Skip to content

Commit 7f4b20e

Browse files
author
Italo Nicola
committed
(wip) Tests: add a larger test
1 parent e70ff02 commit 7f4b20e

File tree

1 file changed

+123
-17
lines changed

1 file changed

+123
-17
lines changed

‎tests/test-backend-ops.cpp‎

Lines changed: 123 additions & 17 deletions
Original file line numberDiff line numberDiff line change
@@ -1154,9 +1154,9 @@ struct test_case {
11541154
double err = nmse(f1.data(), f2.data(), f1.size());
11551155
if (err > ud->max_err) {
11561156
printf("[%s] NMSE = %.9f > %.9f ", ggml_op_desc(t1), err, ud->max_err);
1157-
//for (int i = 0; i < (int) f1.size(); i++) {
1158-
// printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
1159-
//}
1157+
for (int i = 0; i < (int) f1.size(); i++) {
1158+
printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
1159+
}
11601160
printf("\n");
11611161
//exit(1);
11621162
ud->ok = false;
@@ -2953,6 +2953,7 @@ struct test_mul_mat : public test_case {
29532953
return out;
29542954
}
29552955

2956+
#if 0
29562957
void initialize_tensors(ggml_context * ctx) override {
29572958
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
29582959
if (t->type == GGML_TYPE_F32) {
@@ -2976,16 +2977,30 @@ struct test_mul_mat : public test_case {
29762977
for (int64_t r = 0; r < ggml_nrows(t); r++) {
29772978
std::vector<float> data(t->ne[0]);
29782979
for (int i = 0; i < t->ne[0]; i++) {
2979-
data[i] = (1.33f * i + r) / (t->ne[0] + t->ne[1] * 1.2);
2980+
//data[i] = (1.33f * i + r) / (t->ne[0] + t->ne[1] * 1.2);
2981+
data[i] = 1.0f;
29802982
}
29812983
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
29822984
}
29832985
}
2984-
} else {
2986+
} /* else if (t->type == GGML_TYPE_Q8_0) {
2987+
if (t->name[0] == 'a') {
2988+
for (int64_t r = 0; r < ggml_nrows(t); r++) {
2989+
std::vector<uint8_t> data(t->ne[0]);
2990+
for (int i = 0; i < t->ne[0]; i++) {
2991+
//data[i] = (4 * i + r) / (t->ne[0] + t->ne[1] * 0.5);
2992+
//data[i] = std::min(i, 32);
2993+
data[i] = 128;
2994+
}
2995+
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(uint8_t));
2996+
}
2997+
}
2998+
} */ else {
29852999
init_tensor_uniform(t);
29863000
}
29873001
}
29883002
}
3003+
#endif
29893004
};
29903005

29913006
// GGML_OP_MUL_MAT_ID
@@ -3107,6 +3122,56 @@ struct test_out_prod : public test_case {
31073122

31083123
return out;
31093124
}
3125+
3126+
#if 0
3127+
void initialize_tensors(ggml_context * ctx) override {
3128+
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
3129+
if (t->type == GGML_TYPE_F32) {
3130+
if (t->name[0] == 'o') {
3131+
for (int64_t r = 0; r < ggml_nrows(t); r++) {
3132+
std::vector<float> data(t->ne[0]);
3133+
for (int i = 0; i < t->ne[0]; i++) {
3134+
data[i] = -13.0f;
3135+
}
3136+
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
3137+
}
3138+
} else if (t->name[0] == 'a') {
3139+
for (int64_t r = 0; r < ggml_nrows(t); r++) {
3140+
std::vector<float> data(t->ne[0]);
3141+
for (int i = 0; i < t->ne[0]; i++) {
3142+
data[i] = (3.50f * i + r) / (t->ne[0] + t->ne[1] * 0.5);
3143+
}
3144+
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
3145+
}
3146+
} else {
3147+
for (int64_t r = 0; r < ggml_nrows(t); r++) {
3148+
std::vector<float> data(t->ne[0]);
3149+
for (int i = 0; i < t->ne[0]; i++) {
3150+
//data[i] = (1.33f * i + r) / (t->ne[0] + t->ne[1] * 1.2);
3151+
data[i] = i+1;
3152+
}
3153+
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
3154+
}
3155+
}
3156+
} else if (t->type == GGML_TYPE_Q8_0) {
3157+
if (t->name[0] == 'a') {
3158+
for (int64_t r = 0; r < ggml_nrows(t); r++) {
3159+
std::vector<uint8_t> data(t->ne[0]);
3160+
for (int i = 0; i < t->ne[0]; i++) {
3161+
//data[i] = (4 * i + r) / (t->ne[0] + t->ne[1] * 0.5);
3162+
//data[i] = std::min(i, 32);
3163+
data[i] = 100 + i;
3164+
}
3165+
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(uint8_t));
3166+
}
3167+
}
3168+
} else {
3169+
init_tensor_uniform(t);
3170+
}
3171+
}
3172+
}
3173+
#endif
3174+
31103175
};
31113176

31123177
// GGML_OP_SQR
@@ -4856,7 +4921,8 @@ struct test_falcon : public test_llm {
48564921
// ###########################################
48574922
static const ggml_type all_types[] = {
48584923
GGML_TYPE_F32, GGML_TYPE_F16, // GGML_TYPE_BF16,
4859-
// GGML_TYPE_Q4_0, GGML_TYPE_Q4_1,
4924+
GGML_TYPE_Q4_0,
4925+
// GGML_TYPE_Q4_1,
48604926
// GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
48614927
GGML_TYPE_Q8_0,
48624928
// GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
@@ -4871,13 +4937,14 @@ static const ggml_type all_types[] = {
48714937
static const ggml_type base_types[] = {
48724938
GGML_TYPE_F32, GGML_TYPE_F16,
48734939
GGML_TYPE_Q8_0, // for I8MM tests
4874-
// GGML_TYPE_Q4_0,
4940+
GGML_TYPE_Q4_0,
48754941
// GGML_TYPE_Q4_1, // for I8MM tests
48764942
// GGML_TYPE_Q4_K,
48774943
// GGML_TYPE_IQ2_XXS
48784944
};
48794945

