1  #include "ggml.h"
  2  
  3  #include <cmath>
  4  #include <cstdio>
  5  #include <cstdlib>
  6  #include <cassert>
  7  
  8  #define MAX_NARGS 2
  9  
 10  #if defined(__GNUC__)
 11  #pragma GCC diagnostic ignored "-Wdouble-promotion"
 12  #endif
 13  
 14  //
 15  // logging
 16  //
 17  #define GGML_DEBUG 0
 18  #if (GGML_DEBUG >= 1)
 19  #define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
 20  #else
 21  #define GGML_PRINT_DEBUG(...)
 22  #endif
 23  
 24  #if (GGML_DEBUG >= 5)
 25  #define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
 26  #else
 27  #define GGML_PRINT_DEBUG_5(...)
 28  #endif
 29  
 30  #if (GGML_DEBUG >= 10)
 31  #define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
 32  #else
 33  #define GGML_PRINT_DEBUG_10(...)
 34  #endif
 35  
 36  #define GGML_PRINT(...) printf(__VA_ARGS__)
 37  
 38  
 39  static float frand(void) {
 40      return (float)rand()/(float)RAND_MAX;
 41  }
 42  
 43  static struct ggml_tensor * get_random_tensor(
 44      struct ggml_context * ctx0, int ndims, int64_t ne[], float fmin, float fmax
 45  ) {
 46      struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
 47  
 48      switch (ndims) {
 49          case 1:
 50              for (int i0 = 0; i0 < ne[0]; i0++) {
 51                  ((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
 52              }
 53              break;
 54          case 2:
 55              for (int i1 = 0; i1 < ne[1]; i1++) {
 56                  for (int i0 = 0; i0 < ne[0]; i0++) {
 57                      ((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
 58                  }
 59              }
 60              break;
 61          case 3:
 62              for (int i2 = 0; i2 < ne[2]; i2++) {
 63                  for (int i1 = 0; i1 < ne[1]; i1++) {
 64                      for (int i0 = 0; i0 < ne[0]; i0++) {
 65                          ((float *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
 66                      }
 67                  }
 68              }
 69              break;
 70          case 4:
 71              for (int i3 = 0; i3 < ne[3]; i3++) {
 72                  for (int i2 = 0; i2 < ne[2]; i2++) {
 73                      for (int i1 = 0; i1 < ne[1]; i1++) {
 74                          for (int i0 = 0; i0 < ne[0]; i0++) {
 75                              ((float *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
 76                          }
 77                      }
 78                  }
 79              }
 80              break;
 81          default:
 82              assert(false);
 83      }
 84  
 85      return result;
 86  }
 87  
 88  int main(void) {
 89      struct ggml_init_params params = {
 90          /* .mem_size   = */ 1024*1024*1024,
 91          /* .mem_buffer = */ NULL,
 92          /* .no_alloc   = */ false,
 93      };
 94  
 95      struct ggml_context * ctx = ggml_init(params);
 96  
 97      int64_t ne1[4] = {4, 128, 1, 1};
 98      int64_t ne2[4] = {4, 256, 1, 1};
 99      int64_t ne3[4] = {128, 256, 1, 1};
100  
101      struct ggml_tensor * a = get_random_tensor(ctx, 2, ne1, -1, +1);
102      struct ggml_tensor * b = get_random_tensor(ctx, 2, ne2, -1, +1);
103      ggml_set_param(ctx, a);
104      ggml_set_param(ctx, b);
105  
106      struct ggml_tensor * c = get_random_tensor(ctx, 2, ne3, -1, +1);
107  
108      struct ggml_tensor * ab = ggml_mul_mat(ctx, a, b);
109      struct ggml_tensor * d  = ggml_sub(ctx, c, ab);
110      struct ggml_tensor * e  = ggml_sum(ctx, ggml_sqr(ctx, d));
111  
112      struct ggml_cgraph * ge = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, true);
113      ggml_build_forward_expand(ge, e);
114      ggml_graph_reset(ge);
115  
116      ggml_graph_compute_with_ctx(ctx, ge, /*n_threads*/ 1);
117  
118      const float fe = ggml_get_f32_1d(e, 0);
119      printf("%s: e = %.4f\n", __func__, fe);
120  
121      struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
122  
123      ggml_opt(ctx, opt_params, e);
124  
125      ggml_graph_reset(ge);
126  
127      ggml_graph_compute_with_ctx(ctx, ge, /*n_threads*/ 1);
128  
129      const float fe_opt = ggml_get_f32_1d(e, 0);
130      printf("%s: original  e = %.4f\n", __func__, fe);
131      printf("%s: optimized e = %.4f\n", __func__, fe_opt);
132  
133      const bool success = (fe_opt <= fe);
134      assert(success);
135  
136      ggml_free(ctx);
137      return success ? 0 : -1;
138  }
139  // int64_t ne1[4] = {4, 128, 1, 1};
140  // int64_t ne2[4] = {4, 256, 1, 1};;
141  // int64_t ne3[4] = {128, 256, 1, 1};
142  // main: original  e = 25890.9375
143  // main: optimized e = 10094.7031
144  
145  // int64_t ne1[4] = {8, 128, 1, 1};
146  // int64_t ne2[4] = {8, 256, 1, 1};;
147  // int64_t ne3[4] = {128, 256, 1, 1};
148  // main: original  e = 39429.5078
149  // main: optimized e = 9275.8936
150  
151  // int64_t ne1[4] = {16, 128, 1, 1};
152  // int64_t ne2[4] = {16, 256, 1, 1};;
153  // int64_t ne3[4] = {128, 256, 1, 1};
154  // main: original  e = 68371.1328
155  // main: optimized e = 7854.4502
156  
157  
158  // int64_t ne1[4] = {32, 128, 1, 1};
159  // int64_t ne2[4] = {32, 256, 1, 1};;
160  // int64_t ne3[4] = {128, 256, 1, 1};
161  // main: original  e = 126061.1953
162  // main: optimized e = 5451.0166
163  
164  // int64_t ne1[4] = {4, 1024, 1, 1};
165  // int64_t ne2[4] = {4, 2048, 1, 1};;
166  // int64_t ne3[4] = {1024, 2048, 1, 1};
167  // main: original  e = 1620817.8750
168  // main: optimized e = 698387.6875
169  
170  // another run on M1
171  // int64_t ne1[4] = {4, 1024, 1, 1};
172  // int64_t ne2[4] = {4, 2048, 1, 1};;
173  // int64_t ne3[4] = {1024, 2048, 1, 1};
174  // main: original  e = 1629595.6250
175  // main: optimized e = 698169.1250
176  
177  // int64_t ne1[4] = {32, 1024, 1, 1};
178  // int64_t ne2[4] = {32, 2048, 1, 1};;
179  // int64_t ne3[4] = {1024, 2048, 1, 1};
180  // main: original  e = 8146770.5000
181  // main: optimized e = 651119.1250