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>

1 files changed, 12 insertions, 8 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,
     {