Task04 Евгений Свирин ITMO #298

EvgenySvirin · 2024-12-08T11:50:21Z

Локальный вывод

matrix_transpose:

OpenCL devices: Device #0: GPU. NVIDIA GeForce RTX 4060 Laptop GPU. Total memory: 7940 Mb Using device #0: GPU. NVIDIA GeForce RTX 4060 Laptop GPU. Total memory: 7940 Mb Data generated for M=4096, K=4096 [matrix_transpose_naive] GPU: 0.000922783+-1.25289e-06 s GPU: 18181.1 millions/s [matrix_transpose_local_bad_banks] GPU: 0.00070685+-1.91988e-05 s GPU: 23735.2 millions/s [matrix_transpose_local_good_banks] GPU: 0.000704517+-2.50583e-05 s GPU: 23813.8 millions/s

matrix_multiplication
OpenCL devices:
  Device #0: GPU. NVIDIA GeForce RTX 4060 Laptop GPU. Total memory: 7940 Mb
Using device #0: GPU. NVIDIA GeForce RTX 4060 Laptop GPU. Total memory: 7940 Mb
Data generated for M=1024, K=1024, N=1024
CPU: 2.31521+-0 s
CPU: 0.863853 GFlops
[naive, ts=4]
    GPU: 0.00959667+-7.45356e-07 s
    GPU: 208.406 GFlops
    Average difference: 0.000149043%
[naive, ts=8]
    GPU: 0.0047985+-2.87228e-06 s
    GPU: 416.797 GFlops
    Average difference: 0.000149043%
[naive, ts=16]
    GPU: 0.00407633+-4.71405e-07 s
    GPU: 490.637 GFlops
    Average difference: 0.000149043%
[local, ts=4]
    GPU: 0.0106832+-0.00147418 s
    GPU: 187.21 GFlops
    Average difference: 0.000149043%
[local, ts=8]
    GPU: 0.002946+-2.73435e-05 s
    GPU: 678.887 GFlops
    Average difference: 0.000149043%
[local, ts=16]
    GPU: 0.0022445+-8.52834e-05 s
    GPU: 891.067 GFlops
    Average difference: 0.000149043%
[local wpt, ts=4, wpt=2]
    GPU: 0.0123123+-4.89069e-05 s
    GPU: 162.439 GFlops
    Average difference: 0.000149043%
[local wpt, ts=4, wpt=4]
    GPU: 0.0201325+-1.60728e-06 s
    GPU: 99.3419 GFlops
    Average difference: 0.000149043%
[local wpt, ts=8, wpt=2]
    GPU: 0.002337+-5.7735e-07 s
    GPU: 855.798 GFlops
    Average difference: 0.000149043%
[local wpt, ts=8, wpt=4]
    GPU: 0.00327633+-4.1696e-05 s
    GPU: 610.438 GFlops
    Average difference: 0.000149043%
[local wpt, ts=8, wpt=8]
    GPU: 0.00561233+-0.000115237 s
    GPU: 356.358 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=2]
    GPU: 0.00149383+-3.72678e-07 s
    GPU: 1338.84 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=4]
    GPU: 0.00115217+-6.87184e-07 s
    GPU: 1735.86 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=8]
    GPU: 0.000981667+-4.71405e-07 s
    GPU: 2037.35 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=16]
    GPU: 0.00173333+-4.71405e-07 s
    GPU: 1153.85 GFlops
    Average difference: 0.000149043%

Вывод Github CI

matrix_transpose:
OpenCL devices:
  Device #0: CPU. AMD EPYC [7](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:7:8)763 64-Core Processor                . Intel(R) Corporation. Total memory: 15991 Mb
Using device #0: CPU. AMD EPYC 7763 64-Core Processor                . Intel(R) Corporation. Total memory: 15991 Mb
Data generated for M=4096, K=4096
[matrix_transpose_naive]
    GPU: 0.0150901+-[8](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:7:9).66783e-05 s
    GPU: 1111.8 millions/s
[matrix_transpose_local_bad_banks]
    GPU: 0.0278532+-0.00045152 s
    GPU: 602.345 millions/s
[matrix_transpose_local_good_banks]
    GPU: 0.0277682+-0.000530473 s
    GPU: 604.188 millions/s

