Module: TinyNNCuda

Defined in:
lib/toy/ffi/tinynn_cuda.rb

Class Method Summary collapse

Class Method Details

.adam_step(param, grad, m, v, lr, b1, b2, eps, omc1, omc2) ⇒ Object

Adam step via custom CPU kernel. Returns AdamStepResult (param, mom_m, mom_v) — same shape as TinyNN.adam_step.



758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 758

def self.adam_step(param, grad, m, v, lr, b1, b2, eps, omc1, omc2)
  sess = TinyNNCuda.tnn_session_new(1)
  n = param.nrows * param.ncols
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, param.flat[i])
    i = i + 1
  end
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, n + i, grad.flat[i])
    i = i + 1
  end
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, 2 * n + i, m.flat[i])
    i = i + 1
  end
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, 3 * n + i, v.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_adam_step_scratch(sess, n, lr, b1, b2, eps, omc1, omc2)
  new_param = Mat.new(param.nrows, param.ncols)
  new_mom_m = Mat.new(param.nrows, param.ncols)
  new_mom_v = Mat.new(param.nrows, param.ncols)
  i = 0
  while i < n
    new_param.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    new_mom_m.flat[i] = TinyNNCuda.tnn_scratch_get(sess, 2 * n + i)
    new_mom_v.flat[i] = TinyNNCuda.tnn_scratch_get(sess, 3 * n + i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  AdamStepResult.new(new_param, new_mom_m, new_mom_v)
end

.add(a, b) ⇒ Object 



379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 379

def self.add(a, b)
  sess = TinyNNCuda.tnn_session_new(1)
  ta = TinyNNCuda.tnn_input_2d_f32(sess, a.nrows, a.ncols)
  tb = TinyNNCuda.tnn_input_2d_f32(sess, b.nrows, b.ncols)
  tc = TinyNNCuda.tnn_add(sess, ta, tb)
  TinyNNCuda.tnn_realize(sess, tc)
  n = a.nrows * a.ncols
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, a.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta)
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, b.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tb)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(a.nrows, a.ncols)
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.alloc_1d_i32(sess, n) ⇒ Object 



800
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 800

def self.alloc_1d_i32(sess, n);        TinyNNCuda.tnn_input_1d_i32(sess, n); end

.alloc_2d(sess, r, c) ⇒ Object 



799
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 799

def self.alloc_2d(sess, r, c);         TinyNNCuda.tnn_input_2d_f32(sess, r, c); end

.build_add(sess, ta, tb) ⇒ Object 



802
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 802

def self.build_add(sess, ta, tb);      TinyNNCuda.tnn_add(sess, ta, tb); end

.build_gelu(sess, ta) ⇒ Object 



803
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 803

def self.build_gelu(sess, ta);         TinyNNCuda.tnn_gelu(sess, ta); end

.build_matmul(sess, ta, tb) ⇒ Object 



801
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 801

def self.build_matmul(sess, ta, tb);   TinyNNCuda.tnn_matmul(sess, ta, tb); end

.build_rms_norm(sess, x, g, e) ⇒ Object 



806
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 806

def self.build_rms_norm(sess, x, g, e); TinyNNCuda.tnn_rms_norm(sess, x, g, e); end

.build_scale(sess, ta, s) ⇒ Object 



805
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 805

def self.build_scale(sess, ta, s);     TinyNNCuda.tnn_scale(sess, ta, s); end

.build_softmax(sess, ta) ⇒ Object 



804
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 804

def self.build_softmax(sess, ta);      TinyNNCuda.tnn_softmax(sess, ta); end

.compute(sess) ⇒ Object 



808
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 808

def self.compute(sess);                TinyNNCuda.tnn_compute(sess); end

.cross_entropy_grad(logits, targets, n_pred) ⇒ Object

cross_entropy_grad = (softmax(logits) - one_hot(targets)) / n_pred. Composable from TinyNNCuda.softmax + scale + add.



