Add quantized CL MatMul kernels for Lhs NT/T, Rhs NT
Implement OpenCL kernels for batched Matrix Multiplication for the quantized data types QASYMM8 and QASYMM8_SIGNED.
Quantized MatMul is supported with the following MatMul attributes:
* adj_x = false, adj_y = false
* adj_x = true, adj_y = false
We consider native format kernels only. In other words, no reshaping of the operand matrices is done.
Resolves: COMPMID-5921, COMPMID-5922
Change-Id: I99e0f68054a2bd635c60ec2641acc2e7ff398473
Signed-off-by: Omar Al Khatib <omar.alkhatib@arm.com>
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9435
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/cl_kernels/common/mat_mul_quantized.cl b/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
new file mode 100644
index 0000000..c250b4b
--- /dev/null
+++ b/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
@@ -0,0 +1,387 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers.h"
+#include "tile_helpers.h"
+
+#if defined(MAT_MUL_NATIVE_QUANTIZED_NT_NT)
+/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS non-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=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_NT)
+ * @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 are not expected to be efficient
+ *
+ * @param[in] lhs_ptr Pointer to the lhs matrix. Supported data types: QASYMM8_SIGNED/QASYMM8
+ * @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_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 * lhs_stride_y + 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(int, M0, N0, acc);
+ LOOP_UNROLLING(int, i, 0, 1, M0,
+ {
+ acc[i].v = K * ((int)LHS_OFFSET) * ((int)RHS_OFFSET);
+ })
+
+ TILE(int, 1, N0, b_sum);
+ b_sum[0].v = 0;
+
+ TILE(int, 1, M0, a_sum);
+ a_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
+ T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+
+ // Load tile from the rhs tensor in a transposed fashion
+ // in order to use T_MMUL_NT_T macro because only this macro
+ // can utilize dot product instruction for Int8/UInt8 by
+ // directly multiplying the rows of Lhs and Rhs tensors.
+ T_LOAD_TRANSPOSED(DATA_TYPE, K0, N0, 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, K0,
+ {
+ LOOP_UNROLLING(int, j, 0, 1, N0,
+ {
+ b_sum[0].s[j] += (int)b[j].s[i];
+ })
+ })
+
+ lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
+ rhs_offset_first_element_in_bytes += K0 * rhs_stride_y;
+ }
+
+#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
+ T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a);
+
+ // Load tile from the rhs tensor in a transposed fashion.
+ // See the main loop for more explanation
+ T_LOAD_TRANSPOSED(DATA_TYPE, 1, N0, 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, 1,
+ {
+ LOOP_UNROLLING(int, j, 0, 1, N0,
+ {
+ b_sum[0].s[j] += (int)b[j].s[i];
+ })
+ })
+
+ lhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE);
+ rhs_offset_first_element_in_bytes += 1 * rhs_stride_y;
+ }
+#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_NT)
+
+#if defined(MAT_MUL_NATIVE_QUANTIZED_T_NT)
+/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS non-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_NT)
+ * @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 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_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(int, M0, N0, acc);
+ LOOP_UNROLLING(int, i, 0, 1, M0,
+ {
+ acc[i].v = K * ((int)LHS_OFFSET) * ((int)RHS_OFFSET);
+ })
+
+ TILE(int, 1, N0, b_sum);
+ b_sum[0].v = 0;
+
+ TILE(int, 1, M0, a_sum);
+ a_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/rhs tensors 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);
+ T_LOAD_TRANSPOSED(DATA_TYPE, K0, N0, 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, K0,
+ {
+ LOOP_UNROLLING(int, j, 0, 1, N0,
+ {
+ b_sum[0].s[j] += (int)b[j].s[i];
+ })
+ })
+
+ lhs_offset_first_element_in_bytes += K0 * lhs_stride_y;
+ rhs_offset_first_element_in_bytes += K0 * rhs_stride_y;
+ }
+
+#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/rhs tensors 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);
+ T_LOAD_TRANSPOSED(DATA_TYPE, 1, N0, 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, 1,
+ {
+ LOOP_UNROLLING(int, j, 0, 1, N0,
+ {
+ b_sum[0].s[j] += (int)b[j].s[i];
+ })
+ })
+
+ lhs_offset_first_element_in_bytes += 1 * lhs_stride_y;
+ rhs_offset_first_element_in_bytes += 1 * rhs_stride_y;
+ }
+#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_NT)