tide-GPR 0.0.9__py3-none-manylinux_2_28_x86_64.whl

tide/csrc/staggered_grid.h

@@ -0,0 +1,175 @@
+#ifndef STAGGERED_GRID_H
+#define STAGGERED_GRID_H
+
+// FD_PAD is half the stencil width
+#if TIDE_STENCIL == 2
+#define FD_PAD 1
+#elif TIDE_STENCIL == 4
+#define FD_PAD 2
+#elif TIDE_STENCIL == 6
+#define FD_PAD 3
+#elif TIDE_STENCIL == 8
+#define FD_PAD 4
+#else
+#error "TIDE_STENCIL must be 2, 4, 6, or 8"
+#endif
+
+#if TIDE_STENCIL == 2
+// 2nd order accuracy
+#define DIFFY1(F) ((F(0, 0) - F(-1, 0)) * rdy)
+#define DIFFX1(F) ((F(0, 0) - F(0, -1)) * rdx)
+#define DIFFYH1(F) ((F(1, 0) - F(0, 0)) * rdy)
+#define DIFFXH1(F) ((F(0, 1) - F(0, 0)) * rdx)
+
+#elif TIDE_STENCIL == 4
+// 4th order accuracy
+#define DIFFY1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 0) - F(-1, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(1, 0) - F(-2, 0))) * rdy
+
+#define DIFFX1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 0) - F(0, -1)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(0, 1) - F(0, -2))) * rdx
+
+#define DIFFYH1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(1, 0) - F(0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(2, 0) - F(-1, 0))) * rdy
+
+#define DIFFXH1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 1) - F(0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(0, 2) - F(0, -1))) * rdx
+
+#elif TIDE_STENCIL == 6
+// 6th order accuracy
+#define DIFFY1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 0) - F(-1, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(1, 0) - F(-2, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(2, 0) - F(-3, 0))) * rdy
+
+#define DIFFX1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 0) - F(0, -1)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(0, 1) - F(0, -2)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(0, 2) - F(0, -3))) * rdx
+
+#define DIFFYH1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(1, 0) - F(0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(2, 0) - F(-1, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(3, 0) - F(-2, 0))) * rdy
+
+#define DIFFXH1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 1) - F(0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(0, 2) - F(0, -1)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(0, 3) - F(0, -2))) * rdx
+
+#elif TIDE_STENCIL == 8
+// 8th order accuracy
+#define DIFFY1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 0) - F(-1, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(1, 0) - F(-2, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(2, 0) - F(-3, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(3, 0) - F(-4, 0))) * rdy
+
+#define DIFFX1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 0) - F(0, -1)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(0, 1) - F(0, -2)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(0, 2) - F(0, -3)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(0, 3) - F(0, -4))) * rdx
+
+#define DIFFYH1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(1, 0) - F(0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(2, 0) - F(-1, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(3, 0) - F(-2, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(4, 0) - F(-3, 0))) * rdy
+
+#define DIFFXH1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 1) - F(0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(0, 2) - F(0, -1)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(0, 3) - F(0, -2)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(0, 4) - F(0, -3))) * rdx
+
+#endif
+
+/*
+ * Adjoint derivative operators for backward pass
+ * These compute the transpose of the forward derivative operators
+ */
+
+#if TIDE_STENCIL == 2
+#define DIFFY1_ADJ(C, F) \
+  ((C(0, 0) * F(0, 0) - C(1, 0) * F(1, 0)) * rdy)
+
+#define DIFFX1_ADJ(C, F) \
+  ((C(0, 0) * F(0, 0) - C(0, 1) * F(0, 1)) * rdx)
+
+#define DIFFYH1_ADJ(C, F) \
