[DRAFT] Move solve to GPU

[govindchari]

Currently, only the factorization and solves are performed on the GPU. All other operations are done on the CPU. This incurs significant overhead due to cudaMemcpy, so this PR seeks to remove all cudaMemcpy.

Track down all get_data_vectorf() calls. These should ideally return GPU pointers, but currently just returns CPU pointers.

Then write kernels to operate directly on QOCOVectorf rather than directly accessing pointers in memory.

The main code for QOCO should not touch pointers at all and should just use the QOCOVectorf abstraction.

[github-actions]

Download benchmark artifacts

Benchmark Summary

• Baseline solved: 129 problems
• Diff branch solved: 129 problems

Runtime regressions (> 5.0%)

• HS35: diff=0.0000s, baseline=0.0000s, Δ=+70.0%

Runtime improvements (> 5.0%)

• CVXQP1_S: diff=0.0005s, baseline=0.0005s, Δ=-6.7%
• DPKLO1: diff=0.0001s, baseline=0.0001s, Δ=-8.1%
• QADLITTL: diff=0.0004s, baseline=0.0004s, Δ=-7.6%
• QAFIRO: diff=0.0001s, baseline=0.0002s, Δ=-7.2%

[github-actions]

Download benchmark artifacts

Benchmark Summary

• Baseline solved: 129 problems
• Diff branch solved: 129 problems

Runtime regressions (> 5.0%)

• DUALC5: diff=0.0007s, baseline=0.0006s, Δ=+10.9%
• HS268: diff=0.0000s, baseline=0.0000s, Δ=+6.3%
• HS53: diff=0.0000s, baseline=0.0000s, Δ=+6.3%
• HS76: diff=0.0000s, baseline=0.0000s, Δ=+6.7%
• QSHARE1B: diff=0.0023s, baseline=0.0021s, Δ=+6.7%
• S268: diff=0.0000s, baseline=0.0000s, Δ=+6.3%
• ZECEVIC2: diff=0.0000s, baseline=0.0000s, Δ=+8.3%

Runtime improvements (> 5.0%)

• DUALC8: diff=0.0016s, baseline=0.0020s, Δ=-20.9%
• HS35: diff=0.0000s, baseline=0.0000s, Δ=-65.6%
• PRIMALC2: diff=0.0018s, baseline=0.0020s, Δ=-7.0%
• QPTEST: diff=0.0000s, baseline=0.0000s, Δ=-60.7%

[github-actions]

Download benchmark artifacts

Benchmark Summary

• Baseline solved: 129 problems
• Diff branch solved: 8 problems

Differences in solved problems

• Baseline solved additional problems: AUG2DCQP, AUG2DQP, AUG3DCQP, AUG3DQP, BOYD1, BOYD2, CONT-050, CONT-100, CONT-101, CONT-200, CONT-201, CONT-300, CVXQP1_M, CVXQP1_S, CVXQP2_L, CVXQP2_M, CVXQP2_S, CVXQP3_M, CVXQP3_S, DTOC3, DUAL1, DUAL2, DUAL3, DUAL4, DUALC1, DUALC2, DUALC5, DUALC8, EXDATA, GOULDQP2, GOULDQP3, HS118, HS21, HS268, HS35, HS35MOD, HS53, HS76, HUES-MOD, HUESTIS, KSIP, LASER, LISWET2, LISWET3, LISWET4, LISWET5, LISWET6, LISWET7, LOTSCHD, MOSARQP1, MOSARQP2, POWELL20, PRIMAL1, PRIMAL2, PRIMAL3, PRIMAL4, PRIMALC1, PRIMALC2, PRIMALC5, PRIMALC8, Q25FV47, QADLITTL, QAFIRO, QBANDM, QBEACONF, QBORE3D, QBRANDY, QCAPRI, QE226, QETAMACR, QFFFFF80, QFORPLAN, QGFRDXPN, QGROW15, QGROW22, QGROW7, QISRAEL, QPCBLEND, QPCBOEI1, QPCBOEI2, QPCSTAIR, QPILOTNO, QPTEST, QRECIPE, QSC205, QSCAGR25, QSCAGR7, QSCFXM1, QSCFXM2, QSCFXM3, QSCORPIO, QSCRS8, QSCSD1, QSCSD6, QSCSD8, QSCTAP1, QSCTAP2, QSCTAP3, QSEBA, QSHARE1B, QSHARE2B, QSHELL, QSHIP04L, QSHIP04S, QSHIP08L, QSHIP08S, QSHIP12L, QSHIP12S, QSIERRA, QSTAIR, QSTANDAT, S268, STADAT1, STADAT2, STADAT3, STCQP1, STCQP2, TAME, UBH1, VALUES, ZECEVIC2

Runtime regressions (> 5.0%)

• AUG3D: diff=0.0019s, baseline=0.0016s, Δ=+20.9%
• DPKLO1: diff=0.0002s, baseline=0.0001s, Δ=+48.9%
• GENHS28: diff=0.0000s, baseline=0.0000s, Δ=+25.0%
• HS51: diff=0.0000s, baseline=0.0000s, Δ=+66.7%
• HS52: diff=0.0000s, baseline=0.0000s, Δ=+33.3%