Implement OpenCL MatMul for Lhs T Rhs T/NT FP32/16

 - Implement opencl kernel for LHS transposed and RHS non-transposed
 - Implement opencl kernel for LHS transposed and RHS transposed
 - Add validation tests

Resolves: COMPMID-5953, COMPMID-5955
Change-Id: I55589acbffe86c44e29807574975978a1ec09bad
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9345
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 369 insertions, 7 deletions

diff --git a/src/core/CL/cl_kernels/common/mat_mul.cl b/src/core/CL/cl_kernels/common/mat_mul.cl
index 7c74e9d07b..956d37a9d8 100644
--- a/src/core/CL/cl_kernels/common/mat_mul.cl
+++ b/src/core/CL/cl_kernels/common/mat_mul.cl
@@ -29,8 +29,11 @@
  *
  * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it
  *       should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension
+ * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4).
+ * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3)
  * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6)
+ * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_NT_NT)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 > 0
  *  - N0 = 1, 2, 3, 4, 8, 16
@@ -44,14 +47,14 @@
  * @param[in]  lhs_h                             The height of the lhs tensor
  * @param[in]  lhs_n                             Number of the matrices (buffers) in the batch
  * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix
- * @param[in]  rhs_ptr                           Pointer to the rhs matrix. Supported data types: F32/F16
+ * @param[in]  rhs_ptr                           Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr
  * @param[in]  rhs_stride_y                      Stride of the rhs matrix in Y (2nd) dimension (in bytes)
  * @param[in]  rhs_stride_z                      Stride of the rhs tensor in Z (3rd) dimension (in bytes)
  * @param[in]  rhs_w                             The width of the rhs tensor
  * @param[in]  rhs_h                             The height of the rhs tensor
  * @param[in]  rhs_n                             Number of the matrices (buffers) in the batch
  * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix
- * @param[out] dst_ptr                           Pointer to the dst matrix. Supported data types: F32/F16
+ * @param[out] dst_ptr                           Pointer to the dst matrix. Supported data types: same as @p lhs_ptr
  * @param[in]  dst_stride_y                      Stride of the dst matrix in Y (2nd) dimension (in bytes)
  * @param[in]  dst_stride_z                      Stride of the dst tensor in Z (3rd) dimension (in bytes)
  * @param[in]  dst_w                             The width of the dst tensor
@@ -108,6 +111,7 @@ __kernel void mat_mul_native_nt_nt(
     }
 
 #ifdef K % K0 != 0
+    /* Leftover Loop */
     for(; k < K; ++k)
     {
         TILE(DATA_TYPE, M0, 1, a);
@@ -152,8 +156,11 @@ __kernel void mat_mul_native_nt_nt(
  *
  * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it
  *       should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension
+ * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
  * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4).
+ * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3)
  * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6)
+ * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_NT_T)
  * @note Only the following configurations of M0, N0 and K0 are currently supported:
  *  - M0 > 0
  *  - N0 = 1, 2, 3, 4, 8, 16
@@ -167,14 +174,14 @@ __kernel void mat_mul_native_nt_nt(
  * @param[in]  lhs_h                             The height of the lhs tensor
  * @param[in]  lhs_n                             Number of the matrices (buffers) in the batch
  * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix
- * @param[in]  rhs_ptr                           Pointer to the rhs matrix. Supported data types: F32/F16
+ * @param[in]  rhs_ptr                           Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr
  * @param[in]  rhs_stride_y                      Stride of the rhs matrix in Y (2nd) dimension (in bytes)
  * @param[in]  rhs_stride_z                      Stride of the rhs tensor in Z (3rd) dimension (in bytes)
  * @param[in]  rhs_w                             The width of the rhs tensor
  * @param[in]  rhs_h                             The height of the rhs tensor
  * @param[in]  rhs_n                             Number of the matrices (buffers) in the batch
  * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix
- * @param[out] dst_ptr                           Pointer to the dst matrix. Supported data types: F32/F16
+ * @param[out] dst_ptr                           Pointer to the dst matrix. Supported data types: same as @p lhs_ptr
  * @param[in]  dst_stride_y                      Stride of the dst matrix in Y (2nd) dimension (in bytes)
  * @param[in]  dst_stride_z                      Stride of the dst tensor in Z (3rd) dimension (in bytes)
  * @param[in]  dst_w                             The width of the dst tensor
@@ -239,7 +246,7 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER),
             })
         })
         T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, a, bt, acc);
-#else // GPU_ARCH == GPU_ARCH_MIDGARD
+#else  // GPU_ARCH == GPU_ARCH_MIDGARD
         T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, T, a, b, acc);
 #endif // GPU_ARCH == GPU_ARCH_MIDGARD
 
