Implement CL kernel for a native batched matmul Quantized - LHS transposed, RHS transposed

Resolves: [COMPMID-5924]

Signed-off-by: Omar Al Khatib <omar.alkhatib@arm.com>
Change-Id: I9ba657737eb1e3a096c8341ad4ad311571f8edeb
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9454
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 273 insertions, 3 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 c250b4b988..5c931d2fc1 100644
--- a/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
+++ b/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
@@ -385,3 +385,185 @@ __kernel void mat_mul_native_quantized_t_nt(
     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_T_NT)
+
+#if defined(MAT_MUL_NATIVE_QUANTIZED_T_T)
+/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS transposed
+ *
+ * @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_T_T)
+ * @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: 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_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(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 the lhs tensor in a transposed fashion
+        // see mat_mul_native_quantized_nt_nt main loop for more explanation
+        T_LOAD_TRANSPOSED(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+
+        // Load tile from the rhs tensor
+        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, K0,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, M0,
+            {
+                a_sum[0].s[j] += (int)a[j].s[i];
+            })
+        })
+
+        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 * lhs_stride_y;
+        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 the lhs tensor in a transposed fashion
+        // see mat_mul_native_quantized_nt_nt main loop for more explanation
+        T_LOAD_TRANSPOSED(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+
+        // Load tile from the rhs tensor
+        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, 1,
+        {
+            LOOP_UNROLLING(int, j, 0, 1, M0,
+            {
+                a_sum[0].s[j] += (int)a[j].s[i];
+            })
+        })
+
+        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 * lhs_stride_y;
+        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_T_T)
diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp
index e657687887..4612ca35b8 100644
--- a/src/gpu/cl/ClKernelLibrary.cpp
+++ b/src/gpu/cl/ClKernelLibrary.cpp
@@ -325,6 +325,7 @@ const std::map<std::string, std::string> ClKernelLibrary::_kernel_program_map =
     { "mat_mul_native_t_t", "common/mat_mul.cl" },
     { "mat_mul_native_quantized_nt_nt", "common/mat_mul_quantized.cl" },
     { "mat_mul_native_quantized_t_nt", "common/mat_mul_quantized.cl" },
+    { "mat_mul_native_quantized_t_t", "common/mat_mul_quantized.cl" },
     { "max_unpooling_layer_2", "common/unpooling_layer.cl" },
     { "mean_stddev_normalization", "common/mean_stddev_normalization.cl" },
     { "memset", "common/memset.cl" },
diff --git a/tests/validation/CL/MatMulLowpNativeKernel.cpp b/tests/validation/CL/MatMulLowpNativeKernel.cpp
index 5932fa7c21..a0b2a37b4b 100644
--- a/tests/validation/CL/MatMulLowpNativeKernel.cpp
+++ b/tests/validation/CL/MatMulLowpNativeKernel.cpp
@@ -64,12 +64,12 @@ const auto m0_values_nightly_lhs_t  = framework::dataset::make("M0", { 1, 2, 3,
 
 /** N0 values to test --nightly*/
 const auto n0_values_nightly_rhs_nt = framework::dataset::make("N0", { 1, 2, 3, 4, 8, 16 });
-// const auto n0_values_nightly_rhs_t  = framework::dataset::make("N0", { 1, 2, 3, 4, 8 });
+const auto n0_values_nightly_rhs_t  = framework::dataset::make("N0", { 1, 2, 3, 4, 8 });
 
 /** K0 values to test --nightly*/
 const auto k0_values_nightly_lhs_nt_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 8, 16 });
-// const auto k0_values_nightly_rhs_t         = framework::dataset::make("K0", { 1, 2, 3, 4, 8 });
-const auto k0_values_nightly_lhs_t_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 5, 6, 7, 8 });
+const auto k0_values_nightly_rhs_t         = framework::dataset::make("K0", { 1, 2, 3, 4, 8 });
+const auto k0_values_nightly_lhs_t_rhs_nt  = framework::dataset::make("K0", { 1, 2, 3, 4, 5, 6, 7, 8 });
 
