OV - Compare Loops

Loops

▶attention.cpp: 30 - 260.14 %

ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s
Run march=znver5 mprefer-vector-width=256							Run march=znver5 [mprefer-vector-width=512]							Run march=znver5 mprefer-vector-width=512 ffast-math
Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 30-31						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 30-31						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 30-31
27	3.85	3.85	40.96	0	8.33	2.7	30	3.70	3.69	38.73	0	8.33	2.81	30	3.82	3.82	40.69	0	8.33	2.72
62	0.69	0.68	7.29	0	8.33	3.77	47	2.44	2.44	25.58	0	8.33	3.33	47	2.82	2.82	30.00	0	8.33	3.38
52	0.70	0.70	7.50	0	8.33	3.69	69	0.56	0.56	5.87	0	8.33	4.16	69	0.70	0.70	7.45	0	8.33	3.45
43	1.95	1.95	20.74	0	8.33	5.33	63	0.57	0.57	5.97	0	8.33	4.09	63	0.71	0.71	7.55	0	8.33	3.49
57	0.68	0.68	7.23	0	8.33	3.84	57	0.70	0.70	7.34	0	8.33	3.46	57	0.68	0.68	7.23	0	8.33	3.64

Sum on 5 analyzed binary loops (attention-avx512 - 27, attention-avx512 - 62, attention-avx512 - 52, attention-avx512 - 43, attention-avx512 - 57)							Sum on 5 analyzed binary loops (attention-avx512 - 30, attention-avx512 - 47, attention-avx512 - 69, attention-avx512 - 63, attention-avx512 - 57)							Sum on 5 analyzed binary loops (attention-avx512 - 30, attention-avx512 - 47, attention-avx512 - 69, attention-avx512 - 63, attention-avx512 - 57)
Analysis						Count	Analysis						Count	Analysis						Count
Loop Computation Issues							Loop Computation Issues							Loop Computation Issues
Low iteration count						1	Low iteration count						1	Low iteration count						1
Control Flow Issues							Control Flow Issues							Control Flow Issues
Low iteration count						1	Low iteration count						1	Low iteration count						1
Data Access Issues							Data Access Issues							Data Access Issues
Presence of constant non-unit stride data access						1	Presence of constant non-unit stride data access						1	Presence of constant non-unit stride data access						1
Vectorization Roadblocks							Vectorization Roadblocks							Vectorization Roadblocks
Presence of constant non-unit stride data access						1	Presence of constant non-unit stride data access						1	Presence of constant non-unit stride data access						1

▶attention.cpp: 55 - 2.28 %

ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s
Run march=znver5 mprefer-vector-width=256							Run march=znver5 [mprefer-vector-width=512]							Run march=znver5 mprefer-vector-width=512 ffast-math
Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 55-56						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 55-56						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 55-56
38	0.00	0.00	0.05	78.95	20.72	2.5	43	0.01	0.01	0.10	37.5	14.06	2	43	0.01	0.01	0.11	28.57	12.5	2.75
39	0.01	0.01	0.11	42.86	14.29	1	39	0.04	0.04	0.42	83.33	30.21	1.62	39	0.01	0.01	0.16	100	100	13.66
35	0.11	0.12	1.22	81.08	25	2.59	42	0.01	0.01	0.10	83.33	25.35	1.5

Sum on 1 analyzed binary loop (attention-avx512 - 35)							Sum on 1 analyzed binary loop (attention-avx512 - 39)							No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.
Analysis						Count	Analysis						Count	Analysis						Count
Loop Computation Issues							Loop Computation Issues
Presence of expensive FP instructions						1	Presence of expensive FP instructions						1
Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA						1	Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA						1
Presence of a large number of scalar integer instructions						1	Presence of a large number of scalar integer instructions						1
Control Flow Issues							Control Flow Issues
Presence of calls						1	Presence of calls						1
Data Access Issues							Data Access Issues
Presence of indirect access						1	Presence of indirect access						0
More than 10% of the vector loads instructions are unaligned						1	More than 10% of the vector loads instructions are unaligned						1
Presence of special instructions executing on a single port						1	Presence of special instructions executing on a single port						1
More than 20% of the loads are accessing the stack						1	More than 20% of the loads are accessing the stack						1
Vectorization Roadblocks							Vectorization Roadblocks
Presence of calls						1	Presence of calls						1
Presence of indirect access						1	Presence of indirect access						0
Inefficient Vectorization							Inefficient Vectorization
Presence of special instructions executing on a single port						1	Presence of special instructions executing on a single port						1

