Skip to content

Commit e09a800

Browse files
MengqingCaoMengqingCao
authored
cann: Fix ggml_cann_im2col for 1D im2col (#8819)
* fix ggml_cann_im2col for 1D im2col * fix build warning
1 parent 0fbbd88 commit e09a800

File tree

2 files changed

+141
-27
lines changed

2 files changed

+141
-27
lines changed

‎ggml/src/ggml-cann/aclnn_ops.cpp

Lines changed: 138 additions & 27 deletions
Original file line numberDiff line numberDiff line change
@@ -1312,6 +1312,111 @@ aclnnStatus aclnnIm2col(void* workspace, uint64_t workspaceSize,
13121312
#ifdef __cplusplus
13131313
}
13141314
#endif
1315+
1316+
static void ggml_cann_im2col_2d_post_process(ggml_backend_cann_context& ctx,
1317+
ggml_tensor* dst,
1318+
ggml_tensor* src1,
1319+
aclTensor* tmp_cast_tensor,
1320+
aclTensor* tmp_im2col_tensor) {
1321+
// Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
1322+
int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
1323+
size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
1324+
aclTensor* acl_dst =
1325+
ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
1326+
1327+
int64_t permute_dim[] = {0, 2, 1};
1328+
if (src1->type != dst->type) {
1329+
aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
1330+
} else {
1331+
aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
1332+
}
1333+
1334+
// release
1335+
ACL_CHECK(aclDestroyTensor(acl_dst));
1336+
}
1337+
1338+
static void ggml_cann_im2col_1d_post_process(
1339+
ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_tensor* src1,
1340+
aclTensor* tmp_cast_tensor, aclTensor* tmp_im2col_tensor,
1341+
const std::vector<int64_t>& im2col_op_params) {
1342+
// get params
1343+
const int64_t KH = im2col_op_params[0];
1344+
const int64_t KW = im2col_op_params[1];
1345+
const int64_t IW = im2col_op_params[2];
1346+
const int64_t IC = im2col_op_params[3];
1347+
const int64_t N = im2col_op_params[4];
1348+
const int64_t OH = im2col_op_params[5];
1349+
const int64_t OW = im2col_op_params[6];
1350+
const int64_t s0 = im2col_op_params[7];
1351+
const int64_t p0 = im2col_op_params[8];
1352+
const int64_t d0 = im2col_op_params[9];
1353+
const int64_t n_bytes_factor = im2col_op_params[10];
1354+
1355+
// Permute: [N, IC * KH * KW, OW * OH] ->
1356+
// [N, OW * OH * n_bytes_factor, IC * KH * KW]
1357+
aclTensor* tmp_permute_tensor = nullptr;
1358+
ggml_cann_pool_alloc tmp_permute_allocator(ctx.pool());
1359+
tmp_permute_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
1360+
void* tmp_permute_buffer = tmp_permute_allocator.get();
1361+
1362+
int64_t tmp_permute_ne[] = {IC * KH * KW, OW * OH * n_bytes_factor, N};
1363+
size_t tmp_permute_nb[GGML_MAX_DIMS - 1];
1364+
tmp_permute_nb[0] = ggml_type_size(dst->type);
1365+
for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
1366+
tmp_permute_nb[i] = tmp_permute_nb[i - 1] * tmp_permute_ne[i - 1];
1367+
}
1368+
1369+
tmp_permute_tensor = ggml_cann_create_tensor(
1370+
tmp_permute_buffer, ggml_cann_type_mapping(dst->type),
1371+
ggml_type_size(dst->type), tmp_permute_ne, tmp_permute_nb,
1372+
GGML_MAX_DIMS - 1, ACL_FORMAT_ND);
1373+
1374+
int64_t permute_dim[] = {0, 2, 1};
1375+
if (src1->type != dst->type) {
1376+
aclnn_permute(ctx, tmp_cast_tensor, tmp_permute_tensor, permute_dim, 3);
1377+
} else {
1378+
aclnn_permute(ctx, tmp_im2col_tensor, tmp_permute_tensor, permute_dim,
1379+
3);
1380+
}
1381+
1382+
// number of times the kernel moves in W dimension
1383+
const int n_step_w = (IW + 2 * p0 - d0 * (KW - 1) - 1) / s0 + 1;
1384+
size_t offset;
1385+
void *cur_dst_buffer = dst->data, *cur_permute_buffer = tmp_permute_buffer;
1386+
1387+
// memory copy with offset to restore 1D im2col from 2d
1388+
if (IC > 1) {
1389+
offset = IC * KH * KW * n_step_w * ggml_type_size(dst->type);
1390+
size_t size_cpy = KH * KW * ggml_type_size(dst->type);
1391+
1392+
for (int c = 0; c < IC; c++) {
1393+
cur_permute_buffer = (char*)tmp_permute_buffer + offset +
1394+
KH * KW * c * ggml_type_size(dst->type);
1395+
cur_dst_buffer = (char*)dst->data +
1396+
c * KH * KW * n_step_w * ggml_type_size(dst->type);
1397+
1398+
for (int i = 0; i < n_step_w; i++) {
1399+
ACL_CHECK(aclrtMemcpyAsync(
1400+
cur_dst_buffer, size_cpy, cur_permute_buffer, size_cpy,
1401+
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
1402+
cur_dst_buffer =
1403+
(char*)cur_dst_buffer + KH * KW * ggml_type_size(dst->type);
1404+
cur_permute_buffer = (char*)cur_permute_buffer +
1405+
KH * KW * IC * ggml_type_size(dst->type);
1406+
}
1407+
}
1408+
} else {
1409+
offset = KH * KW * n_step_w *
1410+
ggml_type_size(dst->type); // equal to ggml_nbytes(dst)
1411+
ACL_CHECK(aclrtMemcpyAsync(dst->data, offset,
1412+
(char*)tmp_permute_buffer + offset, offset,
1413+
ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
1414+
}
1415+
1416+
// release
1417+
ACL_CHECK(aclDestroyTensor(tmp_permute_tensor));
1418+
}
1419+
13151420
void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
13161421
ggml_tensor* src0 = dst->src[0]; // kernel
13171422
ggml_tensor* src1 = dst->src[1]; // input
@@ -1320,31 +1425,36 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
13201425
GGML_ASSERT(src1->type == GGML_TYPE_F32);
13211426
GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
13221427

