diff options
| author | Gunes Bayir <gunes.bayir@arm.com> | 2021-12-10 16:17:56 +0000 |
|---|---|---|
| committer | Gunes Bayir <gunes.bayir@arm.com> | 2022-07-13 14:47:44 +0000 |
| commit | 4bfc70e31766587c951204c93a127a486e007d0c (patch) | |
| tree | 198b2150c43b14c571c100b8dfa0d3aaa4c968d0 /src/core/CL/cl_kernels/tile_helpers.h | |
| parent | 29cab36ddd73c174bf6b2de453663aa49c1cc576 (diff) | |
| download | ComputeLibrary-4bfc70e31766587c951204c93a127a486e007d0c.tar.gz | |
Add Gemm MMUL Reshaped Only Rhs Support for FP32/FP16
This patch introduces a GEMM routine that is optimized for Arm(R) Mali(TM)-G715 and Arm(R) Mali(TM)-G615
Resolves: COMPMID-5216
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I2e5d7806f5904347185bb3e250f73d73d6669dba
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7914
Reviewed-by: SiCong Li <sicong.li@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels/tile_helpers.h')
| -rw-r--r-- | src/core/CL/cl_kernels/tile_helpers.h | 55 |
1 files changed, 45 insertions, 10 deletions
diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h index 0ce343e3ec..4b6144a22d 100644 --- a/src/core/CL/cl_kernels/tile_helpers.h +++ b/src/core/CL/cl_kernels/tile_helpers.h @@ -970,8 +970,8 @@ #define ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) op##_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) #define ACTIVATION_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) -#define T_ADD(A_VAL, B_VAL) ((A_VAL) + (B_VAL)) -#define T_DIV(A_VAL, B_VAL) ((A_VAL) / (B_VAL)) +#define V_ADD(A_VAL, B_VAL) ((A_VAL) + (B_VAL)) +#define V_DIV(A_VAL, B_VAL) ((A_VAL) / (B_VAL)) /** Element-wise activation for quantized types * @@ -995,6 +995,25 @@ }) \ }) +/** Element-wise addition between two tiles + * + * @note Performs: LHS + RHS = DST + * + * @param[in] DATA_TYPE LHS/RHS/DST data type + * @param[in] M0 Number of LHS rows + * @param[in] N0 Number of LHS columns + * @param[in] lhs LHS tile + * @param[in] rhs Constant RHS tile + * @param[out] dst DST tile + */ +#define T_ADD(DATA_TYPE, M0, N0, lhs, rhs, dst) \ + ({ \ + LOOP_UNROLLING(int, _m0, 0, 1, M0, \ + { \ + dst[_m0].v = lhs[_m0].v + rhs[_m0].v; \ + }) \ + }) + /** Element-wise addition with a constant value * * @note Performs: LHS + constant = DST @@ -1010,15 +1029,31 @@ ({ \ LOOP_UNROLLING(int, _m0, 0, 1, M0, \ { \ - LOOP_UNROLLING(int, _n0, 0, 1, N0, \ - { \ - dst[_m0].s[_n0] = lhs[_m0].s[_n0] + rhs_constant; \ - }) \ + dst[_m0].v = lhs[_m0].v + (DATA_TYPE)rhs_constant; \ }) \ }) -#define T_ELTWISE_BROADCAST_ADD_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(T_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) -#define T_ELTWISE_BROADCAST_DIV_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(T_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) +#define T_ELTWISE_BROADCAST_ADD_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) +#define T_ELTWISE_BROADCAST_DIV_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) + +/** Element-wise scale with a constant value + * + * @note Performs: LHS * constant = DST + * + * @param[in] DATA_TYPE LHS/RHS/DST data type + * @param[in] M0 Number of LHS rows + * @param[in] N0 Number of LHS columns + * @param[in] lhs LHS tile + * @param[in] rhs_constant Constant value + * @param[out] dst DST tile + */ +#define T_SCALE_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \ + ({ \ + LOOP_UNROLLING(int, _m0, 0, 1, M0, \ + { \ + dst[_m0].v = lhs[_m0].v * (DATA_TYPE)rhs_constant; \ + }) \ + }) /** Element-wise operation with RHS broadcasted (RHS has the X dimension only) * @@ -1041,8 +1076,8 @@ }) \ }) -#define T_ELTWISE_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(T_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) -#define T_ELTWISE_DIV(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(T_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) +#define T_ELTWISE_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) +#define T_ELTWISE_DIV(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) /** Element-wise operation between two tiles (LHS and RHS) * |