diff --git a/doc/orgmode/chapters/using.org b/doc/orgmode/chapters/using.org
index 7bf39140f57702191dee06f459a506e2754c254f..86e108b4fc76133408a616fa21c6cc3750b0507f 100644
--- a/doc/orgmode/chapters/using.org
+++ b/doc/orgmode/chapters/using.org
@@ -165,33 +165,58 @@
#+end_src
List of main options that can be used in timing:
- * ~--help~: show usage
- * ~--threads~: Number of CPU workers (default:
- ~_SC_NPROCESSORS_ONLN~)
- * ~--gpus~: number of GPU workers (default: ~0~)
- * ~--n_range=R~: range of N values, with ~R=Start:Stop:Step~
- (default: ~500:5000:500~)
- * ~--m=X~: dimension (M) of the matrices (default: ~N~)
- * ~--k=X~: dimension (K) of the matrices (default: ~1~), useful for
- GEMM algorithm (k is the shared dimension and must be defined
- >1 to consider matrices and not vectors)
- * ~--nrhs=X~: number of right-hand size (default: ~1~)
- * ~--nb=X~: block/tile size. (default: ~128~)
- * ~--ib=X~: inner-blocking/IB size. (default: ~32~)
- * ~--niter=X~: number of iterations performed for each test
- (default: ~1~)
- * ~--rhblk=X~: if X > 0, enable Householder mode for QR and LQ
- factorization. X is the size of each subdomain (default: ~0~)
- * ~--[no]check~: check result (default: ~nocheck~)
- * ~--[no]profile~: print profiling informations (default:
- ~noprofile~)
- * ~--[no]trace~: enable/disable trace generation (default: ~notrace~)
- * ~--[no]dag~: enable/disable DAG generation (default: ~nodag~)
- * ~--[no]inv~: check on inverse (default: ~noinv~)
- * ~--nocpu~: all GPU kernels are exclusively executed on GPUs
- * ~--ooc~: Enable out-of-core (available only with StarPU)
- * ~--bound~: Compare result to area bound (available only with
- StarPU) (default: ~0~)
+ * ~--help~: Show usage
+ * Machine parameters
+ * ~-t x, --threads=x~: Number of CPU workers (default: automatic
+ detection through runtime)
+ * ~-g x, --gpus=x~: Number of GPU workers (default: ~0~)
+ * ~-P x, --P=x~: Rows (P) in the PxQ process grid (default: ~1~)
+ * ~--nocpu~: All GPU kernels are exclusively executed on GPUs
+ * Matrix parameters
+ * ~-m x, --m=X, --M=x~: Dimension (M) of the matrices (default:
+ ~N~)
+ * ~-n x, --n=X, --N=x~: Dimension (N) of the matrices
+ * ~-N R, --n_range=R~: Range of N values to time with
+ ~R=Start:Stop:Step~ (default: ~500:5000:500~)
+ * ~-k x, --k=x, --K=x, --nrhs=x~: Dimension (K) of the matrices
+ or number of right-hand size (default: ~1~). This is useful for
+ GEMM algorithms (k is the shared dimension and must be defined
+ >1 to consider matrices and not vectors)
+ * ~-b x, --nb=x~: NB size. (default: ~320~)
+ * ~-i x, --ib=x~: IB size. (default: ~32~)
+ * Check/prints
+ * ~--niter=X~: Number of iterations performed for each test
+ (default: ~1~)
+ * ~-W, --nowarning~: Do not show warnings
+ * ~-w, --nowarmup~: Cancel the warmup run to pre-load libraries
+ * ~-c, --check~: Check result
+ * ~-C, --inc~: Check on inverse
+ * ~--mode=x~ : Change the xLATMS matrix mode generation for
+ SVD/EVD (default: ~4~). It must be between 0 and 20 included.
+ * Profiling parameters
+ * ~-T, --trace~: Enable trace generation
+ * ~--progress~: Display progress indicator
+ * ~-d, --dag~: Enable DAG generation. Generates a dot_dag_file.dot.