+  ((C(-1, 0) * F(-1, 0) - C(0, 0) * F(0, 0)) * rdy)
+
+#define DIFFXH1_ADJ(C, F) \
+  ((C(0, -1) * F(0, -1) - C(0, 0) * F(0, 0)) * rdx)
+
+#elif TIDE_STENCIL == 4
+#define DIFFY1_ADJ(C, F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (C(0, 0) * F(0, 0) - C(1, 0) * F(1, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (C(-1, 0) * F(-1, 0) - C(2, 0) * F(2, 0))) * rdy
+
+#define DIFFX1_ADJ(C, F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (C(0, 0) * F(0, 0) - C(0, 1) * F(0, 1)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (C(0, -1) * F(0, -1) - C(0, 2) * F(0, 2))) * rdx
+
+#define DIFFYH1_ADJ(C, F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (C(-1, 0) * F(-1, 0) - C(0, 0) * F(0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (C(-2, 0) * F(-2, 0) - C(1, 0) * F(1, 0))) * rdy
+
+#define DIFFXH1_ADJ(C, F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (C(0, -1) * F(0, -1) - C(0, 0) * F(0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (C(0, -2) * F(0, -2) - C(0, 1) * F(0, 1))) * rdx
+
+#elif TIDE_STENCIL == 6
+#define DIFFY1_ADJ(C, F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (C(0, 0) * F(0, 0) - C(1, 0) * F(1, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (C(-1, 0) * F(-1, 0) - C(2, 0) * F(2, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (C(-2, 0) * F(-2, 0) - C(3, 0) * F(3, 0))) * rdy
+
+#define DIFFX1_ADJ(C, F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (C(0, 0) * F(0, 0) - C(0, 1) * F(0, 1)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (C(0, -1) * F(0, -1) - C(0, 2) * F(0, 2)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (C(0, -2) * F(0, -2) - C(0, 3) * F(0, 3))) * rdx
+
+#define DIFFYH1_ADJ(C, F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (C(-1, 0) * F(-1, 0) - C(0, 0) * F(0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (C(-2, 0) * F(-2, 0) - C(1, 0) * F(1, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (C(-3, 0) * F(-3, 0) - C(2, 0) * F(2, 0))) * rdy
+
+#define DIFFXH1_ADJ(C, F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (C(0, -1) * F(0, -1) - C(0, 0) * F(0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (C(0, -2) * F(0, -2) - C(0, 1) * F(0, 1)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (C(0, -3) * F(0, -3) - C(0, 2) * F(0, 2))) * rdx
+
+#elif TIDE_STENCIL == 8
+#define DIFFY1_ADJ(C, F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (C(0, 0) * F(0, 0) - C(1, 0) * F(1, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (C(-1, 0) * F(-1, 0) - C(2, 0) * F(2, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (C(-2, 0) * F(-2, 0) - C(3, 0) * F(3, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (C(-3, 0) * F(-3, 0) - C(4, 0) * F(4, 0))) * rdy
+
+#define DIFFX1_ADJ(C, F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (C(0, 0) * F(0, 0) - C(0, 1) * F(0, 1)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (C(0, -1) * F(0, -1) - C(0, 2) * F(0, 2)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (C(0, -2) * F(0, -2) - C(0, 3) * F(0, 3)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (C(0, -3) * F(0, -3) - C(0, 4) * F(0, 4))) * rdx
+
+#define DIFFYH1_ADJ(C, F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (C(-1, 0) * F(-1, 0) - C(0, 0) * F(0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (C(-2, 0) * F(-2, 0) - C(1, 0) * F(1, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (C(-3, 0) * F(-3, 0) - C(2, 0) * F(2, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (C(-4, 0) * F(-4, 0) - C(3, 0) * F(3, 0))) * rdy
+
+#define DIFFXH1_ADJ(C, F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (C(0, -1) * F(0, -1) - C(0, 0) * F(0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (C(0, -2) * F(0, -2) - C(0, 1) * F(0, 1)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (C(0, -3) * F(0, -3) - C(0, 2) * F(0, 2)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (C(0, -4) * F(0, -4) - C(0, 3) * F(0, 3))) * rdx
+
+#endif
+
+#endif  // STAGGERED_GRID_H

