COMPMID-434 - Port CLGEMM to support 16 bit fixed point

Change-Id: I30aef3c7ecd1ee740c2a7f2ce65a63c7dcd66e49
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79630
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

14 files changed, 270 insertions, 26 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h b/arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h
index 9466b16a91..203e0fc6c4 100644
--- a/arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h
@@ -64,7 +64,7 @@ public:
     CLGEMMInterleave4x4Kernel &operator=(CLGEMMInterleave4x4Kernel &&) = default;
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  input  Input tensor. Data types supported: U8/S8/QS8/U16/S16/F16/U32/S32/F32
+     * @param[in]  input  Input tensor. Data types supported: U8/S8/QS8/U16/S16/QS16/F16/U32/S32/F32
      * @param[out] output Output tensor. Data type supported: same as @p input
      */
     void configure(const ICLTensor *input, ICLTensor *output);
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h
index b3a85a1706..ada67d1eaf 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixAdditionKernel.h
@@ -52,7 +52,7 @@ public:
      *
      * @note The input and output tensors must have the same dimensions
      *
-     * @param[in]      input  Input tensor (Matrix C). Data types supported: QS8/F16/F32
+     * @param[in]      input  Input tensor (Matrix C). Data types supported: QS8/QS16/F16/F32
      * @param[in, out] output Output tensor. If this kernel is used to finalize the GEMM result (alpha * AB + beta * C), output must contain the result obtained by @ref CLGEMMMatrixMultiplyKernel. Data type supported: same as @p input
      * @param[in]      beta   Weight of matrix C
      */
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
index 7625358b8b..dec63e0679 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h
@@ -53,7 +53,7 @@ public:
     CLGEMMMatrixMultiplyKernel &operator=(CLGEMMMatrixMultiplyKernel &&) = default;
     /** Initialise the kernel's input, output and alpha
      *
-     * @param[in]  input0 Input tensor containing the interleaved Matrix A or the vector A. Data types supported: QS8/F16/F32
+     * @param[in]  input0 Input tensor containing the interleaved Matrix A or the vector A. Data types supported: QS8/QS16/F16/F32
      * @param[in]  input1 Input tensor containing the transposed Matrix B if the first input tensor A is not a vector.
      *                    If the output tensor is a vector, input1 must contain the matrix B not reshaped. Data type supported: same as @p input0
      * @param[out] output Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
diff --git a/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h b/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h
index 9657a2af45..0e467aac13 100644
--- a/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h
@@ -70,7 +70,7 @@ class CLGEMMTranspose1xWKernel : public ICLSimple2DKernel
 public:
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  input  Input tensor. Data types supported: U8/S8/QS8/U16/S16/F16/U32/S32/F32
+     * @param[in]  input  Input tensor. Data types supported: U8/S8/QS8/U16/S16/QS16/F16/U32/S32/F32
      * @param[out] output Output tensor. Data type supported: same as @p input
      */
     void configure(const ICLTensor *input, ICLTensor *output);
diff --git a/arm_compute/runtime/CL/functions/CLGEMM.h b/arm_compute/runtime/CL/functions/CLGEMM.h
index 080f497b7b..9207efd68f 100644
--- a/arm_compute/runtime/CL/functions/CLGEMM.h
+++ b/arm_compute/runtime/CL/functions/CLGEMM.h
@@ -57,7 +57,7 @@ public:
      *
      * @note Whilst the first input tensor can be a vector, the second input tensor must be at least a matrix
      *
-     * @param[in]  a      First input tensor  (Matrix or Vector A). Data types supported: QS8/F16/F32
+     * @param[in]  a      First input tensor  (Matrix or Vector A). Data types supported: QS8/QS16/F16/F32
      * @param[in]  b      Second input tensor (Matrix B). Data type supported: same as @p a.
      * @param[in]  c      Third input tensor  (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a.
      * @param[out] output Output tensor. Data type supported: same as @p a
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index dd3531e858..72230435d8 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -158,14 +158,17 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_ma_f16", "gemm.cl" },
     { "gemm_ma_f32", "gemm.cl" },
     { "gemm_ma_qs8", "gemm.cl" },
+    { "gemm_ma_qs16", "gemm.cl" },
     { "gemm_mm_u8", "gemm.cl" },
     { "gemm_mm_f16", "gemm.cl" },
     { "gemm_mm_f32_midgard", "gemm.cl" },
     { "gemm_mm_f32_bifrost", "gemm.cl" },
     { "gemm_mm_qs8", "gemm.cl" },
+    { "gemm_mm_qs16", "gemm.cl" },
     { "gemm_vm_f16", "gemm.cl" },
     { "gemm_vm_f32", "gemm.cl" },
     { "gemm_vm_qs8", "gemm.cl" },
+    { "gemm_vm_qs16", "gemm.cl" },
     { "gemm_lc_vm_f32", "gemm.cl" },
     { "gemm_transpose1x16", "gemm.cl" },
     { "gemm_transpose1x8", "gemm.cl" },
diff --git a/src/core/CL/cl_kernels/fixed_point.h b/src/core/CL/cl_kernels/fixed_point.h
index 32e49c2fad..dcdf840444 100644
--- a/src/core/CL/cl_kernels/fixed_point.h
+++ b/src/core/CL/cl_kernels/fixed_point.h
@@ -35,16 +35,21 @@
 
