Add Bias to MatMul Kernels and add support for use in Fully Connected Layer
Resolves: [COMPMID-6316]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: I08e6bac9e6b46b76978da0dc6a48ccfe3dde5086
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9833
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp
index 02c5754..a0eb3f2 100644
--- a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp
+++ b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp
@@ -100,7 +100,8 @@
{
_type = CLKernelType::GEMM;
}
-Status ClMatMulLowpNativeKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info, const ActivationLayerInfo &act_info)
+Status ClMatMulLowpNativeKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *bias, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+ const ActivationLayerInfo &act_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs, dst);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
@@ -111,24 +112,32 @@
ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.activation() != ActivationFunction::IDENTITY && act_info.activation() != ActivationFunction::RELU
&& act_info.activation() != ActivationFunction::LU_BOUNDED_RELU && act_info.activation() != ActivationFunction::BOUNDED_RELU),
"Activation Function specified is unsupported.");
+ const TensorShape expected_output_shape = misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info);
if(dst->total_size() != 0)
{
- const TensorInfo tensor_info_output = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info));
+ const TensorInfo tensor_info_output = dst->clone()->set_tensor_shape(expected_output_shape);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, dst);
}
+ if(bias != nullptr)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
+ ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+ ARM_COMPUTE_RETURN_ERROR_ON(expected_output_shape[0] != bias->dimension(0));
+ }
+
return Status{};
}
-void ClMatMulLowpNativeKernel::configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+void ClMatMulLowpNativeKernel::configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *bias, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
const ActivationLayerInfo &act_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst, &compile_context, &matmul_kernel_info);
- ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst, matmul_kernel_info);
- ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, dst, matmul_kernel_info));
+ ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst, matmul_kernel_info);
+ ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, bias, dst, matmul_kernel_info));
- // output tensor auto initialization if not yet initialized
+ // dst tensor auto initialization if not yet initialized
auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)));
const int m = dst->dimension(1);
@@ -172,7 +181,8 @@
// Note : Offset is not negated, unlike gemmlowp kernels
build_opts.add_option("-DLHS_OFFSET=" + support::cpp11::to_string(lqinfo.offset));
build_opts.add_option("-DRHS_OFFSET=" + support::cpp11::to_string(rqinfo.offset));
- build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(dqinfo.offset)); // Passed as positive (unlike the above two)
+ build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(dqinfo.offset));
+ build_opts.add_option_if(bias != nullptr, "-DBIAS");
// Floating point boundaries are quantized prior to being passed as arguments.
// Note: We expect the input and output tensors to always adopt a per-tensor quantization approach
@@ -222,17 +232,22 @@
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
- const ICLTensor *lhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
- const ICLTensor *rhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
- ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+ const ICLTensor *lhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
+ const ICLTensor *rhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
+ const ICLTensor *bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
+ ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst);
- ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst);
+ ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst);
unsigned int idx = 0;
Window window_collapsed = window.collapse(ICLKernel::window(), Window::DimZ);
add_3d_tensor_nhw_argument(idx, lhs);
add_3d_tensor_nhw_argument(idx, rhs);
+ if(bias != nullptr)
+ {
+ add_3d_tensor_nhw_argument(idx, bias);
+ }
add_3d_tensor_nhw_argument(idx, dst);
enqueue(queue, *this, window_collapsed, lws_hint());
diff --git a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.h b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.h
index 67d1a66..c908280 100644
--- a/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.h
+++ b/src/gpu/cl/kernels/ClMatMulLowpNativeKernel.h
@@ -45,15 +45,16 @@
/** Initialise the kernel's input and output.
*
* @param[in] compile_context The compile context to be used.
- * @param[in] lhs Input tensor for the LHS matrix. Data type supported: QASYMM8_SIGNED/QASYMM8.
+ * @param[in] lhs Input tensor info for the LHS matrix. Data type supported: QASYMM8_SIGNED/QASYMM8.
* Dimensions above 2 are collapsed onto dimension 2 and represent the batch.
- * @param[in] rhs Input tensor for the RHS matrix. Data type supported: same as @p lhs.
+ * @param[in] rhs Input tensor info for the RHS matrix. Data type supported: same as @p lhs.
* Dimensions above 2 are collapsed onto dimension 2 and represent the batch.
+ * @param[in] bias Bias tensor info. Can be nullptr. Data type supported: S32.
* @param[out] dst Output tensor info. Data type supported: same as @p lhs
* @param[in] matmul_kernel_info Attributes for Batch MatMul Kernel
- * @param[in] act_info Class containing information about fused activation function.
