Add quantized CL MatMul kernel for LHS NT, RHS T

Implement a native kernel for batched Matrix Multiplication for the quantized data types QASYMM8 and QASYMM8_SIGNED and with the MatMul attributes `adj_x = false, adj_y = true`.

Resolves: COMPMID-5923

Change-Id: I477b2dd886edfe83beaba9efc7d6b05ed19f5da4
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9467
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 176 insertions, 0 deletions

diff --git a/src/core/CL/cl_kernels/common/mat_mul_quantized.cl b/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
index 5c931d2fc1..0c3cbca9a6 100644
--- a/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
+++ b/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
@@ -208,6 +208,182 @@ __kernel void mat_mul_native_quantized_nt_nt(
 }
 #endif // defined(MAT_MUL_NATIVE_QUANTIZED_NT_NT)
 
+#if defined(MAT_MUL_NATIVE_QUANTIZED_NT_T)
+/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS non-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=uchar)
+ * @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_QUANTIZED_NT_T)
+ * @note Only the following configurations of M0, N0 and K0 are currently supported:
+ *  - M0 > 0
+ *  - N0 = 1, 2, 3, 4, 8, 16
+ *  - K0 = 1, 2, 3, 4, 8, 16
+ * @note Values > 8 for M0, N0, K0 are not expected to be efficient
+ *
+ * @param[in]  lhs_ptr                           Pointer to the lhs matrix. Supported data types: QASYMM8/QASYMM8_SIGNED
+ * @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_quantized_nt_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 * lhs_stride_y + 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(int, M0, N0, acc);
+    LOOP_UNROLLING(int, i, 0, 1, M0,
+    {
+        acc[i].v = K * ((int)LHS_OFFSET) * ((int)RHS_OFFSET);
+    })
+
+    TILE(int, 1, M0, a_sum);
+    a_sum[0].v = 0;
+
+    TILE(int, 1, N0, b_sum);
+    b_sum[0].v = 0;
+
+    int k;
+    for(k = 0; k <= K - K0; k += K0)
+    {
+        TILE(DATA_TYPE, M0, K0, a);
+        TILE(DATA_TYPE, N0, K0, b);
+
+        LOOP_UNROLLING(int, i, 0, 1, M0,
+        {
+            a[i].v = 0;
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            b[i].v = 0;
+        })
+
+        // Load tile from lhs/rhs tensors
+        T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+        T_LOAD(DATA_TYPE, N0, K0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b);
+
+        T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, K0, NT, T, a, b, acc);
+
+        LOOP_UNROLLING(int, i, 0, 1, M0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, K0,
+            {
+                a_sum[0].s[i] += (int)a[i].s[j];
+            })
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, K0,
+            {
+                b_sum[0].s[i] += (int)b[i].s[j];
+            })
+        })
+
+        lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
+        rhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
+    }
+
+#if ((K % K0) != 0)
+    // Leftover loop
+    for(; k < K; ++k)
+    {
+        TILE(DATA_TYPE, M0, 1, a);
+        TILE(DATA_TYPE, N0, 1, b);
+
+        LOOP_UNROLLING(int, i, 0, 1, M0,
+        {
+            a[i].v = 0;
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            b[i].v = 0;
+        })
+
+        // Load tile from lhs/rhs tensors
+        T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+        T_LOAD(DATA_TYPE, N0, 1, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b);
+
+        T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, 1, NT, T, a, b, acc);
+
+        LOOP_UNROLLING(int, i, 0, 1, M0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, 1,
+            {
+                a_sum[0].s[i] += (int)a[i].s[j];
+            })
+        })
+
+        LOOP_UNROLLING(int, i, 0, 1, N0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, 1,
+            {
+                b_sum[0].s[i] += (int)b[i].s[j];
+            })
+        })
+
+        lhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE);
+        rhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE);
+    }
+#endif // ((K % K0) != 0)
+
+    LOOP_UNROLLING(int, i, 0, 1, M0,
+    {
+        LOOP_UNROLLING(int, j, 0, 1, N0,
+        {
+            acc[i].s[j] += ((int)(RHS_OFFSET)) * a_sum[0].s[i] + ((int)(LHS_OFFSET)) * b_sum[0].s[j];
+        })
+    })
+
+    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;
+
+    // Quantize the tile
+    TILE(DATA_TYPE, M0, N0, accq);
+    T_QUANTIZE8_ASYMMETRIC(int, DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, acc, accq);
+
+    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, accq, indirect_buffer);
+}
+#endif // defined(MAT_MUL_NATIVE_QUANTIZED_NT_T)
+
 #if defined(MAT_MUL_NATIVE_QUANTIZED_T_NT)
 /** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS non-transposed
  *