Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

MAINT: einsum: Optimize the sub function two-operands by using SIMD. #18194

Merged
merged 1 commit into from
Jan 21, 2021
Merged

MAINT: einsum: Optimize the sub function two-operands by using SIMD. #18194

merged 1 commit into from
Jan 21, 2021

Conversation

Qiyu8
Copy link
Member

@Qiyu8 Qiyu8 commented Jan 20, 2021

Introduction

Here is the final part of #17049 , With code reduced by 69%, There has no impact on X86 platform and about 14%~49% increased performance in ARM.

Benchmark

Here is the ASV benchmark result.

SSE2 enabled 

· Creating environments
· Discovering benchmarks
·· Uninstalling from virtualenv-py3.7-Cython
·· Building e4402bd8  for virtualenv-py3.7-Cython
·· Installing e4402bd8  into virtualenv-py3.7-Cython
· Running 28 total benchmarks (2 commits * 1 environments * 14 benchmarks)
[  0.00%] · For numpy commit 2908338b  (round 1/2):
[  0.00%] ·· Building for virtualenv-py3.7-Cython
[  0.00%] ·· Benchmarking virtualenv-py3.7-Cython
[  1.79%] ··· Running (bench_linalg.Einsum.time_einsum_contig_contig--)..............
[ 25.00%] · For numpy commit e4402bd8  (round 1/2):
[ 25.00%] ·· Building for virtualenv-py3.7-Cython
[ 25.00%] ·· Benchmarking virtualenv-py3.7-Cython
[ 26.79%] ··· Running (bench_linalg.Einsum.time_einsum_contig_contig--)..............
[ 50.00%] · For numpy commit e4402bd8  (round 2/2):
[ 50.00%] ·· Benchmarking virtualenv-py3.7-Cython
[ 51.79%] ··· bench_linalg.Einsum.time_einsum_contig_contig                                                                                                                                                   ok
[ 51.79%] ··· =============== =========
                   dtype
              --------------- ---------
               numpy.float32   139±2μs
               numpy.float64   240±5μs
              =============== =========

[ 53.57%] ··· bench_linalg.Einsum.time_einsum_contig_outstride0                                                                                                                                               ok

[ 53.57%] ··· =============== =========

dtype

--------------- ---------

numpy.float32   130±1μs

numpy.float64   229±9μs

=============== =========


[ 55.36%] ··· bench_linalg.Einsum.time_einsum_mul                                                                                                                                                             ok

[ 55.36%] ··· =============== =============

dtype

--------------- -------------

numpy.float32   1.40±0.05ms

numpy.float64   2.74±0.03ms

=============== =============


[ 57.14%] ··· bench_linalg.Einsum.time_einsum_multiply                                                                                                                                                        ok

[ 57.14%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   65.5±0.7μs

numpy.float64    78.9±3μs

=============== ============


[ 58.93%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_contig                                                                                                                                            ok

[ 58.93%] ··· =============== ============

dtype

--------------- ------------

numpy.float32    42.1±1μs

numpy.float64   43.4±0.5μs

=============== ============


[ 60.71%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_outstride0                                                                                                                                        ok

[ 60.71%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   32.0±0.7μs

numpy.float64    33.0±2μs

=============== ============


[ 62.50%] ··· bench_linalg.Einsum.time_einsum_noncon_mul                                                                                                                                                      ok

[ 62.50%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   39.6±2μs

numpy.float64   46.3±3μs

=============== ==========


[ 64.29%] ··· bench_linalg.Einsum.time_einsum_noncon_multiply                                                                                                                                                 ok

[ 64.29%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   67.9±3μs

numpy.float64   78.3±4μs

=============== ==========


[ 66.07%] ··· bench_linalg.Einsum.time_einsum_noncon_outer                                                                                                                                                    ok

[ 66.07%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   10.7±0.1ms

numpy.float64   21.8±0.3ms

=============== ============


[ 67.86%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul                                                                                                                                                  ok

[ 67.86%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   65.3±1μs

numpy.float64   65.9±6μs

=============== ==========


[ 69.64%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul2                                                                                                                                                 ok

[ 69.64%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   57.7±5μs

numpy.float64   63.5±2μs

=============== ==========


[ 71.43%] ··· bench_linalg.Einsum.time_einsum_outer                                                                                                                                                           ok

[ 71.43%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   24.7±0.3ms

numpy.float64   49.9±0.7ms

=============== ============


[ 73.21%] ··· bench_linalg.Einsum.time_einsum_sum_mul                                                                                                                                                         ok

