Skip to content

fix index_put #1337

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
NeosZhang wants to merge 7 commits into
base: main
Choose a base branch
from
Open

fix index_put #1337

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
8f92bfc
fix index_put
NeosZhang Aug 8, 2024
f713c69
fix
NeosZhang Aug 8, 2024
903378a
fix device_config
NeosZhang Aug 8, 2024
d1ea96b
fix check for uint8
NeosZhang Aug 9, 2024
288dd2d
fix
NeosZhang Aug 9, 2024
e0106d9
Merge branch 'main' into zhangqiu/update_index_put
NeosZhang Aug 9, 2024
fa40333
Merge branch 'main' into zhangqiu/update_index_put
NeosZhang Sep 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion adaptor/codegen/gen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -196,7 +196,7 @@ def prepare() -> Tuple[dict, str]:

impl_plugin = options.impl_plugin
base_device = options.base_device
assert(base_device is None or base_device == "" or base_device == "torch", f"invalid base_device:{base_device}")
assert base_device is None or base_device == "" or base_device == "torch", f"invalid base_device:{base_device}"
if base_device == "":
base_device = None
def create_if_not_exist(name):
Expand Down
13 changes: 12 additions & 1 deletion impl/ascend/device_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -911,15 +911,26 @@
para=dict(
accumulate=[Skip(False),],
),
tensor_para=dict(
args=[
{
"ins": ['input'],
"shape": [Skip((16, 4, 4)),],
},
]
),
),

'index_put_bool_indices_value': dict( # llm used
name=['index_put'],
para=dict(
accumulate=[Skip(False),],
),
tensor_para=dict(
args=[
{
"ins": ['input'],
"shape": [Skip((3, 2, 2, 20)),],
"shape": [Skip((3, 2, 2, 20)), Skip((4, 2, 2, 6, 2))],
},
]
),
Expand Down
34 changes: 18 additions & 16 deletions impl/ascend/functions/index.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,8 @@
namespace impl {
namespace ascend {

static std::vector<AscendTensor> castIntIndicesToLongIndices(diopiContextHandle_t ctx, std::vector<AscendTensor>& indices) {
namespace indexProcess {
std::vector<AscendTensor> castIntIndicesToLongIndices(diopiContextHandle_t ctx, std::vector<AscendTensor>& indices) {
std::vector<AscendTensor> result;
for (auto& t : indices) {
if (!t.defined()) {
Expand All @@ -37,7 +38,7 @@ static std::vector<AscendTensor> castIntIndicesToLongIndices(diopiContextHandle_
return result;
}

static void checkIndexTensorTypes(const std::vector<AscendTensor>& indices) {
void checkIndexTensorTypes(const std::vector<AscendTensor>& indices) {
for (const auto& t : indices) {
if (t.defined()) {
diopiDtype_t type = t.dtype();
Expand All @@ -47,7 +48,7 @@ static void checkIndexTensorTypes(const std::vector<AscendTensor>& indices) {
}
}

static AscendTensor nonZeroTensor(diopiContextHandle_t ctx, const AscendTensor& self) {
AscendTensor nonZeroTensor(diopiContextHandle_t ctx, const AscendTensor& self) {
int64_t numELem = self.numel() * self.dim();
std::vector<int64_t> nShape{self.numel(), self.dim()};
std::vector<int64_t> nStride(nShape.size(), 1);
Expand Down Expand Up @@ -86,14 +87,14 @@ static AscendTensor nonZeroTensor(diopiContextHandle_t ctx, const AscendTensor&
return AscendTensor(nzTrans);
}

static std::vector<AscendTensor> expandIndicesTensors(diopiContextHandle_t ctx, const AscendTensor& self, const std::vector<AscendTensor>& indices) {
std::vector<AscendTensor> expandIndicesTensors(diopiContextHandle_t ctx, const AscendTensor& self, const std::vector<AscendTensor>& indices) {
std::vector<AscendTensor> result;
for (auto& t : indices) {
if (!t.defined()) {
result.push_back(t);
} else {
if (t.dtype() == diopi_dtype_uint8 || t.dtype() == diopi_dtype_bool) {
ASCEND_CHECK(t.dtype() != diopi_dtype_uint8,
ASCEND_CHECK(t.dtype() == diopi_dtype_bool,
"indexing with dtype torch.uint8 is now deprecated,"
" please use a dtype torch.bool instead.");
for (uint64_t j = 0; j < static_cast<uint64_t>(t.dim()); j++) {
Expand All @@ -117,7 +118,7 @@ static std::vector<AscendTensor> expandIndicesTensors(diopiContextHandle_t ctx,
return result;
}

