diff options
author | Quincey Koziol <koziol@hdfgroup.org> | 2003-11-07 20:21:45 (GMT) |
---|---|---|
committer | Quincey Koziol <koziol@hdfgroup.org> | 2003-11-07 20:21:45 (GMT) |
commit | 1716cab24b78f7cdd0590da2b30a35320af9456b (patch) | |
tree | b53873ebdd1b17edabecbc90bb90aed14437cb89 | |
parent | 9bd97b4496fa2159289bb7b2b789b2c46473244b (diff) | |
download | hdf5-1716cab24b78f7cdd0590da2b30a35320af9456b.zip hdf5-1716cab24b78f7cdd0590da2b30a35320af9456b.tar.gz hdf5-1716cab24b78f7cdd0590da2b30a35320af9456b.tar.bz2 |
[svn-r7825] Purpose:
Code optimization
Description:
Improved integer & floating-point datatype conversions by removing some
corner cases. Got rid of algorithm which walked the buffer of elements to
convert backwards, switching to algorithm which computes the non-overlapping
space in the buffer and repeatedly converts it in the forward direction.
Platforms tested:
FreeBSD 4.9 (sleipnir)
h5committest
Cray SV1 & T3E (T90 is not accessible)
-rw-r--r-- | release_docs/RELEASE.txt | 2 | ||||
-rw-r--r-- | src/H5Tconv.c | 149 |
2 files changed, 98 insertions, 53 deletions
diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index 7a14a4c..decb346 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -77,6 +77,8 @@ Bug Fixes since HDF5-1.6.1 release Performance ------------- + - More optimizations to inner loops of datatype conversions for + integers and floats which give a 10-50% speedup. QAK - 2003/11/07 - Hoisted invariant 'if/else's out of inner datatype conversion loop for integer and floating-point values, giving about a 20% speedup. QAK - 2003/10/20 diff --git a/src/H5Tconv.c b/src/H5Tconv.c index cab5553..5a66ef2 100644 --- a/src/H5Tconv.c +++ b/src/H5Tconv.c @@ -172,10 +172,6 @@ H5FL_BLK_DEFINE_STATIC(array_seq); * as the destination. Overflows can occur when the destination is * narrower than the source. * - * SU: Generic signed to unsigned conversion where the source is - * the same size or smaller than the destination. Overflow occurs - * when the source value is negative. - * * Ux: Generic conversion for the `Us', `Uu' & `us' cases * Overflow occurs when the source magnitude is too large for the * destination. @@ -196,14 +192,6 @@ H5FL_BLK_DEFINE_STATIC(array_seq); *((DT*)D) = (DT)(*((ST*)S)); \ } -#define H5T_CONV_SU_CORE(S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ - if (*((ST*)S)<0) { \ - if (!H5T_overflow_g || (H5T_overflow_g)(src_id, dst_id, S, D)<0) \ - *((DT*)D) = 0; \ - } else \ - *((DT*)D) = (DT)(*((ST*)S)); \ -} - #define H5T_CONV_Ux_CORE(S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ if (*((ST*)S) > (D_MAX)) { \ if (!H5T_overflow_g || (H5T_overflow_g)(src_id, dst_id, S, D)<0) \ @@ -214,12 +202,20 @@ H5FL_BLK_DEFINE_STATIC(array_seq); #define H5T_CONV_sS(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)<=sizeof(DT)); \ - H5T_CONV(H5T_CONV_xX, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, nelmts-1) \ + H5T_CONV(H5T_CONV_xX, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ +} + +#define H5T_CONV_sU_CORE(S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ + if (*((ST*)S)<0) { \ + if (!H5T_overflow_g || (H5T_overflow_g)(src_id, dst_id, S, D)<0) \ + *((DT*)D) = 0; \ + } else \ + *((DT*)D) = (DT)(*((ST*)S)); \ } #define H5T_CONV_sU(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)<=sizeof(DT)); \ - H5T_CONV(H5T_CONV_SU, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, nelmts-1) \ + H5T_CONV(H5T_CONV_sU, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_uS_CORE(S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ @@ -232,17 +228,17 @@ H5FL_BLK_DEFINE_STATIC(array_seq); #define H5T_CONV_uS(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)<=sizeof(DT)); \ - H5T_CONV(H5T_CONV_uS, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, nelmts-1) \ + H5T_CONV(H5T_CONV_uS, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_uU(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)<=sizeof(DT)); \ - H5T_CONV(H5T_CONV_xX, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, nelmts-1) \ + H5T_CONV(H5T_CONV_xX, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_Ss(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)>=sizeof(DT)); \ - H5T_CONV(H5T_CONV_Xx, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, 0) \ + H5T_CONV(H5T_CONV_Xx, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_Su_CORE(S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ @@ -259,48 +255,65 @@ H5FL_BLK_DEFINE_STATIC(array_seq); #define H5T_CONV_Su(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)>=sizeof(DT)); \ - H5T_CONV(H5T_CONV_Su, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, 0) \ + H5T_CONV(H5T_CONV_Su, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_Us(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)>=sizeof(DT)); \ - H5T_CONV(H5T_CONV_Ux, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, 0) \ + H5T_CONV(H5T_CONV_Ux, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_Uu(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)>=sizeof(DT)); \ - H5T_CONV(H5T_CONV_Ux, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, 0) \ + H5T_CONV(H5T_CONV_Ux, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ +} + +#define H5T_CONV_su_CORE(S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ + /* Assumes memory format of unsigned & signed integers is same */ \ + if (*((ST*)S)<0) { \ + if (!H5T_overflow_g || (H5T_overflow_g)(src_id, dst_id, S, D)<0) \ + *((DT*)D) = 0; \ + } \ } #define H5T_CONV_su(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)==sizeof(DT)); \ - H5T_CONV(H5T_CONV_SU, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, 0) \ + H5T_CONV(H5T_CONV_su, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ +} + +#define H5T_CONV_us_CORE(S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ + /* Assumes memory format of unsigned & signed integers is same */ \ + if (*((ST*)S) > (D_MAX)) { \ + if (!H5T_overflow_g || (H5T_overflow_g)(src_id, dst_id, S, D)<0) \ + *((DT*)D) = (D_MAX); \ + } \ } #define H5T_CONV_us(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)==sizeof(DT)); \ - H5T_CONV(H5T_CONV_Ux, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, 0) \ + H5T_CONV(H5T_CONV_us, long_long, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_fF(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)<=sizeof(DT)); \ - H5T_CONV(H5T_CONV_xX, double, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, nelmts-1) \ + H5T_CONV(H5T_CONV_xX, double, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } #define H5T_CONV_Ff(STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ assert(sizeof(ST)>=sizeof(DT)); \ - H5T_CONV(H5T_CONV_Xx, double, STYPE, DTYPE, ST, DT, D_MIN, D_MAX, 0) \ + H5T_CONV(H5T_CONV_Xx, double, STYPE, DTYPE, ST, DT, D_MIN, D_MAX) \ } /* The main part of every integer hardware conversion macro */ -#define H5T_CONV(GUTS,ATYPE,STYPE,DTYPE,ST,DT,D_MIN,D_MAX,STRT) { \ - hsize_t elmtno; /*element number */ \ +#define H5T_CONV(GUTS,ATYPE,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) { \ + size_t elmtno; /*element number */ \ uint8_t *src, *s; /*source buffer */ \ uint8_t *dst, *d; /*destination buffer */ \ H5T_t *st, *dt; /*data type descriptors */ \ ATYPE aligned; /*aligned type */ \ hbool_t s_mv, d_mv; /*move data to align it? */ \ - size_t s_stride, d_stride; /*src and dst strides */ \ + ssize_t s_stride, d_stride; /*src and dst strides */ \ + size_t safe; /* How many elements are safe to process in each pass */ \ \ switch (cdata->command) { \ case H5T_CONV_INIT: \ @@ -322,21 +335,15 @@ H5FL_BLK_DEFINE_STATIC(array_seq); break; \ \ case H5T_CONV_CONV: \ - /* Initialize pointers */ \ + /* Initialize source & destination strides */ \ if (buf_stride) { \ + assert(buf_stride>=sizeof(ST)); \ + assert(buf_stride>=sizeof(DT)); \ s_stride = d_stride = buf_stride; \ } else { \ s_stride = sizeof(ST); \ d_stride = sizeof(DT); \ } \ - if (STRT) { \ - src = (uint8_t*)buf+(STRT)*s_stride; \ - dst = (uint8_t*)buf+(STRT)*d_stride; \ - s_stride = -s_stride; \ - d_stride = -d_stride; \ - } else { \ - src = dst = buf; \ - } \ \ /* Is alignment required for source or dest? */ \ s_mv = H5T_NATIVE_##STYPE##_ALIGN_g>1 && \ @@ -350,21 +357,57 @@ H5FL_BLK_DEFINE_STATIC(array_seq); CI_INC_SRC(s_mv) \ CI_INC_DST(d_mv) \ \ - if (s_mv && d_mv) { \ - /* Alignment is required for both source and dest */ \ - s = (uint8_t*)&aligned; \ - H5T_CONV_LOOP(PRE_SALIGN,PRE_DALIGN,POST_SALIGN,POST_DALIGN,GUTS,s,d,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ - } else if(s_mv) { \ - /* Alignment is required only for source */ \ - s = (uint8_t*)&aligned; \ - H5T_CONV_LOOP(PRE_SALIGN,PRE_DNOALIGN,POST_SALIGN,POST_DNOALIGN,GUTS,s,dst,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ - } else if(d_mv) { \ - /* Alignment is required only for destination */ \ - H5T_CONV_LOOP(PRE_SNOALIGN,PRE_DALIGN,POST_SNOALIGN,POST_DALIGN,GUTS,src,d,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ - } else { \ - /* Alignment is not required for both source and destination */ \ - H5T_CONV_LOOP(PRE_SNOALIGN,PRE_DNOALIGN,POST_SNOALIGN,POST_DNOALIGN,GUTS,src,dst,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ - } \ + /* The outer loop of the type conversion macro, controlling which */ \ + /* direction the buffer is walked */ \ + while (nelmts>0) { \ + /* Check if we need to go backwards through the buffer */ \ + if(d_stride>s_stride) { \ + /* Compute the number of "safe" destination elements at */ \ + /* the end of the buffer (Those which don't overlap with */ \ + /* any source elements at the beginning of the buffer) */ \ + safe=nelmts-(((nelmts*s_stride)+(d_stride-1))/d_stride); \ + \ + /* If we're down to the last few elements, just wrap up */ \ + /* with a "real" reverse copy */ \ + if(safe<2) { \ + src = (uint8_t*)buf+(nelmts-1)*s_stride; \ + dst = (uint8_t*)buf+(nelmts-1)*d_stride; \ + s_stride = -s_stride; \ + d_stride = -d_stride; \ + \ + safe=nelmts; \ + } /* end if */ \ + else { \ + src = (uint8_t*)buf+(nelmts-safe)*s_stride; \ + dst = (uint8_t*)buf+(nelmts-safe)*d_stride; \ + } /* end else */ \ + } /* end if */ \ + else { \ + /* Single forward pass over all data */ \ + src = dst = buf; \ + safe=nelmts; \ + } /* end else */ \ + \ + /* Perform loop over elements to convert */ \ + if (s_mv && d_mv) { \ + /* Alignment is required for both source and dest */ \ + s = (uint8_t*)&aligned; \ + H5T_CONV_LOOP(PRE_SALIGN,PRE_DALIGN,POST_SALIGN,POST_DALIGN,GUTS,s,d,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ + } else if(s_mv) { \ + /* Alignment is required only for source */ \ + s = (uint8_t*)&aligned; \ + H5T_CONV_LOOP(PRE_SALIGN,PRE_DNOALIGN,POST_SALIGN,POST_DNOALIGN,GUTS,s,dst,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ + } else if(d_mv) { \ + /* Alignment is required only for destination */ \ + H5T_CONV_LOOP(PRE_SNOALIGN,PRE_DALIGN,POST_SNOALIGN,POST_DALIGN,GUTS,src,d,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ + } else { \ + /* Alignment is not required for both source and destination */ \ + H5T_CONV_LOOP(PRE_SNOALIGN,PRE_DNOALIGN,POST_SNOALIGN,POST_DNOALIGN,GUTS,src,dst,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ + } \ + \ + /* Decrement number of elements left to convert */ \ + nelmts-=safe; \ + } /* end while */ \ break; \ \ default: \ @@ -408,9 +451,9 @@ H5FL_BLK_DEFINE_STATIC(array_seq); #define H5T_CONV_LOOP_POST_DNOALIGN(DT) { \ } -/* The inner loop of the type conversion macro */ +/* The inner loop of the type conversion macro, actually converting the elements */ #define H5T_CONV_LOOP(PRE_SALIGN_GUTS,PRE_DALIGN_GUTS,POST_SALIGN_GUTS,POST_DALIGN_GUTS,GUTS,S,D,STYPE,DTYPE,ST,DT,D_MIN,D_MAX) \ - for (elmtno=0; elmtno<nelmts; elmtno++) { \ + for (elmtno=0; elmtno<safe; elmtno++) { \ /* Handle source pre-alignment */ \ H5_GLUE(H5T_CONV_LOOP_,PRE_SALIGN_GUTS)(ST) \ \ |