 TYPE_ALIAS(char, qs8)
 TYPE_ALIAS(short, qs16)
+TYPE_ALIAS(int, qs32)
 
 #define qs8_MIN ((char)CHAR_MIN)
 #define qs8_MAX ((char)CHAR_MAX)
 #define qs16_MIN ((short)SHRT_MIN)
 #define qs16_MAX ((short)SHRT_MAX)
+#define qs32_MIN ((int)INT_MIN)
+#define qs32_MAX ((int)INT_MAX)
 
 #define qu8_MIN ((uchar)0)
 #define qu8_MAX ((uchar)UCHAR_MAX)
 #define qu16_MIN ((ushort)0)
 #define qu16_MAX ((ushort)USHRT_MAX)
+#define qu32_MIN ((uint)0)
+#define qu32_MAX ((uint)UINT_MAX)
 
 #define qs8_TYPE char
 #define qs8x1_TYPE char
@@ -60,6 +65,13 @@ TYPE_ALIAS(short, qs16)
 #define qs16x8_TYPE short8
 #define qs16x16_TYPE short16
 
+#define qs32_TYPE int
+#define qs32x1_TYPE int
+#define qs32x2_TYPE int2
+#define qs32x4_TYPE int4
+#define qs32x8_TYPE int8
+#define qs32x16_TYPE int16
+
 /* All internal constants are represented in the maximum supported fixed point format (QS16),
  * thus we define an additional shift parameter required to convert the constant
  * from the maximum supported format to the require one.
@@ -166,6 +178,7 @@ SUBQ_SAT_IMPL(qs8x16)
     }
 
 MULQ_SAT_IMPL(qs8x16, qs16x16)
+MULQ_SAT_IMPL(qs16x8, qs32x8)
 
 #define MUL_SAT_OP_EXPAND_STR(a, b, type, size, position) mul_sat_##type##x##size((a), (b), (position))
 #define MUL_SAT_OP_EXPAND(a, b, type, size, position) MUL_SAT_OP_EXPAND_STR(a, b, type, size, position)
@@ -186,6 +199,7 @@ MULQ_SAT_IMPL(qs8x16, qs16x16)
 
 MLAQ_SAT_IMPL(qs8x8, qs16x8)
 MLAQ_SAT_IMPL(qs8x16, qs16x16)
+MLAQ_SAT_IMPL(qs16x8, qs32x8)
 
 #define MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mla_sat_##type##x##size((a), (b), (c), (position))
 #define MLA_SAT_OP_EXPAND(a, b, c, type, size, position) MLA_SAT_OP_EXPAND_STR(a, b, c, type, size, position)
@@ -205,6 +219,7 @@ MLAQ_SAT_IMPL(qs8x16, qs16x16)
     }
 
 MLALQ_SAT_IMPL(qs8x8, qs16x8)
+MLALQ_SAT_IMPL(qs16x8, qs32x8)
 