1428+
GGML_TENSOR_BINARY_OP_LOCALS;
1429+
1430+
// aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
1431+
// im2col and do post-processing to restore it to 1D.
1432+
const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
13231433
const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
1324-
const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
1434+
const int32_t s1 = is_2D ? ((const int32_t*)(dst->op_params))[1] : 1;
13251435
const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
1326-
const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
1436+
const int32_t p1 = is_2D ? ((const int32_t*)(dst->op_params))[3] : 1;
13271437
const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
1328-
const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
1329-
const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
1330-
1331-
GGML_TENSOR_BINARY_OP_LOCALS;
1332-
1333-
const int64_t N = is_2D ? ne13 : ne12;
1334-
const int64_t IC = is_2D ? ne12 : ne11;
1438+
const int32_t d1 = is_2D ? ((const int32_t*)(dst->op_params))[5] : 1;
13351439

1336-
const int64_t KH = is_2D ? ne01 : 1;
1440+
const int64_t N = ne13;
1441+
const int64_t IC = ne12;
1442+
const int64_t KH = ne01;
13371443
const int64_t KW = ne00;
1444+
const int64_t IW = ne10;
13381445

13391446
const int64_t OH = is_2D ? ne2 : 1;
13401447
const int64_t OW = ne1;
13411448

13421449
GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
13431450
GGML_ASSERT(nb10 == sizeof(float));
13441451

1345-
// im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH]
1452+
// memory allocated increased to 3x when is_2D == false
1453+
const int64_t n_bytes_factor = is_2D ? 1 : 3;
1454+
1455+
// im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH * n_bytes_factor]
13461456
aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
1347-
int64_t tmp_im2col_ne[] = {OW * OH, IC * KH * KW, N};
1457+
int64_t tmp_im2col_ne[] = {OW * OH * n_bytes_factor, IC * KH * KW, N};
13481458
size_t tmp_im2col_nb[GGML_MAX_DIMS - 1];
13491459

