Update.

diff --git a/fridge b/fridge
new file mode 100644 (file)
index 0000000..5dd85dd
--- /dev/null
+++ b/fridge
@@ -0,0 +1,76 @@
+
+######################################################################
+
+2024 Jan 07 21:37:48 (from mygpt.py)
+
+
+# This is one order of magnitude more complicated than I expected, not
+# elegant, slow, hopefully not buggy
+
+
+def flash_back_time_src(N, H, t0, t1, CL, CH, proba, device):
+    # starting flash backs
+    fb_start = (torch.rand(N, CH, t1 - t0, device=device) <= proba).long()
+    fb_start[:, :, -CL:] = 0
+    fb_start[:, :, :CL] = 0
+
+    # Remove series longer than CL
+    fb_body = fb_start.clone()
+    fb_body[:, :, CL + 1 :] -= fb_start[:, :, : -(CL + 1)]
+    fb_body = fb_body.cumsum(dim=2)
+    fb_start = fb_start * (fb_body == 1)
+
+    # Set a origin source time (starting time of the chunck to copy
+    # here) We set it as the current time minus a multiple of CL to be
+    # consistent with the "rolling" caterpillar
+    t = torch.arange(fb_start.size(2), device=fb_start.device)[None, None, :]
+    src_time = fb_start * (
+        t
+        - CL
+        * (
+            1
+            + (
+                torch.rand(fb_start.size(), device=fb_start.device) * (t // CL - 1)
+            ).long()
+        )
+    )
+    src_time[:, :, CL:] -= src_time.clone()[:, :, :-CL]
+    src_time = src_time.cumsum(dim=2)
+
+    src_head = fb_start * torch.randint(H, fb_start.size(), device=fb_start.device)
+    src_head[:, :, CL:] -= src_head.clone()[:, :, :-CL]
+    src_head = src_head.cumsum(dim=2)
+
+    # combine
+    src_delta = fb_start.clone()
+    src_delta[:, :, CL:] -= fb_start[:, :, :-CL]
+    src_delta = src_delta.cumsum(dim=2)
+    src_delta[:, :, CL:] -= CL * fb_start[:, :, :-CL]
+    src_time += src_delta.cumsum(dim=2) - 1
+
+    return src_time, src_head
+
+
+def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
+    N, H, CH = V.size(0), V.size(1), rec_V.size(1)
+
+    fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device)
+
+    fbt_V = fbt[:, :, :, None]
+    fbh_V = fbh[:, :, :, None]
+    t = fbt_V.clamp(min=0)
+    n = torch.arange(V.size(0), device=V.device)[:, None, None, None]
+    d = torch.arange(V.size(3), device=V.device)[None, None, None, :]
+    q = V[:, :, t0:t1][n, fbh_V, t, d]
+    rec_V[:, :, t0:t1] = q * (fbt_V >= 0) + rec_V[:, :, t0:t1] * (fbt_V < 0)
+
+    fbt_K = fbt[:, :, :, None]
+    fbh_K = fbh[:, :, :, None]
+    t = fbt_K.clamp(min=0)
+    n = torch.arange(K.size(0), device=K.device)[:, None, None, None]
+    d = torch.arange(K.size(3), device=K.device)[None, None, None, :]
+    q = K[:, :, t0:t1][n, fbh_K, t, d]
+    rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0)
+
+
+######################################################################

diff --git a/mygpt.py b/mygpt.py
index b5ac916..eda8685 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -457,77 +457,6 @@ def moving_window(x, dim, win_dim, win_size):
 
 ##############################
 
-# This is one order of magnitude more complicated than I expected, not
-# elegant, slow, hopefully not buggy
-
-
-def flash_back_time_src(N, H, t0, t1, CL, CH, proba, device):
-    # starting flash backs
-    fb_start = (torch.rand(N, CH, t1 - t0, device=device) <= proba).long()
-    fb_start[:, :, -CL:] = 0
-    fb_start[:, :, :CL] = 0
-
-    # Remove series longer than CL
-    fb_body = fb_start.clone()
-    fb_body[:, :, CL + 1 :] -= fb_start[:, :, : -(CL + 1)]
-    fb_body = fb_body.cumsum(dim=2)
-    fb_start = fb_start * (fb_body == 1)
-
-    # Set a origin source time (starting time of the chunck to copy
-    # here) We set it as the current time minus a multiple of CL to be
-    # consistent with the "rolling" caterpillar
-    t = torch.arange(fb_start.size(2), device=fb_start.device)[None, None, :]
-    src_time = fb_start * (
-        t
-        - CL
-        * (
-            1
-            + (
-                torch.rand(fb_start.size(), device=fb_start.device) * (t // CL - 1)
-            ).long()
-        )
-    )
-    src_time[:, :, CL:] -= src_time.clone()[:, :, :-CL]
-    src_time = src_time.cumsum(dim=2)
-
-    src_head = fb_start * torch.randint(H, fb_start.size(), device=fb_start.device)
-    src_head[:, :, CL:] -= src_head.clone()[:, :, :-CL]
-    src_head = src_head.cumsum(dim=2)
-
-    # combine
-    src_delta = fb_start.clone()
-    src_delta[:, :, CL:] -= fb_start[:, :, :-CL]
-    src_delta = src_delta.cumsum(dim=2)
-    src_delta[:, :, CL:] -= CL * fb_start[:, :, :-CL]
-    src_time += src_delta.cumsum(dim=2) - 1
-
-    return src_time, src_head
-
-
-def insert_flash_back(rec_V, V, rec_K, K, t0, t1, CL, proba):
-    N, H, CH = V.size(0), V.size(1), rec_V.size(1)
-
-    fbt, fbh = flash_back_time_src(N, H, t0, t1, CL, CH, proba, rec_V.device)
-
-    fbt_V = fbt[:, :, :, None]
-    fbh_V = fbh[:, :, :, None]
-    t = fbt_V.clamp(min=0)
-    n = torch.arange(V.size(0), device=V.device)[:, None, None, None]
-    d = torch.arange(V.size(3), device=V.device)[None, None, None, :]
-    q = V[:, :, t0:t1][n, fbh_V, t, d]
-    rec_V[:, :, t0:t1] = q * (fbt_V >= 0) + rec_V[:, :, t0:t1] * (fbt_V < 0)
-
-    fbt_K = fbt[:, :, :, None]
-    fbh_K = fbh[:, :, :, None]
-    t = fbt_K.clamp(min=0)
-    n = torch.arange(K.size(0), device=K.device)[:, None, None, None]
-    d = torch.arange(K.size(3), device=K.device)[None, None, None, :]
-    q = K[:, :, t0:t1][n, fbh_K, t, d]
-    rec_K[:, :, t0:t1] = q * (fbt_K >= 0) + rec_K[:, :, t0:t1] * (fbt_K < 0)
-
-
-######################################################################
-
 
 class Caterpillar(nn.Module):
     def __init__(

diff --git a/pscan.py b/pscan.py
index a14f009..0bb0d14 100755 (executable)
--- a/pscan.py
+++ b/pscan.py
@@ -122,6 +122,22 @@ def naive_pscan(A, X, Y_init):
 if __name__ == "__main__":
     import time, sys
 
+    ######################################################################
+
+    N, T, D = 16, 4096, 32
+
+    for r in range(timing.size(0)):
+        A = 0.9 + 0.1 * torch.rand(N, T, dtype=torch.float64).requires_grad_()
+        X = torch.randn(N, T, D, dtype=torch.float64).requires_grad_()
+        Y_init = torch.randn(N, D, dtype=torch.float64).requires_grad_()
+
+        start_time = time.perf_counter()
+        for _ in range(1000):
+            Y = pscan(A, X, Y_init)
+        duration = time.perf_counter() - start_time
+
+    ######################################################################
+
     # A = torch.rand(17, 12, 3)
     # X = torch.rand(17, 12, 3, 11)
     # Y_init = torch.rand(17, 3, 11)
@@ -130,11 +146,12 @@ if __name__ == "__main__":
     # exit(0)
 
     err = 0
+    timing = torch.empty(10)
 
-    for _ in range(100):
-        N, T, D = 2, 112, 3
+    for r in range(timing.size(0)):
+        N, T, D = 2, 1120, 3
 
-        T = torch.randint(10, (1,)).item() + 1
+        # T = torch.randint(10, (1,)).item() + 1
 
         A = 0.9 + 0.1 * torch.rand(N, T, dtype=torch.float64).requires_grad_()
         X = torch.randn(N, T, D, dtype=torch.float64).requires_grad_()
@@ -161,7 +178,9 @@ if __name__ == "__main__":
         for _ in range(1000):
             Y = pscan(A, X, Y_init)
         duration = time.perf_counter() - start_time
+
         print(f"duration {duration}")
+        timing[r] = duration
 
         s = Y.sum()
 
@@ -181,4 +200,4 @@ if __name__ == "__main__":
 
         # print((Y - torch.cat([Y1, Y2], dim=1)).abs().max())
 
-    print(f"{err=}")
+    print(f"err={err:.2e} duration={timing.mean():.2e} (+/- {timing.std():.2e})")