742
743
744
745
746
747
748
749
750
751
752
753
754
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 742

def self.cross_entropy_grad(logits, targets, n_pred)
  oh = Mat.new(logits.nrows, logits.ncols)
  i = 0
  while i < n_pred
    oh.flat[i * logits.ncols + targets[i]] = 1.0
    i = i + 1
  end
  sm = TinyNNCuda.softmax(logits)
  inv_n = 1.0 / n_pred.to_f
  sm_s = TinyNNCuda.scale(sm, inv_n)
  oh_s = TinyNNCuda.scale(oh, -inv_n)
  TinyNNCuda.add(sm_s, oh_s)
end

.download_matmul(sess, tensor, m, n) ⇒ Object 



844
845
846
847
848
849
850
851
852
853
854
855
856
857
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 844

def self.download_matmul(sess, tensor, m, n)
  TinyNNCuda.tnn_download(sess, tensor)
  out = Mat.new(m, n)
  i = 0
  while i < m
    j = 0
    while j < n
      out.flat[i * n + j] = TinyNNCuda.tnn_scratch_get(sess, j * m + i)
      j = j + 1
    end
    i = i + 1
  end
  out
end

.download_row_major(sess, tensor, rows, cols) ⇒ Object 



832
833
834
835
836
837
838
839
840
841
842
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 832

def self.download_row_major(sess, tensor, rows, cols)
  TinyNNCuda.tnn_download(sess, tensor)
  out = Mat.new(rows, cols)
  n = rows * cols
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  out
end

.embed_back(d_out, indices, vocab_size) ⇒ Object

Embedding scatter-add (backward).



675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 675

def self.embed_back(d_out, indices, vocab_size)
  n_idx = indices.length
  sess  = TinyNNCuda.tnn_session_new(1)
  td    = TinyNNCuda.tnn_input_2d_f32(sess, d_out.nrows, d_out.ncols)
  tidx  = TinyNNCuda.tnn_input_1d_i32(sess, n_idx)
  tshape = TinyNNCuda.tnn_input_2d_f32(sess, vocab_size, d_out.ncols)
  tout  = TinyNNCuda.tnn_get_rows_back(sess, td, tidx, tshape)
  TinyNNCuda.tnn_realize(sess, tout)
  nd = d_out.nrows * d_out.ncols
  i = 0
  while i < nd
    TinyNNCuda.tnn_scratch_set(sess, i, d_out.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, td)
  i = 0
  while i < n_idx
    TinyNNCuda.tnn_scratch_set_i32(sess, i, indices[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tidx)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tout)
  out = Mat.new(vocab_size, d_out.ncols)
  n = vocab_size * d_out.ncols
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.embed_lookup(table, indices) ⇒ Object

Embedding lookup: gather rows.



641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 641

def self.embed_lookup(table, indices)
  n_idx = indices.length
  sess  = TinyNNCuda.tnn_session_new(1)
  ttab  = TinyNNCuda.tnn_input_2d_f32(sess, table.nrows, table.ncols)
  tidx  = TinyNNCuda.tnn_input_1d_i32(sess, n_idx)
  tout  = TinyNNCuda.tnn_get_rows(sess, ttab, tidx)
  TinyNNCuda.tnn_realize(sess, tout)
  nt = table.nrows * table.ncols
  i = 0
  while i < nt
    TinyNNCuda.tnn_scratch_set(sess, i, table.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ttab)
  i = 0
  while i < n_idx
    TinyNNCuda.tnn_scratch_set_i32(sess, i, indices[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tidx)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tout)
  out = Mat.new(n_idx, table.ncols)
  n = n_idx * table.ncols
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.ffn_pipeline(h, w1, w2) ⇒ Object

gelu(h * w1) * w2 chained via the persistent CUDA engine.



514
515
516
517
518
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 514

def self.ffn_pipeline(h, w1, w2)
  pre    = TinyNNCuda.matmul(h, w1)
  hidden = TinyNNCuda.gelu(pre)
  TinyNNCuda.matmul(hidden, w2)
end

.gelu(a) ⇒ Object 



410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 410

def self.gelu(a)
  sess = TinyNNCuda.tnn_session_new(1)
  ta = TinyNNCuda.tnn_input_2d_f32(sess, a.nrows, a.ncols)
  tc = TinyNNCuda.tnn_gelu(sess, ta)
  TinyNNCuda.tnn_realize(sess, tc)
  n = a.nrows * a.ncols
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, a.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(a.nrows, a.ncols)
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.gelu_back(x, dh) ⇒ Object

GeLU backward (tanh approx) via custom CPU kernel. Mirrors TinyNN.gelu_back; same scratch-layout protocol.