▶attention.cpp: 52 - 1.32 %

ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s
Run march=znver5 mprefer-vector-width=256							Run march=znver5 [mprefer-vector-width=512]							Run march=znver5 mprefer-vector-width=512 ffast-math
Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 52-53						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 52-53						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 52-53
34	0.00	0.00	0.05	57.14	16.96	2	38	0.01	0.01	0.05	57.14	16.96	2.99	44	0.01	0.01	0.11	100	50	5
33	0.05	0.04	0.48	83.66	26.55	2.78	37	0.01	0.01	0.10	83.78	32.09	1.5	38	0.01	0.01	0.11	57.14	16.96	5.25
40	0.01	0.01	0.11	84.21	21.38	5	44	0.03	0.03	0.31	83.78	26.01	1.25

Sum on 1 analyzed binary loop (attention-avx512 - 33)							Sum on 1 analyzed binary loop (attention-avx512 - 44)							No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.
Analysis						Count	Analysis						Count	Analysis						Count
Loop Computation Issues							Loop Computation Issues
Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA						1	Less than 10% of the FP ADD/SUB/MUL arithmetic operations are performed using FMA						1
Presence of a large number of scalar integer instructions						1	Presence of a large number of scalar integer instructions						1
Control Flow Issues							Control Flow Issues
Presence of calls						1	Presence of calls						1
Data Access Issues							Data Access Issues
Presence of indirect access						1	Presence of indirect access						0
More than 10% of the vector loads instructions are unaligned						1	More than 10% of the vector loads instructions are unaligned						1
Presence of special instructions executing on a single port						1	Presence of special instructions executing on a single port						1
More than 20% of the loads are accessing the stack						1	More than 20% of the loads are accessing the stack						1
Vectorization Roadblocks							Vectorization Roadblocks
Presence of calls						1	Presence of calls						1
Presence of indirect access						1	Presence of indirect access						0
Inefficient Vectorization							Inefficient Vectorization
Presence of special instructions executing on a single port						1	Presence of special instructions executing on a single port						1

▶attention.cpp: 240 - 0.32 %

ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s
Run march=znver5 mprefer-vector-width=256							Run march=znver5 [mprefer-vector-width=512]							Run march=znver5 mprefer-vector-width=512 ffast-math
Loop Source Regions	/usr/lib/gcc/x86_64-redhat-linux/11/../../../../include/c++/11/bits/stl_vector.h: 1046-1046 /home/eoseret/llm-attention/attention.cpp: 240-241						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 240-241						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 240-241
48	0.01	0.01	0.11	38.46	21.15	0	53	0.01	0.01	0.10	100	80	0	53	0.01	0.01	0.11	100	80	0

No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.							No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.							No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.
Analysis						Count	Analysis						Count	Analysis						Count

▶attention.cpp: 47 - 0.21 %

ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s	ASM Loop ID	Max Time Over Threads (s)	Time w.r.t. Wall Time (s)	Cov (%)	Vect. Ratio (%)	Vector Length Use (%)	GFLOP/s
Run march=znver5 mprefer-vector-width=256							Run march=znver5 [mprefer-vector-width=512]							Run march=znver5 mprefer-vector-width=512 ffast-math
Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 47-48						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 47-48						Loop Source Regions	/home/eoseret/llm-attention/attention.cpp: 47-48
41	0.00	0.00	0.05	0	6.25	0	46	0.01	0.01	0.05	100	50	1	46	0.00	0.00	0.05	100	50	3.5
42	0.00	0.00	0.05	100	25	0

No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.							No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.							No Loops Overview analysis found for any assembly loop. More loops can be analyzed using option --summary-loop-count.
Analysis						Count	Analysis						Count	Analysis						Count

Report Configuration

Loops

▶attention.cpp: 30 - 260.14 %

▶attention.cpp: 55 - 2.28 %

▶attention.cpp: 52 - 1.32 %

▶attention.cpp: 240 - 0.32 %

▶attention.cpp: 47 - 0.21 %