13501460
tmp_im2col_nb[0] = ggml_type_size(src1->type);
@@ -1356,8 +1466,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
13561466
// If dst is f16, tmp_buffer is f32, we need alloc src.typesize *
13571467
// dst.elemcount.
13581468
ggml_cann_pool_alloc im2col_allocator(
1359-
ctx.pool(), ggml_nelements(dst) * ggml_element_size(src1));
1469+
ctx.pool(),
1470+
ggml_nelements(dst) * ggml_element_size(src1) * n_bytes_factor);
13601471
void* tmp_im2col_buffer = im2col_allocator.get();
1472+
13611473
aclTensor* tmp_im2col_tensor = ggml_cann_create_tensor(
13621474
tmp_im2col_buffer, ggml_cann_type_mapping(src1->type),
13631475
ggml_type_size(src1->type), tmp_im2col_ne, tmp_im2col_nb,
@@ -1380,8 +1492,9 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
13801492
paddings, strides, tmp_im2col_tensor,
13811493
&workspaceSize, &executor));
13821494

1495+
ggml_cann_pool_alloc workspace_allocator(ctx.pool());
13831496
if (workspaceSize > 0) {
1384-
ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
1497+
workspace_allocator.alloc(workspaceSize);
13851498
workspaceAddr = workspace_allocator.get();
13861499
}
13871500

@@ -1391,9 +1504,10 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
13911504
// Cast if dst is f16.
13921505
aclTensor* tmp_cast_tensor = nullptr;
13931506
ggml_cann_pool_alloc tmp_cast_allocator(ctx.pool());
1507+
void* tmp_cast_buffer = nullptr;
13941508
if (src1->type != dst->type) {
1395-
tmp_cast_allocator.alloc(ggml_nbytes(dst));
1396-
void* tmp_cast_buffer = tmp_cast_allocator.get();
1509+
tmp_cast_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
1510+
tmp_cast_buffer = tmp_cast_allocator.get();
13971511
size_t temp_cast_nb[GGML_MAX_DIMS - 1];
13981512
temp_cast_nb[0] = ggml_type_size(dst->type);
13991513
for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
@@ -1408,24 +1522,21 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
14081522
ggml_cann_type_mapping(dst->type));
14091523
}
14101524

1411-
// Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
1412-
int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
1413-
size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
1414-
aclTensor* acl_dst =
1415-
ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
1416-
1417-
int64_t permute_dim[] = {0, 2, 1};
1418-
if (src1->type != dst->type) {
1419-
aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
1525+
// post-processing
1526+
if (is_2D) {
1527+
ggml_cann_im2col_2d_post_process(ctx, dst, src1, tmp_cast_tensor,
1528+
tmp_im2col_tensor);
14201529
} else {
1421-
aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
1530+
std::vector<int64_t> im2col_op_params = {
1531+
KH, KW, IW, IC, N, OH, OW, s0, p0, d0, n_bytes_factor};
1532+
ggml_cann_im2col_1d_post_process(ctx, dst, src1, tmp_cast_tensor,
1533+
tmp_im2col_tensor, im2col_op_params);
14221534
}
14231535

14241536
// release
14251537
ACL_CHECK(aclDestroyTensor(acl_src1));
14261538
ACL_CHECK(aclDestroyTensor(tmp_im2col_tensor));
14271539
ACL_CHECK(aclDestroyTensor(tmp_cast_tensor));
1428-
ACL_CHECK(aclDestroyTensor(acl_dst));
14291540
ACL_CHECK(aclDestroyIntArray(kernel_size));
14301541
ACL_CHECK(aclDestroyIntArray(dilations));
14311542
ACL_CHECK(aclDestroyIntArray(paddings));

‎tests/test-backend-ops.cpp

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2139,6 +2139,9 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
21392139

21402140
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32));
21412141
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16));
2142+
// test cases for 1D im2col
2143+
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
2144+
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
21422145

21432146
test_cases.emplace_back(new test_conv_transpose_1d());
21442147
test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));

0 commit comments

Comments
 (0)