tide/csrc/staggered_grid_3d.h

@@ -0,0 +1,124 @@
+#ifndef STAGGERED_GRID_3D_H
+#define STAGGERED_GRID_3D_H
+
+// FD_PAD is half the stencil width
+#if TIDE_STENCIL == 2
+#define FD_PAD 1
+#elif TIDE_STENCIL == 4
+#define FD_PAD 2
+#elif TIDE_STENCIL == 6
+#define FD_PAD 3
+#elif TIDE_STENCIL == 8
+#define FD_PAD 4
+#else
+#error "TIDE_STENCIL must be 2, 4, 6, or 8"
+#endif
+
+#if TIDE_STENCIL == 2
+// 2nd order accuracy
+#define DIFFZ1(F) ((F(0, 0, 0) - F(-1, 0, 0)) * rdz)
+#define DIFFY1(F) ((F(0, 0, 0) - F(0, -1, 0)) * rdy)
+#define DIFFX1(F) ((F(0, 0, 0) - F(0, 0, -1)) * rdx)
+#define DIFFZH1(F) ((F(1, 0, 0) - F(0, 0, 0)) * rdz)
+#define DIFFYH1(F) ((F(0, 1, 0) - F(0, 0, 0)) * rdy)
+#define DIFFXH1(F) ((F(0, 0, 1) - F(0, 0, 0)) * rdx)
+
+#elif TIDE_STENCIL == 4
+// 4th order accuracy
+#define DIFFZ1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 0, 0) - F(-1, 0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(1, 0, 0) - F(-2, 0, 0))) * rdz
+
+#define DIFFY1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 0, 0) - F(0, -1, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(0, 1, 0) - F(0, -2, 0))) * rdy
+
+#define DIFFX1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 0, 0) - F(0, 0, -1)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(0, 0, 1) - F(0, 0, -2))) * rdx
+
+#define DIFFZH1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(1, 0, 0) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(2, 0, 0) - F(-1, 0, 0))) * rdz
+
+#define DIFFYH1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 1, 0) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(0, 2, 0) - F(0, -1, 0))) * rdy
+
+#define DIFFXH1(F) \
+  ((TIDE_DTYPE)(9.0/8.0) * (F(0, 0, 1) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-1.0/24.0) * (F(0, 0, 2) - F(0, 0, -1))) * rdx
+
+#elif TIDE_STENCIL == 6
+// 6th order accuracy
+#define DIFFZ1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 0, 0) - F(-1, 0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(1, 0, 0) - F(-2, 0, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(2, 0, 0) - F(-3, 0, 0))) * rdz
+
+#define DIFFY1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 0, 0) - F(0, -1, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(0, 1, 0) - F(0, -2, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(0, 2, 0) - F(0, -3, 0))) * rdy
+
+#define DIFFX1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 0, 0) - F(0, 0, -1)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(0, 0, 1) - F(0, 0, -2)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(0, 0, 2) - F(0, 0, -3))) * rdx
+
+#define DIFFZH1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(1, 0, 0) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(2, 0, 0) - F(-1, 0, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(3, 0, 0) - F(-2, 0, 0))) * rdz
+
+#define DIFFYH1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 1, 0) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(0, 2, 0) - F(0, -1, 0)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(0, 3, 0) - F(0, -2, 0))) * rdy
+
+#define DIFFXH1(F) \
+  ((TIDE_DTYPE)(75.0/64.0) * (F(0, 0, 1) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-25.0/384.0) * (F(0, 0, 2) - F(0, 0, -1)) \
+   + (TIDE_DTYPE)(3.0/640.0) * (F(0, 0, 3) - F(0, 0, -2))) * rdx
+
+#elif TIDE_STENCIL == 8
+// 8th order accuracy
+#define DIFFZ1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 0, 0) - F(-1, 0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(1, 0, 0) - F(-2, 0, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(2, 0, 0) - F(-3, 0, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(3, 0, 0) - F(-4, 0, 0))) * rdz
+
+#define DIFFY1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 0, 0) - F(0, -1, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(0, 1, 0) - F(0, -2, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(0, 2, 0) - F(0, -3, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(0, 3, 0) - F(0, -4, 0))) * rdy
+
+#define DIFFX1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 0, 0) - F(0, 0, -1)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(0, 0, 1) - F(0, 0, -2)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(0, 0, 2) - F(0, 0, -3)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(0, 0, 3) - F(0, 0, -4))) * rdx
+
+#define DIFFZH1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(1, 0, 0) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(2, 0, 0) - F(-1, 0, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(3, 0, 0) - F(-2, 0, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(4, 0, 0) - F(-3, 0, 0))) * rdz
+
+#define DIFFYH1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 1, 0) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(0, 2, 0) - F(0, -1, 0)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(0, 3, 0) - F(0, -2, 0)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(0, 4, 0) - F(0, -3, 0))) * rdy
+
+#define DIFFXH1(F) \
+  ((TIDE_DTYPE)(1225.0/1024.0) * (F(0, 0, 1) - F(0, 0, 0)) \
+   + (TIDE_DTYPE)(-245.0/3072.0) * (F(0, 0, 2) - F(0, 0, -1)) \
+   + (TIDE_DTYPE)(49.0/5120.0) * (F(0, 0, 3) - F(0, 0, -2)) \
+   + (TIDE_DTYPE)(-5.0/7168.0) * (F(0, 0, 4) - F(0, 0, -3))) * rdx
+
+#endif
+
+#endif  // STAGGERED_GRID_3D_H

tide/csrc/storage_utils.c

@@ -0,0 +1,78 @@
+#include "storage_utils.h"
+
+#include <errno.h>
+#include <string.h>
+
+static bool read_exact(FILE* fp, void* dst, size_t nbytes) {
+  size_t total = 0;
+  while (total < nbytes) {
+    size_t n = fread((char*)dst + total, 1, nbytes - total, fp);
+    if (n == 0) {
+      return false;
+    }
+    total += n;
+  }
+  return true;
+}
+
+static bool write_exact(FILE* fp, const void* src, size_t nbytes) {
+  size_t total = 0;
+  while (total < nbytes) {
+    size_t n = fwrite((const char*)src + total, 1, nbytes - total, fp);
+    if (n == 0) {
+      return false;
+    }
+    total += n;
+  }
+  return true;
+}
+
+static void report_io_error(const char* op, int64_t step_idx) {
+  if (errno != 0) {
+    fprintf(stderr, "storage_utils: %s failed at step %lld: %s\n",
+            op, (long long)step_idx, strerror(errno));
+  } else {
+    fprintf(stderr, "storage_utils: %s failed at step %lld\n",
+            op, (long long)step_idx);
+  }
+}
+
+void storage_save_snapshot_cpu(void* store_1, FILE* fp, int64_t storage_mode,
+                               int64_t step_idx, size_t step_bytes_uncomp) {
+  if (storage_mode == STORAGE_NONE) return;
+  if (storage_mode == STORAGE_DISK) {
+    int64_t offset = step_idx * (int64_t)step_bytes_uncomp;
+    errno = 0;
+    if (fseek(fp, offset, SEEK_SET) != 0) {
+      report_io_error("fseek(write)", step_idx);
+      return;
+    }
+    if (!write_exact(fp, store_1, step_bytes_uncomp)) {
+      report_io_error("fwrite", step_idx);
+    }
+  }
+}
+
+void storage_load_snapshot_cpu(void* store_1, FILE* fp, int64_t storage_mode,
+                               int64_t step_idx, size_t step_bytes_uncomp) {
+  if (storage_mode == STORAGE_NONE) return;
+  if (storage_mode == STORAGE_DISK) {
+    int64_t offset = step_idx * (int64_t)step_bytes_uncomp;
+    errno = 0;
+    if (fseek(fp, offset, SEEK_SET) != 0) {
+      report_io_error("fseek(read)", step_idx);
+      memset(store_1, 0, step_bytes_uncomp);
+      return;
+    }
+    if (!read_exact(fp, store_1, step_bytes_uncomp)) {
+      if (feof(fp)) {
+        fprintf(stderr, "storage_utils: unexpected EOF at step %lld\n",
+                (long long)step_idx);
+      } else {
+        report_io_error("fread", step_idx);
+      }
+      clearerr(fp);
+      memset(store_1, 0, step_bytes_uncomp);
+    }
+  }
+}

tide/csrc/storage_utils.cu

@@ -0,0 +1,135 @@
+#include <cuda_runtime.h>
+#include <errno.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "storage_utils.h"
+
+static bool read_exact(FILE* fp, void* dst, size_t nbytes) {
+  size_t total = 0;
+  while (total < nbytes) {
+    size_t n = fread((char*)dst + total, 1, nbytes - total, fp);
+    if (n == 0) {
+      return false;
+    }
+    total += n;
+  }
+  return true;
+}
+
+static bool write_exact(FILE* fp, const void* src, size_t nbytes) {
+  size_t total = 0;
+  while (total < nbytes) {
+    size_t n = fwrite((const char*)src + total, 1, nbytes - total, fp);
+    if (n == 0) {
+      return false;
+    }
+    total += n;
+  }
+  return true;
+}
+
+static void report_io_error(const char* op, int64_t step_idx) {
+  if (errno != 0) {
+    fprintf(stderr, "storage_utils: %s failed at step %lld: %s\n",
+            op, (long long)step_idx, strerror(errno));
+  } else {
+    fprintf(stderr, "storage_utils: %s failed at step %lld\n",
+            op, (long long)step_idx);
+  }
+}
+
+static void report_cuda_error(const char* op, cudaError_t err) {
+  fprintf(stderr, "storage_utils: %s failed: %s\n",
+          op, cudaGetErrorString(err));
+}
+
+extern "C" {
+
+void storage_save_snapshot_gpu(
+    void* store_1, void* store_3, FILE* fp, int64_t storage_mode,
+    int64_t step_idx, size_t shot_bytes_uncomp, size_t n_shots) {
+  if (storage_mode == STORAGE_NONE) return;
+  size_t bytes_to_store = shot_bytes_uncomp * n_shots;
+
+  if (storage_mode == STORAGE_CPU || storage_mode == STORAGE_DISK) {
+    if (storage_mode == STORAGE_DISK) {
+      // Disk mode needs host-visible data immediately for fwrite.
+      cudaError_t err =
+          cudaMemcpy(store_3, store_1, bytes_to_store, cudaMemcpyDeviceToHost);
+      if (err != cudaSuccess) {
+        report_cuda_error("cudaMemcpy(D2H)", err);
+        return;
+      }
+    } else {
+      // CPU mode: avoid blocking the host thread on every step.
+      // Copies are enqueued on the current (default) stream and will be ordered
+      // with subsequent CUDA work in the same stream.
+      cudaError_t err = cudaMemcpyAsync(
+          store_3, store_1, bytes_to_store, cudaMemcpyDeviceToHost, 0);
+      if (err != cudaSuccess) {
+        report_cuda_error("cudaMemcpyAsync(D2H)", err);
+        return;
+      }
+    }
+  }
+  if (storage_mode == STORAGE_DISK) {
+    int64_t offset = step_idx * (int64_t)bytes_to_store;
+    errno = 0;
+    if (fseek(fp, offset, SEEK_SET) != 0) {
+      report_io_error("fseek(write)", step_idx);
+      return;
+    }
+    if (!write_exact(fp, store_3, bytes_to_store)) {
+      report_io_error("fwrite", step_idx);
+    }
+  }
+}
+
+void storage_load_snapshot_gpu(void* store_1, void* store_3, FILE* fp,
+                               int64_t storage_mode, int64_t step_idx,
+                               size_t shot_bytes_uncomp, size_t n_shots) {
+  if (storage_mode == STORAGE_NONE) return;
+  size_t bytes_to_load = shot_bytes_uncomp * n_shots;
+
+  if (storage_mode == STORAGE_DISK) {
+    int64_t offset = step_idx * (int64_t)bytes_to_load;
+    errno = 0;
+    if (fseek(fp, offset, SEEK_SET) != 0) {
+      report_io_error("fseek(read)", step_idx);
+      memset(store_3, 0, bytes_to_load);
+    } else if (!read_exact(fp, store_3, bytes_to_load)) {
+      if (feof(fp)) {
+        fprintf(stderr, "storage_utils: unexpected EOF at step %lld\n",
+                (long long)step_idx);
+      } else {
+        report_io_error("fread", step_idx);
+      }
+      clearerr(fp);
+      memset(store_3, 0, bytes_to_load);
+    }
+  }
+
+  if (storage_mode == STORAGE_CPU || storage_mode == STORAGE_DISK) {
+    if (storage_mode == STORAGE_DISK) {
+      // Disk mode reuses the same pinned host buffer (store_3) for many steps.
+      // Use a synchronous copy to avoid the host overwriting store_3 (fread)
+      // before the device copy has consumed it.
+      cudaError_t err =
+          cudaMemcpy(store_1, store_3, bytes_to_load, cudaMemcpyHostToDevice);
+      if (err != cudaSuccess) {
+        report_cuda_error("cudaMemcpy(H2D)", err);
+        return;
+      }
+    } else {
+      cudaError_t err = cudaMemcpyAsync(
+          store_1, store_3, bytes_to_load, cudaMemcpyHostToDevice, 0);
+      if (err != cudaSuccess) {
+        report_cuda_error("cudaMemcpyAsync(H2D)", err);
+        return;
+      }
+    }
+  }
+}
+}

tide/csrc/storage_utils.h

@@ -0,0 +1,36 @@
+#ifndef STORAGE_UTILS_H
+#define STORAGE_UTILS_H
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+
+#define STORAGE_DEVICE 0
+#define STORAGE_CPU 1
+#define STORAGE_DISK 2
+#define STORAGE_NONE 3
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void storage_save_snapshot_cpu(void* store_1, FILE* fp, int64_t storage_mode,
+                               int64_t step_idx, size_t step_bytes_uncomp);
+
+void storage_save_snapshot_gpu(void* store_1, void* store_3, FILE* fp,
+                               int64_t storage_mode, int64_t step_idx,
+                               size_t shot_bytes_uncomp, size_t n_shots);
+
+void storage_load_snapshot_cpu(void* store_1, FILE* fp, int64_t storage_mode,
+                               int64_t step_idx, size_t step_bytes_uncomp);
+
+void storage_load_snapshot_gpu(void* store_1, void* store_3, FILE* fp,
+                               int64_t storage_mode, int64_t step_idx,
+                               size_t shot_bytes_uncomp, size_t n_shots);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

tide/grid_utils.py

@@ -0,0 +1,31 @@
+"""Grid-related helpers for padding and boundary bookkeeping."""
+
+from typing import Sequence, Union
+
+
+def _normalize_grid_spacing_2d(
+    grid_spacing: Union[float, Sequence[float]],
+) -> list[float]:
+    """Normalize 2D grid spacing to [dy, dx]."""
+    if isinstance(grid_spacing, (int, float)):
+        return [float(grid_spacing), float(grid_spacing)]
+    return list(grid_spacing)
+
+
+def _normalize_pml_width_2d(
+    pml_width: Union[int, Sequence[int]],
+) -> list[int]:
+    """Normalize 2D PML width to [top, bottom, left, right]."""
+    if isinstance(pml_width, int):
+        return [pml_width] * 4
+    pml_width_list = list(pml_width)
+    if len(pml_width_list) == 1:
+        return pml_width_list * 4
+    if len(pml_width_list) == 2:
+        return [
+            pml_width_list[0],
+            pml_width_list[0],
+            pml_width_list[1],
+            pml_width_list[1],
+        ]
+    return pml_width_list