+ * @param[in] act_info (Optional) Class containing information about fused activation function.
*/
- void configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+ void configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *bias, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
const ActivationLayerInfo &act_info = ActivationLayerInfo());
/** Static function to check if given info will lead to a valid configuration
*
@@ -61,7 +62,8 @@
*
* @return a status
*/
- static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+ static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *bias, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+ const ActivationLayerInfo &act_info = ActivationLayerInfo());
// Inherited methods overridden:
void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
diff --git a/src/gpu/cl/kernels/ClMatMulNativeKernel.cpp b/src/gpu/cl/kernels/ClMatMulNativeKernel.cpp
index 205396a..545a5b2 100644
--- a/src/gpu/cl/kernels/ClMatMulNativeKernel.cpp
+++ b/src/gpu/cl/kernels/ClMatMulNativeKernel.cpp
@@ -120,7 +120,8 @@
_type = CLKernelType::GEMM;
}
-Status ClMatMulNativeKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info, const ActivationLayerInfo &act_info)
+Status ClMatMulNativeKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *bias, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+ const ActivationLayerInfo &act_info)
{
ARM_COMPUTE_UNUSED(act_info);
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs, dst);
@@ -130,21 +131,30 @@
ARM_COMPUTE_RETURN_ON_ERROR(validate_input_shapes(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info));
ARM_COMPUTE_RETURN_ON_ERROR(validate_export_to_cl_image(rhs, matmul_kernel_info));
+ const TensorShape expected_output_shape = misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info);
+
if(dst->total_size() != 0)
{
- const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info));
+ const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(expected_output_shape);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, dst);
}
+ if(bias != nullptr)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(bias, lhs);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((bias->num_dimensions() > 1), "Multi dimensional bias is unsupported.");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->dimension(0) != expected_output_shape[0], "First dimension of bias and output tensors must match.");
+ }
+
return Status{};
}
-void ClMatMulNativeKernel::configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+void ClMatMulNativeKernel::configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *bias, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
const ActivationLayerInfo &act_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst, &compile_context, &matmul_kernel_info);
- ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst, matmul_kernel_info);
- ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, dst, matmul_kernel_info));
+ ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst, matmul_kernel_info);
+ ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, bias, dst, matmul_kernel_info));
// dst tensor auto initialization if not yet initialized
auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)));
@@ -176,6 +186,7 @@
build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
build_opts.add_option("-DK=" + support::cpp11::to_string(k));
+ build_opts.add_option_if(bias != nullptr, "-DBIAS");
build_opts.add_option_if_else(_export_rhs_to_cl_image, "-DRHS_TENSOR_TYPE=IMAGE", "-DRHS_TENSOR_TYPE=BUFFER");
// Define values for activation function
@@ -225,11 +236,12 @@
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
- const ICLTensor *lhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
- const ICLTensor *rhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
- ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+ const ICLTensor *lhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
+ const ICLTensor *rhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
+ const ICLTensor *bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); // nullptr if bias is not present
+ ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst);
- ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst);
+ ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst);
unsigned int idx = 0;
Window window_collapsed = window.collapse(ICLKernel::window(), Window::DimZ);
@@ -250,6 +262,10 @@
}
add_3d_tensor_nhw_argument(idx, rhs);
+ if(bias != nullptr)
+ {
+ add_3d_tensor_nhw_argument(idx, bias);
+ }
add_3d_tensor_nhw_argument(idx, dst);
enqueue(queue, *this, window_collapsed, lws_hint());
diff --git a/src/gpu/cl/kernels/ClMatMulNativeKernel.h b/src/gpu/cl/kernels/ClMatMulNativeKernel.h
index 02d8ac3..fe2b787 100644
--- a/src/gpu/cl/kernels/ClMatMulNativeKernel.h
+++ b/src/gpu/cl/kernels/ClMatMulNativeKernel.h
@@ -43,15 +43,16 @@
/** Initialise the kernel's input and output.
*
* @param[in] compile_context The compile context to be used.
- * @param[in] lhs Input tensor for the LHS matrix. Data type supported: F32/F16.
+ * @param[in] lhs Input tensor info for the LHS matrix. Data type supported: F32/F16.
* Dimensions above 2 are collapsed onto dimension 2 and represent the batch.
- * @param[in] rhs Input tensor for the RHS matrix. Data type supported: same as @p lhs.
+ * @param[in] rhs Input tensor info for the RHS matrix. Data type supported: same as @p lhs.