716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 716

def self.gelu_back(x, dh)
  sess = TinyNNCuda.tnn_session_new(1)
  n = x.nrows * x.ncols
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, x.flat[i])
    i = i + 1
  end
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, n + i, dh.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_gelu_back_scratch(sess, n)
  out = Mat.new(x.nrows, x.ncols)
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, 2 * n + i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.matmul(a, b) ⇒ Object 



330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 330

def self.matmul(a, b)
  sess = TinyNNCuda.tnn_session_new(1)   # 1 = prefer CUDA

  ta   = TinyNNCuda.tnn_input_2d_f32(sess, a.nrows, a.ncols)
  tb_t = TinyNNCuda.tnn_input_2d_f32(sess, b.ncols, b.nrows)
  tc   = TinyNNCuda.tnn_matmul(sess, ta, tb_t)
  TinyNNCuda.tnn_realize(sess, tc)

  i = 0
  na = a.nrows * a.ncols
  while i < na
    TinyNNCuda.tnn_scratch_set(sess, i, a.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta)

  bc = b.ncols
  br = b.nrows
  i = 0
  while i < br
    j = 0
    while j < bc
      TinyNNCuda.tnn_scratch_set(sess, j * br + i, b.flat[i * bc + j])
      j = j + 1
    end
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tb_t)

  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)

  out = Mat.new(a.nrows, b.ncols)
  m = a.nrows
  n = b.ncols
  i = 0
  while i < m
    j = 0
    while j < n
      out.flat[i * n + j] = TinyNNCuda.tnn_scratch_get(sess, j * m + i)
      j = j + 1
    end
    i = i + 1
  end

  TinyNNCuda.tnn_session_free(sess)
  out
end

.matmul_t(a, b) ⇒ Object

a * b^T (matches Mat#matmul_t).



521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 521

def self.matmul_t(a, b)
  sess = TinyNNCuda.tnn_session_new(1)
  ta = TinyNNCuda.tnn_input_2d_f32(sess, a.nrows, a.ncols)
  tb = TinyNNCuda.tnn_input_2d_f32(sess, b.nrows, b.ncols)
  tc = TinyNNCuda.tnn_matmul(sess, ta, tb)
  TinyNNCuda.tnn_realize(sess, tc)
  na = a.nrows * a.ncols
  i = 0
  while i < na
    TinyNNCuda.tnn_scratch_set(sess, i, a.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta)
  nb = b.nrows * b.ncols
  i = 0
  while i < nb
    TinyNNCuda.tnn_scratch_set(sess, i, b.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tb)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(a.nrows, b.nrows)
  m = a.nrows
  n = b.nrows
  i = 0
  while i < m
    j = 0
    while j < n
      out.flat[i * n + j] = TinyNNCuda.tnn_scratch_get(sess, j * m + i)
      j = j + 1
    end
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.persistent_free(sess) ⇒ Object 



798
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 798

def self.persistent_free(sess);        TinyNNCuda.tnn_session_free(sess); end

.persistent_new(prefer_cuda) ⇒ Object

—– Persistent-session API (mirrors TinyNN’s; see lib/toy/ffi/tinynn.rb) —–



797
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 797

def self.persistent_new(prefer_cuda);  TinyNNCuda.tnn_session_new(prefer_cuda); end

.realize(sess, r) ⇒ Object 



807
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 807

def self.realize(sess, r);             TinyNNCuda.tnn_realize(sess, r); end

.rms_norm(x, gamma, eps) ⇒ Object 



434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 434

def self.rms_norm(x, gamma, eps)
  sess = TinyNNCuda.tnn_session_new(1)
  tx = TinyNNCuda.tnn_input_2d_f32(sess, x.nrows, x.ncols)
  tg = TinyNNCuda.tnn_input_2d_f32(sess, 1, x.ncols)
  tc = TinyNNCuda.tnn_rms_norm(sess, tx, tg, eps)
  TinyNNCuda.tnn_realize(sess, tc)
  nx = x.nrows * x.ncols
  i = 0
  while i < nx
    TinyNNCuda.tnn_scratch_set(sess, i, x.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tx)
  i = 0
  while i < x.ncols
    TinyNNCuda.tnn_scratch_set(sess, i, gamma[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tg)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(x.nrows, x.ncols)
  i = 0
  while i < nx
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.scale(a, s) ⇒ Object 



489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 489

def self.scale(a, s)
  sess = TinyNNCuda.tnn_session_new(1)
  ta = TinyNNCuda.tnn_input_2d_f32(sess, a.nrows, a.ncols)
  tc = TinyNNCuda.tnn_scale(sess, ta, s)
  TinyNNCuda.tnn_realize(sess, tc)
  n = a.nrows * a.ncols
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, a.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(a.nrows, a.ncols)
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.sgd_step(param, grad, lr) ⇒ Object

SGD: param_new = param - lr * grad. Composed.