 #define MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mlal_sat_##type##x##size((a), (b), (c), (position))
 #define MLAL_SAT_OP_EXPAND(a, b, c, type, size, position) MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position)
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index d25621db64..7ac421b7b6 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -888,7 +888,93 @@ __kernel void gemm_mm_qs8(IMAGE_DECLARATION(src0),
     vstore16(c20_qs8, 0, (__global char *)(offset(&dst, 0, 2)));
     vstore16(c30_qs8, 0, (__global char *)(offset(&dst, 0, 3)));
 }
-#endif /* FIXED_POINT_POSITION */
+
+/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in 16 bit fixed point precision
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_16bit and @ref gemm_transpose1x8 before running the matrix multiplication
+ *
+ * @attention The width of matrix B, the alpha's value and fixed point position need to be passed at compile time using -DWIDTH_MATRIX_B -DALPHA and -DFIXED_POINT_POSITION
+ *
+ * @note: ALPHA must be passed in 16 bit fixed point format
+ *
+ * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS16
+ * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
+ * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
+ * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                         dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                         dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
+ */
+__kernel void gemm_mm_qs16(IMAGE_DECLARATION(src0),
+                           IMAGE_DECLARATION(src1),
+                           IMAGE_DECLARATION(dst))
+{
+    /* src_addr.s0 = address of matrix A */
+    /* src_addr.s1 = address of matrix B */
+
+    /* Compute address for matrix A and B */
+    int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),
+                                                                        (src1_stride_y));
+
+    /* Add offset_first_element_in_bytes */
+    src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
+
+    /* Divide by 2 in order to get the src_addr in unit of short */
+    src_addr = src_addr >> 1;
+
+    /* Compute end row address for matrix B */
+    int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;
+
+    /* Reset accumulators */
+    int8 c00 = 0.0f;
+    int8 c10 = 0.0f;
+    int8 c20 = 0.0f;
+    int8 c30 = 0.0f;
+
+    /* This for loop performs 1 accumulation for each iteration */
+    for(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(4, 8))
+    {
+        /* Load values from matrix A (interleaved) and matrix B (transposed) */
+        short4 a0 = vload4(0, ((__global short *)src0_ptr) + src_addr.s0);
+        short8 b0 = vload8(0, ((__global short *)src1_ptr) + src_addr.s1);
+
+        c00 = mlal_sat_qs16x8(c00, (short8)a0.s0, b0, FIXED_POINT_POSITION);
+        c10 = mlal_sat_qs16x8(c10, (short8)a0.s1, b0, FIXED_POINT_POSITION);
+        c20 = mlal_sat_qs16x8(c20, (short8)a0.s2, b0, FIXED_POINT_POSITION);
+        c30 = mlal_sat_qs16x8(c30, (short8)a0.s3, b0, FIXED_POINT_POSITION);
+    }
+
+    /* Compute destination address */
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Multiply by the weight of matrix product */
+    short8 c00_qs16 = convert_short8_sat(c00);
+    short8 c10_qs16 = convert_short8_sat(c10);
+    short8 c20_qs16 = convert_short8_sat(c20);
+    short8 c30_qs16 = convert_short8_sat(c30);
+
+    c00_qs16 = mul_sat_qs16x8(c00_qs16, (short8)ALPHA, FIXED_POINT_POSITION);
+    c10_qs16 = mul_sat_qs16x8(c10_qs16, (short8)ALPHA, FIXED_POINT_POSITION);
+    c20_qs16 = mul_sat_qs16x8(c20_qs16, (short8)ALPHA, FIXED_POINT_POSITION);
+    c30_qs16 = mul_sat_qs16x8(c30_qs16, (short8)ALPHA, FIXED_POINT_POSITION);
+
+    /* Store 8x4 block */
+    vstore8(c00_qs16, 0, (__global short *)(offset(&dst, 0, 0)));
+    vstore8(c10_qs16, 0, (__global short *)(offset(&dst, 0, 1)));
+    vstore8(c20_qs16, 0, (__global short *)(offset(&dst, 0, 2)));
+    vstore8(c30_qs16, 0, (__global short *)(offset(&dst, 0, 3)));
+}
+#endif // defined(FIXED_POINT_POSITION)
 