+ * ~-p, --profile~: Print profiling informations
+ * HQR parameters
+ * ~-a x, --qr_a=x, --rhblk=x~: Define the size of the local TS
+ trees in housholder reduction trees for QR and LQ
+ factorization. N is the size of each subdomain (default: ~-1~)
+ * ~-l x, --llvl=x~: Tree used for low level reduction inside
+ nodes (default: ~-1~)
+ * ~-L x, --hlvl=x~: Tree used for high level reduction between
+ nodes, only if P > 1 (default: ~-1~). Possible values are -1:
+ Automatic, 0: Flat, 1: Greedy, 2: Fibonacci, 3: Binary, 4:
+ Replicated greedy.
+ * ~-D, --domino~: Enable the domino between upper and lower trees
+ * Advanced options
+ * ~--nobigmat~: Disable single large matrix allocation for
+ multiple tiled allocations
+ * ~-s, --sync~: Enable synchronous calls in wrapper function such
+ as POTRI
+ * ~-o, --ooc~: Enable out-of-core (available only with StarPU)
+ * ~-G, --gemm3m~: Use gemm3m complex method
+ * ~--bound~: Compare result to area bound
List of timing algorithms available:
* LANGE: norms of matrices
diff --git a/doc/texinfo/chapters/using.texi b/doc/texinfo/chapters/using.texi
index 61963c4410973358ad5d0b268bbfeb593075128f..cf83f26e85cd2f7959a3b2ce81cfd96882469bc9 100644
--- a/doc/texinfo/chapters/using.texi
+++ b/doc/texinfo/chapters/using.texi
@@ -83,36 +83,53 @@ copy from LAPACK matrix layout to tile matrix layout are necessary.
List of main options that can be used in timing:
@itemize @bullet
@item @option{--help}: show usage
- @item @option{--threads}: Number of CPU workers (default:
-@option{_SC_NPROCESSORS_ONLN})
- @item @option{--gpus}: number of GPU workers (default: @option{0})
- @item @option{--n_range=R}: range of N values, with
-@option{R=Start:Stop:Step}
-(default: @option{500:5000:500})
- @item @option{--m=X}: dimension (M) of the matrices (default: @option{N})
- @item @option{--k=X}: dimension (K) of the matrices (default: @option{1}),
-useful for GEMM algorithm (k is the shared dimension and must be defined >1 to
-consider matrices and not vectors)
- @item @option{--nrhs=X}: number of right-hand size (default: @option{1})
- @item @option{--nb=X}: block/tile size. (default: @option{128})
- @item @option{--ib=X}: inner-blocking/IB size. (default: @option{32})
- @item @option{--niter=X}: number of iterations performed for each test
-(default: @option{1})
- @item @option{--rhblk=X}: if X > 0, enable Householder mode for QR and LQ
-factorization. X is the size of each subdomain (default: @option{0})
- @item @option{--[no]check}: check result (default: @option{nocheck})
- @item @option{--[no]profile}: print profiling informations (default:
-@option{noprofile})
- @item @option{--[no]trace}: enable/disable trace generation (default:
-@option{notrace})
- @item @option{--[no]dag}: enable/disable DAG generation (default:
-@option{nodag})
- @item @option{--[no]inv}: check on inverse (default: @option{noinv})
- @item @option{--nocpu}: all GPU kernels are exclusively executed on GPUs
- @item @option{--ooc}: Enable out-of-core (available only with StarPU)
- @item @option{--bound}: Compare result to area bound (available only with StarPU)
-(default: @option{0})
- @end itemize
+ @item Machine parameters
+ @itemize @bullet
+ @item @option{-t x, --threads=x}: Number of CPU workers (default: automatic detection through runtime)
+ @item @option{-g x, --gpus=x}: Number of GPU workers (default: @option{0})
+ @item @option{-P x, --P=x}: Rows (P) in the PxQ process grid (deafult: @option{1})
+ @item @option{--nocpu}: All GPU kernels are exclusively executed on GPUs (default: @option{0})
+ @end itemize
+ @item Matrix parameters
+ @itemize @bullet
+ @item @option{-m x, --m=x, --M=x}: Dimension (M) of the matrices (default: @option{N})
+ @item @option{-n x, --n=x, --N=x}: Dimension (N) of the matrices
+ @item @option{-N R, --n_range=R}: Range of N values to time with R=Start:Stop:Step (default: @option{500:5000:500})
+ @item @option{-k x, --k=x, --K=x, --nrhs=x}: Dimension (K) of the matrices or number of right-hand size (default: @option{1}). This is useful for GEMM like algorithms (k is the shared dimension and must be defined >1 to consider matrices and not vectors)
+ @item @option{-b x, --nb=x}: NB size. (default: @option{320})
+ @item @option{-i x, --ib=x}: IB size. (default: @option{32})
+ @end itemize
+ @item Check/prints
+ @itemize @bullet
+ @item @option{--niter=x}: number of iterations performed for each test (default: @option{1})
+ @item @option{-W, --nowarnings}: Do not show warnings
+ @item @option{-w, --nowarmup}: Cancel the warmup run to pre-load libraries
+ @item @option{-c, --check}: Check result
+ @item @option{-C, --inv}: Check on inverse
+ @item @option{--mode=x}: Change the xLATMS matrix mode generation for SVD/EVD (default: @option{4}). It must be between 0 and 20 included.
