Help is available by moving the cursor above any 
symbol or by checking MAQAO website.
[ 4 / 4 ] Application profile is long enough (54.15 s)
To have good quality measurements, it is advised that the application profiling time is greater than 10 seconds.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.55 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.05 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (99.09%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 96.86% of observed threads are actually active
[ 4 / 4 ] CPU activity is good
CPU cores are active 96.86% of time
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (78.51%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (95.84%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (99.98%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 3 / 3 ] Functions mostly use all threads
Functions running on a reduced number of threads (typically sequential code) cover less than 10% of application walltime (0.00%)
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (3.24%) lower than cumulative innermost loop coverage (95.84%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.20%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 78 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 12 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 2 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 2424 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 20.51 % - Vector Length Use: 26.64 % | |
| ►Control Flow Issues | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 2410 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 15.63 % | |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 2387 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 64.41 % - Vector Length Use: 42.29 % | |
| ►Loop Computation Issues | 12 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 2 issues (= instructions) costing 4 points each. | 8 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
[ 4 / 4 ] Application profile is long enough (28.00 s)
To have good quality measurements, it is advised that the application profiling time is greater than 10 seconds.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.77 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.07 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (97.63%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 185.01% of observed threads are actually active
[ 4 / 4 ] CPU activity is good
CPU cores are active 96.15% of time
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (77.32%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (94.70%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (95.57%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 3 / 3 ] Functions mostly use all threads
Functions running on a reduced number of threads (typically sequential code) cover less than 10% of application walltime (4.09%)
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.94%) lower than cumulative innermost loop coverage (94.70%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.23%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 77 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 12 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2424 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 20.51 % - Vector Length Use: 26.64 % | |
| ►Control Flow Issues | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 2410 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 15.63 % | |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 2387 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 64.41 % - Vector Length Use: 42.29 % | |
| ►Loop Computation Issues | 12 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 2 issues (= instructions) costing 4 points each. | 8 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
[ 4 / 4 ] Application profile is long enough (14.96 s)
To have good quality measurements, it is advised that the application profiling time is greater than 10 seconds.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.57 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.03 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (94.54%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 341.00% of observed threads are actually active
[ 4 / 4 ] CPU activity is good
CPU cores are active 94.91% of time
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (74.77%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (91.54%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (94.87%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 3 / 3 ] Functions mostly use all threads
Functions running on a reduced number of threads (typically sequential code) cover less than 10% of application walltime (9.20%)
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.99%) lower than cumulative innermost loop coverage (91.54%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.15%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 74 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 12 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 2424 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 20.51 % - Vector Length Use: 26.64 % | |
| ►Control Flow Issues | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2410 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 15.63 % | |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 2387 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 64.41 % - Vector Length Use: 42.29 % | |
| ►Loop Computation Issues | 12 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 2 issues (= instructions) costing 4 points each. | 8 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
[ 0 / 4 ] Application profile is too short (8.39 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.40 / 3 ] Most of time spent in analyzed modules comes from functions without compilation information
Functions without compilation information (typically not compiled with -g and -grecord-gcc-switches) cumulate 20.00% of the time spent in analyzed modules. Check that -g and -grecord-gcc-switches are present. Remark: if -g and -grecord-gcc-switches are indeed used, this can also be due to some compiler built-in functions (typically math) or statically linked libraries. This warning can be ignored in that case.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.06 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (89.62%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 594.04% of observed threads are actually active
[ 4 / 4 ] CPU activity is good
CPU cores are active 92.20% of time
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (71.10%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (87.06%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (93.67%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (15.48%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.56%) lower than cumulative innermost loop coverage (87.06%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.18%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 71 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 11 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 2424 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 20.51 % - Vector Length Use: 26.64 % | |
| ►Control Flow Issues | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2410 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 15.63 % | |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 2387 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 64.41 % - Vector Length Use: 42.29 % | |
| ►Loop Computation Issues | 12 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 2 issues (= instructions) costing 4 points each. | 8 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
[ 0 / 4 ] Application profile is too short (5.30 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.59 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.06 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (80.04%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 988.34% of observed threads are actually active
[ 3 / 4 ] CPU activity is below 90% (87.51%)
CPU cores are idle more than 10% of time. Threads supposed to run on these cores are probably IO/sync waiting. Some hints: use faster filesystems to read/write data, improve parallel load balancing and/or scheduling.
