COMPMID-411 - Port CLGEMM to support 8 bit fixed point

Change-Id: I6c8bd69ae9715e4d83d128b2162fc15aa5561afb
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78804
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>

7 files changed, 336 insertions, 83 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 081edac8e1..6cf5ce2564 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -157,12 +157,15 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_interleave4x4_32bit", "gemm.cl" },
     { "gemm_ma_f16", "gemm.cl" },
     { "gemm_ma_f32", "gemm.cl" },
+    { "gemm_ma_qs8", "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_vm_f16", "gemm.cl" },
     { "gemm_vm_f32", "gemm.cl" },
+    { "gemm_vm_qs8", "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 2c100c2e28..c0855db056 100644
--- a/src/core/CL/cl_kernels/fixed_point.h
+++ b/src/core/CL/cl_kernels/fixed_point.h
@@ -82,9 +82,9 @@ TYPE_ALIAS(short, qs16)
 
 /* Computes max of fixed point types.
  *
- * @param[in] type is the actual data type.
+ * @param[in] type the actual data type.
  *
- * @return The result of the fixed point vector maximum.
+ * @return The result of the fixed point maximum.
  */
 #define MAXQ_IMPL(type)                          \
     inline type max_##type(type VopA, type VopB) \
@@ -103,9 +103,9 @@ MAXQ_IMPL(qs8x16)
 
 /* Computes saturated addition of fixed point types.
  *
- * @param[in] type is the actual data type.
+ * @param[in] type the actual data type.
  *
- * @return The result of the fixed point vector addition. The result is saturated in case of overflow
+ * @return The result of the fixed point addition. The result is saturated in case of overflow
  */
 #define ADDQ_SAT_IMPL(type)                          \
     inline type add_sat_##type(type VopA, type VopB) \
@@ -124,9 +124,9 @@ ADDQ_SAT_IMPL(qs8x16)
 
 /* Computes saturated subtraction of fixed point types.
  *
- * @param[in] type is the actual data type.
+ * @param[in] type the actual data type.
  *
- * @return The result of the fixed point vector subtraction. The result is saturated in case of overflow
+ * @return The result of the fixed point subtraction. The result is saturated in case of overflow
  */
 #define SUBQ_SAT_IMPL(type)                          \
     inline type sub_sat_##type(type VopA, type VopB) \
@@ -143,12 +143,12 @@ SUBQ_SAT_IMPL(qs8x16)
 #define SUB_SAT_OP_EXPAND_STR(a, b, type, size) sub_sat_##type##x##size((a), (b))
 #define SUB_SAT_OP_EXPAND(a, b, type, size) SUB_SAT_OP_EXPAND_STR(a, b, type, size)
 
-/* Saturate multiply of two fixed point vectors
+/* Saturate multiply of two fixed point numbers
  *
- * @param[in] type  is the actual data type.
- * @param[in] itype is the intermediate data type.
+ * @param[in] type  the actual data type.
+ * @param[in] itype the intermediate data type.
  *
- * @return The result of the fixed point vector subtraction. The result is saturated in case of overflow
+ * @return The result of the fixed point multiplication. The result is saturated in case of overflow
  */
 #define MULQ_SAT_IMPL(type, itype)                                                            \
     inline type mul_sat_##type(type VopA, type VopB, int fixed_point_position)                \
@@ -163,11 +163,50 @@ MULQ_SAT_IMPL(qs8x16, qs16x16)
 #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)
 
