- // Copyright 2021 Tencent
- // SPDX-License-Identifier: BSD-3-Clause
-
- #include "testutil.h"
-
- static int test_gru(int size, int T, int outch, int direction)
- {
- ncnn::Mat a = RandomMat(size, T);
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
-
- std::vector<ncnn::Mat> weights(3);
- weights[0] = RandomMat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomMat(outch * outch * 3 * num_directions);
-
- int ret = test_layer("GRU", pd, weights, a);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_with_hidden(int size, int T, int outch, int direction)
- {
- ncnn::Mat a = RandomMat(size, T);
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
-
- std::vector<ncnn::Mat> weights(3);
- weights[0] = RandomMat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomMat(outch * outch * 3 * num_directions);
-
- // initial hidden state
- ncnn::Mat hidden = RandomMat(outch, num_directions);
-
- std::vector<ncnn::Mat> as(2);
- as[0] = a;
- as[1] = hidden;
-
- int ret = test_layer("GRU", pd, weights, as, 2);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru_with_hidden failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_with_hidden_input(int size, int T, int outch, int direction)
- {
- ncnn::Mat a = RandomMat(size, T);
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
-
- std::vector<ncnn::Mat> weights(3);
- weights[0] = RandomMat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomMat(outch * outch * 3 * num_directions);
-
- // initial hidden state
- ncnn::Mat hidden = RandomMat(outch, num_directions);
-
- std::vector<ncnn::Mat> as(2);
- as[0] = a;
- as[1] = hidden;
-
- int ret = test_layer("GRU", pd, weights, as, 1);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru_with_hidden_input failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_with_hidden_output(int size, int T, int outch, int direction)
- {
- ncnn::Mat a = RandomMat(size, T);
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
-
- std::vector<ncnn::Mat> weights(3);
- weights[0] = RandomMat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomMat(outch * outch * 3 * num_directions);
-
- std::vector<ncnn::Mat> as(1);
- as[0] = a;
-
- int ret = test_layer("GRU", pd, weights, as, 2);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru_with_hidden_output failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_0()
- {
- return 0
- || test_gru(4, 1, 2, 2)
- || test_gru(8, 2, 2, 2)
- || test_gru(16, 8, 7, 2)
- || test_gru(17, 8, 8, 2)
- || test_gru(19, 15, 8, 2)
- || test_gru(5, 16, 16, 2)
- || test_gru(3, 16, 8, 2)
- || test_gru(8, 16, 16, 2)
- || test_gru(31, 3, 31, 2)
- || test_gru(2, 5, 17, 2);
- }
-
- static int test_gru_1()
- {
- return 0
- || test_gru_with_hidden(4, 4, 1, 2)
- || test_gru_with_hidden(8, 2, 2, 2)
- || test_gru_with_hidden(16, 8, 7, 2)
- || test_gru_with_hidden(17, 8, 8, 2)
- || test_gru_with_hidden(19, 15, 8, 2)
- || test_gru_with_hidden(5, 16, 16, 2)
- || test_gru_with_hidden(3, 16, 8, 2)
- || test_gru_with_hidden(2, 5, 79, 2)
- || test_gru_with_hidden(4, 4, 1, 1)
- || test_gru_with_hidden(8, 2, 2, 1)
- || test_gru_with_hidden(16, 8, 7, 1)
- || test_gru_with_hidden(17, 8, 8, 1)
- || test_gru_with_hidden(19, 15, 8, 1)
