COMPMID-2053: Fuse bias addition with CLGEMMMatrixMultiplyReshapedKernel
Change-Id: I5bfd38c94a6fd18a1cba2104f7e1b04e7bef6ec2
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1359
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index 2ac2eb7..7ada14c 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -1042,11 +1042,12 @@
* @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
* @param[in] bias_ptr (Optional)Pointer to the bias reshaped matrix. Supported data type: same as @p lhs_ptr
- * @param[in] bias_stride_x (Optional)Stride of the bias reshaped matrix in X dimension (in bytes)
- * @param[in] bias_step_x (Optional)bias_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] bias_stride_y (Optional)Stride of the bias reshaped matrix in Y dimension (in bytes)
- * @param[in] bias_step_y (Optional)bias_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] bias_offset_first_element_in_bytes (Optional)The offset of the first element in the bias reshaped matrix
+ * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
+ * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
+ * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
+ * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
* @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
* @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
@@ -1055,7 +1056,7 @@
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
* @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
* @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
- * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
+ * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
* @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
@@ -1415,10 +1416,10 @@
* @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes)
* @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
- * @param[in] bias_ptr (Optional) Pointer to the bias reshaped matrix. Supported data type: same as @p lhs_ptr
- * @param[in] bias_stride_x (Optional) Stride of the bias reshaped matrix in X dimension (in bytes)
+ * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
+ * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
* @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] bias_stride_y (Optional) Stride of the bias reshaped matrix in Y dimension (in bytes)
+ * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
* @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
* @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
@@ -1429,7 +1430,7 @@
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
* @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
* @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
- * @param[in] bias_stride_z (Optional)Stride of the bias reshaped matrix in Z dimension (in bytes)
+ * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] lhs_cross_plane_pad (Optional) Bottom paddings for LHS matrix in unit of elements (only if defined REINTERPRET_INPUT_AS_3D)
* @param[in] dst_cross_plane_pad (Optional) Bottom paddings for the output matrix in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
@@ -1804,36 +1805,49 @@
* -# DEPTH_GEMM3D: The depth of the output in case it has to be reinterpreted as a 3D tensor
* (HEIGHT_GEMM3D * DEPTH_GEMM3D) = columns LHS matrix NOT reshaped
*
- * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32
- * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes)
- * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes)
- * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
- * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
- * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes)
- * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes)
- * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
- * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
- * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
- * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
- * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
- * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped.
- * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
- * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
- * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
+ * @param[in] lhs_ptr Pointer to the LHS reshaped matrix. Supported data type: F16/F32
+ * @param[in] lhs_stride_x Stride of the LHS reshaped matrix in X dimension (in bytes)
+ * @param[in] lhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] lhs_stride_y Stride of the LHS reshaped matrix in Y dimension (in bytes)
+ * @param[in] lhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the LHS reshaped matrix
+ * @param[in] rhs_ptr Pointer to the RHS reshaped matrix. Supported data type: same as @p lhs_ptr
+ * @param[in] rhs_stride_x Stride of the RHS reshaped matrix in X dimension (in bytes)
+ * @param[in] rhs_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] rhs_stride_y Stride of the RHS reshaped matrix in Y dimension (in bytes)
+ * @param[in] rhs_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the RHS reshaped matrix
+ * @param[in] bias_ptr (Optional) Pointer to the bias matrix. Supported data type: same as @p lhs_ptr
+ * @param[in] bias_stride_x (Optional) Stride of the bias matrix in X dimension (in bytes)
+ * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] bias_stride_y (Optional) Stride of the bias matrix in Y dimension (in bytes)
+ * @param[in] bias_step_y (Optional) bias_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the bias matrix
+ * @param[out] dst_ptr Pointer to the destination matrix Supported data type: same as @p lhs_ptr
+ * @param[in] dst_stride_x Stride of the destination matrix in X dimension (in bytes)
+ * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ * @param[in] k Number of columns in LHS matrix and rows in RHS matrix not reshaped.
