| 360 | } |
| 361 | |
| 362 | Status CpuGemmLowpMatrixMultiplyCore::validate(const ITensorInfo *a, |
| 363 | const ITensorInfo *b, |
| 364 | const ITensorInfo *c, |
| 365 | const ITensorInfo *output, |
| 366 | const GEMMInfo &gemm_info) |
| 367 | { |
| 368 | ARM_COMPUTE_TRACE_EVENT(ARM_COMPUTE_PROF_CAT_CPU, ARM_COMPUTE_PROF_LVL_CPU, |
| 369 | "CpuGemmLowpMatrixMultiplyCore::validate"); |
| 370 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); |
| 371 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, |
| 372 | DataType::QSYMM8, DataType::QSYMM8_PER_CHANNEL); |
| 373 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(output); |
| 374 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32, DataType::QASYMM8, |
| 375 | DataType::QASYMM8_SIGNED, DataType::F32, DataType::F16); |
| 376 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(c != nullptr && output->data_type() != DataType::F32 && |
| 377 | gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::NONE, |
| 378 | "Bias addition not supported in NEGEMMLowpMatrixMultiplyCore for output S32"); |
| 379 | ARM_COMPUTE_RETURN_ERROR_ON_MSG( |
| 380 | (a)->dimension(0) != (b)->dimension(1), |
| 381 | "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); |
| 382 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); |
| 383 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); |
| 384 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.pretranspose_A(), "Matrix A already pretransposed is not supported"); |
| 385 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.pretranspose_B(), "Matrix B already pretransposed is not supported"); |
| 386 | |
| 387 | if (int8_dequantize_f32_path(a->data_type(), output->data_type())) |
| 388 | { |
| 389 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8_SIGNED); |
| 390 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::QASYMM8_SIGNED); |
| 391 | } |
| 392 | |
| 393 | // When using accumulation(in place summation), for now, the only supported DataType for output is S32. |
| 394 | if (gemm_info.accumulate()) |
| 395 | { |
| 396 | #ifdef __arm__ |
| 397 | ARM_COMPUTE_RETURN_ERROR_MSG("Accumulation is not supported for armv7"); |
| 398 | #endif /* __arm__ */ |
| 399 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE, |
| 400 | "Accumulation is not supported for output QASYMM8/QASYMM8_SIGNED"); |
| 401 | } |
| 402 | |
| 403 | GEMMInfo info = gemm_info; |
| 404 | const ITensorInfo *matrix_a_info = a; |
| 405 | const ITensorInfo *matrix_b_info = b; |
| 406 | |
| 407 | const ITensorInfo *a_to_use = a; |
| 408 | |
| 409 | TensorInfo tmp_a_info{}; |
| 410 | TensorInfo tmp_b_info{}; |
| 411 | TensorInfo mm_result_s32_info{}; |
| 412 | |
| 413 | int32_t a_offset = a->quantization_info().uniform().offset; |
| 414 | int32_t b_offset = b->quantization_info().uniform().offset; |
| 415 | |
| 416 | bool fuse_output_stage = info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE; |
| 417 | if (fuse_output_stage) |
| 418 | { |
| 419 | auto_init_if_empty(mm_result_s32_info, |
nothing calls this directly
no test coverage detected