Skip to content
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

[QualcommQnn] add ops #9538

Merged
merged 3 commits into from
Oct 17, 2022
Merged

[QualcommQnn] add ops #9538

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e9bf67e
[QualcommQnn] add ops
zhupengyang Oct 11, 2022
9e05682
Merge branch 'develop' into qnn_add_ops_2
zhupengyang Oct 12, 2022
9b1e262
test
zhupengyang Oct 12, 2022
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
54 changes: 27 additions & 27 deletions lite/backends/nnadapter/nnadapter/include/nnadapter/operation/lp_normalization.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,33 +17,33 @@
namespace nnadapter {
namespace operation {

#define LP_NORMALIZATION_OPERATION_EXTRACT_INPUTS_OUTPUTS \
auto& input_operands = operation->input_operands; \
auto& output_operands = operation->output_operands; \
auto input_count = input_operands.size(); \
auto output_count = output_operands.size(); \
NNADAPTER_CHECK_EQ(input_count, 4); \
NNADAPTER_CHECK_EQ(output_count, 1); \
/* Input */ \
auto input_operand = input_operands[0]; \
NNADAPTER_VLOG(5) << "input: " << OperandToString(input_operand); \
/* Output */ \
auto output_operand = output_operands[0]; \
NNADAPTER_VLOG(5) << "output: " << OperandToString(output_operand); \
/* Axis */ \
auto axis_operand = input_operands[1]; \
auto axis_count = axis_operand->length / sizeof(int32_t); \
auto axis_data = reinterpret_cast<int32_t*>(axis_operand->buffer); \
for (uint32_t i = 0; i < axis_count; i++) { \
NNADAPTER_VLOG(5) << "axis[" << i << "]=" << axis_data[i]; \
} \
/* P */ \
auto p = *reinterpret_cast<int32_t*>(input_operands[2]->buffer); \
NNADAPTER_VLOG(5) << "p: " << p; \
NNADAPTER_CHECK(p == 1 || p == 2) \
<< "lp normalization only support p = 1 or p = 2."; \
/* Epsilon */ \
auto epsilon = *reinterpret_cast<float*>(input_operands[3]->buffer); \
#define LP_NORMALIZATION_OPERATION_EXTRACT_INPUTS_OUTPUTS \
auto& input_operands = operation->input_operands; \
auto& output_operands = operation->output_operands; \
auto input_count = input_operands.size(); \
auto output_count = output_operands.size(); \
NNADAPTER_CHECK_EQ(input_count, 4); \
NNADAPTER_CHECK_EQ(output_count, 1); \
/* Input */ \
auto input_operand = input_operands[0]; \
NNADAPTER_VLOG(5) << "input: " << OperandToString(input_operand); \
/* Output */ \
auto output_operand = output_operands[0]; \
NNADAPTER_VLOG(5) << "output: " << OperandToString(output_operand); \
/* Axis */ \
auto axis_operand = input_operands[1]; \
auto axis_count = axis_operand->length / sizeof(int32_t); \
auto axis_data = reinterpret_cast<int32_t*>(axis_operand->buffer); \
for (uint32_t i = 0; i < axis_count; i++) { \
NNADAPTER_VLOG(5) << "axis[" << i << "]=" << axis_data[i]; \
} \
/* P */ \
auto p = *reinterpret_cast<int32_t*>(input_operands[2]->buffer); \
NNADAPTER_VLOG(5) << "p: " << p; \
NNADAPTER_CHECK(p == 1 || p == 2) \
<< "lp normalization only support p = 1 or p = 2. Received p is " << p; \
/* Epsilon */ \
auto epsilon = *reinterpret_cast<float*>(input_operands[3]->buffer); \
NNADAPTER_VLOG(5) << "epsilon: " << epsilon;

} // namespace operation
Expand Down
11 changes: 7 additions & 4 deletions ...ackends/nnadapter/nnadapter/include/nnadapter/optimizer/convert_datalayout_nchw_to_nhwc.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,29 +36,33 @@ class NCHW2NHWCDataLayoutConverter {