static aclTensor* createEmptyAclTensor() {
aclTensor* createEmptyAclTensor() {
std::vector<int64_t> nShape{0};
std::vector<int64_t> nStride{1};
int64_t storageSize = 0;
Expand Down Expand Up @@ -152,7 +153,7 @@ static std::vector<AscendTensor> indicesExpandedOutplace(std::vector<AscendTenso
return result;
}

static bool hasContiguousSubspace(std::vector<AscendTensor> indices) { // true if all the non-null tensors are adjacent
bool hasContiguousSubspace(std::vector<AscendTensor> indices) { // true if all the non-null tensors are adjacent
auto isDefined = [](const AscendTensor& tensor) { return tensor.defined(); };
auto isNull = [](const AscendTensor& tensor) { return !tensor.defined(); };
auto start = std::find_if(indices.begin(), indices.end(), isDefined);
Expand All @@ -161,7 +162,7 @@ static bool hasContiguousSubspace(std::vector<AscendTensor> indices) { // true
return it == stop.base();
}

static std::tuple<AscendTensor, std::vector<AscendTensor>> transposeToFront(AscendTensor self, std::vector<AscendTensor> indices) {
std::tuple<AscendTensor, std::vector<AscendTensor>> transposeToFront(AscendTensor self, std::vector<AscendTensor> indices) {
std::vector<int64_t> dims;
std::vector<AscendTensor> transposedIndices;

Expand All @@ -183,7 +184,7 @@ static std::tuple<AscendTensor, std::vector<AscendTensor>> transposeToFront(Asce
return std::make_tuple(self.permute(dims), transposedIndices);
}

static std::vector<int64_t> indexReshape(std::vector<AscendTensor> endIndices, int64_t dimsBefore, int64_t dimsAfter) {
std::vector<int64_t> indexReshape(std::vector<AscendTensor> endIndices, int64_t dimsBefore, int64_t dimsAfter) {
std::vector<int64_t> indexShape;
for (auto& idx : endIndices) {
if (idx.defined()) {
Expand All @@ -201,7 +202,7 @@ static std::vector<int64_t> indexReshape(std::vector<AscendTensor> endIndices, i
return indexShape;
}

static std::vector<int64_t> indexOutputSize(const AscendTensor& self, std::vector<AscendTensor>& indices) {
std::vector<int64_t> indexOutputSize(const AscendTensor& self, std::vector<AscendTensor>& indices) {
std::vector<AscendTensor> midIndices = indicesExpandedOutplace(indices);
while (midIndices.size() < (size_t)self.dim()) {
midIndices.emplace_back(nullptr);
Expand Down Expand Up @@ -254,6 +255,8 @@ static std::vector<int64_t> indexOutputSize(const AscendTensor& self, std::vecto
return outputSize;
}

} // namespace indexProcess

diopiError_t diopiIndex(diopiContextHandle_t ctx, diopiTensorHandle_t* out, diopiConstTensorHandle_t input, diopiConstTensorHandle_t* indices, int64_t nums) {
AscendTensor inputAt(input);
std::vector<AscendTensor> indicesOrigin(nums);
Expand All @@ -263,13 +266,13 @@ diopiError_t diopiIndex(diopiContextHandle_t ctx, diopiTensorHandle_t* out, diop
}
}

std::vector<AscendTensor> indicesList = castIntIndicesToLongIndices(ctx, indicesOrigin);
checkIndexTensorTypes(indicesList);
std::vector<AscendTensor> indicesList = indexProcess::castIntIndicesToLongIndices(ctx, indicesOrigin);
indexProcess::checkIndexTensorTypes(indicesList);

auto indicesExpanded = expandIndicesTensors(ctx, inputAt, indicesList);
auto indicesExpanded = indexProcess::expandIndicesTensors(ctx, inputAt, indicesList);

std::vector<aclTensor*> allDefinedIndices;
auto emptyTensor = createEmptyAclTensor();
auto emptyTensor = indexProcess::createEmptyAclTensor();
for (const auto& idx : indicesExpanded) {
if (idx.defined()) {
allDefinedIndices.push_back(aclnn_adaptor::createAclTensorFromAscendTensor(idx));
Expand All @@ -278,8 +281,7 @@ diopiError_t diopiIndex(diopiContextHandle_t ctx, diopiTensorHandle_t* out, diop
}
}

std::vector<int64_t> outShape = indexOutputSize(inputAt, indicesExpanded);

std::vector<int64_t> outShape = indexProcess::indexOutputSize(inputAt, indicesExpanded);
diopiSize_t outSize = vectorToDiopiSize(outShape);
diopiRequireTensor(ctx, out, &outSize, nullptr, inputAt.dtype(), diopi_device);