[ 73.21%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   55.7±2μs

numpy.float64   59.5±3μs

=============== ==========


[ 75.00%] ··· bench_linalg.Einsum.time_einsum_sum_mul2                                                                                                                                                        ok

[ 75.00%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   55.2±0.9μs

numpy.float64    57.7±2μs

=============== ============


[ 75.00%] · For numpy commit 2908338b  (round 2/2):

[ 75.00%] ·· Building for virtualenv-py3.7-Cython

[ 75.00%] ·· Benchmarking virtualenv-py3.7-Cython

[ 76.79%] ··· bench_linalg.Einsum.time_einsum_contig_contig                                                                                                                                                   ok

[ 76.79%] ··· =============== =========

dtype

--------------- ---------

numpy.float32   149±7μs

numpy.float64   237±4μs

=============== =========


[ 78.57%] ··· bench_linalg.Einsum.time_einsum_contig_outstride0                                                                                                                                               ok

[ 78.57%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   133±8μs

numpy.float64   226±20μs

=============== ==========


[ 80.36%] ··· bench_linalg.Einsum.time_einsum_mul                                                                                                                                                             ok

[ 80.36%] ··· =============== =============

dtype

--------------- -------------

numpy.float32   1.31±0.01ms

numpy.float64   2.66±0.06ms

=============== =============


[ 82.14%] ··· bench_linalg.Einsum.time_einsum_multiply                                                                                                                                                        ok

[ 82.14%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   65.6±4μs

numpy.float64   79.2±4μs

=============== ==========


[ 83.93%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_contig                                                                                                                                            ok

[ 83.93%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   42.2±1μs

numpy.float64   44.8±3μs

=============== ==========


[ 85.71%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_outstride0                                                                                                                                        ok

[ 85.71%] ··· =============== ============

dtype

--------------- ------------

numpy.float32    34.3±5μs

numpy.float64   33.0±0.5μs

=============== ============


[ 87.50%] ··· bench_linalg.Einsum.time_einsum_noncon_mul                                                                                                                                                      ok

[ 87.50%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   38.8±0.7μs

numpy.float64    41.3±2μs

=============== ============


[ 89.29%] ··· bench_linalg.Einsum.time_einsum_noncon_multiply                                                                                                                                                 ok

[ 89.29%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   67.8±3μs

numpy.float64   78.2±4μs

=============== ==========


[ 91.07%] ··· bench_linalg.Einsum.time_einsum_noncon_outer                                                                                                                                                    ok

[ 91.07%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   10.8±0.2ms

numpy.float64   22.1±0.2ms

=============== ============


[ 92.86%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul                                                                                                                                                  ok

[ 92.86%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   59.9±2μs

numpy.float64   66.2±3μs

=============== ==========


[ 94.64%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul2                                                                                                                                                 ok

[ 94.64%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   56.6±0.9μs

numpy.float64    70.6±7μs

=============== ============


[ 96.43%] ··· bench_linalg.Einsum.time_einsum_outer                                                                                                                                                           ok

[ 96.43%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   24.5±0.6ms

numpy.float64   49.2±0.9ms

=============== ============


[ 98.21%] ··· bench_linalg.Einsum.time_einsum_sum_mul                                                                                                                                                         ok

[ 98.21%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   56.5±1μs

numpy.float64   62.9±3μs

=============== ==========


[100.00%] ··· bench_linalg.Einsum.time_einsum_sum_mul2                                                                                                                                                        ok

[100.00%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   55.1±0.8μs

numpy.float64    62.3±5μs

=============== ============


BENCHMARKS NOT SIGNIFICANTLY CHANGED.