* Dimensions above 2 are collapsed onto dimension 2 and represent the batch.
+ * @param[in] bias Bias tensor info for bias matrix. Can be nullptr. Data type supported: same as @p lhs.
* @param[out] dst Output tensor info. Data type supported: same as @p lhs
* @param[in] matmul_kernel_info Attributes for Batch MatMul Kernel
- * @param[in] act_info Specifies activation function to use after Matrix multiplication. Default is Identity function.
+ * @param[in] act_info (Optional) Specifies activation function to use after Matrix multiplication. Default is Identity function.
*/
- void configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+ void configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *bias, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
const ActivationLayerInfo &act_info = ActivationLayerInfo());
/** Static function to check if given info will lead to a valid configuration
*
@@ -59,7 +60,8 @@
*
* @return a status
*/
- static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info, const ActivationLayerInfo &act_info = ActivationLayerInfo());
+ static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *bias, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info,
+ const ActivationLayerInfo &act_info = ActivationLayerInfo());
// Inherited methods overridden:
void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
diff --git a/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.cpp b/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.cpp
index 4630ec0..0efcfb1 100644
--- a/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.cpp
+++ b/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.cpp
@@ -60,9 +60,9 @@
Status validate_matmul_kernel_info(const MatMulKernelInfo &matmul_kernel_info)
{
const bool adj_lhs = matmul_kernel_info.adj_lhs;
- const int m0 = matmul_kernel_info.m0;
- const int n0 = matmul_kernel_info.n0;
- const int k0 = matmul_kernel_info.k0;
+ const int m0 = matmul_kernel_info.m0;
+ const int n0 = matmul_kernel_info.n0;
+ const int k0 = matmul_kernel_info.k0;
// Validate M0
ARM_COMPUTE_RETURN_ERROR_ON_MSG(m0 < 1, "Only positive integers are supported for M0");
@@ -149,7 +149,7 @@
_type = CLKernelType::GEMM;
}
-Status ClMatMulNativeMMULKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info)
+Status ClMatMulNativeMMULKernel::validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *bias, const ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs, dst);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::F32, DataType::F16);
@@ -158,20 +158,29 @@
ARM_COMPUTE_RETURN_ON_ERROR(validate_matmul_kernel_info(matmul_kernel_info));
ARM_COMPUTE_RETURN_ON_ERROR(validate_input_shapes(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info));
+ const TensorShape expected_output_shape = misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info);
+
if(dst->total_size() != 0)
{
- const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info));
+ const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(expected_output_shape);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_dst);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, dst);
}
+ if(bias != nullptr)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((bias->num_dimensions() > 1), "Multi dimensional bias is unsupported.");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->dimension(0) != expected_output_shape[0], "First dimension of bias and output tensors must match.");
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, bias);
+ }
+
return Status{};
}
-void ClMatMulNativeMMULKernel::configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info)
+void ClMatMulNativeMMULKernel::configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *bias, ITensorInfo *dst, const MatMulKernelInfo &matmul_kernel_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst);
- ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst, matmul_kernel_info);
- ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, dst, matmul_kernel_info));
+ ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst, matmul_kernel_info);
+ ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, bias, dst, matmul_kernel_info));
// dst tensor auto initialization if not yet initialized
auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)));
@@ -207,6 +216,7 @@
build_opts.add_option("-DMMUL_M0=" + support::cpp11::to_string(mmul_m0));
build_opts.add_option("-DMMUL_N0=" + support::cpp11::to_string(mmul_n0));
build_opts.add_option("-DMMUL_K0=" + support::cpp11::to_string(mmul_k0));
+ build_opts.add_option_if(bias != nullptr, "-DBIAS");
std::string kernel_name("mat_mul_native_mmul");
kernel_name += matmul_kernel_info.adj_lhs ? "_t" : "_nt";
@@ -239,15 +249,20 @@
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
- const ICLTensor *lhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
- const ICLTensor *rhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
- ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
+ const ICLTensor *lhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
+ const ICLTensor *rhs = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
+ const ICLTensor *bias = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); // nullptr if bias is not present
+ ICLTensor *dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst);
- ARM_COMPUTE_LOG_PARAMS(lhs, rhs, dst);
+ ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst);
unsigned int idx = 0;
add_3d_tensor_nhw_argument(idx, lhs);
add_3d_tensor_nhw_argument(idx, rhs);
+ if(bias != nullptr)
+ {
+ add_3d_tensor_nhw_argument(idx, bias);
+ }
add_3d_tensor_nhw_argument(idx, dst);
// Pass m and n at runtime as signed ints, to ensure results of any subtractions they could be operand in, would still be signed.
diff --git a/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.h b/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.h
index 79f675d..8044897 100644
--- a/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.h
+++ b/src/gpu/cl/kernels/ClMatMulNativeMMULKernel.h
@@ -66,19 +66,20 @@
* - No broadcasting in batch dimensions. I.e. batch dims must be the same across lhs, rhs and dst
*
* @param[in] compile_context The compile context to be used.