+ @end itemize
+ @item Profiling parameters
+ @itemize @bullet
+ @item @option{-T, --trace}: Enable trace generation
+ @item @option{--progress}: Display progress indicator
+ @item @option{-d, --dag}: Enable DAG generation. Generates a dot_dag_file.dot.
+ @item @option{-p, --profile}: Print profiling informations
+ @end itemize
+ @item HQR parameters
+ @itemize @bullet
+ @item @option{-a x, --qr_a=x, --rhblk=x}: Define the size of the local TS trees in housholder reduction trees for QR and LQ factorization. N is the size of each subdomain (default: @option{-1})
+ @item @option{-l x, --llvl=x}: Tree used for low level reduction inside nodes (default: @option{-1})
+ @item @option{-L x, --hlvl=x}: Tree used for high level reduction between nodes, only if P > 1 (default: @option{-1}). Possible values are -1: Automatic, 0: Flat, 1: Greedy, 2: Fibonacci, 3: Binary, 4: Replicated greedy.
+ @item @option{-D, --domino}: Enable the domino between upper and lower trees
+ @end itemize
+ @item Advanced options
+ @itemize @bullet
+ @item @option{--nobigmat}: Disable single large matrix allocation for multiple tiled allocations
+ @item @option{-s, --sync}: Enable synchronous calls in wrapper function such as POTRI
+ @item @option{-o, --ooc}: Enable out-of-core (available only with StarPU)
+ @item @option{-G, --gemm3m}: Use gemm3m complex method
+ @item @option{--bound}: Compare result to area bound
+ @end itemize
List of timing algorithms available:
@itemize @bullet
diff --git a/example/basic_zposv/basic_posv.h b/example/basic_zposv/basic_posv.h
index 04748c34f6b768f9fa2ff4a6bc7aa9206c9b13b5..5f7aa0dc06ece32f9e856083439ab7cf035f7cb7 100644
--- a/example/basic_zposv/basic_posv.h
+++ b/example/basic_zposv/basic_posv.h
@@ -117,7 +117,6 @@ enum iparam_examples {
/* Added for StarPU version */
IPARAM_PROFILE,
IPARAM_PRINT_ERRORS,
- IPARAM_PEAK,
IPARAM_PARALLEL_TASKS,
IPARAM_NO_CPU,
IPARAM_BOUND,
@@ -132,7 +131,6 @@ enum dparam_examples {
IPARAM_XNORM,
IPARAM_RNORM,
IPARAM_AinvNORM,
- IPARAM_ESTIMATED_PEAK,
IPARAM_RES,
/* Begin section for hydra integration tool */
IPARAM_THRESHOLD_CHECK, /* Maximum value accepted for: |Ax-b||/N/eps/(||A||||x||+||b||) */
@@ -179,7 +177,6 @@ static void init_iparam(int iparam[IPARAM_SIZEOF]){
iparam[IPARAM_Q ] = 1;
iparam[IPARAM_PROFILE ] = 0;
iparam[IPARAM_PRINT_ERRORS ] = 0;
- iparam[IPARAM_PEAK ] = 0;
iparam[IPARAM_PARALLEL_TASKS] = 0;
iparam[IPARAM_NO_CPU ] = 0;
iparam[IPARAM_BOUND ] = 0;
@@ -210,7 +207,6 @@ static void print_header(char *prog_name, int * iparam) {
const char *bound_header = iparam[IPARAM_BOUND] ? " thGflop/s" : "";
const char *check_header = iparam[IPARAM_CHECK] ? " ||Ax-b|| ||A|| ||x|| ||b|| ||Ax-b||/N/eps/(||A||||x||+||b||) RETURN" : "";
const char *inverse_header = iparam[IPARAM_INVERSE] ? " ||I-A*Ainv|| ||A|| ||Ainv|| ||Id - A*Ainv||/((||A|| ||Ainv||).N.eps)" : "";