- || test_gru_with_hidden(5, 16, 16, 1)
- || test_gru_with_hidden(3, 16, 8, 1)
- || test_gru_with_hidden(2, 5, 79, 1)
- || test_gru_with_hidden(4, 2, 1, 0)
- || test_gru_with_hidden(8, 2, 2, 0)
- || test_gru_with_hidden(16, 8, 7, 0)
- || test_gru_with_hidden(17, 8, 8, 0)
- || test_gru_with_hidden(19, 15, 8, 0)
- || test_gru_with_hidden(5, 16, 16, 0)
- || test_gru_with_hidden(3, 16, 8, 0)
- || test_gru_with_hidden(2, 5, 17, 0)
-
- || test_gru_with_hidden_input(4, 4, 1, 2)
- || test_gru_with_hidden_input(8, 2, 2, 2)
- || test_gru_with_hidden_input(16, 8, 7, 2)
- || test_gru_with_hidden_input(17, 8, 8, 2)
- || test_gru_with_hidden_input(19, 15, 8, 2)
- || test_gru_with_hidden_input(5, 16, 16, 2)
- || test_gru_with_hidden_input(3, 16, 8, 2)
- || test_gru_with_hidden_input(2, 5, 79, 2)
- || test_gru_with_hidden_input(4, 4, 1, 1)
- || test_gru_with_hidden_input(8, 2, 2, 1)
- || test_gru_with_hidden_input(16, 8, 7, 1)
- || test_gru_with_hidden_input(17, 8, 8, 1)
- || test_gru_with_hidden_input(19, 15, 8, 1)
- || test_gru_with_hidden_input(5, 16, 16, 1)
- || test_gru_with_hidden_input(3, 16, 8, 1)
- || test_gru_with_hidden_input(2, 5, 79, 1)
- || test_gru_with_hidden_input(4, 2, 1, 0)
- || test_gru_with_hidden_input(8, 2, 2, 0)
- || test_gru_with_hidden_input(16, 8, 7, 0)
- || test_gru_with_hidden_input(17, 8, 8, 0)
- || test_gru_with_hidden_input(19, 15, 8, 0)
- || test_gru_with_hidden_input(5, 16, 16, 0)
- || test_gru_with_hidden_input(3, 16, 8, 0)
- || test_gru_with_hidden_input(2, 5, 17, 0)
-
- || test_gru_with_hidden_output(4, 4, 1, 2)
- || test_gru_with_hidden_output(8, 2, 2, 2)
- || test_gru_with_hidden_output(16, 8, 7, 2)
- || test_gru_with_hidden_output(17, 8, 8, 2)
- || test_gru_with_hidden_output(19, 15, 8, 2)
- || test_gru_with_hidden_output(5, 16, 16, 2)
- || test_gru_with_hidden_output(3, 16, 8, 2)
- || test_gru_with_hidden_output(2, 5, 79, 2)
- || test_gru_with_hidden_output(4, 4, 1, 1)
- || test_gru_with_hidden_output(8, 2, 2, 1)
- || test_gru_with_hidden_output(16, 8, 7, 1)
- || test_gru_with_hidden_output(17, 8, 8, 1)
- || test_gru_with_hidden_output(19, 15, 8, 1)
- || test_gru_with_hidden_output(5, 16, 16, 1)
- || test_gru_with_hidden_output(3, 16, 8, 1)
- || test_gru_with_hidden_output(2, 5, 79, 1)
- || test_gru_with_hidden_output(4, 2, 1, 0)
- || test_gru_with_hidden_output(8, 2, 2, 0)
- || test_gru_with_hidden_output(16, 8, 7, 0)
- || test_gru_with_hidden_output(17, 8, 8, 0)
- || test_gru_with_hidden_output(19, 15, 8, 0)
- || test_gru_with_hidden_output(5, 16, 16, 0)
- || test_gru_with_hidden_output(3, 16, 8, 0)
- || test_gru_with_hidden_output(2, 5, 17, 0);
- }
-
- static int test_gru_2()
- {
- return 0
- || test_gru(4, 1, 1, 0)
- || test_gru(8, 2, 2, 0)
- || test_gru(16, 8, 7, 0)
- || test_gru(17, 8, 8, 0)
- || test_gru(19, 15, 8, 0)
- || test_gru(5, 16, 16, 0)
- || test_gru(3, 16, 8, 0)
- || test_gru(8, 16, 16, 0)
- || test_gru(2, 5, 17, 0);
- }
-
- static int test_gru_3()
- {
- return 0
- || test_gru(4, 1, 1, 1)
- || test_gru(8, 2, 2, 1)
- || test_gru(16, 8, 7, 1)
- || test_gru(17, 8, 8, 1)
- || test_gru(19, 15, 8, 1)
- || test_gru(5, 16, 16, 1)
- || test_gru(3, 16, 8, 1)
- || test_gru(8, 16, 16, 1)
- || test_gru(2, 5, 17, 1);
- }
-
