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>

1 files changed, 99 insertions, 0 deletions

diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h
index 872f4c0b57..c9b5370dea 100644
--- a/src/core/CL/cl_kernels/tile_helpers.h
+++ b/src/core/CL/cl_kernels/tile_helpers.h
@@ -536,6 +536,100 @@
         })                                                                                                             \
     })
 
+/** Store a VECTOR variable (e.g. int4, int8, char2 etc.) to a specified column in the TILE object
+ *
+ * @param[in]      VECTOR Vector variable to store
+ * @param[in, out] TILE   Tile variable to store to
+ * @param[in]      WIDTH  Width of the vector variable, also height of the tile (e.g. 2 if char2)
+ * @param[in]      COLUMN Column index of the tile
+ */
+#define COPY_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, WIDTH, COLUMN) COPY_VECTOR_TO_TILE_COLUMN_STR(VECTOR, TILE, WIDTH, COLUMN)
+#define COPY_VECTOR_TO_TILE_COLUMN_STR(VECTOR, TILE, WIDTH, COLUMN) COPY_##WIDTH##_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, COLUMN)
+#define COPY_1_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, COLUMN) \
+    ({                                                      \
+        TILE[0].s[COLUMN] = VECTOR;                         \
+    })
+
+#define COPY_2_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, COLUMN) \
+    ({                                                      \
+        TILE[0].s[COLUMN] = VECTOR.s0;                      \
+        TILE[1].s[COLUMN] = VECTOR.s1;                      \
+    })
+
+#define COPY_3_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, COLUMN) \
+    ({                                                      \
+        TILE[0].s[COLUMN] = VECTOR.s0;                      \
+        TILE[1].s[COLUMN] = VECTOR.s1;                      \
+        TILE[2].s[COLUMN] = VECTOR.s2;                      \
+    })
+
+#define COPY_4_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, COLUMN) \
+    ({                                                      \
+        TILE[0].s[COLUMN] = VECTOR.s0;                      \
+        TILE[1].s[COLUMN] = VECTOR.s1;                      \
+        TILE[2].s[COLUMN] = VECTOR.s2;                      \
+        TILE[3].s[COLUMN] = VECTOR.s3;                      \
+    })
+
+#define COPY_8_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, COLUMN) \
+    ({                                                      \
+        TILE[0].s[COLUMN] = VECTOR.s0;                      \
+        TILE[1].s[COLUMN] = VECTOR.s1;                      \
+        TILE[2].s[COLUMN] = VECTOR.s2;                      \
+        TILE[3].s[COLUMN] = VECTOR.s3;                      \
+        TILE[4].s[COLUMN] = VECTOR.s4;                      \
+        TILE[5].s[COLUMN] = VECTOR.s5;                      \
+        TILE[6].s[COLUMN] = VECTOR.s6;                      \
+        TILE[7].s[COLUMN] = VECTOR.s7;                      \
+    })
+
+#define COPY_16_VECTOR_TO_TILE_COLUMN(VECTOR, TILE, COLUMN) \
+    ({                                                      \
+        TILE[0].s[COLUMN] = VECTOR.s0;                      \
+        TILE[1].s[COLUMN] = VECTOR.s1;                      \
+        TILE[2].s[COLUMN] = VECTOR.s2;                      \
+        TILE[3].s[COLUMN] = VECTOR.s3;                      \
+        TILE[4].s[COLUMN] = VECTOR.s4;                      \
+        TILE[5].s[COLUMN] = VECTOR.s5;                      \
+        TILE[6].s[COLUMN] = VECTOR.s6;                      \
+        TILE[7].s[COLUMN] = VECTOR.s7;                      \
+        TILE[8].s[COLUMN] = VECTOR.s8;                      \
+        TILE[9].s[COLUMN] = VECTOR.s9;                      \
+        TILE[10].s[COLUMN] = VECTOR.sA;                     \
+        TILE[11].s[COLUMN] = VECTOR.sB;                     \
+        TILE[12].s[COLUMN] = VECTOR.sC;                     \
+        TILE[13].s[COLUMN] = VECTOR.sD;                     \
+        TILE[14].s[COLUMN] = VECTOR.sE;                     \
+        TILE[15].s[COLUMN] = VECTOR.sF;                     \
+    })
+
+/** Load SRC_HEIGHT x SRC_WIDTH elements from global memory (tensor), and store them in a SRC_WIDTH x SRC_HEIGHT tile
+ *
+ * @param[in]  DATA_TYPE     Data type
+ * @param[in]  SRC_HEIGHT    Number of source rows, or number of columns of the output tile
+ * @param[in]  SRC_WIDTH     Number of source columns, or number of tile rows
+ * @param[in]  TENSOR_TYPE   Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image).
+ *                           In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16)
+ * @param[in]  TENSOR        Tensor basename
+ * @param[in]  X             Starting X position
+ * @param[in]  Y             Starting Y position
+ * @param[in]  YI_MULTIPLIER Parameter used to multiply the internal row increment (_i).
+ *                           In common cases should be 1 but it becomes useful when we want to load rows which are multiple of STRIDE_Y.
+ *                           (e.g. loading the weights of convolution layer).
+ *                           In this case the address calculation is performed as: (Y + _i * Y_MULTIPLIER) * STRIDE_Y
+ * @param[in]  STRIDE_Y      Stride Y (in bytes) used to load each row.
+ * @param[out] dst           Output tile
+ */
+#define T_LOAD_TRANSPOSED(DATA_TYPE, SRC_HEIGHT, SRC_WIDTH, TENSOR_TYPE, TENSOR, X, Y, YI_MULTIPLIER, STRIDE_Y, dst)     \
+    ({                                                                                                                   \
+        LOOP_UNROLLING(int, _i, 0, 1, SRC_HEIGHT,                                                                        \
+        {                                                                                                                \
+            VEC_DATA_TYPE(DATA_TYPE, SRC_WIDTH)                                                                          \
+                tmp = V_LOAD(DATA_TYPE, SRC_WIDTH, TENSOR_TYPE, TENSOR, X, ((Y) + _i * (int)(YI_MULTIPLIER)), STRIDE_Y); \
+            COPY_VECTOR_TO_TILE_COLUMN(tmp, dst, SRC_WIDTH, _i);                                                         \
+        })                                                                                                               \
+    })
+
 /** Load a tile from global memory (tensor) using an indirect Y index tile
  *
  * @param[in]  DATA_TYPE   Data type
@@ -1259,6 +1353,11 @@
  * @param[in]      lhs           LHS tile
  * @param[in]      rhs           RHS tile
  * @param[in, out] dst           DST tile
+ *
+ * @note For Int8/UInt8 multiplications, we only have T_MMUL_NT_T because we need
+ *       the multiply the rows of Lhs and Rhs tensors to utilize dot product extension.
+ *       Addition of other versions requires dealing with on the fly transposition of
+ *       these tile elements and therefore is not favored.
  */
 #define T_MMUL(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, LHS_LAYOUT, RHS_LAYOUT, lhs, rhs, dst) T_MMUL_##LHS_LAYOUT##_##RHS_LAYOUT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst)
 #define T_MMUL_NT_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_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)