AV2 enabled 

· Creating environments
· Discovering benchmarks
·· Uninstalling from virtualenv-py3.7-Cython
·· Building e4402bd8  for virtualenv-py3.7-Cython
·· Installing e4402bd8  into virtualenv-py3.7-Cython
· Running 28 total benchmarks (2 commits * 1 environments * 14 benchmarks)
[  0.00%] · For numpy commit 2908338b  (round 1/2):
[  0.00%] ·· Building for virtualenv-py3.7-Cython
[  0.00%] ·· Benchmarking virtualenv-py3.7-Cython
[  1.79%] ··· Running (bench_linalg.Einsum.time_einsum_contig_contig--)..............
[ 25.00%] · For numpy commit e4402bd8  (round 1/2):
[ 25.00%] ·· Building for virtualenv-py3.7-Cython
[ 25.00%] ·· Benchmarking virtualenv-py3.7-Cython
[ 26.79%] ··· Running (bench_linalg.Einsum.time_einsum_contig_contig--)..............
[ 50.00%] · For numpy commit e4402bd8  (round 2/2):
[ 50.00%] ·· Benchmarking virtualenv-py3.7-Cython
[ 51.79%] ··· bench_linalg.Einsum.time_einsum_contig_contig                                                                                                                                                   ok
[ 51.79%] ··· =============== =========
                   dtype
              --------------- ---------
               numpy.float32   100±2μs
               numpy.float64   152±6μs
              =============== =========

[ 53.57%] ··· bench_linalg.Einsum.time_einsum_contig_outstride0                                                                                                                                               ok

[ 53.57%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   146±10μs

numpy.float64   244±2μs

=============== ==========


[ 55.36%] ··· bench_linalg.Einsum.time_einsum_mul                                                                                                                                                             ok

[ 55.36%] ··· =============== =============

dtype

--------------- -------------

numpy.float32   1.41±0.02ms

numpy.float64   2.66±0.03ms

=============== =============


[ 57.14%] ··· bench_linalg.Einsum.time_einsum_multiply                                                                                                                                                        ok

[ 57.14%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   79.9±2μs

numpy.float64   87.4±2μs

=============== ==========


[ 58.93%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_contig                                                                                                                                            ok

[ 58.93%] ··· =============== ============

dtype

--------------- ------------

numpy.float32    42.6±2μs

numpy.float64   41.6±0.6μs

=============== ============


[ 60.71%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_outstride0                                                                                                                                        ok

[ 60.71%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   41.0±0.9μs

numpy.float64   40.2±0.5μs

=============== ============


[ 62.50%] ··· bench_linalg.Einsum.time_einsum_noncon_mul                                                                                                                                                      ok

[ 62.50%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   49.2±0.4μs

numpy.float64    48.2±1μs

=============== ============


[ 64.29%] ··· bench_linalg.Einsum.time_einsum_noncon_multiply                                                                                                                                                 ok

[ 64.29%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   76.7±2μs

numpy.float64   83.3±2μs

=============== ==========


[ 66.07%] ··· bench_linalg.Einsum.time_einsum_noncon_outer                                                                                                                                                    ok

[ 66.07%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   10.4±0.1ms

numpy.float64   21.9±0.4ms

=============== ============


[ 67.86%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul                                                                                                                                                  ok

[ 67.86%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   68.6±2μs

numpy.float64   70.7±1μs

=============== ==========


[ 69.64%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul2                                                                                                                                                 ok

[ 69.64%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   75.4±6μs

numpy.float64   71.7±1μs

=============== ==========


[ 71.43%] ··· bench_linalg.Einsum.time_einsum_outer                                                                                                                                                           ok

[ 71.43%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   24.1±0.3ms

numpy.float64   50.2±0.6ms

=============== ============


[ 73.21%] ··· bench_linalg.Einsum.time_einsum_sum_mul                                                                                                                                                         ok

[ 73.21%] ··· =============== ===========

dtype

--------------- -----------

numpy.float32    67.5±2μs

numpy.float64   71.0±10μs

=============== ===========


[ 75.00%] ··· bench_linalg.Einsum.time_einsum_sum_mul2                                                                                                                                                        ok

[ 75.00%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   69.8±4μs

numpy.float64   69.6±2μs

=============== ==========


[ 75.00%] · For numpy commit 2908338b  (round 2/2):

[ 75.00%] ·· Building for virtualenv-py3.7-Cython

[ 75.00%] ·· Benchmarking virtualenv-py3.7-Cython

[ 76.79%] ··· bench_linalg.Einsum.time_einsum_contig_contig                                                                                                                                                   ok

[ 76.79%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   101±5μs

numpy.float64   155±10μs

=============== ==========


[ 78.57%] ··· bench_linalg.Einsum.time_einsum_contig_outstride0                                                                                                                                               ok

[ 78.57%] ··· =============== =========

dtype

--------------- ---------

numpy.float32   146±7μs

numpy.float64   252±7μs

=============== =========


[ 80.36%] ··· bench_linalg.Einsum.time_einsum_mul                                                                                                                                                             ok