- * @param[in] lhs Input tensor for the LHS matrix.
- * @param[in] rhs Input tensor for the RHS matrix.
+ * @param[in] lhs Input tensor info for the LHS matrix.
+ * @param[in] rhs Input tensor info for the RHS matrix.
+ * @param[in] bias Bias tensor info. Can be nullptr. Data type supported: Same as @p lhs.
* @param[out] dst Output tensor info.
* @param[in] matmul_info Attributes for Batch MatMul Kernel
*/
- void configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *dst, const MatMulKernelInfo &matmul_info);
+ void configure(const ClCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *bias, ITensorInfo *dst, const MatMulKernelInfo &matmul_info);
/** Static function to check if given info will lead to a valid configuration
*
* Similar to @ref ClMatMulNativeMMULKernel::configure()
*
* @return a status
*/
- static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *dst, const MatMulKernelInfo &matmul_info);
+ static Status validate(const ITensorInfo *lhs, const ITensorInfo *rhs, const ITensorInfo *bias, const ITensorInfo *dst, const MatMulKernelInfo &matmul_info);
// Inherited methods overridden:
void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override;
diff --git a/src/gpu/cl/operators/ClFullyConnected.cpp b/src/gpu/cl/operators/ClFullyConnected.cpp
index b7ba8b8..0be3f0f 100644
--- a/src/gpu/cl/operators/ClFullyConnected.cpp
+++ b/src/gpu/cl/operators/ClFullyConnected.cpp
@@ -113,22 +113,25 @@
Status validate_mm(const ITensorInfo &src, const ITensorInfo &weights, const ITensorInfo *bias, const ITensorInfo &dst, const FullyConnectedLayerInfo &fc_info)
{
- // If weights are dynamic, data is not batched, and bias is nullptr validate using matmul.
- const bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
- const bool use_matmul = !weights.are_values_constant() && !weights_reshaped && !(dst.dimension(1) > 1) && (bias == nullptr);
+ // Note : If input is dynamic and data is not batched, use matmul, else use gemm
+ const bool transpose_weights = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false;
+ const bool use_matmul = !weights.are_values_constant() && !(dst.dimension(1) > 1);
+ const bool use_dynamic_gemm = !use_matmul && !weights.are_values_constant() && transpose_weights; // use dynamic gemm as fallback for matmul
+ const bool is_quantized = is_data_type_quantized_asymmetric(src.data_type());
if(use_matmul)
{
- MatMulInfo m_info{};
- m_info.adj_rhs(fc_info.transpose_weights);
+ const MatMulInfo m_info = MatMulInfo().adj_rhs(transpose_weights);
- // Note: Currently, shape is [M, B0, B1]
- // LHS is reshaped here to match ClMatMul expectations of batch index in format - [M, 1, B0, B1, .. ]
- TensorInfo lhs_to_use{ src };
- lhs_to_use.set_tensor_shape(get_reshaped_matmul_tensor(src.tensor_shape()));
+ // Note: LHS is reshaped here to match ClMatMul expectations of batch index - From [M, B0, B1] to [M, 1, B0, B1]
+ TensorInfo lhs_to_use = src.clone()->set_tensor_shape(get_reshaped_matmul_tensor(src.tensor_shape()));
- // Operator level validation.
