Skip to content

Commit 927987b

Browse files
hipuddinghipudding
committed
format
1 parent 4457232 commit 927987b

File tree

2 files changed

+97
-88
lines changed

2 files changed

+97
-88
lines changed

‎ggml/src/ggml-cann/aclnn_ops.cpp

Lines changed: 77 additions & 68 deletions
Original file line numberDiff line numberDiff line change
@@ -1330,60 +1330,68 @@ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
13301330
}
13311331

13321332

1333-
static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer, float m, int64_t size, float start, float stop, float step){
1333+
static void aclnn_get_slope_inner(ggml_backend_cann_context& ctx, void* slope_buffer,
1334+
float m, int64_t size, float start, float stop, float step){
13341335
int64_t ne[] = {size};
13351336
size_t nb[] = {sizeof(float)};
13361337

1337-
ggml_cann_pool_alloc arange_allocator(ctx.pool(),size * sizeof(float));
1338-
void* arange_buffer = arange_allocator.get();
1339-
1340-
aclTensor* arange_tensor = ggml_cann_create_tensor(
1341-
arange_buffer, ACL_FLOAT,
1342-
sizeof(float), ne, nb, 1);
1338+
ggml_cann_pool_alloc arange_allocator(ctx.pool(), size * sizeof(float));
1339+
void * arange_buffer = arange_allocator.get();
1340+
1341+
aclTensor * arange_tensor = ggml_cann_create_tensor(
1342+
arange_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
13431343
aclnn_arange(ctx, arange_tensor, start, stop, step, size);
13441344

1345-
aclTensor* slope_tensor = ggml_cann_create_tensor(
1346-
slope_buffer, ACL_FLOAT,
1347-
sizeof(float), ne, nb, 1);
1345+
aclTensor * slope_tensor = ggml_cann_create_tensor(
1346+
slope_buffer, ACL_FLOAT, sizeof(float), ne, nb, 1);
13481347

1349-
aclScalar* sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
1348+
aclScalar * sc = aclCreateScalar(&m, aclDataType::ACL_FLOAT);
13501349

13511350
GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, sc, arange_tensor, slope_tensor);
13521351
ggml_cann_release_resources(ctx, sc, arange_tensor, slope_tensor);
13531352
}
13541353

1355-
static void aclnn_get_slope(ggml_backend_cann_context& ctx, int64_t n_head, void* slope_buffer, float max_bias) {
1356-
const int n_head_log2 = 1u << (uint32_t)floor(log2(n_head));
1354+
static void aclnn_get_slope(ggml_backend_cann_context & ctx, int64_t n_head,
1355+
void * slope_buffer, float max_bias) {
1356+
const int n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
13571357

13581358
float m0 = powf(2.0f, -(max_bias) / n_head_log2);
13591359
float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
13601360

1361-
// const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
1361+
// const float slope = (max_bias > 0.0f) ?
1362+
// h < n_head_log2 ?
1363+
// powf(m0, h + 1) :
1364+
// powf(m1, 2*(h - n_head_log2) + 1) :
1365+
// 1.0f;
13621366
// arange1
13631367
float start = 0 + 1;
1364-
float end = (n_head_log2 - 1) + 1;
1365-
float step = 1;
1368+
float end = (n_head_log2 - 1) + 1;
1369+
float step = 1;
13661370
float count = n_head_log2;
13671371
// end needs to be +1 because aclnn uses a left-closed, right-open interval.
13681372
aclnn_get_slope_inner(ctx, slope_buffer, m0, count, start, end + 1, step);
13691373
if (n_head_log2 < n_head) {
13701374
// arange2
1371-
start = 2*(n_head_log2 - n_head_log2) + 1;
1372-
end = 2*((n_head - 1) - n_head_log2) + 1;
1373-
step = 2;
1375+
start = 2 * (n_head_log2 - n_head_log2) + 1;
1376+
end = 2 * ((n_head - 1) - n_head_log2) + 1;
1377+
step = 2;
13741378
count = n_head - n_head_log2;
1375-
aclnn_get_slope_inner(ctx, (char*)slope_buffer + n_head_log2* sizeof(float), m1, count, start, end + 1, step);
1379+
aclnn_get_slope_inner(
1380+
ctx, (char *) slope_buffer + n_head_log2 * sizeof(float),
1381+
m1, count, start, end + 1, step);
13761382
}
13771383
}
13781384