[ 80.36%] ··· =============== =============

dtype

--------------- -------------

numpy.float32   1.39±0.09ms

numpy.float64   2.63±0.04ms

=============== =============


[ 82.14%] ··· bench_linalg.Einsum.time_einsum_multiply                                                                                                                                                        ok

[ 82.14%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   78.6±2μs

numpy.float64   87.2±3μs

=============== ==========


[ 83.93%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_contig                                                                                                                                            ok

[ 83.93%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   47.4±5μs

numpy.float64   53.9±9μs

=============== ==========


[ 85.71%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_outstride0                                                                                                                                        ok

[ 85.71%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   40.2±2μs

numpy.float64   41.6±4μs

=============== ==========


[ 87.50%] ··· bench_linalg.Einsum.time_einsum_noncon_mul                                                                                                                                                      ok

[ 87.50%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   41.3±3μs

numpy.float64   39.6±2μs

=============== ==========


[ 89.29%] ··· bench_linalg.Einsum.time_einsum_noncon_multiply                                                                                                                                                 ok

[ 89.29%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   77.5±4μs

numpy.float64   86.9±1μs

=============== ==========


[ 91.07%] ··· bench_linalg.Einsum.time_einsum_noncon_outer                                                                                                                                                    ok

[ 91.07%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   10.7±0.2ms

numpy.float64   21.9±0.1ms

=============== ============


[ 92.86%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul                                                                                                                                                  ok

[ 92.86%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   67.9±3μs

numpy.float64   70.6±2μs

=============== ==========


[ 94.64%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul2                                                                                                                                                 ok

[ 94.64%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   68.3±1μs

numpy.float64   69.7±1μs

=============== ==========


[ 96.43%] ··· bench_linalg.Einsum.time_einsum_outer                                                                                                                                                           ok

[ 96.43%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   25.0±0.4ms

numpy.float64   49.1±0.5ms

=============== ============


[ 98.21%] ··· bench_linalg.Einsum.time_einsum_sum_mul                                                                                                                                                         ok

[ 98.21%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   70.0±3μs

numpy.float64   69.2±7μs

=============== ==========


[100.00%] ··· bench_linalg.Einsum.time_einsum_sum_mul2                                                                                                                                                        ok

[100.00%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   69.9±1μs

numpy.float64   69.1±1μs

=============== ==========


BENCHMARKS NOT SIGNIFICANTLY CHANGED.

NEON enabled 

· Creating environments
· Discovering benchmarks
·· Uninstalling from virtualenv-py3.7-Cython.
·· Building ebfed05a  for virtualenv-py3.7-Cython................................................
·· Installing ebfed05a  into virtualenv-py3.7-Cython.
· Running 28 total benchmarks (2 commits * 1 environments * 14 benchmarks)
[  0.00%] · For numpy commit efaf210f  (round 1/2):
[  0.00%] ·· Building for virtualenv-py3.7-Cython..................................................
[  0.00%] ·· Benchmarking virtualenv-py3.7-Cython
[  1.79%] ··· Running (bench_linalg.Einsum.time_einsum_contig_contig--)..............
[ 25.00%] · For numpy commit ebfed05a  (round 1/2):
[ 25.00%] ·· Building for virtualenv-py3.7-Cython..
[ 25.00%] ·· Benchmarking virtualenv-py3.7-Cython
[ 26.79%] ··· Running (bench_linalg.Einsum.time_einsum_contig_contig--)..............
[ 50.00%] · For numpy commit ebfed05a  (round 2/2):
[ 50.00%] ·· Benchmarking virtualenv-py3.7-Cython
[ 51.79%] ··· bench_linalg.Einsum.time_einsum_contig_contig                                                       ok
[ 51.79%] ··· =============== =========
                   dtype               
              --------------- ---------
               numpy.float32   198±2μs 
               numpy.float64   353±5μs 
              =============== =========

[ 53.57%] ··· bench_linalg.Einsum.time_einsum_contig_outstride0                                                   ok

[ 53.57%] ··· =============== =========

dtype

--------------- ---------

numpy.float32   224±2μs

numpy.float64   356±6μs

=============== =========


[ 55.36%] ··· bench_linalg.Einsum.time_einsum_mul                                                                 ok

[ 55.36%] ··· =============== ==========

dtype

--------------- ----------

numpy.float32   487±10μs

numpy.float64   863±20μs

=============== ==========


[ 57.14%] ··· bench_linalg.Einsum.time_einsum_multiply                                                            ok

