sell_kacz_fallback.cpp File Reference

#include "ghost/sell_kacz_fallback.h"
#include "ghost/cpp11_fixes.h"
#include <complex>

Include dependency graph for sell_kacz_fallback.cpp:

Macros
#define	LOOP_IN_CHUNK(row)
#define	FORWARD_LOOP(start, end, MT, VT)
#define	BACKWARD_LOOP(start, end, MT, VT)
#define	LOCK_NEIGHBOUR(tid)

Functions
template<typename MT , typename VT >
static ghost_error	ghost_kacz_fallback_tmpl (ghost_densemat *x, ghost_sparsemat *mat, ghost_densemat *b, ghost_kacz_opts opts)
ghost_error	ghost_kacz_fallback (ghost_densemat *x, ghost_sparsemat *mat, ghost_densemat *b, ghost_kacz_opts opts)

Macro Definition Documentation

#define BACKWARD_LOOP	(		start,
			end,
			MT,
			VT
	)

Value:

{\
    start_rem = start%CHUNKHEIGHT; \
    start_chunk = start/CHUNKHEIGHT; \
    end_chunk = end/CHUNKHEIGHT; \
    end_rem = end%CHUNKHEIGHT; \
    chunk = 0; \
    rowinchunk = 0; \
    idx=0, row=0; \
    for(rowinchunk=start_rem; rowinchunk>=MAX(0,(end_chunk-start_chunk)*CHUNKHEIGHT+end_rem+1); --rowinchunk) { \
        row = rowinchunk + (start_chunk)*CHUNKHEIGHT; \
        chunk = start_chunk;\
        LOOP_IN_CHUNK(row) \
    } \
    _Pragma("omp parallel for private(chunk, rowinchunk, idx, row)") \
    for (chunk=start_chunk-1; chunk>end_chunk; --chunk){ \
        for(rowinchunk=CHUNKHEIGHT-1; rowinchunk>=0; --rowinchunk) { \
            row = rowinchunk + chunk*CHUNKHEIGHT; \
            LOOP_IN_CHUNK(row) \
        } \
    } \
    if(start_chunk>end_chunk) { \
        for(rowinchunk=CHUNKHEIGHT-1; rowinchunk>end_rem; --rowinchunk) { \
            row = rowinchunk + (end_chunk)*CHUNKHEIGHT; \
            chunk = end_chunk; \
            LOOP_IN_CHUNK(row) \
        } \
    } \
}\

#define FORWARD_LOOP	(		start,
			end,
			MT,
			VT
	)

Value:

{\
    start_rem = start%CHUNKHEIGHT; \
    start_chunk = start/CHUNKHEIGHT; \
    end_chunk = end/CHUNKHEIGHT; \
    end_rem = end%CHUNKHEIGHT; \
    chunk = 0; \
    rowinchunk = 0; \
    idx=0, row=0; \
    for(rowinchunk=start_rem; rowinchunk<MIN(CHUNKHEIGHT,(end_chunk-start_chunk)*CHUNKHEIGHT+end_rem); ++rowinchunk) { \
        row = rowinchunk + (start_chunk)*CHUNKHEIGHT; \
        chunk = start_chunk;\
        LOOP_IN_CHUNK(row); \
    } \
    _Pragma("omp parallel for private(chunk, rowinchunk, idx, row)") \
    for (chunk=start_chunk+1; chunk<end_chunk; ++chunk){ \
        for(rowinchunk=0; rowinchunk<CHUNKHEIGHT; ++rowinchunk) { \
            row = rowinchunk + chunk*CHUNKHEIGHT; \
            LOOP_IN_CHUNK(row); \
        } \
    } \
    if(start_chunk<end_chunk) { \
        for(rowinchunk=0; rowinchunk<end_rem; ++rowinchunk) { \
            row = rowinchunk + (end_chunk)*CHUNKHEIGHT; \
            chunk = end_chunk;\
            LOOP_IN_CHUNK(row); \
        } \
    }\
}\

#define LOCK_NEIGHBOUR

(

tid

)

Value:

if(tid == 0) \
    flag[0] = zone+1; \
if(tid == nthreads-1) \
    flag[nthreads+1] = zone+1; \
    flag[tid+1] = zone+1; \
    _Pragma("omp flush") \
    \
    if(opts.direction == GHOST_KACZ_DIRECTION_FORWARD) {\
        while(flag[tid+2]<zone+1){ \
            _Pragma("omp flush") \
        } \
    } else { \
        while(flag[tid]<zone+1 ){ \
            _Pragma("omp flush") \
        } \
    }

#define LOOP_IN_CHUNK

(

row

)

Function Documentation

ghost_error ghost_kacz_fallback	(	ghost_densemat *	x,
		ghost_sparsemat *	mat,
		ghost_densemat *	b,
		ghost_kacz_opts	opts
	)

Here is the call graph for this function:

template<typename MT , typename VT >

static ghost_error ghost_kacz_fallback_tmpl ( ghost_densemat *  x, ghost_sparsemat *  mat, ghost_densemat *  b, ghost_kacz_opts  opts  )

static

Here is the call graph for this function: