Using `ggml_cont`

ggml_cont_permute.cpp

#include "ggml-cpu.h"
#include "ggml.h"
#include <iostream>

void print_tensor(struct ggml_tensor *tensor) {
  for (size_t i = 0; i < tensor->ne[3]; i++) {
    for (size_t k = 0; k < tensor->ne[2]; k++) {
      for (size_t j = 0; j < tensor->ne[1]; j++) {
        for (size_t l = 0; l < tensor->ne[0]; l++) {
          std::cout << ggml_get_i32_nd(tensor, l, j, k, i) << " ";
        }
        std::cout << "\n";
      }
      std::cout << "\n";
    }
    std::cout << "\n";
  }
}

int main (int argc, char *argv[]) {
  struct ggml_init_params params = {
    1024 * ggml_tensor_overhead(),
    nullptr,
    false
  };
  struct ggml_context *ctx = ggml_init(params);
  // Creates an array with values 0, 1, 2, ..., 15
  const int N = 16;
  int values[N] = { 0 };
  for (int i = 0; i < N; i++)
    values[i] = i;
  struct ggml_tensor *tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
  for (int i = 0; i < N; i++)
    ggml_set_i32_1d(tensor, i, values[i]);
  struct ggml_tensor *t = ggml_reshape_4d(ctx, tensor, 2, 2, 4, 1);
  std::cout << "Original tensor\n--------------\n";
  print_tensor(t);
  // 0 -> 1, 1 -> 2, 2 -> 0, 3 -> 3
  struct ggml_tensor *permuted_t = ggml_permute(ctx, t, 1, 2, 0, 3);
  std::cout << "Permuted tensor\n--------------\n";
  std::cout << "New Shape: " << permuted_t->ne[0] << " x " << permuted_t->ne[1] << " x " << permuted_t->ne[2] << " x " << permuted_t->ne[3] << "\n";
  GGML_ASSERT(permuted_t->ne[0] == 4);
  GGML_ASSERT(permuted_t->ne[1] == 2);
  GGML_ASSERT(permuted_t->ne[2] == 2);
  GGML_ASSERT(permuted_t->ne[3] == 1);
  print_tensor(permuted_t);
  struct ggml_tensor *cont_permuted_t = ggml_cont(ctx, permuted_t);
  struct ggml_tensor *a = ggml_reshape_2d(ctx, cont_permuted_t, 2, 8);
  GGML_ASSERT(a->ne[0] == 2); 
  GGML_ASSERT(a->ne[1] == 8);
  GGML_ASSERT(a->ne[2] == 1);
  GGML_ASSERT(a->ne[3] == 1);
  std::cout << "All zeros since we haven't computed anything\n";
  print_tensor(a); // NOTE: Should get all zeros!
  // Build computational graph
  struct ggml_cgraph *gf = ggml_new_graph(ctx);
  ggml_build_forward_expand(gf, a);
  ggml_graph_compute_with_ctx(ctx, gf, 1);
  // Try printing the tensor
  std::cout << "Now-contiguous tensor after we perform the computation\n";
  print_tensor(a);
  ggml_free(ctx);
  return 0;
}

I created this because I was confused on ggml_cont and why it (seemingly) was not working as intended. I wrote a short article accompanying this gist here