From 00f81b6f31c011fc601a266320f9a68b77383583 Mon Sep 17 00:00:00 2001 From: Daniel Llorens Date: Fri, 13 Feb 2015 18:42:27 +0100 Subject: [PATCH] Speed up for multi-arg cases of scm_ramap functions This patch results in a 20%-40% speedup in the > 1 argument cases of the following microbenchmarks: (define A (make-shared-array #0(1) (const '()) #e1e7)) ; 1, 2, 3 arguments. (define a 0) ,time (array-for-each (lambda (b) (set! a (+ a b))) A) (define a 0) ,time (array-for-each (lambda (b c) (set! a (+ a b c))) A A) (define a 0) ,time (array-for-each (lambda (b c d) (set! a (+ a b c d))) A A A) (define A (make-shared-array (make-array 1) (const '()) #e1e7)) (define B (make-shared-array #0(1) (const '()) #e1e7)) ; 1, 2, 3 arguments. ,time (array-map! A + B) ,time (array-map! A + B B) ,time (array-map! A + B B B) * libguile/array-map.c (scm_ramap): Note on cproc arguments. (rafill): Assume that dst's lbnd is 0. (racp): Assume that src's lbnd is 0. (ramap): Assume that ra0's lbnd is 0. When there're more than two arguments, compute the array handles before the loop. Allocate the arg list once and reuse it in the loop. (rafe): Do as in ramap(), when there's more than one argument. (AREF, ASET): Remove. --- libguile/array-map.c | 136 ++++++++++++++++++++---------------- libguile/array-map.h | 2 +- test-suite/tests/ramap.test | 4 +- 3 files changed, 77 insertions(+), 65 deletions(-) diff --git a/libguile/array-map.c b/libguile/array-map.c index 587df023b..9caded8a3 100644 --- a/libguile/array-map.c +++ b/libguile/array-map.c @@ -1,5 +1,5 @@ /* Copyright (C) 1996, 1998, 2000, 2001, 2004, 2005, 2006, 2008, 2009, - * 2010, 2011, 2012, 2013, 2014 Free Software Foundation, Inc. + * 2010, 2011, 2012, 2013, 2014, 2015 Free Software Foundation, Inc. * * This library is free software; you can redistribute it and/or * modify it under the terms of the GNU Lesser General Public License @@ -47,18 +47,6 @@ /* The WHAT argument for `scm_gc_malloc ()' et al. */ static const char vi_gc_hint[] = "array-indices"; -static SCM -AREF (SCM v, size_t pos) -{ - return scm_c_array_ref_1 (v, pos); -} - -static void -ASET (SCM v, size_t pos, SCM val) -{ - scm_c_array_set_1_x (v, val, pos); -} - static SCM make1array (SCM v, ssize_t inc) { @@ -99,6 +87,10 @@ cindk (SCM ra, ssize_t *ve, int kend) #define LBND(ra, k) SCM_I_ARRAY_DIMS (ra)[k].lbnd #define UBND(ra, k) SCM_I_ARRAY_DIMS (ra)[k].ubnd + +/* scm_ramapc() always calls cproc with rank-1 arrays created by + make1array. cproc (rafe, ramap, rafill, racp) can assume that the + dims[0].lbnd of these arrays is always 0. */ int scm_ramapc (void *cproc_ptr, SCM data, SCM ra0, SCM lra, const char *what) { @@ -167,7 +159,7 @@ scm_ramapc (void *cproc_ptr, SCM data, SCM ra0, SCM lra, const char *what) scm_misc_error (what, "array shape mismatch: ~S", scm_list_1 (ra0)); va1 = make1array (ra1, 1); - if (LBND (ra0, 0) < LBND (va1, 0) || UBND (ra0, 0) > UBND (va1, 0)) + if (LBND (ra0, 0) < 0 /* LBND (va1, 0) */ || UBND (ra0, 0) > UBND (va1, 0)) scm_misc_error (what, "array shape mismatch: ~S", scm_list_1 (ra0)); } *plva = scm_cons (va1, SCM_EOL); @@ -224,14 +216,12 @@ scm_ramapc (void *cproc_ptr, SCM data, SCM ra0, SCM lra, const char *what) static int rafill (SCM dst, SCM fill) { + size_t n = SCM_I_ARRAY_DIMS (dst)->ubnd + 1; + size_t i = SCM_I_ARRAY_BASE (dst); + ssize_t inc = SCM_I_ARRAY_DIMS (dst)->inc; scm_t_array_handle h; - size_t n, i; - ssize_t inc; - scm_array_get_handle (SCM_I_ARRAY_V (dst), &h); - i = SCM_I_ARRAY_BASE (dst); - inc = SCM_I_ARRAY_DIMS (dst)->inc; - n = (SCM_I_ARRAY_DIMS (dst)->ubnd - SCM_I_ARRAY_DIMS (dst)->lbnd + 1); dst = SCM_I_ARRAY_V (dst); + scm_array_get_handle (dst, &h); for (; n-- > 0; i += inc) h.vset (h.vector, i, fill); @@ -255,19 +245,17 @@ SCM_DEFINE (scm_array_fill_x, "array-fill!", 2, 0, 0, static int racp (SCM src, SCM dst) { - scm_t_array_handle h_s, h_d; - size_t n, i_s, i_d; + size_t i_s, i_d, n; ssize_t inc_s, inc_d; - + scm_t_array_handle h_s, h_d; dst = SCM_CAR (dst); i_s = SCM_I_ARRAY_BASE (src); i_d = SCM_I_ARRAY_BASE (dst); + n = (SCM_I_ARRAY_DIMS (src)->ubnd + 1); inc_s = SCM_I_ARRAY_DIMS (src)->inc; inc_d = SCM_I_ARRAY_DIMS (dst)->inc; - n = (SCM_I_ARRAY_DIMS (src)->ubnd - SCM_I_ARRAY_DIMS (src)->lbnd + 1); src = SCM_I_ARRAY_V (src); dst = SCM_I_ARRAY_V (dst); - scm_array_get_handle (src, &h_s); scm_array_get_handle (dst, &h_d); @@ -310,44 +298,56 @@ SCM_DEFINE (scm_array_copy_x, "array-copy!", 2, 0, 0, static int ramap (SCM ra0, SCM proc, SCM ras) { + size_t i0 = SCM_I_ARRAY_BASE (ra0); + ssize_t inc0 = SCM_I_ARRAY_DIMS (ra0)->inc; + size_t n = SCM_I_ARRAY_DIMS (ra0)->ubnd + 1; scm_t_array_handle h0; - size_t n, i0; - ssize_t i, inc0; - i0 = SCM_I_ARRAY_BASE (ra0); - inc0 = SCM_I_ARRAY_DIMS (ra0)->inc; - i = SCM_I_ARRAY_DIMS (ra0)->lbnd; - n = SCM_I_ARRAY_DIMS (ra0)->ubnd - i + 1; ra0 = SCM_I_ARRAY_V (ra0); scm_array_get_handle (ra0, &h0); + if (scm_is_null (ras)) for (; n--; i0 += inc0) h0.vset (h0.vector, i0, scm_call_0 (proc)); else { SCM ra1 = SCM_CAR (ras); + size_t i1 = SCM_I_ARRAY_BASE (ra1); + ssize_t inc1 = SCM_I_ARRAY_DIMS (ra1)->inc; scm_t_array_handle h1; - size_t i1; - ssize_t inc1; - i1 = SCM_I_ARRAY_BASE (ra1); - inc1 = SCM_I_ARRAY_DIMS (ra1)->inc; - ras = SCM_CDR (ras); ra1 = SCM_I_ARRAY_V (ra1); scm_array_get_handle (ra1, &h1); + ras = SCM_CDR (ras); if (scm_is_null (ras)) for (; n--; i0 += inc0, i1 += inc1) h0.vset (h0.vector, i0, scm_call_1 (proc, h1.vref (h1.vector, i1))); else { - ras = scm_vector (ras); - for (; n--; i0 += inc0, i1 += inc1, ++i) + scm_t_array_handle *hs; + size_t restn = scm_ilength (ras); + + SCM args = SCM_EOL; + SCM *p = &args; + SCM **sa = scm_gc_malloc (sizeof(SCM *) * restn, vi_gc_hint); + for (size_t k = 0; k < restn; ++k) { - SCM args = SCM_EOL; - unsigned long k; - for (k = scm_c_vector_length (ras); k--;) - args = scm_cons (AREF (scm_c_vector_ref (ras, k), i), args); - h0.vset (h0.vector, i0, - scm_apply_1 (proc, h1.vref (h1.vector, i1), args)); + *p = scm_cons (SCM_UNSPECIFIED, SCM_EOL); + sa[k] = SCM_CARLOC (*p); + p = SCM_CDRLOC (*p); } + + hs = scm_gc_malloc (sizeof(scm_t_array_handle) * restn, vi_gc_hint); + for (size_t k = 0; k < restn; ++k, ras = scm_cdr (ras)) + scm_array_get_handle (scm_car (ras), hs+k); + + for (ssize_t i = 0; n--; i0 += inc0, i1 += inc1, ++i) + { + for (size_t k = 0; k < restn; ++k) + *(sa[k]) = scm_array_handle_ref (hs+k, i*hs[k].dims[0].inc); + h0.vset (h0.vector, i0, scm_apply_1 (proc, h1.vref (h1.vector, i1), args)); + } + + for (size_t k = 0; k < restn; ++k) + scm_array_handle_release (hs+k); } scm_array_handle_release (&h1); } @@ -384,30 +384,44 @@ SCM_DEFINE (scm_array_map_x, "array-map!", 2, 0, 1, static int rafe (SCM ra0, SCM proc, SCM