[j] = size = q[i].n;
+
+        if (q[i].reduce != 0) {
+            m = rb_funcall(INT2FIX(1),id_shift_left,1,INT2FIX(j));
+            na2->base.reduce = rb_funcall(m,'|',1,na2->base.reduce);
+        }
+
+        // array index
+        if (q[i].idx != NULL) {
+            index = q[i].idx;
+            SDX_SET_INDEX(na2->stridx[j],index);
+            q[i].idx = NULL;
+            for (k=0; k<size; k++) {
+                index[k] = index[k] * stride1;
+            }
+        } else {
+            beg  = q[i].beg;
+            step = q[i].step;
+            na2->offset += stride1*beg;
+            SDX_SET_STRIDE(na2->stridx[j], stride1*step);
+        }
+        j++;
+        total *= size;
+    }
+    na2->base.size = total;
+}
+
+
+static void
+na_index_aref_naview(narray_view_t *na1, narray_view_t *na2,
+                     na_index_arg_t *q, ssize_t elmsz, int ndim, int keep_dim)
+{
+    int i, j;
+    ssize_t total=1;
+
+    for (i=j=0; i<ndim; i++) {
+        stridx_t sdx1 = na1->stridx[q[i].orig_dim];
+        ssize_t size;
+
+        // numeric index -- trim dimension
+        if (!keep_dim && q[i].n==1 && q[i].step==0) {
+            if (SDX_IS_INDEX(sdx1)) {
+                na2->offset += SDX_GET_INDEX(sdx1)[q[i].beg];
+            } else {
+                na2->offset += SDX_GET_STRIDE(sdx1)*q[i].beg;
+            }
+            continue;
+        }
+
+        na2->base.shape[j] = size = q[i].n;
+
+        if (q[i].reduce != 0) {
+            VALUE m = rb_funcall(INT2FIX(1),id_shift_left,1,INT2FIX(j));
+            na2->base.reduce = rb_funcall(m,'|',1,na2->base.reduce);
+        }
+
+        if (q[i].orig_dim >= na1->base.ndim) {
+            // new dimension
+            SDX_SET_STRIDE(na2->stridx[j], elmsz);
+        }
+        else if (q[i].idx != NULL && SDX_IS_INDEX(sdx1)) {
+            // index <- index
+            int k;
+            size_t *index = q[i].idx;
+            SDX_SET_INDEX(na2->stridx[j], index);
+            q[i].idx = NULL;
+
+            for (k=0; k<size; k++) {
+                index[k] = SDX_GET_INDEX(sdx1)[index[k]];
+            }
+        }
+        else if (q[i].idx != NULL && SDX_IS_STRIDE(sdx1)) {
+            // index <- step
+            ssize_t stride1 = SDX_GET_STRIDE(sdx1);
+            size_t *index = q[i].idx;
+            SDX_SET_INDEX(na2->stridx[j],index);
+            q[i].idx = NULL;
+
+            if (stride1<0) {
+                size_t  last;
+                int k;
+                stride1 = -stride1;
+                last = na1->base.shape[q[i].orig_dim] - 1;
+                if (na2->offset < last * stride1) {
+                    rb_raise(rb_eStandardError,"bug: negative offset");
+                }
+                na2->offset -= last * stride1;
+                for (k=0; k<size; k++) {
+                    index[k] = (last - index[k]) * stride1;
+                }
+            } else {
+                int k;
+                for (k=0; k<size; k++) {
+                    index[k] = index[k] * stride1;
+                }
+            }
+        }
+        else if (q[i].idx == NULL && SDX_IS_INDEX(sdx1)) {
+            // step <- index
+            int k;
+            size_t beg  = q[i].beg;
+            ssize_t step = q[i].step;
+            size_t *index = ALLOC_N(size_t, size);
+            SDX_SET_INDEX(na2->stridx[j],index);
+            for (k=0; k<size; k++) {
+                index[k] = SDX_GET_INDEX(sdx1)[beg+step*k];
+            }
+        }
+        else if (q[i].idx == NULL && SDX_IS_STRIDE(sdx1)) {
+            // step <- step
+            size_t beg  = q[i].beg;
+            ssize_t step = q[i].step;
+            ssize_t stride1 = SDX_GET_STRIDE(sdx1);
+            na2->offset += stride1*beg;
+            SDX_SET_STRIDE(na2->stridx[j], stride1*step);
+        }
+
+        j++;
+        total *= size;
+    }
+    na2->base.size = total;
+}
+
+
+static int
+na_ndim_new_narray(int ndim, const na_index_arg_t *q)
+{
+    int i, ndim_new=0;
+    for (i=0; i<ndim; i++) {
+        if (q[i].n>1 || q[i].step!=0) {
+            ndim_new++;
+        }
+    }
+    return ndim_new;
+}
+
+typedef struct {
+    VALUE args, self, store;
+    int ndim;
+    na_index_arg_t *q; // multi-dimensional index args
+    narray_t *na1;
+    int keep_dim;
+    size_t pos; // offset position for 0-dimensional narray. 0-dimensional array does not use q.
