feat: Enable QASYMM8_SIGNED to floating-point assembly dequantization

arm_compute/runtime/NEON/functions/NEConvolutionLayer.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2021, 2023-2025 Arm Limited.
+ * Copyright (c) 2018-2021, 2023-2026 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -104,6 +104,7 @@ class NEConvolutionLayer : public IFunction
      * |QASYMM8        |QSYMM8_PER_CHANNEL |S32    |QASYMM8        |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32    |QASYMM8_SIGNED |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |F32    |F32            |
+     * |QASYMM8_SIGNED |QASYMM8_SIGNED     |F16    |F16            |
      * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32    |QASYMM8_SIGNED |
      *
      * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -112,10 +113,11 @@ class NEConvolutionLayer : public IFunction
      * @param[in]  weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
      *                              Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL or QASYMM8_SIGNED if input is QASYMM8/QASYMM8_SIGNED.
      * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                              Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type and
-     *                              for F32 dequantization the bias must be F32.
+     *                              Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     *                              For floating-point dequantization the bias must match @p output.
      * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                              Data types supported: Same as @p input.
+     *                              Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                              listed above.
      * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in]  weights_info     Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
      *                              tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input.
@@ -142,12 +144,12 @@ class NEConvolutionLayer : public IFunction
      *                             Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
      * @param[in] weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
      *                             Data type supported:Same as @p input, also could be QSYMM8_PER_CHANNEL or QASYMM8_SIGNED if input is QASYMM8/QASYMM8_SIGNED.
-     *                             Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
      * @param[in] biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                             Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type and
-     *                             for F32 dequantization the bias must be F32.
+     *                             Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     *                             For floating-point dequantization the bias must match @p output.
      * @param[in] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                             Data types supported: Same as @p input.
+     *                             Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                             listed above.
      * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in] weights_info     Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
      *                             tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input.
@@ -177,7 +179,8 @@ class NEConvolutionLayer : public IFunction
      * @param[in] weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
      *                             Data type supported:Same as @p input, also could be QSYMM8_PER_CHANNEL or QASYMM8_SIGNED if input is QASYMM8/QASYMM8_SIGNED.
      * @param[in] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                             Data types supported: Same as @p input.
+     *                             Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                             listed above.
      * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in] weights_info     Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
      *                             tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input.

arm_compute/runtime/NEON/functions/NEGEMMConv2d.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021, 2024-2025 Arm Limited.
+ * Copyright (c) 2020-2021, 2024-2026 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -75,6 +75,8 @@ class NEGEMMConv2d : public IFunction
      * |:--------------|:--------------|:--------------|:--------------|
      * |QASYMM8        |QASYMM8        |S32            |QASYMM8        |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32            |QASYMM8_SIGNED |
+     * |QASYMM8_SIGNED |QASYMM8_SIGNED |F32            |F32            |
+     * |QASYMM8_SIGNED |QASYMM8_SIGNED |F16            |F16            |
      * |F16            |F16            |F16            |F16            |
      * |F32            |F32            |F32            |F32            |
      * |BFLOAT16       |BFLOAT16       |BFLOAT16       |BFLOAT16       |
@@ -86,8 +88,10 @@ class NEGEMMConv2d : public IFunction
      *                     Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in]  biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                     Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     *                     For floating-point dequantization the bias must match @p output.
      * @param[out] output  Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                     Data types supported: Same as @p input.
+     *                     Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                     listed above.
      * @param[in]  info    Convolution layer descriptor
      */
     void
@@ -101,8 +105,10 @@ class NEGEMMConv2d : public IFunction
      *                    Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in] biases  Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                    Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     *                    For floating-point dequantization the bias must match @p output.
      * @param[in] output  Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                    Data types supported: Same as @p input.
+     *                    Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                    listed above.
      * @param[in] info    Contains padding and stride information described in @ref PadStrideInfo.
      *
      * @return a status

arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2025 Arm Limited.
+ * Copyright (c) 2017-2026 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -79,6 +79,7 @@ class NEGEMMConvolutionLayer : public IFunction
      * |QASYMM8        |QSYMM8_PER_CHANNEL |S32      |QASYMM8        |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32      |QASYMM8_SIGNED |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |F32      |F32            |
+     * |QASYMM8_SIGNED |QASYMM8_SIGNED     |F16      |F16            |
      * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32      |QASYMM8_SIGNED |
      *
      * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -88,9 +89,10 @@ class NEGEMMConvolutionLayer : public IFunction
      *                              Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                              Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     *                              For F32 dequantization the bias must be F32.
+     *                              For floating-point dequantization the bias must match @p output.
      * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                              Data types supported: Same as @p input.
+     *                              Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                              listed above.
      * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in]  weights_info     Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
      *                              tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input.
@@ -119,9 +121,10 @@ class NEGEMMConvolutionLayer : public IFunction
      *                             Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in] biases           Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                             Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     *                             For F32 dequantization the bias must be F32.
+     *                             For floating-point dequantization the bias must match @p output.
      * @param[in] output           Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                             Data types supported: Same as @p input.
+     *                             Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                             listed above.
      * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in] weights_info     Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
      *                             tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input.