- ARM_COMPUTE_RETURN_ON_ERROR(ClMatMul::validate(&lhs_to_use, &weights, &dst, m_info, fc_info.activation_info));
+ const GPUTarget gpu_target = CLScheduler::get().target();
+ std::unique_ptr<cl_matmul::IClMatMulNativeKernelConfig> t = cl_matmul::ClMatMulNativeKernelConfigurationFactory::create(gpu_target);
+ const MatMulKernelInfo kernel_info = t->configure(&lhs_to_use, &weights, m_info);
+
+ return is_quantized ? kernels::ClMatMulLowpNativeKernel::validate(&lhs_to_use, &weights, bias, &dst, kernel_info, fc_info.activation_info) :
+ kernels::ClMatMulNativeKernel::validate(&lhs_to_use, &weights, bias, &dst, kernel_info, fc_info.activation_info);
}
else
{
@@ -137,7 +140,7 @@
const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
false, // is_b_reshaped
- true, // reshape_b_only_on_first_run
+ !use_dynamic_gemm, // reshape_b_only_on_first_run
0, // depth_output_gemm3d
false, // reinterpret_input_as_3d
fc_info.retain_internal_weights, // retain_internal_weights
@@ -147,7 +150,7 @@
true, // broadcast_bias
ActivationLayerInfo()); // activation_info
- if(is_data_type_quantized_asymmetric(src.data_type()))
+ if(is_quantized)
{
const UniformQuantizationInfo iq_info = src.quantization_info().uniform();
const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
@@ -191,35 +194,33 @@
void ClFullyConnected::configure_mm(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst,
const FullyConnectedLayerInfo &fc_info)
{
- // If weights are dynamic, configure matmul operator - else use gemm
+ // If weights are dynamic and matmul is supported use matmul, else use gemm
if(_use_matmul)
{
- // Transpose RHS as _are_weights_reshaped == false when mat_mul is used.
- const MatMulInfo mat_info = MatMulInfo().adj_rhs(fc_info.transpose_weights);
+ // Specify whether transpose weights is necessary in matmul info
+ const MatMulInfo mat_info = MatMulInfo().adj_rhs(_transpose_weights);
// Note: MatMul does not need offset negation unlike gemm
// 1. Change shape when calling matmul to fit batch expectations.
- _lhs_to_use = *src->clone();
- _lhs_to_use.set_tensor_shape(get_reshaped_matmul_tensor(_lhs_to_use.tensor_shape())); // Collapse all dims > 2 into final dimension.
- _is_quantized = is_data_type_quantized_asymmetric(_lhs_to_use.data_type());
+ _lhs_to_use = src->clone()->set_tensor_shape(get_reshaped_matmul_tensor(_lhs_to_use.tensor_shape()));
- // 2. Call kernel for matmul directly.
+ // 2. Use heuristics to get kernel info object
const GPUTarget gpu_target = CLScheduler::get().target();
std::unique_ptr<cl_matmul::IClMatMulNativeKernelConfig> kernel_config = cl_matmul::ClMatMulNativeKernelConfigurationFactory::create(gpu_target);
+ MatMulKernelInfo kernel_info = kernel_config->configure(src, weights, mat_info);
- // Configure relevant matmul kernel
- MatMulKernelInfo kernel_info = kernel_config->configure(src, weights, mat_info);
+ // 3. Configure relevant matmul kernel
if(_is_quantized)
{
_matmul_lowp_native_kernel = std::make_unique<kernels::ClMatMulLowpNativeKernel>();
_matmul_lowp_native_kernel->set_target(gpu_target);
- _matmul_lowp_native_kernel->configure(compile_context, src, weights, dst, kernel_info, fc_info.activation_info);
+ _matmul_lowp_native_kernel->configure(compile_context, src, weights, bias, dst, kernel_info, fc_info.activation_info);
}
else
{
_matmul_native_kernel = std::make_unique<kernels::ClMatMulNativeKernel>();
_matmul_native_kernel->set_target(gpu_target);
- _matmul_native_kernel->configure(compile_context, src, weights, dst, kernel_info, fc_info.activation_info);
+ _matmul_native_kernel->configure(compile_context, src, weights, bias, dst, kernel_info, fc_info.activation_info);
}
}
else
@@ -230,7 +231,7 @@
const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
false, // is_b_reshaped
- !_dynamic_weights, // reshape_b_only_on_first_run
+ !_dynamic_gemm, // reshape_b_only_on_first_run
0, // depth_output_gemm3d
false, // reinterpret_input_as_3d
fc_info.retain_internal_weights, // retain_internal_weights
@@ -269,7 +270,8 @@
void ClFullyConnected::configure_conv_fc(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst,
const FullyConnectedLayerInfo &fc_info)
{
- ARM_COMPUTE_ERROR_ON((weights->dimension((_use_matmul) ? 0 : 1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
+ // MatMul fuses transpose operation, so we use the first dimension for comparison where appropriate.
+ ARM_COMPUTE_ERROR_ON((weights->dimension((_use_matmul && _transpose_weights) ? 0 : 1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
// If the fully connected layer is called after a convolution layer, the input tensor must be linearized
@@ -288,8 +290,8 @@
void ClFullyConnected::configure_fc_fc(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst,
const FullyConnectedLayerInfo &fc_info)
{
- // Compare first dimension when using matmul, as it performs transpose operation
- ARM_COMPUTE_ERROR_ON(src->dimension(0) != weights->dimension((_use_matmul) ? 0 : 1));
+ // MatMul fuses transpose operation, so we use the first dimension for comparison where appropriate.