 #ifdef WIDTH_VECTOR_A
 /** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1)
@@ -1111,9 +1197,87 @@ __kernel void gemm_vm_qs8(IMAGE_DECLARATION(src0),
     /* Store 16 values */
     vstore16(acc_qs8, 0, (__global char *)(offset(&dst, 0, 0)));
 }
-#endif /* FIXED_POINT_POSITION */
-#endif /* WIDTH_VECTOR_A */
-#endif /* WIDTH_MATRIX_B && ALPHA */
+
+/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1) in 16 bit fixed point
+ *
+ * @attention The width of vector A, the width of matrix B, the alpha's value and the fixed point position need to be passed at compile time using -DWIDTH_VECTOR_A -DWIDTH_MATRIX_B, -DALPHA and -DFIXED_POINT_POSITION
+ *
+ * @attention The input vector A and matrix B must not be reshaped
+ *
+ * @note: ALPHA must be passed in 16 bit fixed point format
+ *
+ * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS16
+ * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
+ * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
+ * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                         dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                         dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination matrix
+ */
+__kernel void gemm_vm_qs16(IMAGE_DECLARATION(src0),
+                           IMAGE_DECLARATION(src1),
+                           IMAGE_DECLARATION(dst))
+{
+    int idx = get_global_id(0) * 8;
+
+    /* Compute the address for the vector A and matrix B */
+    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));
+    src_addr.s1 += idx * sizeof(short);
+
+    int end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * sizeof(short));
+
+    /* Reset accumulator */
+    int8 acc0 = 0;
+
+    /* This for loop performs 4 accumulations per iteration */
+    for(; src_addr.s0 <= (end_row_vec_a - 4 * sizeof(short)); src_addr += (int2)(4 * sizeof(short), 4 * src1_stride_y))
+    {
+        short4 a0 = vload4(0, (__global short *)(src0_ptr + src_addr.s0));
+        short8 b0 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y));
+        short8 b1 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y));
+        short8 b2 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y));
+        short8 b3 = vload8(0, (__global short *)(src1_ptr + src_addr.s1 + 3 * src1_stride_y));
+
+        acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s0, b0, FIXED_POINT_POSITION);
+        acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s1, b1, FIXED_POINT_POSITION);
+        acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s2, b2, FIXED_POINT_POSITION);
+        acc0 = mlal_sat_qs16x8(acc0, (short8)a0.s3, b3, FIXED_POINT_POSITION);
+    }
+
+    /* Left-over accumulations */
+    for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(sizeof(short), src1_stride_y))
+    {
+        short  a0 = *((__global short *)(src0_ptr + src_addr.s0));
+        short8 b0 = vload8(0, (__global short *)(src1_ptr + src_addr.s1));
+
+        acc0 = mlal_sat_qs16x8(acc0, (short8)a0, b0, FIXED_POINT_POSITION);
+    }
+
+    /* Compute destination address */
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Multiply by the weight of matrix product */
+    short8 acc_qs16 = convert_short8_sat(acc0);
+
+    acc_qs16 = mul_sat_qs16x8(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION);
+
+    /* Store 8 values */
+    vstore8(acc_qs16, 0, (__global short *)(offset(&dst, 0, 0)));
+}
+#endif /* defined(FIXED_POINT_POSITION) */
+#endif /* defined(WIDTH_VECTOR_A) */
+#endif /* defined(WIDTH_MATRIX_B) && defined(ALPHA) */
 
 #ifdef BETA
 /** This OpenCL kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
@@ -1229,8 +1393,47 @@ __kernel void gemm_ma_qs8(IMAGE_DECLARATION(src),
     /* Store final result in axb matrix */
     vstore16(out, 0, (__global char *)dst.ptr);
 }
-#endif /* FIXED_POINT_POSITION */
-#endif /* BETA */
+
+/** This OpenCL kernel performs the in-place matrix addition between 2 matrices in 16 bit fixed point taking into account that the second matrix might be weighted by a scalar value beta:
+ *
+ * @attention The beta's value and the fixed point position need to be passed at compile time using -DBETA and -DFIXED_POINT_POSITION
+ *
+ * @note: BETA must be passed in 16 bit fixed point format
+ *
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: QS16
+ * @param[in]  src_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination matrix
+ */
+__kernel void gemm_ma_qs16(IMAGE_DECLARATION(src),
+                           IMAGE_DECLARATION(dst))
+{
+    /* Compute source and destination addresses */
+    Image src = CONVERT_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Load values from A x B */
+    short8 alpha_ab = vload8(0, (__global short *)dst.ptr);
+
+    /* Load values from Matrix C */
+    short8 c = vload8(0, (__global short *)src.ptr);
+
+    /* Computes alpha * axb + beta * c */
+    short8 out = mla_sat_qs16x8(alpha_ab, (short8)BETA, c, FIXED_POINT_POSITION);
+
+    /* Store final result in axb matrix */
+    vstore8(out, 0, (__global short *)dst.ptr);
+}
+#endif /* defined(FIXED_POINT_POSITION) */
+#endif /* defined(BETA) */
 
