| 255 | | Following discussions with the previewer, it was decided that low-memory option should be the best approach but the slight deterioration in performance over the original code may be down to the over-zealous replacement of temporary scalars within the second 3D loop. On reflection there are also opportunities to reduce the number of floating point operations and load and store instructions within the first 3D loop. |
| 256 | | |
| 257 | | Although significant variation between identical runs on the NOC cluster means the evidence is not conclusive; this second version of the low-memory option does appear to improve on the original code and is certainly no worse whilst using less storage. Here are the tables with the new results added. Graphs are shown below the code differences. |
| | 255 | Following discussions with the previewer, it was decided that low-memory option should be the best approach but the slight deterioration in performance over the original code may be down to the over-zealous replacement of temporary scalars within the second 3D loop. On reflection there are also opportunities to reduce the number of floating point operations and load and store instructions within the first 3D loop. Importantly, there are also opportunities to avoid some expensive operations by performing some calculations in log space. Here are the mathematically equivalent alternatives: |
| | 256 | |
| | 257 | {{{#!diff |
| | 258 | --- traqsr.F90 2020-05-19 10:28:06.858457146 +0100 |
| | 259 | +++ LOMEM3/traqsr.F90 2020-05-15 15:44:11.652736539 +0100 |
| | 260 | @@ -111,7 +111,6 @@ |
| | 261 | REAL(wp) :: zzc0, zzc1, zzc2, zzc3 ! - - |
| | 262 | REAL(wp) :: zz0 , zz1 , ze3t, zlui ! - - |
| | 263 | REAL(wp) :: zCb, zCmax, zze, zpsi, zpsimax, zdelpsi, zCtot, zCze |
| | 264 | - REAL(wp) :: zlogze, zlogCtot, zlogCze |
| | 265 | REAL(wp) :: zlogc |
| | 266 | REAL(wp), ALLOCATABLE, DIMENSION(:,:) :: ze0, ze1, ze2, ze3 |
| | 267 | REAL(wp), ALLOCATABLE, DIMENSION(:,:,:) :: ztrdt, zetot, ztmp3d |
| | 268 | @@ -168,21 +167,18 @@ |
| | 269 | ! Separation in R-G-B depending of the surface Chl |
| | 270 | DO_3D_00_00 ( 1, nksr + 1 ) |
| | 271 | zchl = MIN( 10. , MAX( 0.03, sf_chl(1)%fnow(ji,jj,1) ) ) |
| | 272 | + zCze = 1.12 * zchl**0.803 |
| | 273 | + zCtot = 40.6 * zchl**0.459 |
| | 274 | zlogc = LOG( zchl ) |
| | 275 | - zlogCze = 0.113328685307 + 0.803 * zlogc ! log(zCze = 1.12 * zchl**0.803) |
| | 276 | - zlogCtot= 3.703768066608 + 0.459 * zlogc ! log(zCtot = 40.6 * zchl**0.459) |
| | 277 | ! |
| | 278 | zCb = 0.768 + zlogc * ( 0.087 - zlogc * ( 0.179 + zlogc * 0.025 ) ) |
| | 279 | zCmax = 0.299 - zlogc * ( 0.289 - zlogc * 0.579 ) |
| | 280 | zpsimax = 0.6 - zlogc * ( 0.640 - zlogc * ( 0.021 + zlogc * 0.115 ) ) |
| | 281 | zdelpsi = 0.710 + zlogc * ( 0.159 + zlogc * 0.021 ) |
| | 282 | ! |
| | 283 | - zlogze = 6.34247346942 - 0.746 * zlogCtot ! log(zze = 568.2 * zCtot**(-0.746)) |
| | 284 | - IF( zlogze > 4.62497281328 ) zlogze = 5.298317366548 - 0.293 * zlogCtot |
| | 285 | - ! log(IF( zze > 102. ) zze = 200.0 * zCtot**(-0.293)) |
| | 286 | - zze = EXP( zlogze ) |
| | 287 | - zpsi = gdepw(ji,jj,jk,Kmm) / zze |
| | 288 | - zCze = EXP( zlogCze ) |
| | 289 | + zze = 568.2 * zCtot**(-0.746) |
| | 290 | + IF( zze > 102. ) zze = 200.0 * zCtot**(-0.293) |
| | 291 | + zpsi = gdepw(ji,jj,jk,Kmm) / zze |
| | 292 | ! |
| | 293 | ! NB. make sure zchl value is such that: zchl = MIN( 10. , MAX( 0.03, zchl ) ) |
| | 294 | zchl = MIN( 10. , MAX( 0.03, zCze * ( zCb + zCmax * EXP( -( (zpsi - zpsimax) / zdelpsi )**2 ) ) ) ) |
| | 295 | }}} |
| | 296 | |
| | 297 | Despite significant variation between identical runs on the NOC cluster there is evidence that this second version of the low-memory option improves on the original code and is certainly no worse whilst using less storage. Here are the tables with the new results added. Graphs are shown below the code differences. |