1379-
static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask, ggml_tensor* dst, void* dst_ptr, float max_bias) {
1385+
static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask,
1386+
ggml_tensor* dst, void* dst_ptr, float max_bias) {
13801387
void* slope_buffer = nullptr;
13811388
void* bias_buffer = nullptr;
13821389

13831390
int64_t n_heads = dst->ne[2];
13841391
ggml_cann_pool_alloc slope_allocator(ctx.pool(), n_heads * sizeof(float));
13851392
slope_buffer = slope_allocator.get();
1386-
ggml_cann_pool_alloc bias_allocator(ctx.pool(), ggml_nelements(dst) * ggml_element_size(dst));
1393+
ggml_cann_pool_alloc bias_allocator(
1394+
ctx.pool(), ggml_nelements(dst) * ggml_element_size(dst));
13871395
bias_buffer = bias_allocator.get();
13881396

13891397
if (max_bias > 0.0f) {
@@ -1396,44 +1404,46 @@ static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask, g
13961404
int64_t nr3 = dst->ne[3] / mask->ne[3];
13971405

13981406
// broadcast the mask across rows
1399-
int64_t mask_ne[] = {mask->ne[0], dst->ne[1], mask->ne[2], 1, mask->ne[3], 1};
1400-
size_t mask_nb[GGML_MAX_DIMS + 2];
1401-
mask_nb[0] = mask->nb[0];
1402-
mask_nb[1] = mask->nb[1];
1403-
mask_nb[2] = mask->nb[2];
1404-
mask_nb[3] = mask->nb[2];
1405-
mask_nb[4] = mask->nb[3];
1406-
mask_nb[5] = mask->nb[3];
1407-
1408-
// ne2 and ne3 may be integer multiples of the mask.
1409-
int64_t dst_ne[] = {dst->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], nr3};
1410-
size_t dst_nb[GGML_MAX_DIMS + 2];
1411-
dst_nb[0] = ggml_element_size(dst);
1412-
for(int i = 1;i<GGML_MAX_DIMS + 2;i++) {
1413-
dst_nb[i] = dst_nb[i-1]* dst_ne[i-1];
1414-
}
1407+
int64_t mask_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], 1, mask->ne[3], 1 };
1408+
size_t mask_nb[] = {
1409+
mask_nb[0] = mask->nb[0], mask_nb[1] = mask->nb[1], mask_nb[2] = mask->nb[2],
1410+
mask_nb[3] = mask->nb[2], mask_nb[4] = mask->nb[3], mask_nb[5] = mask->nb[3]
1411+
};
1412+
1413+
int64_t dst_ne[] = { dst->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], nr3 };
1414+
size_t dst_nb[] = {
1415+
dst_nb[0] = dst->nb[0], dst_nb[1] = dst->nb[1], dst_nb[2] = dst->nb[2],
1416+
dst_nb[3] = dst->nb[2], dst_nb[4] = dst->nb[3], dst_nb[5] = dst->nb[3]
1417+
};
14151418

14161419
// slope is a 1 dim tensor, slope.ne2 == dst.ne2
1417-
int64_t slope_ne[] = {1, 1, mask->ne[2], nr2, 1, 1};
1418-
size_t slope_nb[GGML_MAX_DIMS + 2];
1420+
int64_t slope_ne[] = { 1, 1, mask->ne[2], nr2, 1, 1 };
1421+
size_t slope_nb[GGML_MAX_DIMS + 2];
14191422
slope_nb[0] = sizeof(float);
1420-
for(int i = 1;i<GGML_MAX_DIMS + 2;i++) {
1421-
slope_nb[i] = slope_nb[i-1]* slope_ne[i-1];
1423+
for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
1424+
slope_nb[i] = slope_nb[i - 1] * slope_ne[i - 1];
14221425
}
14231426