- #if NCNN_INT8
- static void RandomizeA(ncnn::Mat& m, float absmax)
- {
- absmax = ncnn::float16_to_float32(ncnn::float32_to_float16(absmax));
- absmax = ncnn::bfloat16_to_float32(ncnn::float32_to_bfloat16(absmax));
-
- const int h = m.h;
- float* p = m;
- for (int i = 0; i < h; i++)
- {
- float* p = m.row(i);
- for (int j = 0; j < m.w; j++)
- {
- p[j] = RandomFloat(-absmax, absmax);
-
- // drop 0.45 ~ 0.55
- float v = p[j] * (127.f / absmax);
- float vv = fabs(v - (int)v);
-
- float hp = ncnn::float16_to_float32(ncnn::float32_to_float16(p[j]));
- float hv = hp * (127.f / absmax);
- float hvv = fabs(hv - (int)hv);
-
- float bp = ncnn::bfloat16_to_float32(ncnn::float32_to_bfloat16(p[j]));
- float bv = bp * (127.f / absmax);
- float bvv = fabs(bv - (int)bv);
-
- while ((vv > 0.45f && vv < 0.55f) || (hvv > 0.45f && hvv < 0.55f) || (bvv > 0.45f && bvv < 0.55f))
- {
- p[j] = RandomFloat(-absmax, absmax);
- v = p[j] * (127.f / absmax);
- vv = fabs(v - (int)v);
-
- hp = ncnn::float16_to_float32(ncnn::float32_to_float16(p[j]));
- hv = hp * (127.f / absmax);
- hvv = fabs(hv - (int)hv);
-
- bp = ncnn::bfloat16_to_float32(ncnn::float32_to_bfloat16(p[j]));
- bv = bp * (127.f / absmax);
- bvv = fabs(bv - (int)bv);
- }
- }
- }
-
- // set random a and b
- m.row(RandomInt(0, h - 1))[RandomInt(0, m.w - 1)] = -absmax;
- m.row(RandomInt(0, h - 1))[RandomInt(0, m.w - 1)] = absmax;
- }
-
- static int test_gru_int8(int size, int T, int outch, int direction)
- {
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
- pd.set(8, 2); // int8_scale_term
-
- std::vector<ncnn::Mat> weights(5);
- weights[0] = RandomS8Mat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomS8Mat(outch * outch * 3 * num_directions);
- weights[3] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
- weights[4] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
-
- ncnn::Mat a(size, T);
- RandomizeA(a, 10.f);
-
- int ret = test_layer("GRU", pd, weights, a);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru_int8 failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_int8_with_hidden(int size, int T, int outch, int direction)
- {
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
- pd.set(8, 2); // int8_scale_term
-
- std::vector<ncnn::Mat> weights(5);
- weights[0] = RandomS8Mat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomS8Mat(outch * outch * 3 * num_directions);
- weights[3] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
- weights[4] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
-
- ncnn::Mat a(size, T);
- RandomizeA(a, 10.f);
-
- // initial hidden state
- ncnn::Mat hidden(outch, num_directions);
- RandomizeA(hidden, 10.f);
-
- std::vector<ncnn::Mat> as(2);
- as[0] = a;
- as[1] = hidden;
-
- int ret = test_layer("GRU", pd, weights, as, 2);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru_int8_with_hidden failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_int8_with_hidden_input(int size, int T, int outch, int direction)
- {
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
- pd.set(8, 2); // int8_scale_term
-
- std::vector<ncnn::Mat> weights(5);
- weights[0] = RandomS8Mat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomS8Mat(outch * outch * 3 * num_directions);