710
711
712
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 710

def self.sgd_step(param, grad, lr)
  TinyNNCuda.add(param, TinyNNCuda.scale(grad, -lr))
end

.softmax(a) ⇒ Object 



465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 465

def self.softmax(a)
  sess = TinyNNCuda.tnn_session_new(1)
  ta = TinyNNCuda.tnn_input_2d_f32(sess, a.nrows, a.ncols)
  tc = TinyNNCuda.tnn_softmax(sess, ta)
  TinyNNCuda.tnn_realize(sess, tc)
  n = a.nrows * a.ncols
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, a.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(a.nrows, a.ncols)
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.softmax_back(a_softmax, dy) ⇒ Object

Per-row softmax backward.



609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 609

def self.softmax_back(a_softmax, dy)
  sess = TinyNNCuda.tnn_session_new(1)
  tdy = TinyNNCuda.tnn_input_2d_f32(sess, dy.nrows, dy.ncols)
  ta  = TinyNNCuda.tnn_input_2d_f32(sess, a_softmax.nrows, a_softmax.ncols)
  tc  = TinyNNCuda.tnn_softmax_back(sess, tdy, ta)
  TinyNNCuda.tnn_realize(sess, tc)
  n = dy.nrows * dy.ncols
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, dy.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tdy)
  i = 0
  while i < n
    TinyNNCuda.tnn_scratch_set(sess, i, a_softmax.flat[i])
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(a_softmax.nrows, a_softmax.ncols)
  i = 0
  while i < n
    out.flat[i] = TinyNNCuda.tnn_scratch_get(sess, i)
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.stage_transposed_and_upload(sess, target, b) ⇒ Object

Alias to match the CPU module’s name; used by feed_forward_ffi.



828
829
830
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 828

def self.stage_transposed_and_upload(sess, target, b)
  TinyNNCuda.upload_transposed(sess, target, b)
end

.t_matmul(a, b) ⇒ Object

a^T * b (matches Mat#t_matmul). Both uploaded transposed.



560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 560

def self.t_matmul(a, b)
  sess = TinyNNCuda.tnn_session_new(1)
  ta_t = TinyNNCuda.tnn_input_2d_f32(sess, a.ncols, a.nrows)
  tb_t = TinyNNCuda.tnn_input_2d_f32(sess, b.ncols, b.nrows)
  tc = TinyNNCuda.tnn_matmul(sess, ta_t, tb_t)
  TinyNNCuda.tnn_realize(sess, tc)
  ar = a.nrows
  ac = a.ncols
  i = 0
  while i < ar
    j = 0
    while j < ac
      TinyNNCuda.tnn_scratch_set(sess, j * ar + i, a.flat[i * ac + j])
      j = j + 1
    end
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, ta_t)
  br = b.nrows
  bc = b.ncols
  i = 0
  while i < br
    j = 0
    while j < bc
      TinyNNCuda.tnn_scratch_set(sess, j * br + i, b.flat[i * bc + j])
      j = j + 1
    end
    i = i + 1
  end
  TinyNNCuda.tnn_upload(sess, tb_t)
  TinyNNCuda.tnn_compute(sess)
  TinyNNCuda.tnn_download(sess, tc)
  out = Mat.new(a.ncols, b.ncols)
  m = a.ncols
  n = b.ncols
  i = 0
  while i < m
    j = 0
    while j < n
      out.flat[i * n + j] = TinyNNCuda.tnn_scratch_get(sess, j * m + i)
      j = j + 1
    end
    i = i + 1
  end
  TinyNNCuda.tnn_session_free(sess)
  out
end

.upload_int_array(sess, tensor, indices) ⇒ Object

Upload an Array<Int> to a 1D int32 tensor in one FFI call.



815
816
817
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 815

def self.upload_int_array(sess, tensor, indices)
  TinyNNCuda.tnn_upload_from_int_array(sess, tensor, indices, indices.length)
end

.upload_row_major(sess, tensor, mat) ⇒ Object 



810
811
812
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 810

def self.upload_row_major(sess, tensor, mat)
  TinyNNCuda.tnn_upload_from_float_array(sess, tensor, mat.flat, mat.nrows * mat.ncols)
end

.upload_transposed(sess, tensor, mat) ⇒ Object 



819
820
821
822
823
824
825
 
# File 'lib/toy/ffi/tinynn_cuda.rb', line 819

def self.upload_transposed(sess, tensor, mat)
  # Chunked in C — works for tensors larger than the 16 MiB scratch
  # buffer (Qwen2.5-0.5B's ffn_* are 17 MB; the per-element path
  # silently truncated at the 4M float boundary).
  TinyNNCuda.tnn_upload_transposed_f64(sess, tensor, mat.flat,
                                        mat.nrows, mat.ncols)
end