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (63.51%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (77.79%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (92.19%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (24.59%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.25%) lower than cumulative innermost loop coverage (77.79%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.19%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 63 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 10 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 2410 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 15.63 % | |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 2424 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 20.51 % - Vector Length Use: 26.64 % | |
| ►Control Flow Issues | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2387 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 64.41 % - Vector Length Use: 42.29 % | |
| ►Loop Computation Issues | 12 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 2 issues (= instructions) costing 4 points each. | 8 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
[ 0 / 4 ] Application profile is too short (4.72 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.61 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.04 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (74.32%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 1355.49% of observed threads are actually active
[ 3 / 4 ] CPU activity is below 90% (84.03%)
CPU cores are idle more than 10% of time. Threads supposed to run on these cores are probably IO/sync waiting. Some hints: use faster filesystems to read/write data, improve parallel load balancing and/or scheduling.
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (61.15%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (72.55%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (91.98%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (29.49%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (1.77%) lower than cumulative innermost loop coverage (72.55%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.22%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 61 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 8 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 2410 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 15.63 % | |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 2424 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 20.51 % - Vector Length Use: 26.64 % | |
| ►Control Flow Issues | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2387 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 64.41 % - Vector Length Use: 42.29 % | |
| ►Loop Computation Issues | 12 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 2 issues (= instructions) costing 4 points each. | 8 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Several paths (2 paths) - Simplify control structure or force the compiler to use masked instructions. There are 2 issues ( = paths) costing 1 point each. | 2 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
[ 0 / 4 ] Application profile is too short (4.40 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.59 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.02 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (71.62%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 1739.25% of observed threads are actually active
[ 3 / 4 ] CPU activity is below 90% (82.31%)
CPU cores are idle more than 10% of time. Threads supposed to run on these cores are probably IO/sync waiting. Some hints: use faster filesystems to read/write data, improve parallel load balancing and/or scheduling.
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (59.56%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (69.48%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (92.03%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (29.73%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.13%) lower than cumulative innermost loop coverage (69.48%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.14%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 59 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 7 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 84 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 19.42 % | |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 1432 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 13.17 % | |
| ►Loop Computation Issues | 8 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 1 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Loop 2424 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 20.51 % - Vector Length Use: 26.64 % | |
| ►Control Flow Issues | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 13 | |
| ○ | [SA] Too many paths (7 paths) - Simplify control structure. There are 7 issues ( = paths) costing 1 point each with a malus of 4 points. | 11 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
[ 0 / 4 ] Application profile is too short (4.14 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.45 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.02 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (68.61%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 2093.02% of observed threads are actually active
[ 3 / 4 ] CPU activity is below 90% (80.07%)
CPU cores are idle more than 10% of time. Threads supposed to run on these cores are probably IO/sync waiting. Some hints: use faster filesystems to read/write data, improve parallel load balancing and/or scheduling.
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (56.88%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (66.52%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (92.11%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (32.98%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.09%) lower than cumulative innermost loop coverage (66.52%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.15%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 56 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 7 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 84 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 19.42 % | |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 1432 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 13.17 % | |
| ►Loop Computation Issues | 8 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 1 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Loop 2410 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 15.63 % | |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
[ 0 / 4 ] Application profile is too short (4.27 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.88 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.03 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (66.59%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 2416.64% of observed threads are actually active
[ 3 / 4 ] CPU activity is below 90% (77.27%)
CPU cores are idle more than 10% of time. Threads supposed to run on these cores are probably IO/sync waiting. Some hints: use faster filesystems to read/write data, improve parallel load balancing and/or scheduling.
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (54.80%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (64.09%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (92.44%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (32.93%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.50%) lower than cumulative innermost loop coverage (64.09%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.13%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 54 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 7 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 84 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 0.00 % - Vector Length Use: 19.42 % | |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 1432 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 13.17 % | |
| ►Loop Computation Issues | 8 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 1 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 1753 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 100.00 % - Vector Length Use: 100.00 % | |
| ►Data Access Issues | 8 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 4 issues ( = data accesses) costing 2 point each. | 8 |
| ►Vectorization Roadblocks | 8 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 4 issues ( = data accesses) costing 2 point each. | 8 |
[ 0 / 4 ] Application profile is too short (4.37 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.32 / 3 ] Most of time spent in analyzed modules comes from functions without compilation information
Functions without compilation information (typically not compiled with -g and -grecord-gcc-switches) cumulate 22.58% of the time spent in analyzed modules. Check that -g and -grecord-gcc-switches are present. Remark: if -g and -grecord-gcc-switches are indeed used, this can also be due to some compiler built-in functions (typically math) or statically linked libraries. This warning can be ignored in that case.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.02 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (64.55%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 2818.45% of observed threads are actually active
[ 3 / 4 ] CPU activity is below 90% (76.19%)
CPU cores are idle more than 10% of time. Threads supposed to run on these cores are probably IO/sync waiting. Some hints: use faster filesystems to read/write data, improve parallel load balancing and/or scheduling.