- weights[3] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
- weights[4] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
-
- ncnn::Mat a(size, T);
- RandomizeA(a, 10.f);
-
- // initial hidden state
- ncnn::Mat hidden(outch, num_directions);
- RandomizeA(hidden, 10.f);
-
- std::vector<ncnn::Mat> as(2);
- as[0] = a;
- as[1] = hidden;
-
- int ret = test_layer("GRU", pd, weights, as, 1);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru_int8_with_hidden_input failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_int8_with_hidden_output(int size, int T, int outch, int direction)
- {
- int num_directions = direction == 2 ? 2 : 1;
-
- ncnn::ParamDict pd;
- pd.set(0, outch);
- pd.set(1, outch * size * 3 * num_directions);
- pd.set(2, direction);
- pd.set(8, 2); // int8_scale_term
-
- std::vector<ncnn::Mat> weights(5);
- weights[0] = RandomS8Mat(outch * size * 3 * num_directions);
- weights[1] = RandomMat(outch * 4 * num_directions);
- weights[2] = RandomS8Mat(outch * outch * 3 * num_directions);
- weights[3] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
- weights[4] = RandomMat(outch * 3 * num_directions, 100.f, 200.f);
-
- ncnn::Mat a(size, T);
- RandomizeA(a, 10.f);
-
- std::vector<ncnn::Mat> as(1);
- as[0] = a;
-
- int ret = test_layer("GRU", pd, weights, as, 2);
- if (ret != 0)
- {
- fprintf(stderr, "test_gru_int8_with_hidden_output failed size=%d T=%d outch=%d direction=%d\n", size, T, outch, direction);
- }
-
- return ret;
- }
-
- static int test_gru_4()
- {
- return 0
- || test_gru_int8(4, 1, 2, 2)
- || test_gru_int8(8, 2, 2, 2)
- || test_gru_int8(16, 8, 7, 2)
- || test_gru_int8(17, 8, 8, 2)
- || test_gru_int8(19, 15, 8, 2)
- || test_gru_int8(5, 16, 16, 2)
- || test_gru_int8(3, 16, 8, 2)
- || test_gru_int8(8, 16, 16, 2)
- || test_gru_int8(31, 3, 31, 2)
- || test_gru_int8(2, 5, 17, 2);
- }
-
- static int test_gru_5()
- {
- return 0
- || test_gru_int8_with_hidden(4, 4, 1, 2)
- || test_gru_int8_with_hidden(8, 2, 2, 2)
- || test_gru_int8_with_hidden(16, 8, 7, 2)
- || test_gru_int8_with_hidden(17, 8, 8, 2)
- || test_gru_int8_with_hidden(19, 15, 8, 2)
- || test_gru_int8_with_hidden(5, 16, 16, 2)
- || test_gru_int8_with_hidden(3, 16, 8, 2)
- || test_gru_int8_with_hidden(2, 5, 79, 2)
- || test_gru_int8_with_hidden(4, 4, 1, 1)
- || test_gru_int8_with_hidden(8, 2, 2, 1)
- || test_gru_int8_with_hidden(16, 8, 7, 1)
- || test_gru_int8_with_hidden(17, 8, 8, 1)
- || test_gru_int8_with_hidden(19, 15, 8, 1)
- || test_gru_int8_with_hidden(5, 16, 16, 1)
- || test_gru_int8_with_hidden(3, 16, 8, 1)
- || test_gru_int8_with_hidden(2, 5, 79, 1)
- || test_gru_int8_with_hidden(4, 2, 1, 0)
- || test_gru_int8_with_hidden(8, 2, 2, 0)
- || test_gru_int8_with_hidden(16, 8, 7, 0)
- || test_gru_int8_with_hidden(17, 8, 8, 0)
- || test_gru_int8_with_hidden(19, 15, 8, 0)
- || test_gru_int8_with_hidden(5, 16, 16, 0)
- || test_gru_int8_with_hidden(3, 16, 8, 0)
- || test_gru_int8_with_hidden(2, 5, 17, 0)
-
- || test_gru_int8_with_hidden_input(4, 4, 1, 2)
- || test_gru_int8_with_hidden_input(8, 2, 2, 2)
- || test_gru_int8_with_hidden_input(16, 8, 7, 2)
- || test_gru_int8_with_hidden_input(17, 8, 8, 2)
- || test_gru_int8_with_hidden_input(19, 15, 8, 2)