- const char *peak_header = iparam[IPARAM_PEAK] ? " (% of peak) peak" : "";
#if defined(CHAMELEON_SIMULATION)
double eps = 0.;
#else
@@ -235,8 +231,8 @@ static void print_header(char *prog_name, int * iparam) {
iparam[IPARAM_IB],
eps );
- printf( "# M N K/NRHS seconds Gflop/s Deviation%s%s%s\n",
- bound_header, peak_header, iparam[IPARAM_INVERSE] ? inverse_header : check_header);
+ printf( "# M N K/NRHS seconds Gflop/s Deviation%s%s\n",
+ bound_header, iparam[IPARAM_INVERSE] ? inverse_header : check_header);
printf( "# %5.0d %5.0d %5.0d\n", iparam[IPARAM_N], iparam[IPARAM_N], iparam[IPARAM_K]);
return;
}
diff --git a/runtime/starpu/include/runtime_codelet_profile.h b/runtime/starpu/include/runtime_codelet_profile.h
index 88fa98da4802e603e196ae80195fc57e76a327eb..c9cbb5099c15d0896f7e0097c593a25ca78b4072 100644
--- a/runtime/starpu/include/runtime_codelet_profile.h
+++ b/runtime/starpu/include/runtime_codelet_profile.h
@@ -67,53 +67,11 @@
} \
} \
} \
- void estimate_##name##_sustained_peak(double *res_peak) \
- { \
- /* We use a heuristic where we assume that all GPUs have some work, and \
- * that some CPUs may not have been involved. This may not be \
- * applicable to small problems where only a subset of the processing \
- * units are used. We assume that all CPUs are the same, so we multiply \
- * the best performance obtained on a CPU by the number of CPUs, and we \
- * add this to the sum of the performance obtained by the GPUs. */ \
- double peak = 0.0; \
- double best_cpu = 0.0; \
- unsigned ncpus = 0; \
- \
- unsigned worker; \
- for (worker = 0; worker < starpu_worker_get_count(); worker++) \
- { \
- unsigned cpu_worker = (starpu_worker_get_type(worker) == STARPU_CPU_WORKER); \
- if (cpu_worker) \
- ncpus++; \
- \
- if (name##_perf[worker].n > 0) \
- { \
- long n = name##_perf[worker].n; \
- double sum = name##_perf[worker].sum; \
- double avg = sum / n; \
- \
- if (cpu_worker) \
- { \
- if (avg > best_cpu) \
- best_cpu = avg; \
- } \
- else \
- { \
- peak += avg; \
- } \
- } \
- } \
- \
- peak += ncpus * best_cpu; \
- \
- *res_peak = peak; \
- } \
#define CHAMELEON_CL_CB_HEADER(name) \
extern struct starpu_perfmodel*cl_##name##_save; \
extern struct starpu_perfmodel cl_##name##_fake; \
void cl_##name##_callback(); \
- void profiling_display_##name##_info(void); \
- void estimate_##name##_sustained_peak(double *res)
+ void profiling_display_##name##_info(void);
#endif /* __CODELET_PROFILE_H__ */
diff --git a/timing/timing.c b/timing/timing.c
index ca5f3d1b7e646f46f4f086963a1c32b5e8805950..093ade83cec701cb3880aaa4ee36e91fe161880f 100644
--- a/timing/timing.c
+++ b/timing/timing.c
@@ -221,34 +221,30 @@ Test(int64_t n, int *iparam) {
if ( MORSE_My_Mpi_Rank() == 0) {
printf( "%9.3f %9.2f +-%7.2f ", sumt/niter, gflops, sd);