matrix_multiplication
OpenCL devices:
  Device #0: CPU. AMD EPYC [7](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:8:8)763 64-Core Processor                . Intel(R) Corporation. Total memory: 15991 Mb
Using device #0: CPU. AMD EPYC 7763 64-Core Processor                . Intel(R) Corporation. Total memory: 15991 Mb
Data generated for M=1024, K=1024, N=1024
CPU: 6.43314+-0 s
CPU: 0.310[8](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:8:9)9 GFlops
[naive, ts=4]
    GPU: 0.256356+-0.00128988 s
    GPU: 7.80166 GFlops
    Average difference: 0.00014[9](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:8:10)043%
[naive, ts=8]
    GPU: 0.262318+-0.00330994 s
    GPU: 7.62432 GFlops
    Average difference: 0.000149043%
[naive, ts=16]
    GPU: 0.271294+-0.00518451 s
    GPU: 7.37207 GFlops
    Average difference: 0.000149043%
[local, ts=4]
    GPU: 0.551742+-0.00170809 s
    GPU: 3.62488 GFlops
    Average difference: 0.000149043%
[local, ts=8]
    GPU: 0.145986+-0.000258419 s
    GPU: 13.7 GFlops
    Average difference: 0.000149043%
[local, ts=16]
    GPU: 0.09[10](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:8:11)955+-0.000269942 s
    GPU: 21.955 GFlops
    Average difference: 0.000149043%
[local wpt, ts=4, wpt=2]
    GPU: 0.529485+-0.00160939 s
    GPU: 3.77725 GFlops
    Average difference: 0.000149043%
[local wpt, ts=4, wpt=4]
    GPU: 0.460945+-0.000407938 s
    GPU: 4.33891 GFlops
    Average difference: 0.000149043%
[local wpt, ts=8, wpt=2]
    GPU: 0.134053+-0.000816249 s
    GPU: 14.9195 GFlops
    Average difference: 0.000149043%
[local wpt, ts=8, wpt=4]
    GPU: 0.148363+-0.000794369 s
    GPU: 13.4805 GFlops
    Average difference: 0.000149043%
[local wpt, ts=8, wpt=8]
    GPU: 0.153269+-0.000191614 s
    GPU: 13.0489 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=2]
    GPU: 0.078937+-0.000759654 s
    GPU: 25.3367 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=4]
    GPU: 0.089023+-0.000379301 s
    GPU: 22.4661 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=8]
    GPU: 0.0806555+-0.00109715 s
    GPU: 24.7968 GFlops
    Average difference: 0.000149043%
[local wpt, ts=16, wpt=16]
    GPU: 0.0780773+-0.000503[12](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:8:13)3 s
    GPU: 25.6156 GFlops
    Average difference: 0.000[14](https://github.com/EvgenySvirin/GPGPUTasks2024/actions/runs/12221487152/job/34090582750#step:8:15)9043%

simiyutin · 2024-12-10T23:17:46Z

src/cl/matrix_transpose.cl

+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    as_t[i * M + j] = tile[local_j][local_i];


нет смысла в таком использовании локальной памяти, что записали то и считали

что в итоге приводит к некоалесд записи

Да, я там строчку со стобцом перепутал. Столбцы дожны быть TYLE_SIZE + 1.

Но теперь сделал большой рефактор (попытка сделать коалесед чтение и запись).

Локальный вывод

matrix_transpose:
OpenCL devices: Device #0: GPU. NVIDIA GeForce RTX 4060 Laptop GPU. Total memory: 7940 Mb Using device #0: GPU. NVIDIA GeForce RTX 4060 Laptop GPU. Total memory: 7940 Mb Data generated for M=4096, K=4096 [matrix_transpose_naive] GPU: 0.00080885+-1.58981e-06 s GPU: 20742.1 millions/s [matrix_transpose_local_bad_banks] GPU: 0.000732367+-0.00014281 s GPU: 22908.2 millions/s [matrix_transpose_local_good_banks] GPU: 0.00063305+-7.89478e-05 s GPU: 26502.2 millions/s

Вывод Github CI

matrix_transpose:
OpenCL devices: Device #0: CPU. AMD EPYC [7](https://github.com/GPGPUCourse/GPGPUTasks2024/actions/runs/12333311266/job/34422166124#step:7:8)763 64-Core Processor . Intel(R) Corporation. Total memory: 15991 Mb Using device #0: CPU. AMD EPYC 7763 64-Core Processor . Intel(R) Corporation. Total memory: 15991 Mb Data generated for M=4096, K=4096 [matrix_transpose_naive] GPU: 0.01752[8](https://github.com/GPGPUCourse/GPGPUTasks2024/actions/runs/12333311266/job/34422166124#step:7:9)4+-0.00137051 s GPU: 957.144 millions/s [matrix_transpose_local_bad_banks] GPU: 0.0135164+-0.000128746 s GPU: 1241.25 millions/s [matrix_transpose_local_good_banks] GPU: 0.0137078+-5.88634e-05 s GPU: 1223.[9](https://github.com/GPGPUCourse/GPGPUTasks2024/actions/runs/12333311266/job/34422166124#step:7:10)2 millions/s

… 1 instead of tyle rows

simiyutin · 2024-12-17T13:37:44Z

Все хорошо, задача зачтена, 8/10 баллов 👍

EvgenySvirin added 2 commits December 7, 2024 22:40

add matrix_transpose solution

30cb63b

add matrix_multiplcation solution

cfa26bc

EvgenySvirin force-pushed the task04 branch from c708d8c to cfa26bc Compare December 8, 2024 12:16

simiyutin reviewed Dec 10, 2024

View reviewed changes

EvgenySvirin added 2 commits December 14, 2024 20:41

fix shift for memory banks

b1e747d

refactor: now coaleced write and read, fix: tyle cols now TYLE_SIZE +…

d1ebb3e

… 1 instead of tyle rows

EvgenySvirin force-pushed the task04 branch from e56db7f to d1ebb3e Compare December 14, 2024 22:17

simiyutin closed this Dec 17, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Task04 Евгений Свирин ITMO #298

Task04 Евгений Свирин ITMO #298

EvgenySvirin commented Dec 8, 2024

simiyutin Dec 10, 2024

simiyutin Dec 10, 2024

EvgenySvirin Dec 14, 2024 •

edited

Loading

simiyutin commented Dec 17, 2024


		barrier(CLK_LOCAL_MEM_FENCE);

		as_t[i * M + j] = tile[local_j][local_i];

Task04 Евгений Свирин ITMO #298

Task04 Евгений Свирин ITMO #298

Conversation

EvgenySvirin commented Dec 8, 2024

simiyutin Dec 10, 2024

Choose a reason for hiding this comment

simiyutin Dec 10, 2024

Choose a reason for hiding this comment

EvgenySvirin Dec 14, 2024 • edited Loading

Choose a reason for hiding this comment

simiyutin commented Dec 17, 2024

EvgenySvirin Dec 14, 2024 •

edited

Loading