Add Fused Activation to OpenCL MatMul

- Added fused activation to MatMul function interface
- Added fused activation to CL backend
- Includes tests for supported Activation Functions in MatMul

Resolves: [COMPMID-6192]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: Ie103212b600b60699eaf6a6394d609e6e1f5aba6
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/522465
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9714
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 14 insertions, 1 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 0c3cbca9a6..bd415bb4a7 100644
--- a/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
+++ b/src/core/CL/cl_kernels/common/mat_mul_quantized.cl
@@ -23,6 +23,7 @@
  */
 #include "helpers.h"
 #include "tile_helpers.h"
+#include "activation_float_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
@@ -32,6 +33,7 @@
  * @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 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 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:
@@ -194,6 +196,8 @@ __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);
@@ -216,6 +220,7 @@ __kernel void mat_mul_native_quantized_nt_nt(
  * @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 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 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:
@@ -315,7 +320,7 @@ __kernel void mat_mul_native_quantized_nt_t(
         rhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE);
     }
 
-#if ((K % K0) != 0)
+#if((K % K0) != 0)
     // Leftover loop
     for(; k < K; ++k)
     {
@@ -370,6 +375,8 @@ __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);
@@ -392,6 +399,7 @@ __kernel void mat_mul_native_quantized_nt_t(
  * @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 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 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:
@@ -548,6 +556,8 @@ __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);
@@ -570,6 +580,7 @@ __kernel void mat_mul_native_quantized_t_nt(
  * @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 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 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:
@@ -731,6 +742,8 @@ __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);