- if (iparam[IPARAM_BOUND])
+ if (iparam[IPARAM_BOUND]) {
printf(" %9.2f", sumgf_upper/niter);
-
- if ( iparam[IPARAM_PEAK] )
- {
- if (dparam[IPARAM_ESTIMATED_PEAK]<0.0f)
- printf(" n/a n/a ");
- else
- printf(" %5.2f%% %9.2f ", 100.0f*(gflops/dparam[IPARAM_ESTIMATED_PEAK]), dparam[IPARAM_ESTIMATED_PEAK]);
}
if ( iparam[IPARAM_CHECK] ){
hres = ( dparam[IPARAM_RES] / n / eps / (dparam[IPARAM_ANORM] * dparam[IPARAM_XNORM] + dparam[IPARAM_BNORM] ) > dparam[IPARAM_THRESHOLD_CHECK] );
- if (hres)
+ if (hres) {
printf( "%8.5e %8.5e %8.5e %8.5e %8.5e FAILURE",
dparam[IPARAM_RES], dparam[IPARAM_ANORM], dparam[IPARAM_XNORM], dparam[IPARAM_BNORM],
dparam[IPARAM_RES] / n / eps / (dparam[IPARAM_ANORM] * dparam[IPARAM_XNORM] + dparam[IPARAM_BNORM] ));
- else
+ }
+ else {
printf( "%8.5e %8.5e %8.5e %8.5e %8.5e SUCCESS",
dparam[IPARAM_RES], dparam[IPARAM_ANORM], dparam[IPARAM_XNORM], dparam[IPARAM_BNORM],
dparam[IPARAM_RES] / n / eps / (dparam[IPARAM_ANORM] * dparam[IPARAM_XNORM] + dparam[IPARAM_BNORM] ));
+ }
}
- if ( iparam[IPARAM_INVERSE] )
+ if ( iparam[IPARAM_INVERSE] ) {
printf( " %8.5e %8.5e %8.5e %8.5e",
dparam[IPARAM_RNORM], dparam[IPARAM_ANORM], dparam[IPARAM_AinvNORM],
dparam[IPARAM_RNORM] /((dparam[IPARAM_ANORM] * dparam[IPARAM_AinvNORM])*n*eps));
+ }
printf("\n");
@@ -387,7 +383,6 @@ show_help(char *prog_name) {
" -s, --sync Enable synchronous calls in wrapper function such as POTRI\n"
" -o, --ooc Enable out-of-core (available only with StarPU)\n"
" -G, --gemm3m Use gemm3m complex method\n"
- //" --peak ?\n"todo
" --bound Compare result to area bound\n"
"\n");
}
@@ -398,7 +393,6 @@ print_header(char *prog_name, int * iparam) {
const char *bound_header = iparam[IPARAM_BOUND] ? " thGflop/s" : "";
const char *check_header = iparam[IPARAM_CHECK] ? " ||Ax-b|| ||A|| ||x|| ||b|| ||Ax-b||/N/eps/(||A||||x||+||b||) RETURN" : "";
const char *inverse_header = iparam[IPARAM_INVERSE] ? " ||I-A*Ainv|| ||A|| ||Ainv|| ||Id - A*Ainv||/((||A|| ||Ainv||).N.eps)" : "";
- const char *peak_header = iparam[IPARAM_PEAK] ? " (% of peak) peak" : "";
#if defined(CHAMELEON_SIMULATION)
_PREC eps = 0.;
#else
@@ -431,8 +425,8 @@ print_header(char *prog_name, int * iparam) {
iparam[IPARAM_IB],
eps );
- printf( "# M N K/NRHS seconds Gflop/s Deviation%s%s%s\n",
- bound_header, peak_header, iparam[IPARAM_INVERSE] ? inverse_header : check_header);
+ printf( "# M N K/NRHS seconds Gflop/s Deviation%s%s\n",
+ bound_header, iparam[IPARAM_INVERSE] ? inverse_header : check_header);
return;
}
@@ -482,7 +476,6 @@ static struct option long_options[] =
{"sync", no_argument, 0, 's'},
{"ooc", no_argument, 0, 'o'},
{"gemm3m", no_argument, 0, 'G'},
- {"peak", no_argument, 0, '4'},
{"bound", no_argument, 0, '5'},
{0, 0, 0, 0}
};
@@ -496,7 +489,7 @@ set_iparam_default(int *iparam){
