Fix nightly failures in MatMulLowpNativeKernel when using bounded activation functions

- Added checks for supported activation functions in MatMulLowpKernel validate
- Replaced incorrect float activation macro with quantized implementation in mat_mul_quantized

Resolves: [COMPMID-6339]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: I15661f14877f1d3305644e6473feb5482a67e773
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/532858
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9855
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 26 insertions, 20 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 8cf857dd84..7029af2188 100644
--- a/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
+++ b/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
@@ -34,6 +34,7 @@
  * @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 fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output with the relu and bounded relu operations.
+ * @note The value of 0 in quantized format is equivalent to the quantization offset of the output data. This should be passed with -DZERO_POINT
  * @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:
@@ -196,12 +197,12 @@ __kernel void mat_mul_native_quantized_nt_nt(
     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;
 
-    T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc);
-
     // 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);
 
+    T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_POINT, A_VAL, B_VAL, accq, accq);
+
     TILE(int, M0, 1, indirect_buffer);
     LOOP_UNROLLING(int, _i, 0, 1, M0,
     {
@@ -221,6 +222,7 @@ __kernel void mat_mul_native_quantized_nt_nt(
  * @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 fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output bounded activation functions.
+ * @note The value of 0 in quantized format is equivalent to the quantization offset of the output data. This should be passed with -DZERO_POINT
  * @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:
@@ -375,12 +377,12 @@ __kernel void mat_mul_native_quantized_nt_t(
     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;
 
-    T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc);
-
     // 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);
 
+    T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_POINT, A_VAL, B_VAL, accq, accq);
+
     TILE(int, M0, 1, indirect_buffer);
     LOOP_UNROLLING(int, _i, 0, 1, M0,
     {
@@ -400,6 +402,7 @@ __kernel void mat_mul_native_quantized_nt_t(
  * @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 fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output with the relu and bounded relu operations.
+ * @note The value of 0 in quantized format is equivalent to the quantization offset of the output data. This should be passed with -DZERO_POINT
  * @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:
@@ -556,12 +559,12 @@ __kernel void mat_mul_native_quantized_t_nt(
     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;
 
-    T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc);
-
     // 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);
 
+    T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_POINT, A_VAL, B_VAL, accq, accq);
+
     TILE(int, M0, 1, indirect_buffer);
     LOOP_UNROLLING(int, _i, 0, 1, M0,
     {
@@ -581,6 +584,7 @@ __kernel void mat_mul_native_quantized_t_nt(
  * @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 fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output with the relu and bounded relu operations.
+ * @note The value of 0 in quantized format is equivalent to the quantization offset of the output data. This should be passed with -DZERO_POINT
  * @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:
@@ -742,11 +746,11 @@ __kernel void mat_mul_native_quantized_t_t(
     const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0;
 
     // Quantize the tile
-    T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc);
-
     TILE(DATA_TYPE, M0, N0, accq);
     T_QUANTIZE8_ASYMMETRIC(int, DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, acc, accq);
 
+    T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_POINT, A_VAL, B_VAL, accq, accq);
+
     TILE(int, M0, 1, indirect_buffer);
     LOOP_UNROLLING(int, _i, 0, 1, M0,
     {
diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h
index 85bd59afd4..8129606277 100644
--- a/src/core/CL/cl_kernels/tile_helpers.h
+++ b/src/core/CL/cl_kernels/tile_helpers.h
@@ -1144,19 +1144,21 @@
         })                                                                                     \
     })
 
+
+// NOTE : A_VAL and B_VAL should be quantized values (using same quantization info as x)
 // RELU Activation
-#define relu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (max((DATA_TYPE)ZERO_VALUE, x))
+#define relu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x) (max((DATA_TYPE)ZERO_POINT, x))
 // Bounded RELU Activation
-#define brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)ZERO_VALUE, x)))
+#define brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)ZERO_POINT, x)))
 // Lower Upper Bounded RELU Activation
-#define lu_brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL))
+#define lu_brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL))
 // Hard Swish Activation
-#define hard_swish_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x * ((min(max((DATA_TYPE)(x + (DATA_TYPE)3.f), (DATA_TYPE)0.f), (DATA_TYPE)6.f)) * (DATA_TYPE)0.166666667f))
+#define hard_swish_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x) (x * ((min(max((DATA_TYPE)(x + (DATA_TYPE)3.f), (DATA_TYPE)0.f), (DATA_TYPE)6.f)) * (DATA_TYPE)0.166666667f))
 // Identity Activation
-#define identity_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x)
+#define identity_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x) (x)
 
-#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 ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x) op##_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x)
+#define ACTIVATION_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x) ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_POINT, A_VAL, B_VAL, x)
 
 #define V_ADD(A_VAL, B_VAL) ((A_VAL) + (B_VAL))
 #define V_SUB(A_VAL, B_VAL) ((A_VAL) - (B_VAL))
@@ -1171,17 +1173,17 @@
  * @param[in]  M0              Number of SRC/DST rows
  * @param[in]  N0              Number of SRC/DST columns
  * @param[in]  ACTIVATION_TYPE Activation type
- * @param[in]  ZERO_VALUE      The zero value to consider in the computation
- * @param[in]  A_VAL           A value used for the activation (e.g. tanh_op, brelu,..)
- * @param[in]  B_VAL           B value used for the activation (e.g. tanh_op, brelu,..)
+ * @param[in]  ZERO_POINT      The zero value to consider in the computation
+ * @param[in]  A_VAL           Quantized A value used for the activation (e.g. tanh_op, brelu,..)
+ * @param[in]  B_VAL           Quantized B value used for the activation (e.g. tanh_op, brelu,..)
  * @param[out] src             SRC tile
  * @param[out] dst             DST tile
  */
-#define T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_VALUE, A_VAL, B_VAL, src, dst)               \
+#define T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_POINT, A_VAL, B_VAL, src, dst)               \
     ({ \
         LOOP_UNROLLING(int, _m0, 0, 1, M0, \
         { \
-            dst[_m0].v = ACTIVATION_QUANTIZED(ACTIVATION_TYPE, DATA_TYPE, N0, ZERO_VALUE, A_VAL, B_VAL, src[_m0].v); \
+            dst[_m0].v = ACTIVATION_QUANTIZED(ACTIVATION_TYPE, DATA_TYPE, N0, ZERO_POINT, A_VAL, B_VAL, src[_m0].v); \
         })                                                                                          \
     })