-/** Saturate division of two fixed point vectors
+/* Saturate multiply-accumulate
+ *
+ * @param[in] type  the actual data type.
+ * @param[in] itype the intermediate data type.
+ *
+ * @return The result of the fixed point multiply-accumulate. The result is saturated in case of overflow
+ */
+#define MLAQ_SAT_IMPL(type, itype)                                                                                 \
+    type mla_sat_##type(type VopA, type VopB, type VopC, int fixed_point_position)                                 \
+    {                                                                                                              \
+        itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \
+        return add_sat(VopA, CONVERT_SAT(res >> (itype)fixed_point_position, type));                               \
+    }
+
+MLAQ_SAT_IMPL(qs8x8, qs16x8)
+MLAQ_SAT_IMPL(qs8x16, qs16x16)
+
+#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)
+
+/* Saturate multiply-accumulate long
+ *
+ * @param[in] type  the actual data type.
+ * @param[in] itype the intermediate data type.
+ *
+ * @return The result of the fixed point multiply-accumulate long. The result is saturated in case of overflow
+ */
+#define MLALQ_SAT_IMPL(type, itype)                                                                                \
+    itype mlal_sat_##type(itype VopA, type VopB, type VopC, int fixed_point_position)                              \
+    {                                                                                                              \
+        itype res = mad_sat(CONVERT(VopB, itype), CONVERT(VopC, itype), (itype)(1 << (fixed_point_position - 1))); \
+        return add_sat(VopA, res >> (itype)fixed_point_position);                                                  \
+    }
+
+MLALQ_SAT_IMPL(qs8x8, qs16x8)
+
+#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)
+
+/** Saturate division of two fixed point numbers
   *
-  * @param[in] stype is the actual scalar data type.
-  * @param[in] type  is the actual data type.
-  * @param[in] itype is the intermediate data type.
+  * @param[in] stype the actual scalar data type.
+  * @param[in] type  the actual data type.
+  * @param[in] itype the intermediate data type.
   *
   * @return The result of the fixed point division. The result is saturated in case of overflow
   */
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index 9bec8d5d92..796b343bda 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -21,17 +21,18 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
+#include "fixed_point.h"
 #include "helpers.h"
 
 /** This OpenCL kernel computes the "vector" 1x4 transposition of input matrix
  *
- * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U32/S32/F32
  * @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: F32
+ * @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)
@@ -57,13 +58,13 @@ __kernel void gemm_transpose1x4(IMAGE_DECLARATION(src),
 
 /** This OpenCL kernel computes the "vector" 1x8 transposition of input matrix
  *
- * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F16
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U16/S16/QS16/F16
  * @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: F16
+ * @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)
@@ -89,13 +90,13 @@ __kernel void gemm_transpose1x8(IMAGE_DECLARATION(src),
 
 /** This OpenCL kernel computes the "vector" 1x16 transposition of input matrix
  *
- * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U8
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U8/S8/QS8
  * @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: U8
+ * @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)
@@ -127,7 +128,7 @@ __kernel void gemm_transpose1x16(IMAGE_DECLARATION(src),
  * @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: U32/S32/F32
+ * @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)
@@ -162,13 +163,13 @@ __kernel void gemm_interleave4x4_32bit(IMAGE_DECLARATION(src),
 
 /** This OpenCL kernel reshapes the input matrix transposing each 4x4 block and interleaving the values
  *
- * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U16/S16/F16
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U16/S16/QS16/F16
  * @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: U16/S16/F16
+ * @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)
@@ -203,13 +204,13 @@ __kernel void gemm_interleave4x4_16bit(IMAGE_DECLARATION(src),
 
 /** This OpenCL kernel reshapes the input matrix transposing each 4x4 block and interleaving the values
  *
- * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U8/S8
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U8/S8/QS8
  * @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: U8/S8
+ * @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)
@@ -250,13 +251,13 @@ __kernel void gemm_interleave4x4_8bit(IMAGE_DECLARATION(src),
  *
  * @note The data type must be passed at compile time -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
  *
- * @param[in, out] accum_ptr                            Pointer to the accumulate tensor. Supported data type: F32
+ * @param[in, out] accum_ptr                            Pointer to the accumulate tensor. Supported data type: U8/S8/QS8/U16/S16/F16/U32/S32/F32
  * @param[in]      accum_stride_x                       Stride of the accmulate tensor in X dimension (in bytes)
  * @param[in]      accum_step_x                         accum_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]      accum_stride_y                       Stride of the accumlulate tensor in Y dimension (in bytes)
  * @param[in]      accum_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]      accum_offset_first_element_in_bytes  The offset of the first element in the accumulate tensor
- * @param[in]      biases_ptr                           Pointer to the biases vector. Same as input.
+ * @param[in]      biases_ptr                           Pointer to the biases vector. Same as @p accum_ptr
  * @param[in]      biases_stride_x                      Stride of the destination tensor in X dimension (in bytes)
  * @param[in]      biases_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]      biases_offset_first_element_in_bytes The offset of the first element in the destination tensor
@@ -282,7 +283,7 @@ __kernel void gemm_accumulate_biases(
 
