COMPMID-3069: Resolve Bfloat16 validation issues.

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: Ic1c610cc01064071a7ff8b69ea598c4b0ddbe0ff
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2953
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 18 insertions, 4 deletions

diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index 873145de12..2bd459a389 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -175,17 +175,22 @@ Status NEGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso
     const bool is_c_bias = gemm_info.reshape_b_only_on_first_run();
 
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(a);
+    ARM_COMPUTE_RETURN_ERROR_ON_CPU_BF16_UNSUPPORTED(a);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::BFLOAT16, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, b);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
+    if(a->data_type() != DataType::BFLOAT16)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, output);
+    }
 
     if(c != nullptr && !is_c_bias)
     {
         ARM_COMPUTE_RETURN_ERROR_ON(gemm_info.depth_output_gemm3d() != 0);
         ARM_COMPUTE_RETURN_ERROR_ON(gemm_info.reinterpret_input_as_3d());
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(c, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != c->dimension(1), "The C matrix must have the same number of rows as the matrix A");
         ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != c->dimension(0), "The C matrix must have the same number of columns as the matrix B");
     }
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index f6dc3a8f43..a41d23f8d7 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -418,6 +418,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
 
     const bool append_bias  = false;
     const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
+    const bool is_bf16      = data_type == DataType::BFLOAT16;
     bool       skip_im2col  = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
 
     // Get convolved dimensions
@@ -463,6 +464,10 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
         {
             ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
         }
+        else if(is_bf16)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32);
+        }
         else
         {
             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
diff --git a/tests/validation/reference/GEMMLowp.cpp b/tests/validation/reference/GEMMLowp.cpp
index 61617c8aae..36d86d1532 100644
--- a/tests/validation/reference/GEMMLowp.cpp
+++ b/tests/validation/reference/GEMMLowp.cpp
@@ -26,6 +26,8 @@
 #include "arm_compute/core/Types.h"
 #include "tests/validation/reference/UtilsQuantizedAsymm.h"
 
+#include "support/ToolchainSupport.h"
+
 #include <limits>
 
 namespace arm_compute
@@ -152,7 +154,7 @@ void quantize_down_scale_by_float(const SimpleTensor<TIn> *in, const SimpleTenso
         const float_t multiplier = (is_per_channel) ? result_real_multiplier[i % cols_in] : result_real_multiplier[0];
 
         float_t result_f = static_cast<float_t>(result) * multiplier + static_cast<float_t>(result_offset);
-        result           = static_cast<TIn>(std::round(result_f));
+        result           = static_cast<TIn>(support::cpp11::round(result_f));
 
         // Bounded ReLu
         if(min != max)
diff --git a/tests/validation/reference/QLSTMLayerNormalization.cpp b/tests/validation/reference/QLSTMLayerNormalization.cpp
index 6764a81617..90d59b93ad 100644
--- a/tests/validation/reference/QLSTMLayerNormalization.cpp
+++ b/tests/validation/reference/QLSTMLayerNormalization.cpp
@@ -28,6 +28,8 @@
 #include "PixelWiseMultiplication.h"
 #include "src/core/utils/quantization/AsymmHelpers.cpp"
 
+#include "support/ToolchainSupport.h"
+
 namespace arm_compute
 {
 namespace test
@@ -79,7 +81,7 @@ SimpleTensor<int16_t> qlstm_layer_normalization(const SimpleTensor<int16_t> &src
 
     for(int i = 0; i < output.num_elements(); i++)
     {
-        const auto output_val_s32 = static_cast<int32_t>(std::round(output_float[i] * std::pow(2, 12)));
+        const auto output_val_s32 = static_cast<int32_t>(support::cpp11::round(output_float[i] * std::pow(2, 12)));
         output[i]                 = utility::clamp<int32_t, int16_t>(output_val_s32, std::numeric_limits<int16_t>::min());
     }