1424-
aclTensor* acl_slope = ggml_cann_create_tensor(slope_buffer, ACL_FLOAT, sizeof(float), slope_ne, slope_nb, GGML_MAX_DIMS + 2);
1425-
aclTensor* acl_mask = ggml_cann_create_tensor(mask, mask_ne, mask_nb, GGML_MAX_DIMS + 2);
1426-
aclTensor* acl_dst = ggml_cann_create_tensor(dst_ptr, ggml_cann_type_mapping(dst->type),
1427-
ggml_type_size(dst->type), dst_ne, dst_nb, GGML_MAX_DIMS + 2);
1428-
1427+
aclTensor * acl_slope = ggml_cann_create_tensor(
1428+
slope_buffer, ACL_FLOAT, sizeof(float),
1429+
slope_ne, slope_nb, GGML_MAX_DIMS + 2);
1430+
aclTensor * acl_mask = ggml_cann_create_tensor(
1431+
mask, mask_ne, mask_nb, GGML_MAX_DIMS + 2);
1432+
aclTensor * acl_dst = ggml_cann_create_tensor(
1433+
dst_ptr, ggml_cann_type_mapping(dst->type),
1434+
ggml_type_size(dst->type), dst_ne, dst_nb,
1435+
GGML_MAX_DIMS + 2);
1436+
14291437
if (max_bias > 0.0f) {
1430-
int64_t bias_ne[] = {mask->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], 1};
1431-
size_t bias_nb[GGML_MAX_DIMS + 2];
1438+
int64_t bias_ne[] = { mask->ne[0], dst->ne[1], mask->ne[2], nr2, mask->ne[3], 1 };
1439+
size_t bias_nb[GGML_MAX_DIMS + 2];
14321440
bias_nb[0] = sizeof(float);
1433-
for(int i = 1;i<GGML_MAX_DIMS + 2;i++) {
1434-
bias_nb[i] = bias_nb[i-1]* bias_ne[i-1];
1441+
for (int i = 1; i < GGML_MAX_DIMS + 2; i++) {
1442+
bias_nb[i] = bias_nb[i - 1] * bias_ne[i - 1];
14351443
}
1436-
aclTensor* bias_tensor = ggml_cann_create_tensor(bias_buffer, ACL_FLOAT, sizeof(float), bias_ne, bias_nb, GGML_MAX_DIMS + 2);
1444+
aclTensor * bias_tensor = ggml_cann_create_tensor(
1445+
bias_buffer, ACL_FLOAT, sizeof(float),
1446+
bias_ne, bias_nb, GGML_MAX_DIMS + 2);
14371447

14381448
aclnn_mul(ctx, acl_slope, acl_mask, bias_tensor);
14391449
aclnn_add(ctx, acl_dst, bias_tensor);
@@ -1444,7 +1454,7 @@ static void aclnn_add_alibi(ggml_backend_cann_context& ctx, ggml_tensor* mask, g
14441454
ggml_cann_release_resources(ctx, acl_slope, acl_mask, acl_dst);
14451455
}
14461456

1447-
void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
1457+
void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
14481458
ggml_cann_dup(ctx, dst);
14491459
}
14501460

@@ -1462,31 +1472,31 @@ void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
14621472
* @param acl_dst The destination tensor where the softmax results will be
14631473
* stored.
14641474
*/
1465-
static void aclnn_softmax(ggml_backend_cann_context& ctx, aclTensor* acl_src,
1466-
int64_t dim, aclTensor* acl_dst) {
1475+
static void aclnn_softmax(ggml_backend_cann_context & ctx,
1476+
aclTensor * acl_src, int64_t dim, aclTensor * acl_dst) {
14671477
GGML_CANN_CALL_ACLNN_OP(ctx, Softmax, acl_src, dim, acl_dst);
14681478
}
14691479

1470-
void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
1471-
ggml_tensor* src0 = dst->src[0];
1472-
ggml_tensor* src1 = dst->src[1]; // mask
1480+
void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
1481+
ggml_tensor * src0 = dst->src[0];
1482+
ggml_tensor * src1 = dst->src[1]; // mask
14731483