 #if(defined WIDTH_MATRIX_B)
 /** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1)
- *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_u8 and @ref gemm_transpose1x16_u8 before running the matrix multiplication
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_8bit and @ref gemm_transpose1x16 before running the matrix multiplication
  *
  * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B
  *
@@ -292,13 +293,13 @@ __kernel void gemm_accumulate_biases(
  * @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 formats: U8
+ * @param[in]  src1_ptr                           Pointer to the source matrix. Supported formats: 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 formats: U8
+ * @param[out] dst_ptr                            Pointer to the destination matrix Supported formats: 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)
@@ -388,7 +389,7 @@ __kernel void gemm_mm_u8(IMAGE_DECLARATION(src0),
 
 #if(defined WIDTH_MATRIX_B && defined ALPHA)
 /** This OpenCL kernel is optimised for Midgard. It computes the matrix multiplication between matrix A (src0) and matrix B (src1)
- *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_f32 and @ref gemm_transpose1x4_f32 before running the matrix multiplication
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication
  *
  * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA
  *
@@ -398,13 +399,13 @@ __kernel void gemm_mm_u8(IMAGE_DECLARATION(src0),
  * @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: F32
+ * @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: F32
+ * @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)
@@ -487,7 +488,7 @@ __kernel void gemm_mm_f32_midgard(IMAGE_DECLARATION(src0),
 }
 
 /** This OpenCL kernel is optimised for Bifrost. It computes the matrix multiplication between matrix A (src0) and matrix B (src1)
- *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_f32 and @ref gemm_transpose1x4_f32 before running the matrix multiplication
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication
  *
  * @attention The width of matrix B and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA
  *
@@ -497,13 +498,13 @@ __kernel void gemm_mm_f32_midgard(IMAGE_DECLARATION(src0),
  * @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: F32
+ * @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: F32
+ * @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)
@@ -697,7 +698,7 @@ __kernel void gemm_mm_f32_bifrost(IMAGE_DECLARATION(src0),
 }
 
 /** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1)
- *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_f16 and @ref gemm_transpose1x8_f16 before running the matrix multiplication
+ *  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 and the alpha's value need to be passed at compile time using -DWIDTH_MATRIX_B and -DALPHA
  *
@@ -707,13 +708,13 @@ __kernel void gemm_mm_f32_bifrost(IMAGE_DECLARATION(src0),
  * @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: F16
+ * @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: F16
+ * @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)
@@ -795,6 +796,100 @@ __kernel void gemm_mm_f16(IMAGE_DECLARATION(src0),
     vstore8(c30, 0, (__global half *)(offset(&dst, 0, 3)));
 }
 