+} na_aref_md_data_t;
+
+static na_index_arg_t*
+na_allocate_index_args(int ndim)
+{
+    na_index_arg_t *q;
+    int i;
+    if (ndim == 0) return NULL;
+
+    q = ALLOC_N(na_index_arg_t, ndim);
+    for (i=0; i<ndim; i++) {
+        q[i].idx = NULL;
+    }
+    return q;
+}
+
+static
+VALUE na_aref_md_protected(VALUE data_value)
+{
+    na_aref_md_data_t *data = (na_aref_md_data_t*)(data_value);
+    VALUE self = data->self;
+    VALUE args = data->args;
+    VALUE store = data->store;
+    int ndim = data->ndim;
+    na_index_arg_t *q = data->q;
+    narray_t *na1 = data->na1;
+    int keep_dim = data->keep_dim;
+
+    int ndim_new;
+    VALUE view;
+    narray_view_t *na2;
+    ssize_t elmsz;
+
+    na_index_parse_args(args, na1, q, ndim);
+
+    if (na_debug_flag) print_index_arg(q,ndim);
+
+    if (keep_dim) {
+        ndim_new = ndim;
+    } else {
+        ndim_new = na_ndim_new_narray(ndim, q);
+    }
+    view = na_s_allocate_view(CLASS_OF(self));
+
+    na_copy_flags(self, view);
+    GetNArrayView(view,na2);
+
+    na_alloc_shape((narray_t*)na2, ndim_new);
+
+    na2->stridx = ALLOC_N(stridx_t,ndim_new);
+
+    elmsz = nary_element_stride(self);
+
+    switch(na1->type) {
+    case NARRAY_DATA_T:
+    case NARRAY_FILEMAP_T:
+        if (ndim == 0) {
+            na2->offset = data->pos;
+            na2->base.size = 1;
+        } else {
+            na_index_aref_nadata((narray_data_t *)na1,na2,q,elmsz,ndim,keep_dim);
+        }
+        na2->data = self;
+        break;
+    case NARRAY_VIEW_T:
+        if (ndim == 0) {
+            na2->offset = ((narray_view_t *)na1)->offset + data->pos;
+            na2->data = ((narray_view_t *)na1)->data;
+            na2->base.size = 1;
+        } else {
+            na2->offset = ((narray_view_t *)na1)->offset;
+            na2->data = ((narray_view_t *)na1)->data;
+            na_index_aref_naview((narray_view_t *)na1,na2,q,elmsz,ndim,keep_dim);
+        }
+        break;
+    }
+    if (store) {
+        na_get_pointer_for_write(store); // allocate memory
+        na_store(na_flatten_dim(store,0),view);
+        return store;
+    }
+    return view;
+}
+
+static VALUE
+na_aref_md_ensure(VALUE data_value)
+{
+    na_aref_md_data_t *data = (na_aref_md_data_t*)(data_value);
+    int i;
+    for (i=0; i<data->ndim; i++) {
+        xfree(data->q[i].idx);
+    }
+    if (data->q) xfree(data->q);
+    return Qnil;
+}
+
+static VALUE
+na_aref_md(int argc, VALUE *argv, VALUE self, int keep_dim, int result_nd, size_t pos)
+{
+    VALUE args; // should be GC protected
+    narray_t *na1;
+    na_aref_md_data_t data;
+    VALUE store = 0;
+    VALUE idx;
+    narray_t *nidx;
+
+    GetNArray(self,na1);
+
+    args = rb_ary_new4(argc,argv);
+
+    if (argc == 1 && result_nd == 1) {
+        idx = argv[0];
+        if (rb_obj_is_kind_of(idx, rb_cArray)) {
+            idx = rb_apply(cumo_cNArray,id_bracket,idx);
+        }
+        if (rb_obj_is_kind_of(idx, cumo_cNArray)) {
+            GetNArray(idx,nidx);
+            if (NA_NDIM(nidx)>1) {
+                store = nary_new(CLASS_OF(self),NA_NDIM(nidx),NA_SHAPE(nidx));
+                idx = na_flatten(idx);
+                RARRAY_ASET(args,0,idx);
+            }
+        }
+        // flatten should be done only for narray-view with non-uniform stride.