[ 57.14%] ··· =============== ===========

dtype

--------------- -----------

numpy.float32   115±0.5μs

numpy.float64    141±2μs

=============== ===========


[ 58.93%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_contig                                                ok

[ 58.93%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   73.5±0.5μs

numpy.float64   74.0±0.7μs

=============== ============


[ 60.71%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_outstride0                                            ok

[ 60.71%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   59.6±0.2μs

numpy.float64   60.5±0.6μs

=============== ============


[ 62.50%] ··· bench_linalg.Einsum.time_einsum_noncon_mul                                                          ok

[ 62.50%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   70.1±0.4μs

numpy.float64   72.0±0.5μs

=============== ============


[ 64.29%] ··· bench_linalg.Einsum.time_einsum_noncon_multiply                                                     ok

[ 64.29%] ··· =============== =========

dtype

--------------- ---------

numpy.float32   115±2μs

numpy.float64   140±2μs

=============== =========


[ 66.07%] ··· bench_linalg.Einsum.time_einsum_noncon_outer                                                        ok

[ 66.07%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   2.73±0.1ms

numpy.float64   6.22±0.2ms

=============== ============


[ 67.86%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul                                                      ok

[ 67.86%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   99.3±0.4μs

numpy.float64   107±0.8μs

=============== ============


[ 69.64%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul2                                                     ok

[ 69.64%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   99.0±0.3μs

numpy.float64    107±1μs

=============== ============


[ 71.43%] ··· bench_linalg.Einsum.time_einsum_outer                                                               ok

[ 71.43%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   10.8±0.4ms

numpy.float64   22.1±0.3ms

=============== ============


[ 73.21%] ··· bench_linalg.Einsum.time_einsum_sum_mul                                                             ok

[ 73.21%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   94.5±0.4μs

numpy.float64   98.1±0.3μs

=============== ============


[ 75.00%] ··· bench_linalg.Einsum.time_einsum_sum_mul2                                                            ok

[ 75.00%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   94.3±0.7μs

numpy.float64   98.7±0.4μs

=============== ============


[ 75.00%] · For numpy commit efaf210f  (round 2/2):

[ 75.00%] ·· Building for virtualenv-py3.7-Cython..

[ 75.00%] ·· Benchmarking virtualenv-py3.7-Cython

[ 76.79%] ··· bench_linalg.Einsum.time_einsum_contig_contig                                                       ok

[ 76.79%] ··· =============== =========

dtype

--------------- ---------

numpy.float32   204±6μs

numpy.float64   355±5μs

=============== =========


[ 78.57%] ··· bench_linalg.Einsum.time_einsum_contig_outstride0                                                   ok

[ 78.57%] ··· =============== =========

dtype

--------------- ---------

numpy.float32   226±6μs

numpy.float64   356±4μs

=============== =========


[ 80.36%] ··· bench_linalg.Einsum.time_einsum_mul                                                                 ok

[ 80.36%] ··· =============== =============

dtype

--------------- -------------

numpy.float32     789±30μs

numpy.float64   1.00±0.01ms

=============== =============


[ 82.14%] ··· bench_linalg.Einsum.time_einsum_multiply                                                            ok

[ 82.14%] ··· =============== ===========

dtype

--------------- -----------

numpy.float32   117±0.2μs

numpy.float64    145±3μs

=============== ===========


[ 83.93%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_contig                                                ok

[ 83.93%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   74.2±0.4μs

numpy.float64   76.0±0.6μs

=============== ============


[ 85.71%] ··· bench_linalg.Einsum.time_einsum_noncon_contig_outstride0                                            ok

[ 85.71%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   59.4±0.3μs

numpy.float64   60.5±0.3μs

=============== ============


[ 87.50%] ··· bench_linalg.Einsum.time_einsum_noncon_mul                                                          ok

[ 87.50%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   72.0±0.2μs

numpy.float64   72.9±0.2μs

=============== ============


[ 89.29%] ··· bench_linalg.Einsum.time_einsum_noncon_multiply                                                     ok

[ 89.29%] ··· =============== ===========

dtype

--------------- -----------

numpy.float32   115±0.5μs

numpy.float64   140±0.8μs

=============== ===========


[ 91.07%] ··· bench_linalg.Einsum.time_einsum_noncon_outer                                                        ok

[ 91.07%] ··· =============== =============

dtype

--------------- -------------

numpy.float32   5.35±0.07ms

numpy.float64    7.07±0.2ms

=============== =============


[ 92.86%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul                                                      ok

