| 96 | } |
| 97 | |
| 98 | Status CpuMatMul::validate(const ITensorInfo *lhs, |
| 99 | const ITensorInfo *rhs, |
| 100 | const ITensorInfo *dst, |
| 101 | const MatMulInfo &info, |
| 102 | const CpuMatMulSettings &settings, |
| 103 | const ActivationLayerInfo &act_info) |
| 104 | { |
| 105 | ARM_COMPUTE_TRACE_EVENT(ARM_COMPUTE_PROF_CAT_CPU, ARM_COMPUTE_PROF_LVL_CPU, "CpuMatMul::validate"); |
| 106 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, rhs, dst); |
| 107 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::F32, DataType::F16, DataType::BFLOAT16, |
| 108 | DataType::QASYMM8, DataType::QASYMM8_SIGNED); |
| 109 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs->are_values_constant(), "LHS Tensor must be dynamic."); |
| 110 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs->are_values_constant(), "RHS Tensor must be dynamic."); |
| 111 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(lhs); |
| 112 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(lhs); |
| 113 | |
| 114 | const auto adj_lhs = info.adj_lhs(); |
| 115 | const auto adj_rhs = info.adj_rhs(); |
| 116 | |
| 117 | const ITensorInfo *lhs_to_use = lhs; |
| 118 | const ITensorInfo *rhs_to_use = rhs; |
| 119 | TensorInfo lhs_transposed{}; |
| 120 | TensorInfo rhs_transposed{}; |
| 121 | |
| 122 | auto gemm_info = AsmGemmInfo(); |
| 123 | gemm_info.activation_info = act_info; |
| 124 | gemm_info.fast_mode = settings.fast_math(); |
| 125 | gemm_info.fixed_format = settings.fixed_format(); |
| 126 | |
| 127 | // Validate and then permute a/b |
| 128 | if (adj_lhs) |
| 129 | { |
| 130 | auto_init_if_empty(lhs_transposed, |
| 131 | lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_transposed_shape(*lhs))); |
| 132 | ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuTransposeKernel::validate(lhs_to_use, &lhs_transposed)); |
| 133 | // Assign lhs_to_use pointer to use transposed TensorInfo |
| 134 | lhs_to_use = &lhs_transposed; |
| 135 | } |
| 136 | if (adj_rhs) |
| 137 | { |
| 138 | auto_init_if_empty(rhs_transposed, |
| 139 | rhs->clone()->set_tensor_shape(misc::shape_calculator::compute_transposed_shape(*rhs))); |
| 140 | ARM_COMPUTE_RETURN_ON_ERROR(cpu::kernels::CpuTransposeKernel::validate(rhs_to_use, &rhs_transposed)); |
| 141 | // Assign rhs_to_use pointer to use transposed TensorInfo |
| 142 | rhs_to_use = &rhs_transposed; |
| 143 | } |
| 144 | |
| 145 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_to_use->dimension(0) != rhs_to_use->dimension(1), |
| 146 | "The product AB is defined only if the number of columns in A is equal to the " |
| 147 | "number of rows in B (after transpose)"); |
| 148 | |
| 149 | // Iterate over dimensions to be collapsed in operator - check dimensions are equivalent between tensors |
| 150 | for (unsigned int i = 2; i < Coordinates::num_max_dimensions; i++) |
| 151 | { |
| 152 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_to_use->dimension(i) != rhs_to_use->dimension(i), |
| 153 | "Broadcasting in Batch dimension is unsupported by this operator."); |
| 154 | } |
| 155 |
nothing calls this directly
no test coverage detected