+ ARM_COMPUTE_ERROR_ON(src->dimension(0) != weights->dimension((_use_matmul && _transpose_weights) ? 0 : 1));
// Configure matrix multiply kernel
configure_mm(compile_context, src, weights, bias, dst, fc_info);
@@ -304,20 +306,18 @@
ARM_COMPUTE_ERROR_THROW_ON(ClFullyConnected::validate(src, weights, biases, dst, fc_info));
ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, fc_info);
- _are_weights_converted = true;
- _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
- _is_fc_after_conv = true;
- _is_quantized = is_data_type_quantized_asymmetric(src->data_type());
- _is_prepared = fc_info.retain_internal_weights;
- _weights_to_use = TensorInfo(*weights);
- _weights_to_use_idx = ACL_SRC_1;
+ _transpose_weights = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false;
+ _is_fc_after_conv = true;
+ _is_quantized = is_data_type_quantized_asymmetric(src->data_type());
+ _is_prepared = fc_info.retain_internal_weights;
+ _weights_to_use = TensorInfo(*weights);
+ _weights_to_use_idx = ACL_SRC_1;
// When using dynamic weights - use matmul kernels.
- // Note: We don't appear to support dynamic weights with pre-reshaped RHS.
- // Note: No matmul with biases for the moment.
+ // Note: MatMul does not support broadcasting batch dimension, and therefore is disabled if fc is batched. Gemm is used as fallback.
const bool is_batched_fc_layer = dst->dimension(1) > 1;
- _dynamic_weights = !weights->are_values_constant() && !_are_weights_reshaped;
- _use_matmul = _dynamic_weights && !is_batched_fc_layer && (biases == nullptr);
+ _use_matmul = !weights->are_values_constant() && !is_batched_fc_layer;
+ _dynamic_gemm = !weights->are_values_constant() && _transpose_weights && !_use_matmul;
// With the Fully Connected layer we can have 4 different cases:
// 1) Convolution layer -> Fully Connected layer without batches
@@ -339,9 +339,8 @@
ITensorInfo *weights_used = weights;
- // Reshape weights if needed
- // Not needed when matmul is in use - MatMul has transpose RHS flags.
- if(!_are_weights_reshaped && !_use_matmul)
+ // Reshape weights if needed - Not needed when matmul is in use as matmul fuses transpose op.
+ if(_transpose_weights && !_use_matmul)
{
// Reshape the weights
_reshape_weights = std::make_unique<ClTranspose>();
@@ -361,9 +360,9 @@
src->tensor_shape(),
fc_info.weights_trained_layout);
- weights_used = &_converted_weights;
- _weights_to_use_idx = offset_int_vec(ConvertedWeights);
- _are_weights_converted = false;
+ weights_used = &_converted_weights;
+ _weights_to_use_idx = offset_int_vec(ConvertedWeights);
+ _run_convert_weights = true;
}
if(_is_fc_after_conv)
@@ -398,11 +397,11 @@
// Keep all the auxiliary tensors in case of dynamic weights as they are recalculated every time
_aux_mem[TransposedWeights] = MemoryInfo(
offset_int_vec(TransposedWeights),
- _dynamic_weights ? MemoryLifetime::Temporary : MemoryLifetime::Prepare,
+ _dynamic_gemm ? MemoryLifetime::Temporary : MemoryLifetime::Prepare,
_reshaped_weights.total_size());
_aux_mem[ConvertedWeights] = MemoryInfo(
offset_int_vec(ConvertedWeights),
- _dynamic_weights ? MemoryLifetime::Temporary : MemoryLifetime::Prepare,
+ _dynamic_gemm ? MemoryLifetime::Temporary : MemoryLifetime::Prepare,
_converted_weights.total_size());
}
else
@@ -413,11 +412,11 @@
_aux_mem[TransposedWeights] = MemoryInfo(
offset_int_vec(TransposedWeights),
- _dynamic_weights ? MemoryLifetime::Temporary : transposed_wei_lft,
+ _dynamic_gemm ? MemoryLifetime::Temporary : transposed_wei_lft,
_reshaped_weights.total_size());
_aux_mem[ConvertedWeights] = MemoryInfo(
offset_int_vec(ConvertedWeights),
- _dynamic_weights ? MemoryLifetime::Temporary : converted_wei_lft,
+ _dynamic_gemm ? MemoryLifetime::Temporary : converted_wei_lft,
_converted_weights.total_size());
}
}
@@ -434,19 +433,17 @@
ARM_COMPUTE_RETURN_ERROR_ON(fc_info.activation_info.enabled() && is_data_type_quantized(src->data_type()) && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::RELU
&& fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU);
- const bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
- bool is_fc_after_conv = true;
+ const bool transpose_weights = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false;
+ bool is_fc_after_conv = true;
// When using dynamic weights - use matmul kernels.