[ 92.86%] ··· =============== ===========

dtype

--------------- -----------

numpy.float32   101±0.6μs

numpy.float64   109±0.5μs

=============== ===========


[ 94.64%] ··· bench_linalg.Einsum.time_einsum_noncon_sum_mul2                                                     ok

[ 94.64%] ··· =============== ===========

dtype

--------------- -----------

numpy.float32    101±1μs

numpy.float64   109±0.7μs

=============== ===========


[ 96.43%] ··· bench_linalg.Einsum.time_einsum_outer                                                               ok

[ 96.43%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   17.7±0.3ms

numpy.float64   25.8±0.5ms

=============== ============


[ 98.21%] ··· bench_linalg.Einsum.time_einsum_sum_mul                                                             ok

[ 98.21%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   95.5±0.5μs

numpy.float64   101±0.8μs

=============== ============


[100.00%] ··· bench_linalg.Einsum.time_einsum_sum_mul2                                                            ok

[100.00%] ··· =============== ============

dtype

--------------- ------------

numpy.float32   95.5±0.8μs

numpy.float64   98.8±0.5μs

=============== ============


   before           after         ratio
 [efaf210f]       [ebfed05a]
 <master>         <einsum-twooperands>



    

  • 

    1.00±0.01ms         863±20μs     0.86  bench_linalg.Einsum.time_einsum_mul(<class 'numpy.float64'>)

    

    • 

  • 

     25.8±0.5ms       22.1±0.3ms     0.86  bench_linalg.Einsum.time_einsum_outer(<class 'numpy.float64'>)

    

    • 

  • 

       789±30μs         487±10μs     0.62  bench_linalg.Einsum.time_einsum_mul(<class 'numpy.float32'>)

    

    • 

  • 

     17.7±0.3ms       10.8±0.4ms     0.61  bench_linalg.Einsum.time_einsum_outer(<class 'numpy.float32'>)

    

    • 

  • 

    5.35±0.07ms       2.73±0.1ms     0.51  bench_linalg.Einsum.time_einsum_noncon_outer(<class 'numpy.float32'>)

    

    • 



SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY.

PERFORMANCE INCREASED.

System Info

  Arm x86
Hardware KunPeng  
Processor ARMv8 2.6GMHZ 8 processors Intel(R) Xeon(R) Gold 6161 CPU @ 2.20GHz
OS Linux ecs-9d50 4.19.36-vhulk1905.1.0.h276.eulerosv2r8.aarch64 Windows Server 2008 R2 Enterprise
Compiler gcc (GCC) 7.3.0 MSVC14.06

@Qiyu8 Qiyu8 added 01 - Enhancement component: SIMD Issues in SIMD (fast instruction sets) code or machinery labels Jan 20, 2021
@eric-wieser eric-wieser changed the title Optimize the sub function two-operands by using SIMD. MAINT: einsum: Optimize the sub function two-operands by using SIMD. Jan 20, 2021
@seiko2plus seiko2plus self-assigned this Jan 20, 2021
@seiko2plus seiko2plus self-requested a review January 20, 2021 22:47
seiko2plus
seiko2plus approved these changes Jan 20, 2021
View reviewed changes
Copy link
Member

@seiko2plus seiko2plus left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Well done, Thank you Chunlin!

@seiko2plus
Copy link
Member

The new code improves the accuracy due to the use of FMA too. we will have to dispatch AVX2&FMA3 and AVX512F in runtime.

@charris charris merged commit b91f3c0 into numpy:master Jan 21, 2021
@charris
Copy link
Member

charris commented Jan 21, 2021

Thanks Chunlin.

@Qiyu8 Qiyu8 deleted the einsum-twooperands branch January 21, 2021 01:11
@Qiyu8
Copy link
Member Author

Qiyu8 commented Jan 21, 2021

The dispatching solution for non-UFunc is not recommend in the past, If it is acceptable now, then we can have a discussion on the mailing list.

@ZiqiChai
Copy link

ZiqiChai commented Mar 2, 2021

This is exactly what I'm looking for right now! And I just found it here! Thanks, Chunlin.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@seiko2plus seiko2plus seiko2plus approved these changes

Assignees

@seiko2plus seiko2plus

Labels
01 - Enhancement 03 - Maintenance component: SIMD Issues in SIMD (fast instruction sets) code or machinery
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
4 participants
@Qiyu8 @seiko2plus @charris @ZiqiChai