48804946
static const ggml_type other_types[] = {
4947+
GGML_TYPE_Q4_0,
48814948
// GGML_TYPE_Q4_1,
48824949
// GGML_TYPE_Q5_0, GGML_TYPE_Q5_1,
48834950
GGML_TYPE_Q8_0,
@@ -5231,13 +5298,58 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
52315298
test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4));
52325299
test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 128, 4));
52335300

5234-
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 32, 2, 2, {1, 1}, {1, 1}));
5235-
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 4096*2, 256, 1024, {1, 1}, {1, 1}));
5236-
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 4096*20, 256, 1024, {1, 1}, {1, 1}));
5237-
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 4096*30, 256, 1024, {1, 1}, {1, 1}));
5301+
test_cases.emplace_back(new test_out_prod(GGML_TYPE_F32, GGML_TYPE_F32, 4096*20, 256, 1024, {1, 1}, {1, 1}));
5302+
test_cases.emplace_back(new test_out_prod(GGML_TYPE_Q8_0, GGML_TYPE_F32, 4096*20, 256, 1024, {1, 1}, {1, 1}));
5303+
test_cases.emplace_back(new test_out_prod(GGML_TYPE_F32, GGML_TYPE_F32, 1024, 256, 4096*40, {1, 1}, {1, 1}));
5304+
test_cases.emplace_back(new test_out_prod(GGML_TYPE_Q8_0, GGML_TYPE_F32, 1024, 256, 4096*40, {1, 1}, {1, 1}));
52385305

5306+
#if 0
5307+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 2, 2, 32, {1, 1}, {1, 1}));
5308+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 4, 4, 32, {1, 1}, {1, 1}));
5309+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 2, 2, 32, {1, 1}, {1, 1}));
5310+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 4, 4, 32, {1, 1}, {1, 1}));
52395311

5312+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 4096*20, 256, 1024, {1, 1}, {1, 1}));
5313+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 4096*20, 256, 1024, {1, 1}, {1, 1}));
5314+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 1024, 256, 4096*20, {1, 1}, {1, 1}));
5315+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 1024, 256, 4096*20, {1, 1}, {1, 1}));
5316+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 1024, 256, 4096*40, {1, 1}, {1, 1}));
5317+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 1024, 256, 4096*40, {1, 1}, {1, 1}));
5318+
#endif
5319+
52405320
#if 0
5321+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 2, 2, 1024, {1, 1}, {1, 1}));
5322+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 1024, 2, 2, {1, 1}, {1, 1}));
5323+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 2, 1024, 2, {1, 1}, {1, 1}));
5324+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 2, 1, 4096, {1, 1}, {1, 1}));
5325+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 1, 2, 4096, {1, 1}, {1, 1}));
5326+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 2, 2, 4096, {1, 1}, {1, 1}));
5327+
5328+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 2, 2, 32, {1, 1}, {1, 1}));
5329+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 4, 4, 32, {1, 1}, {1, 1}));
5330+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 4, 4, 64, {1, 1}, {1, 1}));
5331+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 8, 8, 128, {1, 1}, {1, 1}));
5332+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 32, 2, 2, {1, 1}, {1, 1}));
5333+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 16, 2, {1, 1}, {1, 1}));
5334+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 64, 2, {1, 1}, {1, 1}));
5335+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 16, 4, {1, 1}, {1, 1}));
5336+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 64, 4, {1, 1}, {1, 1}));
5337+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 4, 16, 32, {1, 1}, {1, 1}));
5338+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 4, 32, {1, 1}, {1, 1}));
5339+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 4, 16, 64, {1, 1}, {1, 1}));
5340+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 4, 64, {1, 1}, {1, 1}));
5341+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 32, 32, 32, {1, 1}, {1, 1}));
5342+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 32, 64, 64, {1, 1}, {1, 1}));
5343+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 32, 64, {1, 1}, {1, 1}));
5344+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 64, 32, {1, 1}, {1, 1}));
5345+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 64, 64, {1, 1}, {1, 1}));
5346+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 128, 64, 2, {1, 1}, {1, 1}));
5347+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 256, 128, 32, {1, 1}, {1, 1}));
5348+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 256, 128, 64, {1, 1}, {1, 1}));
5349+
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 512, 256, 1024, {1, 1}, {1, 1}));
5350+
#endif
5351+
5352+
#if 1
52415353
for (ggml_type type_a : all_types) {
52425354
for (int i = 1; i < 10; ++i) {
52435355
test_cases.emplace_back(new test_mul_mat(type_a, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1}));
@@ -5329,12 +5441,6 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
53295441
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1, 1}, {4, 1}, {0, 2, 1, 3}));
53305442
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67, {1, 1}, {4, 1}, {0, 2, 1, 3}));
53315443

5332-
// XXX
5333-
//test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 151936, 256, 1024, {1, 1}, {1, 1}));
5334-
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 128, 256, 1024, {1, 1}, {1, 1}));
5335-
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 4096*8, 256, 1024, {1, 1}, {1, 1}));
5336-
test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 4096*16, 256, 1024, {1, 1}, {1, 1}));
5337-
53385444
for (auto bs : {1,2,4,8}) {
53395445
for (auto nr : {1,4}) {
53405446
for (uint32_t m = 0; m < 2; ++m) {

0 commit comments

Comments
 (0)