ras) { - ssize_t i = SCM_I_ARRAY_DIMS (ra0)->lbnd; - size_t n = SCM_I_ARRAY_DIMS (ra0)->ubnd - i + 1; - + size_t i0 = SCM_I_ARRAY_BASE (ra0); + ssize_t inc0 = SCM_I_ARRAY_DIMS (ra0)->inc; + size_t n = SCM_I_ARRAY_DIMS (ra0)->ubnd + 1; scm_t_array_handle h0; - size_t i0; - ssize_t inc0; - i0 = SCM_I_ARRAY_BASE (ra0); - inc0 = SCM_I_ARRAY_DIMS (ra0)->inc; ra0 = SCM_I_ARRAY_V (ra0); scm_array_get_handle (ra0, &h0); + if (scm_is_null (ras)) for (; n--; i0 += inc0) scm_call_1 (proc, h0.vref (h0.vector, i0)); else { - ras = scm_vector (ras); - for (; n--; i0 += inc0, ++i) + scm_t_array_handle *hs; + size_t restn = scm_ilength (ras); + + SCM args = SCM_EOL; + SCM *p = &args; + SCM **sa = scm_gc_malloc (sizeof(SCM *) * restn, vi_gc_hint); + for (size_t k = 0; k < restn; ++k) { - SCM args = SCM_EOL; - unsigned long k; - for (k = scm_c_vector_length (ras); k--;) - args = scm_cons (AREF (scm_c_vector_ref (ras, k), i), args); + *p = scm_cons (SCM_UNSPECIFIED, SCM_EOL); + sa[k] = SCM_CARLOC (*p); + p = SCM_CDRLOC (*p); + } + + hs = scm_gc_malloc (sizeof(scm_t_array_handle) * restn, vi_gc_hint); + for (size_t k = 0; k < restn; ++k, ras = scm_cdr (ras)) + scm_array_get_handle (scm_car (ras), hs+k); + + for (ssize_t i = 0; n--; i0 += inc0, ++i) + { + for (size_t k = 0; k < restn; ++k) + *(sa[k]) = scm_array_handle_ref (hs+k, i*hs[k].dims[0].inc); scm_apply_1 (proc, h0.vref (h0.vector, i0), args); } + + for (size_t k = 0; k < restn; ++k) + scm_array_handle_release (hs+k); } scm_array_handle_release (&h0); return 1; @@ -445,15 +459,12 @@ static void array_index_map_n (SCM ra, SCM proc) { scm_t_array_handle h; - size_t i; int k, kmax = SCM_I_ARRAY_NDIM (ra) - 1; - ssize_t *vi; - SCM **si; SCM args = SCM_EOL; SCM *p = &args; - vi = scm_gc_malloc_pointerless (sizeof(ssize_t) * (kmax + 1), vi_gc_hint); - si = scm_gc_malloc_pointerless (sizeof(SCM *) * (kmax + 1), vi_gc_hint); + ssize_t *vi = scm_gc_malloc_pointerless (sizeof(ssize_t) * (kmax + 1), vi_gc_hint); + SCM **si = scm_gc_malloc_pointerless (sizeof(SCM *) * (kmax + 1), vi_gc_hint); for (k = 0; k <= kmax; k++) { @@ -471,6 +482,7 @@ array_index_map_n (SCM ra, SCM proc) { if (k == kmax) { + size_t i; vi[kmax] = SCM_I_ARRAY_DIMS (ra)[kmax].lbnd; i = cindk (ra, vi, kmax+1); for (; vi[kmax] <= SCM_I_ARRAY_DIMS (ra)[kmax].ubnd; ++vi[kmax]) diff --git a/libguile/array-map.h b/libguile/array-map.h index e7431b176..cb18a628a 100644 --- a/libguile/array-map.h +++ b/libguile/array-map.h @@ -4,7 +4,7 @@ #define SCM_ARRAY_MAP_H /* Copyright (C) 1995, 1996, 1997, 2000, 2006, 2008, 2009, 2010, - * 2011, 2013 Free Software Foundation, Inc. + * 2011, 2013, 2015 Free Software Foundation, Inc. * * This library is free software; you can redistribute it and/or * modify it under the terms of the GNU Lesser General Public License diff --git a/test-suite/tests/ramap.test b/test-suite/tests/ramap.test index c8eaf96eb..bd8a434bd 100644 --- a/test-suite/tests/ramap.test +++ b/test-suite/tests/ramap.test @@ -453,11 +453,11 @@ (with-test-prefix "3 sources" (pass-if-equal "noncompact arrays 1" - '((3 3 3) (2 2 2)) + '((3 1 3) (2 0 2)) (let* ((a #2((0 1) (2 3))) (l '()) (rec (lambda args (set! l (cons args l))))) - (array-for-each rec (array-row a 1) (array-row a 1) (array-row a 1)) + (array-for-each rec (array-row a 1) (array-row a 0) (array-row a 1)) l)) (pass-if-equal "noncompact arrays 2"