+#if(defined FIXED_POINT_POSITION)
+/** This OpenCL kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1) in 8 bit fixed point precision
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_8bit and @ref gemm_transpose1x16 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 8 bit fixed point format
+ *
+ * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS8
+ * @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_qs8(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));
+
+    /* Compute end row address for matrix B */
+    int end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;
+
+    /* Reset accumulators */
+    short8 c00 = 0.0f;
+    short8 c10 = 0.0f;
+    short8 c20 = 0.0f;
+    short8 c30 = 0.0f;
+    short8 c01 = 0.0f;
+    short8 c11 = 0.0f;
+    short8 c21 = 0.0f;
+    short8 c31 = 0.0f;
+
+    /* This for loop performs 1 accumulation for each iteration */
+    for(; src_addr.s1 <= (end_row_mtx_b - 16); src_addr += (int2)(4, 16))
+    {
+        /* Load values from matrix A (interleaved) and matrix B (transposed) */
+        char4  a0 = vload4(0, ((__global char *)src0_ptr) + src_addr.s0);
+        char16 b0 = vload16(0, ((__global char *)src1_ptr) + src_addr.s1);
+
+        c00 = mlal_sat_qs8x8(c00, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION);
+        c10 = mlal_sat_qs8x8(c10, (char8)a0.s1, b0.s01234567, FIXED_POINT_POSITION);
+        c20 = mlal_sat_qs8x8(c20, (char8)a0.s2, b0.s01234567, FIXED_POINT_POSITION);
+        c30 = mlal_sat_qs8x8(c30, (char8)a0.s3, b0.s01234567, FIXED_POINT_POSITION);
+
+        c01 = mlal_sat_qs8x8(c01, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION);
+        c11 = mlal_sat_qs8x8(c11, (char8)a0.s1, b0.s89ABCDEF, FIXED_POINT_POSITION);
+        c21 = mlal_sat_qs8x8(c21, (char8)a0.s2, b0.s89ABCDEF, FIXED_POINT_POSITION);
+        c31 = mlal_sat_qs8x8(c31, (char8)a0.s3, b0.s89ABCDEF, FIXED_POINT_POSITION);
+    }
+
+    /* Compute destination address */
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Multiply by the weight of matrix product */
+    char16 c00_qs8 = convert_char16_sat((short16)(c00, c01));
+    char16 c10_qs8 = convert_char16_sat((short16)(c10, c11));
+    char16 c20_qs8 = convert_char16_sat((short16)(c20, c21));
+    char16 c30_qs8 = convert_char16_sat((short16)(c30, c31));
+
+    c00_qs8 = mul_sat_qs8x16(c00_qs8, (char16)ALPHA, FIXED_POINT_POSITION);
+    c10_qs8 = mul_sat_qs8x16(c10_qs8, (char16)ALPHA, FIXED_POINT_POSITION);
+    c20_qs8 = mul_sat_qs8x16(c20_qs8, (char16)ALPHA, FIXED_POINT_POSITION);
+    c30_qs8 = mul_sat_qs8x16(c30_qs8, (char16)ALPHA, FIXED_POINT_POSITION);
+
+    /* Store 16x4 block */
+    vstore16(c00_qs8, 0, (__global char *)(offset(&dst, 0, 0)));
+    vstore16(c10_qs8, 0, (__global char *)(offset(&dst, 0, 1)));
+    vstore16(c20_qs8, 0, (__global char *)(offset(&dst, 0, 2)));
+    vstore16(c30_qs8, 0, (__global char *)(offset(&dst, 0, 3)));
+}
+#endif // (defined FIXED_POINT_POSITION)
+
 #if(defined WIDTH_VECTOR_A)
 /** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1)
  *
@@ -808,13 +903,13 @@ __kernel void gemm_mm_f16(IMAGE_DECLARATION(src0),
  * @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: F32
+ * @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: F32
+ * @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)
@@ -874,13 +969,13 @@ __kernel void gemm_vm_f32(IMAGE_DECLARATION(src0),
  * @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: F16
+ * @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: F16
+ * @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)
@@ -931,6 +1026,92 @@ __kernel void gemm_vm_f16(IMAGE_DECLARATION(src0),
 