@@ -276,7 +283,7 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER),
             bt[0].s[i] = b[i].s[0];
         })
         T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, a, bt, acc);
-#else // GPU_ARCH == GPU_ARCH_MIDGARD
+#else  // GPU_ARCH == GPU_ARCH_MIDGARD
         T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, T, a, b, acc);
 #endif // GPU_ARCH == GPU_ARCH_MIDGARD
 
@@ -296,4 +303,323 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER),
 
     T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, acc, indirect_buffer);
 }
-#endif // defined(MAT_MUL_NATIVE_NT_T)
\ No newline at end of file
+#endif // defined(MAT_MUL_NATIVE_NT_T)
+
+#if defined(MAT_MUL_NATIVE_T_NT)
+/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS non-transposed - buffer only
+ *
+ * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it
+ *       should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension
+ * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
+ * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4).
+ * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3)
+ * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6)
+ * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_T_NT)
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 = 1, 2, 3, 4, 8, 16
+ *  - N0 = 1, 2, 3, 4, 8, 16
+ *  - K0 > 0
+ * * @note Values > 8 for M0, and K0 are not expected to be efficient
+ *
+ * @param[in]  lhs_ptr                           Pointer to the lhs matrix. Supported data types: F32/F16
+ * @param[in]  lhs_stride_y                      Stride of the lhs matrix in Y (2nd) dimension (in bytes)
+ * @param[in]  lhs_stride_z                      Stride of the lhs tensor in Z (3rd) dimension (in bytes)
+ * @param[in]  lhs_w                             The width of the lhs tensor
+ * @param[in]  lhs_h                             The height of the lhs tensor
+ * @param[in]  lhs_n                             Number of the matrices (buffers) in the batch
+ * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix
+ * @param[in]  rhs_ptr                           Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr
+ * @param[in]  rhs_stride_y                      Stride of the rhs matrix in Y (2nd) dimension (in bytes)
+ * @param[in]  rhs_stride_z                      Stride of the rhs tensor in Z (3rd) dimension (in bytes)
+ * @param[in]  rhs_w                             The width of the rhs tensor
+ * @param[in]  rhs_h                             The height of the rhs tensor
+ * @param[in]  rhs_n                             Number of the matrices (buffers) in the batch
+ * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix
+ * @param[out] dst_ptr                           Pointer to the dst matrix. Supported data types: same as @p lhs_ptr
+ * @param[in]  dst_stride_y                      Stride of the dst matrix in Y (2nd) dimension (in bytes)
+ * @param[in]  dst_stride_z                      Stride of the dst tensor in Z (3rd) dimension (in bytes)
+ * @param[in]  dst_w                             The width of the dst tensor
+ * @param[in]  dst_h                             The height of the dst tensor
+ * @param[in]  dst_n                             Number of the matrices (buffers) in the batch
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the dst matrix
+ */
+__kernel void mat_mul_native_t_nt(
+    TENSOR3D_T(lhs, BUFFER),
+    TENSOR3D_T(rhs, BUFFER),
+    TENSOR3D_T(dst, BUFFER))
+{
+    const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0);
+    const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0);
+    const uint z = GET_SPATIAL_IDX(2, 1, 0);
+
+    // Compute LHS/RHS/DST matrix address
+    lhs_offset_first_element_in_bytes += y * sizeof(DATA_TYPE) + z * lhs_stride_z;
+    rhs_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + z * rhs_stride_z;
+    dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z;
+
+    // Initialize the accumulators
+    TILE(DATA_TYPE, M0, N0, acc);
+
+    LOOP_UNROLLING(int, i, 0, 1, M0,
+    {
+        acc[i].v = 0.f;
+    })
+
+    int k;
+    for(k = 0; k <= K - K0; k += K0)
+    {
+        TILE(DATA_TYPE, K0, M0, a);
+        TILE(DATA_TYPE, K0, N0, b);
+
+        LOOP_UNROLLING(int, i, 0, 1, K0,
+        {
+            a[i].v = 0.f;
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, K0,
+        {
+            b[i].v = 0.f;
+        })
+
+        // Load tile from the lhs/rhs tensors
+        T_LOAD(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+        T_LOAD(DATA_TYPE, K0, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b);
+
+#if GPU_ARCH == GPU_ARCH_MIDGARD
+        // For explanation, see mat_mul_native_nt_t
+        TILE(DATA_TYPE, M0, K0, at);
+        LOOP_UNROLLING(int, i, 0, 1, K0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, M0,
+            {
+                at[j].s[i] = a[i].s[j];
+            })
+        })
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, at, b, acc);
+#else  // GPU_ARCH == GPU_ARCH_MIDGARD
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, T, NT, a, b, acc);
+#endif // GPU_ARCH == GPU_ARCH_MIDGARD
+
+        lhs_offset_first_element_in_bytes += K0 * lhs_stride_y;
+        rhs_offset_first_element_in_bytes += K0 * rhs_stride_y;
+    }
+
+#ifdef K % K0 != 0
+    /* Leftover Loop */
+    for(; k < K; ++k)
+    {
+        TILE(DATA_TYPE, 1, M0, a);
+        TILE(DATA_TYPE, 1, N0, b);
+
+        LOOP_UNROLLING(int, i, 0, 1, 1,
+        {
+            a[i].v = 0.f;
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, 1,
+        {
+            b[i].v = 0.f;
+        })
+
+        // Load tile from the lhs/rhs tensors
+        T_LOAD(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+        T_LOAD(DATA_TYPE, 1, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b);
+
+#if GPU_ARCH == GPU_ARCH_MIDGARD
+        // For explanation, see mat_mul_native_nt_t
+        TILE(DATA_TYPE, M0, 1, at);
+        LOOP_UNROLLING(int, j, 0, 1, M0,
+        {
+            at[j].s[0] = a[0].s[j];
+        })
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, at, b, acc);
+#else  // GPU_ARCH == GPU_ARCH_MIDGARD
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, T, NT, a, b, acc);
+#endif // GPU_ARCH == GPU_ARCH_MIDGARD
+
+        lhs_offset_first_element_in_bytes += 1 * lhs_stride_y;
+        rhs_offset_first_element_in_bytes += 1 * rhs_stride_y;
+    }
+#endif // K % K0 != 0
+
+    const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0;
+    const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0;
+
+    TILE(int, M0, 1, indirect_buffer);
+    LOOP_UNROLLING(int, _i, 0, 1, M0,
+    {
+        indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond));
+    });
+
+    T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, acc, indirect_buffer);
+}
+#endif // defined(MAT_MUL_NATIVE_T_NT)