+ * @param[in] lhs_stride_z Stride of the LHS reshaped matrix in Z dimension (in bytes)
+ * @param[in] rhs_stride_z Stride of the RHS reshaped matrix in Z dimension (in bytes)
+ * @param[in] bias_stride_z (Optional) Stride of the bias matrix in Z dimension (in bytes)
+ * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_cross_plane_pad (Optional) Bottom paddings in unit of elements (only if defined REINTERPRET_OUTPUT_AS_3D)
*/
__kernel void gemm_mm_reshaped_lhs_nt_rhs_t(IMAGE_DECLARATION(lhs),
IMAGE_DECLARATION(rhs),
+#if defined(BETA)
+ IMAGE_DECLARATION(bias),
+#endif // defined(BETA)
IMAGE_DECLARATION(dst),
uint k,
uint lhs_stride_z,
uint rhs_stride_z,
+#if defined(BETA)
+ uint bias_stride_z,
+#endif //defined(BETA)
uint dst_stride_z
#if defined(REINTERPRET_OUTPUT_AS_3D)
,
@@ -1892,8 +1906,8 @@
// Initialize the accumulators
REPEAT_VAR_INIT_TO_CONST(M0, VEC_DATA_TYPE(DATA_TYPE, N0), c, 0); //VEC_DATA_TYPE(DATA_TYPE, N0) c0=0,c1=0,c2=0,... c(M0-1)=0;
- REPEAT_VAR_INIT_TO_CONST(8, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
- REPEAT_VAR_INIT_TO_CONST(16, uint, zrhs, 0);
+ REPEAT_VAR_INIT_TO_CONST(M0, uint, zlhs, 0); //uint zlhs0=0,zlhs1=0,zlhs2=0,... zlhs7=0;
+ REPEAT_VAR_INIT_TO_CONST(16, uint, zero, 0);
for(int i = 0; i < k; i += K0)
{
@@ -1910,7 +1924,7 @@
LOAD_BLOCK(M0, K0, DATA_TYPE, a, lhs_addr, 0, LHS_STEP_X * sizeof(DATA_TYPE), zlhs);
// Load values from RHS matrix
- LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_addr, 0, RHS_STEP_X * sizeof(DATA_TYPE), zrhs);
+ LOAD_BLOCK(N0, K0, DATA_TYPE, b, rhs_addr, 0, RHS_STEP_X * sizeof(DATA_TYPE), zero);
// Accumulate
ARM_DOT_K0XN0(a0, b, c0);
@@ -1942,7 +1956,7 @@
__global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * dst_stride_y);
- REPEAT_VAR_INIT_TO_CONST(8, uint, zout, 0); //uint zout0=0,zout1=0,zout2=0,... zout7=0;
+ REPEAT_VAR_INIT_TO_CONST(M0, uint, zout, 0);
#if defined(REINTERPRET_OUTPUT_AS_3D)
@@ -1964,8 +1978,39 @@
SCALE_BLOCK(M0, DATA_TYPE, c, ALPHA);
#endif // defined(ALPHA)
+ // Add beta*bias
+#if defined(BETA)
+#if defined(BROADCAST_BIAS)
+ __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE));
+
+ LOAD_BLOCK(1, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
+
+#ifndef UNIT_BETA
+ SCALE_BLOCK(1, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+ // c = c + bias[broadcasted]
+ ADD_BLOCK_BROADCAST(M0, c, bias0);
+
+#else // defined(BROADCAST_BIAS)
+ __global uchar *bias_addr = bias_ptr + bias_offset_first_element_in_bytes + (get_global_id(0) * (uint)N0 * sizeof(DATA_TYPE)) + (get_global_id(1) * (uint)M0 * bias_stride_y) + get_global_id(
+ 2) * bias_stride_z;
+
+ LOAD_BLOCK(M0, N0, DATA_TYPE, bias, bias_addr, 0, bias_stride_y, zero);
+
+#ifndef UNIT_BETA
+ SCALE_BLOCK(M0, DATA_TYPE, bias, BETA);
+#endif // UNIT_BIAS
+
+ // c = c + bias
+ ADD_BLOCK(M0, c, bias);
+
+#endif // defined(BROADCAST_BIAS)
+#endif // defined(BETA)
+
// Store output block
STORE_BLOCK(M0, N0, DATA_TYPE, c, dst_addr, dst_stride_y, zout);
+
#undef LHS_BLOCK_SIZE
#undef LHS_OFFSET_X
#undef LHS_STEP_X
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp
index 59afa47..4436726 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp
@@ -56,8 +56,9 @@
{
using ElementsProcessed = Steps;
-Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
- const GEMMReshapeInfo &gemm_info)
+Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info,
+ const GEMMRHSMatrixInfo &rhs_info,
+ const GEMMReshapeInfo &gemm_info)
{
ARM_COMPUTE_UNUSED(alpha);
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input0, input1, output);
@@ -86,6 +87,22 @@
tensor_shape1.set(0, n);
tensor_shape1.set(1, k);
+ if(input2 != nullptr && !(helpers::float_ops::is_zero(beta)))
+ {
+ const int input2_dim0 = static_cast<int>(input2->dimension(0));
+ const int input2_dim1 = static_cast<int>(input2->dimension(1));
+
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input2, input1);
+ if(gemm_info.broadcast_bias())
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim1 != 1 || input2_dim0 != n), "Incorrect dimension of bias matrix which is to be broadcasted");
+ }
+ else
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG((input2_dim0 != n || input2_dim1 != m), "Incorrect dimension of bias matrix");
+ }
+ }
+
const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
@@ -105,7 +122,8 @@
return Status{};
}
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info,
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const GEMMLHSMatrixInfo &lhs_info,
+ const GEMMRHSMatrixInfo &rhs_info,
const GEMMReshapeInfo &gemm_info, ElementsProcessed &num_elements_processed)
{
unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
@@ -152,8 +170,24 @@
ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x),
output->dimension(1) + bottom_pad);
- window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
- update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor
+ if(input2 != nullptr)
+ {
+ const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
+
+ const int bias_processed_per_iteration_y = gemm_info.broadcast_bias() ? 1 : num_elems_processed_per_iteration_y;
+
+ AccessWindowStatic input2_access(input2, 0, 0,
+ ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x),
+ ceil_to_multiple(input2->dimension(1), bias_processed_per_iteration_y));
+