     vstore8(acc, 0, (__global half *)(offset(&dst, 0, 0)));
 }
+
+#if(defined FIXED_POINT_POSITION)
+/** This OpenCL kernel computes the vector by matrix multiplication between the vector A (src0) and matrix B (src1) in 8 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 8 bit fixed point format
+ *
+ * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: QS8
+ * @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_qs8(IMAGE_DECLARATION(src0),
+                          IMAGE_DECLARATION(src1),
+                          IMAGE_DECLARATION(dst))
+{
+    int idx = get_global_id(0) * 16;
+
+    /* 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;
+
+    int end_row_vec_a = src_addr.s0 + WIDTH_VECTOR_A;
+
+    short8 acc0 = 0;
+    short8 acc1 = 0;
+
+    /* This for loop performs 4 accumulations per iteration */
+    for(; src_addr.s0 <= (end_row_vec_a - 4); src_addr += (int2)(4, 4 * src1_stride_y))
+    {
+        char4  a0 = vload4(0, (__global char *)(src0_ptr + src_addr.s0));
+        char16 b0 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y));
+        char16 b1 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y));
+        char16 b2 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y));
+        char16 b3 = vload16(0, (__global char *)(src1_ptr + src_addr.s1 + 3 * src1_stride_y));
+
+        acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION);
+        acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s1, b1.s01234567, FIXED_POINT_POSITION);
+        acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s2, b2.s01234567, FIXED_POINT_POSITION);
+        acc0 = mlal_sat_qs8x8(acc0, (char8)a0.s3, b3.s01234567, FIXED_POINT_POSITION);
+
+        acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION);
+        acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s1, b1.s89ABCDEF, FIXED_POINT_POSITION);
+        acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s2, b2.s89ABCDEF, FIXED_POINT_POSITION);
+        acc1 = mlal_sat_qs8x8(acc1, (char8)a0.s3, b3.s89ABCDEF, FIXED_POINT_POSITION);
+    }
+
+    /* Left-over accumulations */
+    for(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y))
+    {
+        char   a0 = *((__global char *)(src0_ptr + src_addr.s0));
+        char16 b0 = vload16(0, (__global char *)(src1_ptr + src_addr.s1));
+
+        acc0 = mlal_sat_qs8x8(acc0, (char8)a0, b0.s01234567, FIXED_POINT_POSITION);
+        acc1 = mlal_sat_qs8x8(acc1, (char8)a0, b0.s89ABCDEF, FIXED_POINT_POSITION);
+    }
+
+    /* Compute destination address */
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Multiply by the weight of matrix product */
+    char16 acc_qs8 = convert_char16_sat((short16)(acc0, acc1));
+
+    acc_qs8 = mul_sat_qs8x16(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION);
+
+    /* Store 16 values */
+    vstore16(acc_qs8, 0, (__global char *)(offset(&dst, 0, 0)));
+}
+#endif /* #if(defined FIXED_POINT_POSITION) */
 #endif /* (defined WIDTH_VECTOR_A) */
 #endif /* (defined WIDTH_MATRIX_B && defined ALPHA) */
 
@@ -945,7 +1126,7 @@ __kernel void gemm_vm_f16(IMAGE_DECLARATION(src0),
  * @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: F32
+ * @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)
@@ -974,13 +1155,15 @@ __kernel void gemm_ma_f32(IMAGE_DECLARATION(src),
 
 /** 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:
  *
+ * @attention The beta's value need to be passed at compile time using -DBETA
+ *
  * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F16
  * @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: F16
+ * @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)
@@ -1006,6 +1189,47 @@ __kernel void gemm_ma_f16(IMAGE_DECLARATION(src),
     /* Store final result in axb matrix */
     vstore8(out, 0, (__global half *)dst.ptr);
 }
+
+#if(defined FIXED_POINT_POSITION)
+/** This OpenCL kernel performs the in-place matrix addition between 2 matrices in 8 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 8 bit fixed point format
+ *
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: QS8
+ * @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_qs8(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 */
+    char16 alpha_ab = vload16(0, (__global char *)dst.ptr);
+
+    /* Load values from Matrix C */
+    char16 c = vload16(0, (__global char *)src.ptr);
+
+    /* Computes alpha * axb + beta * c */
+    char16 out = mla_sat_qs8x16(alpha_ab, (char16)BETA, c, FIXED_POINT_POSITION);
+
+    /* Store final result in axb matrix */
+    vstore16(out, 0, (__global char *)dst.ptr);
+}
+#endif /* #if(defined FIXED_POINT_POSITION) */
 #endif /* (defined BETA) */
 