+        if (na1->ndim > 1) {
+            self = na_flatten(self);
+            GetNArray(self,na1);
+        }
+    }
+
+    data.args = args;
+    data.self = self;
+    data.store = store;
+    data.ndim = result_nd;
+    data.q = na_allocate_index_args(result_nd);
+    data.na1 = na1;
+    data.keep_dim = keep_dim;
+
+    switch(na1->type) {
+    case NARRAY_DATA_T:
+        data.pos = pos;
+        break;
+    case NARRAY_FILEMAP_T:
+        data.pos = pos; // correct? I have never used..
+        break;
+    case NARRAY_VIEW_T:
+        {
+            narray_view_t *nv;
+            GetNArrayView(self,nv);
+            // pos obtained by na_get_result_dimension adds view->offset.
+            data.pos = pos - nv->offset;
+        }
+        break;
+    }
+
+    return rb_ensure(na_aref_md_protected, (VALUE)&data, na_aref_md_ensure, (VALUE)&data);
+}
+
+
+/* method: [](idx1,idx2,...,idxN) */
+VALUE
+na_aref_main(int nidx, VALUE *idx, VALUE self, int keep_dim, int result_nd, size_t pos)
+{
+    na_index_arg_to_internal_order(nidx, idx, self);
+
+    if (nidx==0) {
+        return rb_funcall(self,id_dup,0);
+    }
+    if (nidx==1) {
+        if (CLASS_OF(*idx)==cumo_cBit) {
+            return rb_funcall(*idx,id_mask,1,self);
+        }
+    }
+    return na_aref_md(nidx, idx, self, keep_dim, result_nd, pos);
+}
+
+
+/* method: slice(idx1,idx2,...,idxN) */
+static VALUE na_slice(int argc, VALUE *argv, VALUE self)
+{
+    int result_nd;
+    size_t pos;
+
+    result_nd = na_get_result_dimension(self, argc, argv, 0, &pos);
+    return na_aref_main(argc, argv, self, 1, result_nd, pos);
+}
+
+
+static int
+check_index_count(int argc, int na_ndim, int count_new, int count_rest)
+{
+    int result_nd = na_ndim + count_new;
+
+    switch(count_rest) {
+    case 0:
+        if (count_new == 0 && argc == 1) return 1;
+        if (argc == result_nd) return result_nd;
+        rb_raise(rb_eIndexError,"# of index(=%i) should be "
+                 "equal to ndim(=%i)",argc,na_ndim);
+        break;
+    case 1:
+        if (argc-1 <= result_nd) return result_nd;
+        rb_raise(rb_eIndexError,"# of index(=%i) > ndim(=%i) with :rest",
+                 argc,na_ndim);
+        break;
+    }
+    return -1;
+}
+
+int
+na_get_result_dimension(VALUE self, int argc, VALUE *argv, ssize_t stride, size_t *pos_idx)
+{
+    int i, j;
+    int count_new=0;
+    int count_rest=0;
+    int count_else=0;
+    ssize_t x, s, m, pos, *idx;
+    narray_t *na;
+    narray_view_t *nv;
+    stridx_t sdx;
+    VALUE a;
+
+    GetNArray(self,na);
+    if (na->size == 0) {
+        rb_raise(rb_eRuntimeError, "cannot get index of empty array");
+        return -1;
+    }
+    idx = ALLOCA_N(ssize_t, argc);
+    for (i=j=0; i<argc; i++) {
+        a = argv[i];
+        switch(TYPE(a)) {
+        case T_FIXNUM:
+            idx[j++] = FIX2LONG(a);
+            break;
+        case T_BIGNUM:
+        case T_FLOAT:
+            idx[j++] = NUM2SSIZET(a);
+            break;
+        case T_FALSE:
+        case T_SYMBOL:
+            if (a==sym_rest || a==sym_tilde || a==Qfalse) {
+                argv[i] = Qfalse;
+                count_rest++;
+                break;