iparam[IPARAM_THRDNBR ] = -1;
iparam[IPARAM_THRDNBR_SUBGRP] = 1;
iparam[IPARAM_M ] = -1;
- iparam[IPARAM_N ] = 500;
+ iparam[IPARAM_N ] = -1;
iparam[IPARAM_K ] = 1;
iparam[IPARAM_LDA ] = -1;
iparam[IPARAM_LDB ] = -1;
@@ -611,7 +604,6 @@ parse_arguments(int *_argc, char ***_argv, int *iparam, int *start, int *stop, i
case 's' : iparam[IPARAM_ASYNC ] = 0; break;
case 'o' : iparam[IPARAM_OOC ] = 1; break;
case 'G' : iparam[IPARAM_GEMM3M ] = 1; break;
- case '4' : iparam[IPARAM_PEAK ] = 1; break;
case '5' : iparam[IPARAM_BOUND ] = 1; break;
case 'h' :
case '?' :
@@ -624,7 +616,8 @@ parse_arguments(int *_argc, char ***_argv, int *iparam, int *start, int *stop, i
int
main(int argc, char *argv[]) {
- int i, m, mx, nx;
+ int i, m, n, mx, nx;
+ int status;
int nbnode = 1;
int start = 500;
int stop = 5000;
@@ -644,6 +637,7 @@ main(int argc, char *argv[]) {
}
#endif
+ n = iparam[IPARAM_N];
m = iparam[IPARAM_M];
mx = iparam[IPARAM_MX];
nx = iparam[IPARAM_NX];
@@ -709,26 +703,41 @@ main(int argc, char *argv[]) {
if (step < 1) step = 1;
- int status = Test( -1, iparam ); /* print header */
+ status = Test( -1, iparam ); /* print header */
if (status != MORSE_SUCCESS) return status;
- for (i = start; i <= stop; i += step)
- {
- if ( nx > 0 ) {
- iparam[IPARAM_M] = i;
- iparam[IPARAM_N] = chameleon_max(1, i/nx);
- } else if ( mx > 0 ) {
- iparam[IPARAM_M] = chameleon_max(1, i/mx);
- iparam[IPARAM_N] = i;
- } else {
- if ( m == -1 )
+ if ( n == -1 ){
+ for (i = start; i <= stop; i += step)
+ {
+ if ( nx > 0 ) {
iparam[IPARAM_M] = i;
- iparam[IPARAM_N] = i;
+ iparam[IPARAM_N] = chameleon_max(1, i/nx);
+ }
+ else if ( mx > 0 ) {
+ iparam[IPARAM_M] = chameleon_max(1, i/mx);
+ iparam[IPARAM_N] = i;
+ }
+ else {
+ if ( m == -1 ) {
+ iparam[IPARAM_M] = i;
+ }
+ iparam[IPARAM_N] = i;
+ }
+ status = Test( iparam[IPARAM_N], iparam );
+ if (status != MORSE_SUCCESS) {
+ return status;
+ }
+ success += status;
+ }
+ }
+ else {
+ if ( m == -1 ) {
+ iparam[IPARAM_M] = n;
}
- int status = Test( iparam[IPARAM_N], iparam );
+ iparam[IPARAM_N] = n;
+ status = Test( iparam[IPARAM_N], iparam );
if (status != MORSE_SUCCESS) return status;
success += status;
}
-
MORSE_Finalize();
return success;
}
diff --git a/timing/timing.h b/timing/timing.h
index 9eb12fdbcd6164f1ead0c590e83468287ebe0348..ef7357b3dfadfa8678fce12fab1110c29acd99fd 100644
--- a/timing/timing.h
+++ b/timing/timing.h
@@ -57,7 +57,6 @@ enum iparam_timing {
/* Added for StarPU version */
IPARAM_PROFILE,
IPARAM_PRINT_WARNINGS,
- IPARAM_PEAK,
IPARAM_PARALLEL_TASKS,
IPARAM_NO_CPU,
IPARAM_BOUND,
@@ -80,7 +79,6 @@ enum dparam_timing {
IPARAM_XNORM,
IPARAM_RNORM,
IPARAM_AinvNORM,
- IPARAM_ESTIMATED_PEAK,
IPARAM_RES,
/* Begin section for hydra integration tool */
IPARAM_THRESHOLD_CHECK, /* Maximum value accepted for: |Ax-b||/N/eps/(||A||||x||+||b||) */