+ window_changed = update_window_and_padding(win, input0_access, input1_access, input2_access) || // window used by the execute_window_loop
+ update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor
+ }
+ else
+ {
+ window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
+ update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor
+ }
output_access.set_valid_region(win_out, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
@@ -169,23 +203,28 @@
} // namespace
CLGEMMMatrixMultiplyReshapedKernel::CLGEMMMatrixMultiplyReshapedKernel()
- : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1), _use_dummy_work_items(false)
+ : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1), _use_dummy_work_items(false), _add_bias(false),
+ _broadcast_bias(false)
{
}
-void CLGEMMMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+void CLGEMMMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta,
+ const GEMMLHSMatrixInfo &lhs_info,
const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), alpha, lhs_info, rhs_info, gemm_info));
+ ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr ? input2->info() : nullptr), output->info(), alpha, beta, lhs_info, rhs_info, gemm_info));
_input0 = input0;
_input1 = input1;
+ _input2 = helpers::float_ops::is_zero(beta) ? nullptr : input2;
_output = output;
_reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0);
_k = gemm_info.k();
_use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
+ _add_bias = _input2 != nullptr;
+ _broadcast_bias = gemm_info.broadcast_bias();
// Check if we need to slide the matrix B
const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
@@ -194,7 +233,7 @@
ElementsProcessed num_elements_processed{};
// Configure kernel window
- auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
+ auto win_config = validate_and_configure_window(input0->info(), input1->info(), input2 != nullptr ? input2->info() : nullptr, output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
ICLKernel::configure_internal(win_config.second);
@@ -202,9 +241,12 @@
CLBuildOptions build_opts;
build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
+ build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
+ build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
+ build_opts.add_option_if(gemm_info.broadcast_bias(), "-DBROADCAST_BIAS");
build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
@@ -227,6 +269,8 @@
// Set config_id for enabling LWS tuning
_config_id = kernel_name;
_config_id += "_";
+ _config_id += (_add_bias ? "add_bias_" : "");
+ _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
_config_id += (_reinterpret_output_as_3d ? "3do_" : "");
_config_id += lower_string(string_from_data_type(input0->info()->data_type()));
_config_id += "_";
@@ -253,13 +297,15 @@
_config_id += support::cpp11::to_string(rhs_info.interleave);
}
-Status CLGEMMMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, const GEMMLHSMatrixInfo &lhs_info,
+Status CLGEMMMatrixMultiplyReshapedKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
+ const GEMMLHSMatrixInfo &lhs_info,
const GEMMRHSMatrixInfo &rhs_info, const GEMMReshapeInfo &gemm_info)
{
ElementsProcessed num_elements_processed{};
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, output, alpha, lhs_info, rhs_info, gemm_info));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info));
ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
input1->clone().get(),
+ input2 != nullptr ? input2->clone().get() : nullptr,
output->clone().get(),
lhs_info,
rhs_info,
@@ -290,7 +336,15 @@
if(_reinterpret_output_as_3d)
{
// Pass bottom paddings to the kernel if the output has to be reinterpreted as 3D tensor
- const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 4;
+ unsigned int idx0;
+ if(_add_bias)
+ {
+ idx0 = 4 * num_arguments_per_2D_tensor() + 5;
+ }
+ else
+ {
+ idx0 = 3 * num_arguments_per_2D_tensor() + 4;
+ }
const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
_kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
}
@@ -308,10 +362,18 @@
unsigned int idx = 0;
add_2D_tensor_argument(idx, _input0, slice);
add_2D_tensor_argument(idx, _input1, slice_b);
+ if(_add_bias)
+ {
+ add_2D_tensor_argument(idx, _input2, slice);
+ }
add_2D_tensor_argument(idx, _output, slice);
_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));
_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
+ if(_add_bias)
+ {
+ _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2]));
+ }
_kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
}
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
index 99b3d20..d952de8 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -250,7 +250,7 @@
CLBuildOptions build_opts;
build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
- build_opts.add_option_if(!(helpers::float_ops::is_zero(beta)) && _input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
+ build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
@@ -277,6 +277,8 @@
// Set config_id for enabling LWS tuning
_config_id = kernel_name;
_config_id += "_";
+ _config_id += (_add_bias ? "add_bias_" : "");
+ _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
_config_id += (_reinterpret_input_as_3d ? "3di_" : "");
_config_id += (_reinterpret_output_as_3d ? "3do_" : "");
_config_id += lower_string(string_from_data_type(input0->info()->data_type()));