- // Note: MatMul does not support broadcasting or biases so fallback with batched cases or when biases != nullptr.
- // Note: Pre-Shaped RHS is a deprecated use case and is therefore not supported with matmul.
- const bool dynamic_weights = !weights->are_values_constant() && !weights_reshaped;
+ // Note: MatMul does not support broadcasting so fallback with batched cases.
const bool is_batched_fc_layer = dst->dimension(1) > 1;
- const bool use_matmul = dynamic_weights && !is_batched_fc_layer && (biases == nullptr);
+ const bool use_matmul = !weights->are_values_constant() && !is_batched_fc_layer;
const ITensorInfo &flatten_src = TensorInfo(src->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(src)).set_data_layout(DataLayout::NCHW));
const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
- const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone());
+ const ITensorInfo &converted_weights = transpose_weights ? TensorInfo(*reshaped_weights.clone()) : TensorInfo(weights->clone()->set_is_resizable(true).reset_padding());
// With the Fully Connected layer we can have 4 different cases:
// 1) Convolution layer -> Fully Connected layer without batches
@@ -482,7 +479,8 @@
is_fc_after_conv = src->num_dimensions() > 1;
}
- if(!weights_reshaped && !use_matmul)
+ // Transpose kernel does not run when matmul is supported as matmul fuses transpose op.
+ if(transpose_weights && !use_matmul)
{
// Validate reshape weights kernel
ARM_COMPUTE_RETURN_ON_ERROR(ClTranspose::validate(weights, &reshaped_weights));
@@ -502,14 +500,9 @@
if(is_fc_after_conv)
{
// Fully Connected layer after a Convolution Layer without batches
- if(use_matmul)
- {
- ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(0) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
- }
- else
- {
- ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
- }
+ // K Index of matrix multiplication. MatMul performs transpose in kernel, so index is 0 when matmul and transpose enabled
+ const int weight_idx = (use_matmul && transpose_weights) ? 0 : 1;
+ ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(weight_idx) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
// Validate flatten kernel
ARM_COMPUTE_RETURN_ON_ERROR(ClFlatten::validate(src, &flatten_src));
@@ -518,7 +511,9 @@
else
{
// Fully Connected layer after a Fully Connected Layer without batches
- ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != weights_to_use->dimension((use_matmul) ? 0 : 1));
+ // K Index of matrix multiplication. MatMul performs transpose in kernel, so index is 0 when matmul and transpose enabled
+ const int weight_idx = (use_matmul && transpose_weights) ? 0 : 1;
+ ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != weights_to_use->dimension(weight_idx));
}
// Validate matrix multiply kernel
@@ -533,7 +528,7 @@
#ifdef ARM_COMPUTE_ASSERTS_ENABLED
++_asrt_run_count;
- ARM_COMPUTE_ERROR_ON(_dynamic_weights && _asrt_prepare_count != _asrt_run_count);
+ ARM_COMPUTE_ERROR_ON(_dynamic_gemm && _asrt_prepare_count != _asrt_run_count);
#endif // ARM_COMPUTE_ASSERTS_ENABLED
auto src = tensors.get_const_tensor(ACL_SRC_0);
@@ -584,11 +579,12 @@
void ClFullyConnected::prepare(ITensorPack &tensors)
{
- if(!_is_prepared || _dynamic_weights)
+ // Note : Running prepare() each run when _use_matmul is true is unnecessary unless weights conversion is needed.
+ if(!_is_prepared || _dynamic_gemm || (_use_matmul && _run_convert_weights))
{
#ifdef ARM_COMPUTE_ASSERTS_ENABLED
++_asrt_prepare_count;
- ARM_COMPUTE_ERROR_ON(!_dynamic_weights && _asrt_prepare_count > 1);
+ ARM_COMPUTE_ERROR_ON(!_dynamic_gemm && !_use_matmul && _asrt_prepare_count > 1);
#endif // ARM_COMPUTE_ASSERTS_ENABLED
auto weights = tensors.get_const_tensor(ACL_SRC_1);
@@ -599,8 +595,8 @@
// Pointer to current weights
const ITensor *cur_weights = weights;
- // Reshape of the weights if needed
- if(!_are_weights_reshaped && !_use_matmul)
+ // Reshape weights if needed. Disabled when matmul kernels are enabled as matmul fuses transpose.