private:
// Operation converters
void ConvertActivation(core::Operation* operation);
void ConvertAdaptivePool2D(core::Operation* operation);
void ConvertArgMinMax(core::Operation* operation);
void ConvertBatchNormalization(core::Operation* operation);
void ConvertCast(core::Operation* operation);
void ConvertChannelShuffle(core::Operation* operation);
void ConvertClip(core::Operation* operation);
void ConvertComparisons(core::Operation* operation);
void ConvertConcat(core::Operation* operation);
void ConvertCumSum(core::Operation* operation);
void ConvertDequantize(core::Operation* operation);
void ConvertElementwise(core::Operation* operation);
void ConvertPool2D(core::Operation* operation);
void ConvertConcat(core::Operation* operation);
void ConvertFill(core::Operation* operation);
void ConvertFillLike(core::Operation* operation);
void ConvertFlatten(core::Operation* operation);
void ConvertFullyConnected(core::Operation* operation);
void ConvertGather(core::Operation* operation);
void ConvertGelu(core::Operation* operation);
void ConvertInstanceNormalization(core::Operation* operation);
void ConvertLayerNormalization(core::Operation* operation);
void ConvertLeakyRelu(core::Operation* operation);
void ConvertLpNormalization(core::Operation* operation);
void ConvertActivation(core::Operation* operation);
void ConvertMatMul(core::Operation* operation);
void ConvertPad(core::Operation* operation);
void ConvertPow(core::Operation* operation);
void ConvertPool2D(core::Operation* operation);
void ConvertPrelu(core::Operation* operation);
void ConvertQuantize(core::Operation* operation);
void ConvertReduce(core::Operation* operation);
void ConvertReshape(core::Operation* operation);
Expand All @@ -73,7 +77,6 @@ class NCHW2NHWCDataLayoutConverter {
void ConvertStack(core::Operation* operation);
void ConvertTile(core::Operation* operation);
void ConvertTranspose(core::Operation* operation);
void ConvertMatMul(core::Operation* operation);
void ConvertUnsqueeze(core::Operation* operation);
void ConvertUnstack(core::Operation* operation);

Expand Down
108 changes: 108 additions & 0 deletions lite/backends/nnadapter/nnadapter/src/optimizer/convert_datalayout_nchw_to_nhwc.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,6 +171,46 @@ void NCHW2NHWCDataLayoutConverter::ConvertAdaptivePool2D(
SetOperationLayout(operation);
}

void NCHW2NHWCDataLayoutConverter::ConvertArgMinMax(
core::Operation* operation) {
auto& input_operands = operation->input_operands;
auto& output_operands = operation->output_operands;
auto input_count = input_operands.size();
auto output_count = output_operands.size();
NNADAPTER_CHECK_EQ(input_count, 4);
NNADAPTER_CHECK_EQ(output_count, 1);
auto input_operand = input_operands[0];
auto input_dimensions_count = input_operand->type.dimensions.count;
auto axis = reinterpret_cast<int32_t*>(input_operands[1]->buffer);
if (*axis < 0) {
*axis += input_dimensions_count;
}
auto keepdim = *reinterpret_cast<bool*>(input_operands[2]->buffer);
auto output_operand = output_operands[0];
auto input_permutation = GetPermutation(input_operand);
if (keepdim) {
// Recalculate the axis according to the dimorder vector of the input
// operand
*axis = TransposeAxis(*axis, input_permutation);
TransposeOperand(output_operand, input_permutation);
SetPermutation(output_operand, input_permutation);
} else {
// Force to restore the dimorder vector of the input operand
auto transpose_input_permutation = InversePermutation(input_permutation);
if (!IsIdentityPermutation(transpose_input_permutation)) {
auto transpose_input_operand = AppendTransposeOperation(
model_, input_operand, transpose_input_permutation);
UpdateOperationInputOperands(
{operation}, input_operand, transpose_input_operand);
SetPermutation(transpose_input_operand,
IdentityPermutation(input_dimensions_count));
}
int output_dimensions_count = output_operand->type.dimensions.count;
SetPermutation(output_operand,
IdentityPermutation(output_dimensions_count));
}
}

void NCHW2NHWCDataLayoutConverter::ConvertBatchNormalization(
core::Operation* operation) {
auto& input_operands = operation->input_operands;
Expand Down Expand Up @@ -704,6 +744,38 @@ void NCHW2NHWCDataLayoutConverter::ConvertPow(core::Operation* operation) {
SetPermutation(output_operand, input_permutation);
}

void NCHW2NHWCDataLayoutConverter::ConvertPrelu(core::Operation* operation) {
auto& input_operands = operation->input_operands;
auto& output_operands = operation->output_operands;
auto input_count = input_operands.size();
auto output_count = output_operands.size();
NNADAPTER_CHECK_EQ(input_count, 2);
NNADAPTER_CHECK_EQ(output_count, 1);
auto input_operand = input_operands[0];
auto output_operand = output_operands[0];
auto input_permutation = GetPermutation(input_operand);
auto slope_size = input_operands[1]->type.dimensions.data[0];
if (slope_size == 1) {
// The input and output operands share the same dimorder vector
TransposeOperand(output_operand, input_permutation);
SetPermutation(output_operand, input_permutation);
} else {
// Force to apply the dimorder vector of NCHW2NHWC conversion
auto transpose_input_permutation =
MultiplyPermutation(InversePermutation(input_permutation), kNCHW2NHWC);
if (!IsIdentityPermutation(transpose_input_permutation)) {
auto transpose_input_operand = AppendTransposeOperation(
model_, input_operand, transpose_input_permutation);
UpdateOperationInputOperands(
{operation}, input_operand, transpose_input_operand);
SetPermutation(transpose_input_operand, kNCHW2NHWC);
}
TransposeOperand(output_operand, kNCHW2NHWC);
SetPermutation(output_operand, kNCHW2NHWC);
SetOperationLayout(operation);
}
}