- || test_gru_int8_with_hidden_input(5, 16, 16, 2)
- || test_gru_int8_with_hidden_input(3, 16, 8, 2)
- || test_gru_int8_with_hidden_input(2, 5, 79, 2)
- || test_gru_int8_with_hidden_input(4, 4, 1, 1)
- || test_gru_int8_with_hidden_input(8, 2, 2, 1)
- || test_gru_int8_with_hidden_input(16, 8, 7, 1)
- || test_gru_int8_with_hidden_input(17, 8, 8, 1)
- || test_gru_int8_with_hidden_input(19, 15, 8, 1)
- || test_gru_int8_with_hidden_input(5, 16, 16, 1)
- || test_gru_int8_with_hidden_input(3, 16, 8, 1)
- || test_gru_int8_with_hidden_input(2, 5, 79, 1)
- || test_gru_int8_with_hidden_input(4, 2, 1, 0)
- || test_gru_int8_with_hidden_input(8, 2, 2, 0)
- || test_gru_int8_with_hidden_input(16, 8, 7, 0)
- || test_gru_int8_with_hidden_input(17, 8, 8, 0)
- || test_gru_int8_with_hidden_input(19, 15, 8, 0)
- || test_gru_int8_with_hidden_input(5, 16, 16, 0)
- || test_gru_int8_with_hidden_input(3, 16, 8, 0)
- || test_gru_int8_with_hidden_input(2, 5, 17, 0)
-
- || test_gru_int8_with_hidden_output(4, 4, 1, 2)
- || test_gru_int8_with_hidden_output(8, 2, 2, 2)
- || test_gru_int8_with_hidden_output(16, 8, 7, 2)
- || test_gru_int8_with_hidden_output(17, 8, 8, 2)
- || test_gru_int8_with_hidden_output(19, 15, 8, 2)
- || test_gru_int8_with_hidden_output(5, 16, 16, 2)
- || test_gru_int8_with_hidden_output(3, 16, 8, 2)
- || test_gru_int8_with_hidden_output(2, 5, 79, 2)
- || test_gru_int8_with_hidden_output(4, 4, 1, 1)
- || test_gru_int8_with_hidden_output(8, 2, 2, 1)
- || test_gru_int8_with_hidden_output(16, 8, 7, 1)
- || test_gru_int8_with_hidden_output(17, 8, 8, 1)
- || test_gru_int8_with_hidden_output(19, 15, 8, 1)
- || test_gru_int8_with_hidden_output(5, 16, 16, 1)
- || test_gru_int8_with_hidden_output(3, 16, 8, 1)
- || test_gru_int8_with_hidden_output(2, 5, 79, 1)
- || test_gru_int8_with_hidden_output(4, 2, 1, 0)
- || test_gru_int8_with_hidden_output(8, 2, 2, 0)
- || test_gru_int8_with_hidden_output(16, 8, 7, 0)
- || test_gru_int8_with_hidden_output(17, 8, 8, 0)
- || test_gru_int8_with_hidden_output(19, 15, 8, 0)
- || test_gru_int8_with_hidden_output(5, 16, 16, 0)
- || test_gru_int8_with_hidden_output(3, 16, 8, 0)
- || test_gru_int8_with_hidden_output(2, 5, 17, 0);
- }
-
- static int test_gru_6()
- {
- return 0
- || test_gru_int8(4, 1, 1, 0)
- || test_gru_int8(8, 2, 2, 0)
- || test_gru_int8(16, 8, 7, 0)
- || test_gru_int8(17, 8, 8, 0)
- || test_gru_int8(19, 15, 8, 0)
- || test_gru_int8(5, 16, 16, 0)
- || test_gru_int8(3, 16, 8, 0)
- || test_gru_int8(8, 16, 16, 0)
- || test_gru_int8(2, 5, 17, 0);
- }
-
- static int test_gru_7()
- {
- return 0
- || test_gru_int8(4, 1, 1, 1)
- || test_gru_int8(8, 2, 2, 1)
- || test_gru_int8(16, 8, 7, 1)
- || test_gru_int8(17, 8, 8, 1)
- || test_gru_int8(19, 15, 8, 1)
- || test_gru_int8(5, 16, 16, 1)
- || test_gru_int8(3, 16, 8, 1)
- || test_gru_int8(8, 16, 16, 1)
- || test_gru_int8(2, 5, 17, 1);
- }
- #endif
-
- int main()
- {
- SRAND(7767517);
-
- #if NCNN_INT8
- return 0
- || test_gru_0()
- || test_gru_1()
- || test_gru_2()
- || test_gru_3()
- || test_gru_4()
- || test_gru_5()
- || test_gru_6()
- || test_gru_7();
- #else
- return 0
- || test_gru_0()
- || test_gru_1()
- || test_gru_2()
- || test_gru_3();
- #endif
- }
|