+ if(_transpose_weights && !_use_matmul)
{
// Run reshape weights kernel and mark weights as unused
ITensorPack transpose_pack{ { ACL_SRC, weights }, { ACL_DST, reshaped_weights.get() } };
@@ -611,7 +607,7 @@
}
// Convert weights if needed
- if(!_are_weights_converted)
+ if(_run_convert_weights)
{
ITensorPack convert_pack{ { ACL_SRC, cur_weights }, { ACL_DST, converted_weights.get() } };
_convert_weights->run(convert_pack);
@@ -623,8 +619,8 @@
ITensorPack gemm_pack = tensors;
gemm_pack.add_const_tensor(ACL_SRC_1, cur_weights);
- // Prepare GEMM prepare and release unused weights (If not using matmul)
- if(!_use_matmul)
+ // Prepare GEMM prepare and release unused weights
+ if(_dynamic_gemm || !_use_matmul)
{
if(!_is_quantized)
{
diff --git a/src/gpu/cl/operators/ClFullyConnected.h b/src/gpu/cl/operators/ClFullyConnected.h
index 9a5ba40..5a71bd2 100644
--- a/src/gpu/cl/operators/ClFullyConnected.h
+++ b/src/gpu/cl/operators/ClFullyConnected.h
@@ -132,17 +132,18 @@
TensorInfo _flattened_src{};
TensorInfo _converted_weights{};
TensorInfo _reshaped_weights{};
- TensorInfo _lhs_to_use{};
+ TensorInfo _lhs_to_use{};
TensorInfo _weights_to_use{};
int _weights_to_use_idx{ ACL_SRC_1 };
- bool _are_weights_converted{ true };
- bool _are_weights_reshaped{ true };
+ bool _run_convert_weights{ false };
+ bool _transpose_weights{ false };
+ bool _dynamic_gemm{ false };
+ bool _use_matmul{ false };
+
bool _is_fc_after_conv{ true };
bool _is_quantized{ false };
bool _is_prepared{ false };
- bool _dynamic_weights{ false };
- bool _use_matmul{ false };
#ifdef ARM_COMPUTE_ASSERTS_ENABLED
int _asrt_run_count {};
diff --git a/src/gpu/cl/operators/ClMatMul.cpp b/src/gpu/cl/operators/ClMatMul.cpp
index c453761..49d1412 100644
--- a/src/gpu/cl/operators/ClMatMul.cpp
+++ b/src/gpu/cl/operators/ClMatMul.cpp
@@ -61,8 +61,8 @@
const bool is_quantized = is_data_type_quantized_asymmetric(lhs->data_type());
- return is_quantized ? ClMatMulLowpNativeKernel::validate(lhs, rhs, dst, kernel_info, act_info) :
- ClMatMulNativeKernel::validate(lhs, rhs, dst, kernel_info, act_info);
+ return is_quantized ? ClMatMulLowpNativeKernel::validate(lhs, rhs, nullptr /* bias */, dst, kernel_info, act_info) :
+ ClMatMulNativeKernel::validate(lhs, rhs, nullptr /* bias */, dst, kernel_info, act_info);
}
void ClMatMul::configure(const CLCompileContext &compile_context, ITensorInfo *lhs, ITensorInfo *rhs, ITensorInfo *dst, const MatMulInfo &matmul_info, const ActivationLayerInfo &act_info)
@@ -86,14 +86,14 @@
_matmul_lowp_native_kernel->set_target(gpu_target);
// Configure the low-precision native matrix multiply kernel
- _matmul_lowp_native_kernel->configure(compile_context, lhs, rhs, dst, kernel_info, act_info);
+ _matmul_lowp_native_kernel->configure(compile_context, lhs, rhs, nullptr /* bias */, dst, kernel_info, act_info);
}
else
{
_matmul_native_kernel->set_target(gpu_target);
// Configure the native matrix multiply kernel
- _matmul_native_kernel->configure(compile_context, lhs, rhs, dst, kernel_info, act_info);
+ _matmul_native_kernel->configure(compile_context, lhs, rhs, nullptr /* bias */, dst, kernel_info, act_info);
}
}