Expand Down
62 changes: 58 additions & 4 deletions impl/ascend/functions/index_put.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,18 +9,72 @@

namespace impl {
namespace ascend {

namespace indexProcess {
extern std::vector<AscendTensor> castIntIndicesToLongIndices(diopiContextHandle_t ctx, std::vector<AscendTensor>& indices);
extern void checkIndexTensorTypes(const std::vector<AscendTensor>& indices);
extern std::vector<AscendTensor> expandIndicesTensors(diopiContextHandle_t ctx, const AscendTensor& self, const std::vector<AscendTensor>& indices);
extern aclTensor* createEmptyAclTensor();
} // namespace indexProcess

diopiError_t diopiIndexPut(diopiContextHandle_t ctx, diopiTensorHandle_t out, diopiConstTensorHandle_t input, diopiConstTensorHandle_t values,
diopiConstTensorHandle_t* indices, int64_t indicesCounts, bool accumulate) {
diopiCopyInp(ctx, input, out);
std::vector<diopiConstTensorHandle_t> indicesVec(indices, indices + indicesCounts);
DIOPI_ASCEND_CALL_ACLNN(aclnnIndexPutImpl, ctx, out, indicesVec, values, accumulate, false);
AscendTensor inputAt(input);
AscendTensor valuesAt(values);
if (inputAt.numel() == 0 || valuesAt.numel() == 0) {
return diopiSuccess;
}
std::vector<AscendTensor> indicesOrigin(indicesCounts);
for (int64_t i = 0; i < indicesCounts; i++) {
if (indices[i] != nullptr) {
indicesOrigin[i] = AscendTensor(indices[i]);
}
}
std::vector<AscendTensor> indicesList = indexProcess::castIntIndicesToLongIndices(ctx, indicesOrigin);
indexProcess::checkIndexTensorTypes(indicesList);
auto indicesExpanded = indexProcess::expandIndicesTensors(ctx, inputAt, indicesList);
std::vector<aclTensor*> allDefinedIndices;
auto emptyTensor = indexProcess::createEmptyAclTensor();
for (const auto& idx : indicesExpanded) {
if (idx.defined()) {
allDefinedIndices.push_back(aclnn_adaptor::createAclTensorFromAscendTensor(idx));
} else {
allDefinedIndices.push_back(emptyTensor);
}
}

DIOPI_ASCEND_CALL_ACLNN(aclnnIndexPutImpl, ctx, out, allDefinedIndices, values, accumulate, false);
return diopiSuccess;
}

diopiError_t diopiIndexPutInp(diopiContextHandle_t ctx, diopiTensorHandle_t input, diopiConstTensorHandle_t values, diopiConstTensorHandle_t* indices,
int64_t indicesCounts, bool accumulate) {
std::vector<diopiConstTensorHandle_t> indicesVec(indices, indices + indicesCounts);
DIOPI_ASCEND_CALL_ACLNN(aclnnIndexPutImpl, ctx, input, indicesVec, values, accumulate, false);
AscendTensor inputAt(input);
AscendTensor valuesAt(values);
if (inputAt.numel() == 0 || valuesAt.numel() == 0) {
return diopiSuccess;
}
std::vector<AscendTensor> indicesOrigin(indicesCounts);
for (int64_t i = 0; i < indicesCounts; i++) {
if (indices[i] != nullptr) {
indicesOrigin[i] = AscendTensor(indices[i]);
}
}
std::vector<AscendTensor> indicesList = indexProcess::castIntIndicesToLongIndices(ctx, indicesOrigin);
indexProcess::checkIndexTensorTypes(indicesList);
auto indicesExpanded = indexProcess::expandIndicesTensors(ctx, inputAt, indicesList);
std::vector<aclTensor*> allDefinedIndices;
auto emptyTensor = indexProcess::createEmptyAclTensor();
for (const auto& idx : indicesExpanded) {
if (idx.defined()) {
allDefinedIndices.push_back(aclnn_adaptor::createAclTensorFromAscendTensor(idx));
} else {
allDefinedIndices.push_back(emptyTensor);
}
}

DIOPI_ASCEND_CALL_ACLNN(aclnnIndexPutImpl, ctx, input, allDefinedIndices, values, accumulate, false);
return diopiSuccess;
}

Expand Down