+
+#if defined(MAT_MUL_NATIVE_T_T)
+/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS transposed - buffer only
+ *
+ * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it
+ *       should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension
+ * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float)
+ * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4).
+ * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3)
+ * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6)
+ * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_T_NT)
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 = 1, 2, 3, 4, 8, 16
+ *  - N0 = 1, 2, 3, 4, 8, 16
+ *  - K0 = 1, 2, 3, 4, 8, 16
+ * @note Values > 8 for M0, N0 and K0 are not expected to be efficient
+ *
+ * @param[in]  lhs_ptr                           Pointer to the lhs matrix. Supported data types: F32/F16
+ * @param[in]  lhs_stride_y                      Stride of the lhs matrix in Y (2nd) dimension (in bytes)
+ * @param[in]  lhs_stride_z                      Stride of the lhs tensor in Z (3rd) dimension (in bytes)
+ * @param[in]  lhs_w                             The width of the lhs tensor
+ * @param[in]  lhs_h                             The height of the lhs tensor
+ * @param[in]  lhs_n                             Number of the matrices (buffers) in the batch
+ * @param[in]  lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix
+ * @param[in]  rhs_ptr                           Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr
+ * @param[in]  rhs_stride_y                      Stride of the rhs matrix in Y (2nd) dimension (in bytes)
+ * @param[in]  rhs_stride_z                      Stride of the rhs tensor in Z (3rd) dimension (in bytes)
+ * @param[in]  rhs_w                             The width of the rhs tensor
+ * @param[in]  rhs_h                             The height of the rhs tensor
+ * @param[in]  rhs_n                             Number of the matrices (buffers) in the batch
+ * @param[in]  rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix
+ * @param[out] dst_ptr                           Pointer to the dst matrix. Supported data types: same as @p lhs_ptr
+ * @param[in]  dst_stride_y                      Stride of the dst matrix in Y (2nd) dimension (in bytes)
+ * @param[in]  dst_stride_z                      Stride of the dst tensor in Z (3rd) dimension (in bytes)
+ * @param[in]  dst_w                             The width of the dst tensor
+ * @param[in]  dst_h                             The height of the dst tensor
+ * @param[in]  dst_n                             Number of the matrices (buffers) in the batch
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the dst matrix
+ */
+__kernel void mat_mul_native_t_t(
+    TENSOR3D_T(lhs, BUFFER),
+    TENSOR3D_T(rhs, BUFFER),
+    TENSOR3D_T(dst, BUFFER))
+{
+    const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0);
+    const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0);
+    const uint z = GET_SPATIAL_IDX(2, 1, 0);
+
+    // Compute LHS/RHS/DST matrix address
+    lhs_offset_first_element_in_bytes += y * sizeof(DATA_TYPE) + z * lhs_stride_z;
+    rhs_offset_first_element_in_bytes += x * rhs_stride_y + z * rhs_stride_z;
+    dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z;
+
+    // Initialize the accumulators
+    TILE(DATA_TYPE, M0, N0, acc);
+
+    LOOP_UNROLLING(int, i, 0, 1, M0,
+    {
+        acc[i].v = 0.f;
+    })
+
+    int k;
+    for(k = 0; k <= K - K0; k += K0)
+    {
+        TILE(DATA_TYPE, K0, M0, a);
+        TILE(DATA_TYPE, N0, K0, b);
+
+        LOOP_UNROLLING(int, i, 0, 1, K0,
+        {
+            a[i].v = 0.f;
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            b[i].v = 0.f;
+        })
+
+        // Load tile from the lhs/rhs tensors
+        T_LOAD(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+        T_LOAD(DATA_TYPE, N0, K0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b);
+
+#if GPU_ARCH == GPU_ARCH_MIDGARD
+        // For explanation, see mat_mul_native_nt_t
+        TILE(DATA_TYPE, M0, K0, at);
+        TILE(DATA_TYPE, K0, N0, bt);
+
+        LOOP_UNROLLING(int, i, 0, 1, K0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, M0,
+            {
+                at[j].s[i] = a[i].s[j];
+            })
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, K0,
+            {
+                bt[j].s[i] = b[i].s[j];
+            })
+        })
+
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, at, bt, acc);
+#else  // GPU_ARCH == GPU_ARCH_MIDGARD
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, T, T, a, b, acc);
+#endif // GPU_ARCH == GPU_ARCH_MIDGARD
+
+        lhs_offset_first_element_in_bytes += K0 * lhs_stride_y;
+        rhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
+    }
+
+#ifdef K % K0 != 0
+    /* Leftover Loop */
+    for(; k < K; ++k)
+    {
+        TILE(DATA_TYPE, 1, M0, a);
+        TILE(DATA_TYPE, N0, 1, b);
+
+        LOOP_UNROLLING(int, i, 0, 1, 1,
+        {
+            a[i].v = 0.f;
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            b[i].v = 0.f;
+        })
+
+        // Load tile from the lhs/rhs tensors
+        T_LOAD(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+        T_LOAD(DATA_TYPE, N0, 1, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b);
+
+#if GPU_ARCH == GPU_ARCH_MIDGARD
+        // For explanation, see mat_mul_native_nt_t
+        TILE(DATA_TYPE, M0, 1, at);
+        TILE(DATA_TYPE, 1, N0, bt);
+
+        LOOP_UNROLLING(int, j, 0, 1, M0,
+        {
+            at[j].s[0] = a[0].s[j];
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            bt[0].s[i] = b[i].s[0];
+        })
+
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, at, bt, acc);
+#else  // GPU_ARCH == GPU_ARCH_MIDGARD
+        T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, T, T, a, b, acc);
+#endif // GPU_ARCH == GPU_ARCH_MIDGARD
+
+        lhs_offset_first_element_in_bytes += 1 * lhs_stride_y;
+        rhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE);
+    }
+#endif // K % K0 != 0
+
+    const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0;
+    const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0;
+
+    TILE(int, M0, 1, indirect_buffer);
+    LOOP_UNROLLING(int, _i, 0, 1, M0,
+    {
+        indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond));
+    });
+
+    T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, acc, indirect_buffer);
+}
+#endif // defined(MAT_MUL_NATIVE_T_T)
diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h
index 5d397ad333..872f4c0b57 100644
--- a/src/core/CL/cl_kernels/tile_helpers.h
+++ b/src/core/CL/cl_kernels/tile_helpers.h
@@ -1297,6 +1297,42 @@
         })                                                                                                               \
     }
 