1474-
aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
1475-
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
1484+
aclTensor * acl_src0 = ggml_cann_create_tensor(src0);
1485+
aclTensor * acl_dst = ggml_cann_create_tensor(dst);
14761486

1477-
float scale = 1.0f;
1487+
float scale = 1.0f;
14781488
float max_bias = 0.0f;
14791489

1480-
memcpy(&scale, (float*)dst->op_params + 0, sizeof(float));
1481-
memcpy(&max_bias, (float*)dst->op_params + 1, sizeof(float));
1490+
memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
1491+
memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
14821492

14831493
// input mul scale
1484-
aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT);
1494+
aclScalar * acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT);
14851495
ggml_cann_pool_alloc src_tensor_allocator(ctx.pool(), ggml_nbytes(src0));
14861496
void* src_tensor_buffer = src_tensor_allocator.get();
14871497
aclTensor* softmax_tensor = ggml_cann_create_tensor(
1488-
src_tensor_buffer, ggml_cann_type_mapping(src0->type), ggml_element_size(src0), src0->ne,
1489-
src0->nb, GGML_MAX_DIMS);
1498+
src_tensor_buffer, ggml_cann_type_mapping(src0->type),
1499+
ggml_element_size(src0), src0->ne, src0->nb,GGML_MAX_DIMS);
14901500

14911501
aclnn_muls(ctx, acl_src0, scale, softmax_tensor, false);
14921502

@@ -1496,8 +1506,7 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
14961506
}
14971507
// softmax
14981508
aclnn_softmax(ctx, softmax_tensor, 3, acl_dst);
1499-
ggml_cann_release_resources(ctx, acl_src0, acl_dst,
1500-
acl_scale, softmax_tensor);
1509+
ggml_cann_release_resources(ctx, acl_src0, acl_dst, acl_scale, softmax_tensor);
15011510
}
15021511

15031512
/**

‎tests/test-backend-ops.cpp

Lines changed: 20 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -5461,26 +5461,26 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
54615461
test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, false));
54625462
test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, true));
54635463

5464-
// for(uint32_t Cout : {1, 9}){
5465-
// for(uint32_t Cin : {1, 7}){
5466-
// for(uint32_t K : {1, 3, 1337}){
5467-
// for(uint32_t L : {1, 2, 13}){
5468-
// for(uint32_t s0: {1, 2, 3}){
5469-
// test_cases.emplace_back(new test_conv_transpose_1d({L,Cin,1,1}, {K,Cout,Cin,1}, s0, 0, 1));
5470-
// }
5471-
// }
5472-
// }
5473-
// }
5474-
// }
5475-
5476-
// test_cases.emplace_back(new test_conv_transpose_1d());
5477-
// test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
5478-
// test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1));
5479-
// test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 1, 0, 1));
5480-
// test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 2, 0, 1));
5481-
// test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 1, 0, 1));
5482-
// test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1));
5483-
// test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1));
5464+
for(uint32_t Cout : {1, 9}){
5465+
for(uint32_t Cin : {1, 7}){
5466+
for(uint32_t K : {1, 3, 1337}){
5467+
for(uint32_t L : {1, 2, 13}){
5468+
for(uint32_t s0: {1, 2, 3}){
5469+
test_cases.emplace_back(new test_conv_transpose_1d({L,Cin,1,1}, {K,Cout,Cin,1}, s0, 0, 1));
5470+
}
5471+
}
5472+
}
5473+
}
5474+
}
5475+
5476+
test_cases.emplace_back(new test_conv_transpose_1d());
5477+
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
5478+
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1));
5479+
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 1, 0, 1));
5480+
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 2, 0, 1));
5481+
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,2,2,1}, 1, 0, 1));
5482+
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1));
5483+
test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1));
54845484

54855485
test_cases.emplace_back(new test_conv_transpose_2d({3, 2, 3, 1}, {2, 2, 1, 3}, 1));
54865486
test_cases.emplace_back(new test_conv_transpose_2d({10, 10, 9, 1}, {3, 3, 1, 9}, 2));

0 commit comments

Comments
 (0)