+            } else if (a==sym_new || a==sym_minus) {
+                argv[i] = sym_new;
+                count_new++;
+            }
+            // not break
+        default:
+            count_else++;
+        }
+    }
+
+    if (count_rest > 1) {
+        rb_raise(rb_eIndexError,"multiple rest-dimension is not allowd");
+    }
+    if (count_else != 0) {
+        return check_index_count(argc, na->ndim, count_new, count_rest);
+    }
+
+    switch(na->type) {
+    case NARRAY_VIEW_T:
+        GetNArrayView(self,nv);
+        pos = nv->offset;
+        if (j == na->ndim) {
+            for (i=j-1; i>=0; i--) {
+                x = na_range_check(idx[i], na->shape[i], i);
+                sdx = nv->stridx[i];
+                if (SDX_IS_INDEX(sdx)) {
+                    pos += SDX_GET_INDEX(sdx)[x];
+                } else {
+                    pos += SDX_GET_STRIDE(sdx)*x;
+                }
+            }
+            *pos_idx = pos;
+        }
+        else if (argc==1 && j==1) {
+            x = na_range_check(idx[0], na->size, 0);
+            for (i=na->ndim-1; i>=0; i--) {
+                s = na->shape[i];
+                m = x % s;
+                x = x / s;
+                sdx = nv->stridx[i];
+                if (SDX_IS_INDEX(sdx)) {
+                    pos += SDX_GET_INDEX(sdx)[m];
+                } else {
+                    pos += SDX_GET_STRIDE(sdx)*m;
+                }
+            }
+            *pos_idx = pos;
+        } else {
+            return check_index_count(argc, na->ndim, count_new, count_rest);
+        }
+        break;
+    default:
+        if (!stride) {
+            stride = nary_element_stride(self);
+        }
+        if (argc==1 && j==1) {
+            x = na_range_check(idx[0], na->size, 0);
+            *pos_idx = stride * x;
+        }
+        else if (j == na->ndim) {
+            pos = 0;
+            for (i=j-1; i>=0; i--) {
+                x = na_range_check(idx[i], na->shape[i], i);
+                pos += stride * x;
+                stride *= na->shape[i];
+            }
+            *pos_idx = pos;
+        } else {
+            return check_index_count(argc, na->ndim, count_new, count_rest);
+        }
+    }
+    return 0;
+}
+
+
+void
+Init_cumo_nary_index()
+{
+    rb_define_method(cNArray, "slice", na_slice, -1);
+
+    sym_ast        = ID2SYM(rb_intern("*"));
+    sym_all        = ID2SYM(rb_intern("all"));
+    sym_minus      = ID2SYM(rb_intern("-"));
+    sym_new        = ID2SYM(rb_intern("new"));
+    sym_reverse    = ID2SYM(rb_intern("reverse"));
+    sym_plus       = ID2SYM(rb_intern("+"));
+    //sym_reduce   = ID2SYM(rb_intern("reduce"));
+    sym_sum        = ID2SYM(rb_intern("sum"));
+    sym_tilde      = ID2SYM(rb_intern("~"));
+    sym_rest       = ID2SYM(rb_intern("rest"));
+    id_beg         = rb_intern("begin");
+    id_end         = rb_intern("end");
+    id_exclude_end = rb_intern("exclude_end?");
+    id_each        = rb_intern("each");
+    id_step        = rb_intern("step");
+    id_dup         = rb_intern("dup");
+    id_bracket     = rb_intern("[]");
+    id_shift_left  = rb_intern("<<");
+    id_mask        = rb_intern("mask");
+}