arm_compute/runtime/experimental/operators/CpuGemmConv2d.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021-2025 Arm Limited.
+ * Copyright (c) 2021-2026 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -78,6 +78,7 @@ class CpuGemmConv2d : public IOperator
      * |QASYMM8        |QSYMM8_PER_CHANNEL |S32      |QASYMM8        |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32      |QASYMM8_SIGNED |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |F32      |F32            |
+     * |QASYMM8_SIGNED |QASYMM8_SIGNED     |F16      |F16            |
      * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32      |QASYMM8_SIGNED |
      *
      * @param[in]  src              Source tensor info. 3 lower dimensions represent a single input [width, height, IFM],
@@ -87,9 +88,10 @@ class CpuGemmConv2d : public IOperator
      *                              Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in]  biases           Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                              Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
-     *                              For F32 dequantization the bias must be F32.
+     *                              For floating-point dequantization the bias must match @p dst.
      * @param[out] dst              Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                              Data types supported: Same as @p input.
+     *                              Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                              listed above.
      * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
      * @param[in]  weights_info     Specifies if the weights tensor has been reshaped with CpuWeightsReshapeKernel. If this is not part of the fully connected layer the weights
      *                              tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input.

arm_compute/runtime/experimental/operators/CpuGemmDirectConv2d.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2024-2025 Arm Limited.
+ * Copyright (c) 2024-2026 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -71,6 +71,8 @@ class CpuGemmDirectConv2d : public IOperator
      * |:--------------|:--------------|:--------------|:--------------|
      * |QASYMM8        |QASYMM8        |S32            |QASYMM8        |
      * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32            |QASYMM8_SIGNED |
+     * |QASYMM8_SIGNED |QASYMM8_SIGNED |F32            |F32            |
+     * |QASYMM8_SIGNED |QASYMM8_SIGNED |F16            |F16            |
      * |F16            |F16            |F16            |F16            |
      * |F32            |F32            |F32            |F32            |
      * |BFLOAT16       |BFLOAT16       |BFLOAT16       |BFLOAT16       |
@@ -82,8 +84,10 @@ class CpuGemmDirectConv2d : public IOperator
      *                    Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32.
      * @param[in] biases  Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
      *                    Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
+     *                    For floating-point dequantization the bias must match @p dst.
      * @param[in] dst     Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
-     *                    Data types supported: Same as @p input.
+     *                    Data types supported: Same as @p input, except for the QASYMM8_SIGNED to floating-point output path
+     *                    listed above.
      * @param[in] info    Contains padding and stride information described in @ref PadStrideInfo.
      */
     void configure(const ITensorInfo *src,

docs/user_guide/operator_list.dox

@@ -611,6 +611,7 @@ where N = batches, C = channels, H = height, W = width, D = depth
     <tr><td>QASYMM8<td>QSYMM8_PER_CHANNEL<td>S32<td>QASYMM8
     <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>S32<td>QASYMM8_SIGNED
     <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>F32<td>F32
+    <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>F16<td>F16
     <tr><td>QASYMM8_SIGNED<td>QSYMM8_PER_CHANNEL<td>S32<td>QASYMM8_SIGNED
     </table>
 <tr>
@@ -1687,6 +1688,8 @@ where N = batches, C = channels, H = height, W = width, D = depth
     <tr><th>src0<th>src1<th>src2<th>dst
     <tr><td>QASYMM8<td>QASYMM8<td>S32<td>QASYMM8
     <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>S32<td>QASYMM8_SIGNED
+    <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>F32<td>F32
+    <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>F16<td>F16
     <tr><td>F16<td>F16<td>F16<td>F16
     <tr><td>F32<td>F32<td>F32<td>F32
     <tr><td>BFLOAT16<td>BFLOAT16<td>BFLOAT16<td>BFLOAT16
@@ -1714,6 +1717,7 @@ where N = batches, C = channels, H = height, W = width, D = depth
     <tr><td>QASYMM8<td>QSYMM8_PER_CHANNEL<td>S32<td>QASYMM8
     <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>S32<td>QASYMM8_SIGNED
     <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>F32<td>F32
+    <tr><td>QASYMM8_SIGNED<td>QASYMM8_SIGNED<td>F16<td>F16
     <tr><td>QASYMM8_SIGNED<td>QSYMM8_PER_CHANNEL<td>S32<td>QASYMM8_SIGNED
     </table>
 <tr>