void NCHW2NHWCDataLayoutConverter::ConvertQuantize(core::Operation* operation) {
auto& input_operands = operation->input_operands;
auto& output_operands = operation->output_operands;
Expand Down Expand Up @@ -921,6 +993,32 @@ void NCHW2NHWCDataLayoutConverter::ConvertFlatten(core::Operation* operation) {
SetPermutation(output_operand, IdentityPermutation(output_dimensions_count));
}

void NCHW2NHWCDataLayoutConverter::ConvertInstanceNormalization(
core::Operation* operation) {
auto& input_operands = operation->input_operands;
auto& output_operands = operation->output_operands;
auto input_count = input_operands.size();
auto output_count = output_operands.size();
NNADAPTER_CHECK_EQ(input_count, 4);
NNADAPTER_CHECK_EQ(output_count, 1);
auto input_operand = input_operands[0];
auto output_operand = output_operands[0];
// Force to apply the dimorder vector of NCHW2NHWC conversion
auto input_permutation = GetPermutation(input_operand);
auto transpose_input_permutation =
MultiplyPermutation(InversePermutation(input_permutation), kNCHW2NHWC);
if (!IsIdentityPermutation(transpose_input_permutation)) {
auto transpose_input_operand = AppendTransposeOperation(
model_, input_operand, transpose_input_permutation);
UpdateOperationInputOperands(
{operation}, input_operand, transpose_input_operand);
SetPermutation(transpose_input_operand, kNCHW2NHWC);
}
TransposeOperand(output_operand, kNCHW2NHWC);
SetPermutation(output_operand, kNCHW2NHWC);
SetOperationLayout(operation);
}

void NCHW2NHWCDataLayoutConverter::ConvertLayerNormalization(
core::Operation* operation) {
auto& input_operands = operation->input_operands;
Expand Down Expand Up @@ -1297,6 +1395,10 @@ void NCHW2NHWCDataLayoutConverter::Apply(core::Model* model) {
case NNADAPTER_ADAPTIVE_AVERAGE_POOL_2D:
ConvertAdaptivePool2D(operation);
break;
case NNADAPTER_ARG_MAX:
case NNADAPTER_ARG_MIN:
ConvertArgMinMax(operation);
break;
case NNADAPTER_AVERAGE_POOL_2D:
case NNADAPTER_MAX_POOL_2D:
ConvertPool2D(operation);
Expand Down Expand Up @@ -1354,6 +1456,9 @@ void NCHW2NHWCDataLayoutConverter::Apply(core::Model* model) {
case NNADAPTER_GELU:
ConvertGelu(operation);
break;
case NNADAPTER_INSTANCE_NORMALIZATION:
ConvertInstanceNormalization(operation);
break;
case NNADAPTER_LAYER_NORMALIZATION:
ConvertLayerNormalization(operation);
break;
Expand All @@ -1372,6 +1477,9 @@ void NCHW2NHWCDataLayoutConverter::Apply(core::Model* model) {
case NNADAPTER_POW:
ConvertPow(operation);
break;
case NNADAPTER_PRELU:
ConvertPrelu(operation);
break;
case NNADAPTER_QUANTIZE:
ConvertQuantize(operation);
break;
Expand Down
5 changes: 5 additions & 0 deletions lite/backends/nnadapter/nnadapter/src/optimizer/convert_quantization_symm_to_asymm.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,6 +101,8 @@ NNADAPTER_EXPORT void ConvertQuantizationSymmToAsymm(core::Model* model) {
ConvertOperandSymmToAsymm(output_operands[0], 128);
} break;
case NNADAPTER_ABS:
case NNADAPTER_ARG_MAX:
case NNADAPTER_ARG_MIN:
case NNADAPTER_AVERAGE_POOL_2D:
case NNADAPTER_BATCH_NORMALIZATION:
case NNADAPTER_CAST:
Expand All @@ -114,11 +116,14 @@ NNADAPTER_EXPORT void ConvertQuantizationSymmToAsymm(core::Model* model) {
case NNADAPTER_GELU:
case NNADAPTER_HARD_SIGMOID:
case NNADAPTER_HARD_SWISH:
case NNADAPTER_INSTANCE_NORMALIZATION:
case NNADAPTER_LAYER_NORMALIZATION:
case NNADAPTER_LEAKY_RELU:
case NNADAPTER_LOG:
case NNADAPTER_LP_NORMALIZATION:
case NNADAPTER_MAX_POOL_2D:
case NNADAPTER_PAD:
case NNADAPTER_PRELU:
case NNADAPTER_REDUCE_MAX:
case NNADAPTER_REDUCE_MEAN:
case NNADAPTER_REDUCE_SUM:
Expand Down
Loading