diff options
Diffstat (limited to 'reference_model/src')
| -rw-r--r-- | reference_model/src/generate/generate_dot_product.cc | 59 | ||||
| -rw-r--r-- | reference_model/src/generate/generate_utils.cc | 1 | ||||
| -rw-r--r-- | reference_model/src/ops/tensor_ops.cc | 45 |
3 files changed, 100 insertions, 5 deletions
diff --git a/reference_model/src/generate/generate_dot_product.cc b/reference_model/src/generate/generate_dot_product.cc index 7337969..117d49d 100644 --- a/reference_model/src/generate/generate_dot_product.cc +++ b/reference_model/src/generate/generate_dot_product.cc @@ -963,6 +963,63 @@ bool generateFFT2D(const TosaReference::GenerateConfig& cfg, return true; } +//---------------------------------------------------------------------------// +// RFFT2D // +//---------------------------------------------------------------------------// + +template <typename DataType> +bool generateRFFT2DReal(const TosaReference::GenerateConfig& cfg, + TosaReference::IDotProductGenerator& generator, + DataType* data, + size_t size) +{ + const int64_t T = TosaReference::numElementsFromShape(cfg.shape); + const uint32_t H = cfg.shape[1]; + const uint32_t W = cfg.shape[2]; + + for (int64_t t = 0; t < T; ++t) + { + uint32_t x = t % W; + uint32_t y = (t / W) % H; + uint32_t k = y * W + x; + + data[t] = static_cast<DataType>(generator(k)); + } + return true; +} + +bool generateRFFT2D(const TosaReference::GenerateConfig& cfg, + TosaReference::IDotProductGenerator& generator, + void* data, + size_t size) +{ + if (cfg.shape.size() != 3) + { + WARNING("[Generator][DP][RFFT2D] Tensor shape expected 3 dimensions."); + return false; + } + + switch (cfg.dataType) + { + case DType::DType_FP32: { + float* outData = reinterpret_cast<float*>(data); + switch (cfg.inputPos) + { + case 0: + return generateRFFT2DReal(cfg, generator, outData, size); + default: + WARNING("[Generator][DP][RFFT2D] Invalid input tensor slot position to operator."); + return false; + } + break; + } + default: + WARNING("[Generator][DP][RFFT2D] Only supports FP32."); + return false; + } + + return true; +} } // namespace namespace TosaReference @@ -1003,6 +1060,8 @@ bool generateDotProduct(const GenerateConfig& cfg, void* data, size_t size) return generateConv3D(cfg, *generator, data, size); case tosa::Op_FFT2D: return generateFFT2D(cfg, *generator, data, size); + case tosa::Op_RFFT2D: + return generateRFFT2D(cfg, *generator, data, size); default: WARNING("[Generator][DP] Unsupported operator."); return false; diff --git a/reference_model/src/generate/generate_utils.cc b/reference_model/src/generate/generate_utils.cc index c495fb6..cf5308b 100644 --- a/reference_model/src/generate/generate_utils.cc +++ b/reference_model/src/generate/generate_utils.cc @@ -78,6 +78,7 @@ NLOHMANN_JSON_SERIALIZE_ENUM(Op, { Op::Op_REDUCE_PRODUCT, "REDUCE_PRODUCT" }, { Op::Op_REDUCE_SUM, "REDUCE_SUM" }, { Op::Op_REVERSE, "REVERSE" }, + { Op::Op_RFFT2D, "RFFT2D" }, { Op::Op_SCATTER, "SCATTER" }, { Op::Op_SELECT, "SELECT" }, { Op::Op_SIGMOID, "SIGMOID" }, diff --git a/reference_model/src/ops/tensor_ops.cc b/reference_model/src/ops/tensor_ops.cc index 8d8dac7..dd66f79 100644 --- a/reference_model/src/ops/tensor_ops.cc +++ b/reference_model/src/ops/tensor_ops.cc @@ -1820,6 +1820,9 @@ int OpRFFT2d<Dtype>::eval() int32_t out_imag_height = out_imag->getShape()[1]; int32_t out_imag_width = out_imag->getShape()[2]; + int32_t half_in_height = in_height / 2; + int32_t half_in_width = in_width / 2; + // Check Tosa Level auto tosa_level = g_func_config.tosa_level; LEVEL_CHECK(in_height <= tosa_level.MAX_KERNEL, "H should be smaller than or equal to MAX_KERNEL"); @@ -1831,7 +1834,8 @@ int OpRFFT2d<Dtype>::eval() in_batch, in_height, in_width, out_real_batch, out_real_height, out_real_width, out_imag_batch, out_imag_height, out_imag_width); - OutEigenType sum_real, sum_imag, a; + OutEigenType sum_real, sum_imag; + OutEigenType a, a_cos, a_sin, v_ir; TIn in_val = this->in->getTensor(); @@ -1853,10 +1857,41 @@ int OpRFFT2d<Dtype>::eval() { for (int ix = 0; ix < in_width; ix++) { - // Use explicit cast to ensure intermmediate calculations are completed using OutEigenType - a = 2 * M_PI * ((iy * (OutEigenType)oy) / in_height + (ix * (OutEigenType)ox) / in_width); - sum_real += in_val(n, iy, ix) * cos(a); - sum_imag += -in_val(n, iy, ix) * sin(a); + OutEigenType val = in_val(n, iy, ix); + // Perform the periodic calculation in integer maths to keep + // the accuracy of the co-efficients similar for FP32 normal + // and FP64 precise mode + int32_t ay = (static_cast<int64_t>(iy) * static_cast<int64_t>(oy)) % in_height; + int32_t ax = (static_cast<int64_t>(ix) * static_cast<int64_t>(ox)) % in_width; + + // Use explicit cast to ensure intermediate calculations are completed using OutEigenType + a = 2 * M_PI * ((OutEigenType)ay / in_height + (OutEigenType)ax / in_width); + + // Calculate weight values (co-efficients) + a_cos = cos(a); + a_sin = sin(a); + + if (g_func_config.abs_mode) + { + // Bounded op - Use abs weight values + a_cos = std::abs(a_cos); + a_sin = std::abs(a_sin); + // Bounded op - Use abs real value for imaginary calc + v_ir = val; + } + else + { + // Normal op - Use negative real value for imaginary calc + v_ir = -val; + } + sum_real += val * a_cos; + // Imaginary values with locations (0,0), (0,W/2), (H/2,0) and (H/2,W/2) are zero. + // But due to sin(M_PI) not returning 0 because of M_PI being approximate, only + // add to the imaginary sum when not processing these locations. + if ((ay % (half_in_height)) + (ax % (half_in_width)) > 0) + { + sum_imag += v_ir * a_sin; + } } } this->out_real->getTensor()(n, oy, ox) = sum_real; |