 #ifdef WIDTH_VECTOR_A
 /** This OpenCL kernel computes the vector by matrix multiplication between each row of A (src0) and matrix B (src1) used for locally connected layer
diff --git a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
index 3850c4d2cd..5b6e0ec6af 100644
--- a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
+++ b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
@@ -43,7 +43,8 @@ CLGEMMInterleave4x4Kernel::CLGEMMInterleave4x4Kernel()
 
 void CLGEMMInterleave4x4Kernel::configure(const ICLTensor *input, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16,
+                                                  DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape output_shape = input->info()->tensor_shape();
diff --git a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
index 5883dd698b..d1cdd7dc61 100644
--- a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
@@ -43,7 +43,7 @@ CLGEMMMatrixAdditionKernel::CLGEMMMatrixAdditionKernel()
 
 void CLGEMMMatrixAdditionKernel::configure(const ICLTensor *input, ICLTensor *output, float beta)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != output->info()->dimension(0));
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != output->info()->dimension(1));
@@ -53,8 +53,19 @@ void CLGEMMMatrixAdditionKernel::configure(const ICLTensor *input, ICLTensor *ou
     const unsigned int num_elems_processed_per_iteration = max_cl_vector_width / data_size_from_type(input->info()->data_type());
 
     std::ostringstream ma_arguments;
-    ma_arguments << "-DBETA=" << (input->info()->data_type() == DataType::QS8 ? scvt_qs8_f32(beta, input->info()->fixed_point_position()) : beta) << " ";
-    ma_arguments << "-DFIXED_POINT_POSITION=" << input->info()->fixed_point_position();
+    if(is_data_type_fixed_point(input->info()->data_type()))
+    {
+        ma_arguments << "-DBETA=" << (input->info()->data_type() == DataType::QS8 ?
+                                      scvt_qs8_f32(beta, input->info()->fixed_point_position()) :
+                                      scvt_qs16_f32(beta, input->info()->fixed_point_position()))
+                     << " ";
+        ma_arguments << "-DFIXED_POINT_POSITION=" << input->info()->fixed_point_position();
+    }
+    else
+    {
+        ma_arguments << "-DBETA=" << beta;
+    }
+
     std::set<std::string> build_opts;
     build_opts.emplace(ma_arguments.str());
 
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 7c5b3d7866..2d6b83a97d 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -50,7 +50,7 @@ CLGEMMMatrixMultiplyKernel::CLGEMMMatrixMultiplyKernel()
 
 void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
 
@@ -74,8 +74,18 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
 
     std::ostringstream mm_arguments;
     mm_arguments << "-DWIDTH_MATRIX_B=" << input1->info()->dimension(0) << " ";
-    mm_arguments << "-DALPHA=" << (input0->info()->data_type() == DataType::QS8 ? scvt_qs8_f32(alpha, input0->info()->fixed_point_position()) : alpha) << " ";
-    mm_arguments << "-DFIXED_POINT_POSITION=" << input0->info()->fixed_point_position() << " ";
+    if(is_data_type_fixed_point(input0->info()->data_type()))
+    {
+        mm_arguments << "-DALPHA=" << (input0->info()->data_type() == DataType::QS8 ?
+                                       scvt_qs8_f32(alpha, input0->info()->fixed_point_position()) :
+                                       scvt_qs16_f32(alpha, input0->info()->fixed_point_position()))
+                     << " ";
+        mm_arguments << "-DFIXED_POINT_POSITION=" << input0->info()->fixed_point_position() << " ";
+    }
+    else
+    {
+        mm_arguments << "-DALPHA=" << alpha << " ";
+    }
     std::set<std::string> build_opts;
 
     // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
index ecef7e1774..73c8429055 100644
--- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
@@ -40,7 +40,8 @@ using namespace arm_compute;
 
 void CLGEMMTranspose1xWKernel::configure(const ICLTensor *input, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16,
+                                                  DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape  output_shape{ input->info()->tensor_shape() };
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index 6d22825694..935e856333 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -45,7 +45,7 @@ CLGEMM::CLGEMM()
 
 void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
 
     if(c != nullptr)
diff --git a/tests/validation/CL/GEMM.cpp b/tests/validation/CL/GEMM.cpp
index f79d84f271..a9b35a8f62 100644
--- a/tests/validation/CL/GEMM.cpp
+++ b/tests/validation/CL/GEMM.cpp
@@ -50,7 +50,7 @@ using namespace arm_compute::test::validation;
 namespace
 {
 const float tolerance_f32 = 1e-03f; /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-const float tolerance_qs8 = 1.0f;   /**< Tolerance value for comparing reference's output against implementation's output for DataType::QS8 */
+const float tolerance_q   = 1.0f;   /**< Tolerance value for comparing reference's output against implementation's output for fixed point data types */
 
 CLTensor compute_gemm(const TensorShape &src_shape1, const TensorShape &src_shape2, const TensorShape &src_shape3,
                       const TensorShape &out_shape, float alpha, float beta, DataType dt, int fixed_point_position = 0)
@@ -104,7 +104,7 @@ BOOST_AUTO_TEST_SUITE(GEMM)
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly"))
 BOOST_DATA_TEST_CASE(Configuration,
-                     SmallGEMMDataset() * boost::unit_test::data::make({ DataType::F32, DataType::QS8 }),
+                     SmallGEMMDataset() * boost::unit_test::data::make({ DataType::F32, DataType::QS8, DataType::QS16 }),
                      gemm_set, dt)
 {
     // Set fixed point position data type allowed
@@ -169,7 +169,7 @@ BOOST_AUTO_TEST_SUITE_END()
 
 BOOST_AUTO_TEST_SUITE(Quantized)
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
-BOOST_DATA_TEST_CASE(SmallGEMM, SmallGEMMDataset() * boost::unit_test::data::make(DataType::QS8) * boost::unit_test::data::xrange(4, 7),
+BOOST_DATA_TEST_CASE(SmallGEMM, SmallGEMMDataset() * boost::unit_test::data::make({ DataType::QS8, DataType::QS16 }) * boost::unit_test::data::xrange(4, 7),
                      gemm_set, dt, fixed_point_position)
 {
     // Compute reference
@@ -179,11 +179,11 @@ BOOST_DATA_TEST_CASE(SmallGEMM, SmallGEMMDataset() * boost::unit_test::data::mak
     CLTensor dst = compute_gemm(gemm_set.shape_a, gemm_set.shape_b, gemm_set.shape_c, gemm_set.shape_d, gemm_set.alpha, gemm_set.beta, dt, fixed_point_position);
 
     // Validate output
-    validate(CLAccessor(dst), ref_dst, tolerance_qs8);
+    validate(CLAccessor(dst), ref_dst, tolerance_q);
 }
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
-BOOST_DATA_TEST_CASE(LargeGEMM, LargeGEMMDataset() * boost::unit_test::data::make(DataType::QS8) * boost::unit_test::data::xrange(4, 7),
+BOOST_DATA_TEST_CASE(LargeGEMM, LargeGEMMDataset() * boost::unit_test::data::make({ DataType::QS8, DataType::QS16 }) * boost::unit_test::data::xrange(4, 7),
                      gemm_set, dt, fixed_point_position)
 {
     // Compute reference
@@ -193,7 +193,7 @@ BOOST_DATA_TEST_CASE(LargeGEMM, LargeGEMMDataset() * boost::unit_test::data::mak
     CLTensor dst = compute_gemm(gemm_set.shape_a, gemm_set.shape_b, gemm_set.shape_c, gemm_set.shape_d, gemm_set.alpha, gemm_set.beta, dt, fixed_point_position);
 
     // Validate output
-    validate(CLAccessor(dst), ref_dst, tolerance_qs8);
+    validate(CLAccessor(dst), ref_dst, tolerance_q);
 }
 BOOST_AUTO_TEST_SUITE_END()