 TEST_SUITE(CL)
 TEST_SUITE(MatMulLowpNativeKernel)
@@ -234,6 +234,30 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulLowpNativeKernelFixture<int8_t>, framew
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_quant);
 }
+FIXTURE_DATA_TEST_CASE(RunTiny_T_T, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(),
+                                                                                                                  framework::dataset::make("TransposeA", { true })),
+                                                                                                                  framework::dataset::make("TransposeB", { true })),
+                                                                                                                  m0_values_precommit),
+                                                                                                                  n0_values_precommit),
+                                                                                                                  k0_values_precommit),
+                                                                                                                  framework::dataset::make("ExportRhsToCLImage", { false })),
+                                                                                                                  framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_quant);
+}
+FIXTURE_DATA_TEST_CASE(RunSmall_T_T, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+                                                                                                                   framework::dataset::make("TransposeA", { true })),
+                                                                                                                   framework::dataset::make("TransposeB", { true })),
+                                                                                                                   m0_values_precommit),
+                                                                                                                   n0_values_precommit),
+                                                                                                                   k0_values_precommit),
+                                                                                                                   framework::dataset::make("ExportRhsToCLImage", { false })),
+                                                                                                                   framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_quant);
+}
 FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY,
                        combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                framework::dataset::make("TransposeA", { false })),
@@ -260,6 +284,19 @@ FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulLowpNativeKernelFixture<in
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_quant);
 }
+FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+                                                                               framework::dataset::make("TransposeA", { true })),
+                                                                       framework::dataset::make("TransposeB", { true })),
+                                                               m0_values_nightly_lhs_t),
+                                                       n0_values_nightly_rhs_t),
+                                               k0_values_nightly_rhs_t),
+                                       framework::dataset::make("ExportRhsToCLImage", { false })),
+                               framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_quant);
+}
 // Running High Dimensional test is enough for qasymm8_signed, because we're stressing the number of dimensions, not data type or M0/N0/K0
 // It's a good idea to test for each Lhs/Rhs T/NT combinations because they're different CL kernels
 FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL,
@@ -275,6 +312,19 @@ FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulLowpNativeKernelFixture<int8_
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_quant);
 }
+FIXTURE_DATA_TEST_CASE(RunHighDimensional_T_T, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL,
+                       combine(combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulDataset(),
+                                                                               framework::dataset::make("TransposeA", { true })),
+                                                                       framework::dataset::make("TransposeB", { true })),
+                                                               framework::dataset::make("M0", { 2 })),
+                                                       framework::dataset::make("N0", { 2 })),
+                                               framework::dataset::make("K0", { 2 })),
+                                       framework::dataset::make("ExportRhsToCLImage", { false })),
+                               framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_quant);
+}
 TEST_SUITE_END() // QASYMM8_SIGNED
 
 TEST_SUITE(QASYMM8)
@@ -302,6 +352,30 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulLowpNativeKernelFixture<uint8_t>, frame
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_quant);
 }
+FIXTURE_DATA_TEST_CASE(RunTiny_T_T, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(),
+                                                                                                                   framework::dataset::make("TransposeA", { true })),
+                                                                                                                   framework::dataset::make("TransposeB", { true })),
+                                                                                                                   m0_values_precommit),
+                                                                                                                   n0_values_precommit),
+                                                                                                                   k0_values_precommit),
+                                                                                                                   framework::dataset::make("ExportRhsToCLImage", { false })),
+                                                                                                                   framework::dataset::make("DataType", DataType::QASYMM8)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_quant);
+}
+FIXTURE_DATA_TEST_CASE(RunSmall_T_T, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+                                                                                                                    framework::dataset::make("TransposeA", { true })),
+                                                                                                                    framework::dataset::make("TransposeB", { true })),
+                                                                                                                    m0_values_precommit),
+                                                                                                                    n0_values_precommit),
+                                                                                                                    k0_values_precommit),
+                                                                                                                    framework::dataset::make("ExportRhsToCLImage", { false })),
+                                                                                                                    framework::dataset::make("DataType", DataType::QASYMM8)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_quant);
+}
 FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
                        combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
                                                                                framework::dataset::make("TransposeA", { false })),
@@ -328,6 +402,19 @@ FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulLowpNativeKernelFixture<ui
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_quant);
 }
+FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+                                                                               framework::dataset::make("TransposeA", { true })),
+                                                                       framework::dataset::make("TransposeB", { true })),
+                                                               m0_values_nightly_lhs_t),
+                                                       n0_values_nightly_rhs_t),
+                                               k0_values_nightly_rhs_t),
+                                       framework::dataset::make("ExportRhsToCLImage", { false })),
+                               framework::dataset::make("DataType", DataType::QASYMM8)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_quant);
+}
 TEST_SUITE_END() // QASYMM8
 TEST_SUITE_END() // Quantized
 TEST_SUITE_END() // MatMulLowpNativeKernel