+#define T_MMUL_T_NT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_NT_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_NT_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_NT_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)                       \
+    {                                                                                     \
+        LOOP_UNROLLING(int, _m, 0, 1, M0,                                                 \
+        {                                                                                 \
+            LOOP_UNROLLING(int, _n, 0, 1, N0,                                             \
+            {                                                                             \
+                LOOP_UNROLLING(int, _k, 0, 1, K0,                                         \
+                {                                                                         \
+                    dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_k].s[_m]), (DST_DATA_TYPE)(rhs[_k].s[_n]), dst[_m].s[_n]); \
+                })                                                                        \
+            })                                                                            \
+        })                                                                                \
+    }
+
+#define T_MMUL_T_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_T_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_T_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_T_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
+#define T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)                       \
+    {                                                                                     \
+        LOOP_UNROLLING(int, _m, 0, 1, M0,                                                 \
+        {                                                                                 \
+            LOOP_UNROLLING(int, _n, 0, 1, N0,                                             \
+            {                                                                             \
+                LOOP_UNROLLING(int, _k, 0, 1, K0,                                         \
+                {                                                                         \
+                    dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_k].s[_m]), (DST_DATA_TYPE)(rhs[_n].s[_k]), dst[_m].s[_n]); \
+                })                                                                        \
+            })                                                                            \
+        })                                                                                \
+    }
+
 #define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)                            \
     ({ \
         LOOP_UNROLLING(int, _m, 0, 1, M0, \