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (53.15%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (62.11%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (92.61%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (35.09%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.45%) lower than cumulative innermost loop coverage (62.11%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.15%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 53 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 7 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 84 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 0.00 % - Vector Length Use: 19.42 % | |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Loop 1432 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 13.17 % | |
| ►Loop Computation Issues | 8 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ○ | [SA] Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA - Reorganize arithmetic expressions to exhibit potential for FMA. This issue costs 4 points. | 4 |
| ►Control Flow Issues | 1 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ►Data Access Issues | 2 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Vectorization Roadblocks | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 1 issues ( = data accesses) costing 2 point each. | 2 |
| ►Loop 1753 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 100.00 % - Vector Length Use: 100.00 % | |
| ►Data Access Issues | 8 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 4 issues ( = data accesses) costing 2 point each. | 8 |
| ►Vectorization Roadblocks | 8 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 4 issues ( = data accesses) costing 2 point each. | 8 |
[ 0 / 4 ] Application profile is too short (4.38 s)
If the overall application profiling time is less than 10 seconds, many of the measurements at function or loop level will very likely be under the measurement quality threshold (0,1 seconds). Rerun to increase runtime duration: for example use a larger dataset or include a repetition loop.
[ 3 / 3 ] Optimization level option is correctly used
[ 2.55 / 3 ] Most of time spent in analyzed modules comes from functions compiled with -g and -fno-omit-frame-pointer
-g option gives access to debugging informations, such are source locations. -fno-omit-frame-pointer improve the accuracy of callchains found during the application profiling.
[ 2 / 3 ] Security settings from the host restrict profiling. Some metrics will be missing or incomplete.
Current value for kernel.perf_event_paranoid is 2. If possible, set it to 1 or check with your system administrator which flag can be used to achieve this.
[ 0 / 3 ] Compilation of some functions is not optimized for the target processor
Architecture specific options are needed to produce efficient code for a specific processor ( -mcpu=native ).
[ 2 / 2 ] Application is correctly profiled ("Others" category represents 0.01 % of the execution time)
To have a representative profiling, it is advised that the category "Others" represents less than 20% of the execution time in order to analyze as much as possible of the user code
[ 1 / 1 ] Lstopo present. The Topology lstopo report will be generated.
[ 0 / 0 ] Fastmath not used
Consider to add ffast-math to compilation flags (or replace -O3 with -Ofast) to unlock potential extra speedup by relaxing floating-point computation consistency. Warning: floating-point accuracy may be reduced and the compliance to IEEE/ISO rules/specifications for math functions will be relaxed, typically 'errno' will no longer be set after calling some math functions.
[ 4 / 4 ] Enough time of the experiment time spent in analyzed loops (64.68%)
If the time spent in analyzed loops is less than 30%, standard loop optimizations will have a limited impact on application performances.
[ 4 / 4 ] Threads activity is good
On average, more than 3123.87% of observed threads are actually active
[ 3 / 4 ] CPU activity is below 90% (73.85%)
CPU cores are idle more than 10% of time. Threads supposed to run on these cores are probably IO/sync waiting. Some hints: use faster filesystems to read/write data, improve parallel load balancing and/or scheduling.
[ 4 / 4 ] Loop profile is not flat
At least one loop coverage is greater than 4% (53.56%), representing an hotspot for the application
[ 4 / 4 ] Enough time of the experiment time spent in analyzed innermost loops (62.05%)
If the time spent in analyzed innermost loops is less than 15%, standard innermost loop optimizations such as vectorisation will have a limited impact on application performances.
[ 4 / 4 ] Affinity is good (92.86%)
Threads are not migrating to CPU cores: probably successfully pinned
[ 3 / 3 ] Less than 10% (0.00%) is spend in BLAS1 operations
It could be more efficient to inline by hand BLAS1 operations
[ 0 / 3 ] Too many functions do not use all threads
Functions running on a reduced number of threads (typically sequential code) cover at least 10% of application walltime (32.05%). Check both "Max Inclusive Time Over Threads" and "Nb Threads" in Functions or Loops tabs and consider parallelizing sequential regions or improving parallelization of regions running on a reduced number of threads
[ 3 / 3 ] Cumulative Outermost/In between loops coverage (2.62%) lower than cumulative innermost loop coverage (62.05%)
Having cumulative Outermost/In between loops coverage greater than cumulative innermost loop coverage will make loop optimization more complex
[ 2 / 2 ] Less than 10% (0.00%) is spend in BLAS2 operations
BLAS2 calls usually could make a poor cache usage and could benefit from inlining.
[ 2 / 2 ] Less than 10% (0.09%) is spend in Libm/SVML (special functions)
| Loop ID | Analysis | Penalty Score |
|---|---|---|
| ►Loop 2423 - libggml-cpu.so | Execution Time: 53 % - Vectorization Ratio: 32.79 % - Vector Length Use: 30.39 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 2 issues ( = data accesses) costing 2 point each. | 4 |
| ►Loop 2308 - libggml-cpu.so | Execution Time: 6 % - Vectorization Ratio: 50.56 % - Vector Length Use: 33.98 % | |
| ►Data Access Issues | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Vectorization Roadblocks | 40 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 20 issues ( = data accesses) costing 2 point each. | 40 |
| ►Loop 84 - libggml-cpu.so | Execution Time: 1 % - Vectorization Ratio: 0.00 % - Vector Length Use: 19.42 % | |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2309 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 57.69 % - Vector Length Use: 59.01 % | |
| ►Control Flow Issues | 2 | |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Data Access Issues | 44 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ►Vectorization Roadblocks | 46 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 22 issues ( = data accesses) costing 2 point each. | 44 |
| ○ | [SA] Non innermost loop (Outermost) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 2426 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 28.14 % - Vector Length Use: 27.69 % | |
| ►Data Access Issues | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Vectorization Roadblocks | 10 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 5 issues ( = data accesses) costing 2 point each. | 10 |
| ►Loop 901 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 96.00 % - Vector Length Use: 97.00 % | |
| ►Data Access Issues | 32 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 16 issues ( = data accesses) costing 2 point each. | 32 |
| ►Vectorization Roadblocks | 32 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 16 issues ( = data accesses) costing 2 point each. | 32 |
| ►Loop 1746 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 15.28 % - Vector Length Use: 38.09 % | |
| ►Loop Computation Issues | 2 | |
| ○ | [SA] Presence of a large number of scalar integer instructions - Simplify loop structure, perform loop splitting or perform unroll and jam. This issue costs 2 points. | 2 |
| ►Control Flow Issues | 3 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Vectorization Roadblocks | 1003 | |
| ○ | [SA] Presence of calls - Inline either by compiler or by hand and use SVML for libm calls. There are 1 issues (= calls) costing 1 point each. | 1 |
| ○ | [SA] Too many paths (at least 1000 paths) - Simplify control structure. There are at least 1000 issues ( = paths) costing 1 point. | 1000 |
| ○ | [SA] Non innermost loop (InBetween) - Collapse loop with innermost ones. This issue costs 2 points. | 2 |
| ►Loop 1753 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 100.00 % - Vector Length Use: 100.00 % | |
| ►Data Access Issues | 8 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 4 issues ( = data accesses) costing 2 point each. | 8 |
| ►Vectorization Roadblocks | 8 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 4 issues ( = data accesses) costing 2 point each. | 8 |
| ►Loop 2409 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 0.00 % - Vector Length Use: 25.00 % | |
| ►Data Access Issues | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Vectorization Roadblocks | 4 | |
| ○ | [SA] Presence of indirect accesses - Use array restructuring or gather instructions to lower the cost. There are 1 issues ( = indirect data accesses) costing 4 point each. | 4 |
| ►Loop 908 - libggml-cpu.so | Execution Time: 0 % - Vectorization Ratio: 68.18 % - Vector Length Use: 79.76 % | |
| ►Loop Computation Issues | 4 | |
| ○ | [SA] Presence of expensive FP instructions - Perform hoisting, change algorithm, use SVML or proper numerical library or perform value profiling (count the number of distinct input values). There are 1 issues (= instructions) costing 4 points each. | 4 |
| ►Data Access Issues | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |
| ►Vectorization Roadblocks | 6 | |
| ○ | [SA] Presence of constant non unit stride data access - Use array restructuring, perform loop interchange or use gather instructions to lower a bit the cost. There are 3 issues ( = data accesses) costing 2 point each. | 6 |