 #if(defined WIDTH_VECTOR_A)
@@ -1021,7 +1245,7 @@ __kernel void gemm_ma_f16(IMAGE_DECLARATION(src),
  * @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: F32
+ * @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)
@@ -1029,7 +1253,7 @@ __kernel void gemm_ma_f16(IMAGE_DECLARATION(src),
  * @param[in]  src1_stride_z                      Stride of the source matrix in Z dimension (in bytes)
  * @param[in]  src1_step_z                        src_stride_z * number of elements along Z 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: F32
+ * @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)
diff --git a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
index 343838f2f9..5883dd698b 100644
--- a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
@@ -28,6 +28,7 @@
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/OpenCL.h"
 #include "arm_compute/core/Error.h"
+#include "arm_compute/core/FixedPoint.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
@@ -40,10 +41,9 @@ CLGEMMMatrixAdditionKernel::CLGEMMMatrixAdditionKernel()
 {
 }
 
-void CLGEMMMatrixAdditionKernel::configure(const ICLTensor *input, ICLTensor *output, const float beta)
+void CLGEMMMatrixAdditionKernel::configure(const ICLTensor *input, ICLTensor *output, float beta)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, 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));
@@ -53,7 +53,8 @@ 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=" << beta;
+    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();
     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 d7388e8579..7c5b3d7866 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -25,12 +25,12 @@
 
 #include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/AccessWindowTranspose.h"
-
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/OpenCL.h"
 #include "arm_compute/core/Error.h"
+#include "arm_compute/core/FixedPoint.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
@@ -50,10 +50,10 @@ 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::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
+
     if(output->info()->dimension(1) == 1)
     {
         ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1));
@@ -74,7 +74,8 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
 
     std::ostringstream mm_arguments;
     mm_arguments << "-DWIDTH_MATRIX_B=" << input1->info()->dimension(0) << " ";
-    mm_arguments << "-DALPHA=" << alpha << " ";
+    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() << " ";
     std::set<std::string> build_opts;
 
     // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
@@ -98,7 +99,9 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
 
         update_window_and_padding(win, input0_access, input1_access, output_access);
 
-        output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape()));
+        Coordinates coord;
+        coord.set_num_dimensions(output->info()->num_dimensions());
+        output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape()));
 
         ICLKernel::configure(win);
     }
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index 07b19421d6..6d22825694 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -45,13 +45,11 @@ 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::F32, DataType::F16);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::F32, DataType::F16);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32, DataType::F16);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
 
     if(c != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
         ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
         ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix C");
@@ -59,7 +57,6 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *
         ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != output->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix");
     }
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
     ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
 
     // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped
@@ -73,25 +70,14 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *
         shape_tmp_a.set(0, a->info()->dimension(0) * 4);
         shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f));
 
-        if(DataType::F32 == a->info()->data_type())
-        {
-            shape_tmp_b.set(0, b->info()->dimension(1) * 4);
-            shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 4.0f));
-        }
-        else if(DataType::F16 == a->info()->data_type())
-        {
-            shape_tmp_b.set(0, b->info()->dimension(1) * 8);
-            shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 8.0f));
-        }
-        else
-        {
-            ARM_COMPUTE_ERROR("DataType not supported");
-        }
-
-        TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type());
+        const unsigned int transpose_w = max_cl_vector_width / data_size_from_type(b->info()->data_type());
+        shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w);
+        shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w)));
+
+        TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position());
         _tmp_a.allocator()->init(info_a);
 
-        TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type());
+        TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), b->info()->fixed_point_position());
         _tmp_b.allocator()->init(info_b);
 
         // Configure interleave kernel
diff --git a/src/runtime/NEON/functions/NEGEMM.cpp b/src/runtime/NEON/functions/NEGEMM.cpp
index 15d5f4effb..730735590d 100644
--- a/src/runtime/NEON/functions/NEGEMM.cpp
+++ b/src/runtime/NEON/functions/NEGEMM.cpp
@@ -44,12 +44,10 @@ NEGEMM::NEGEMM()
 void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *d, float alpha, float beta)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16, DataType::QS8);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::F32, DataType::F16, DataType::QS8);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(d, 1, DataType::F32, DataType::F16, DataType::QS8);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, d);
 
     if(c != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16, DataType::QS8);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
         ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
         ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix B");
@@ -57,7 +55,6 @@ void NEGEMM::configure(const ITensor *a, const ITensor *b, const ITensor *c, ITe
         ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != d->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix");
     }
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, d);
     ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
 
     // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped