Remove OpenCL padding: CLPixelWiseMultiplicationKernel
- Change kernel's vec_size to 16 / sizeof(output)
- Change ICLKernel.cpp to handle broadcast without padding
Resolve COMPMID-3913
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Change-Id: I03e884b250ef5784dc109bff8cf2c96b345d119f
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5450
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/core/CL/cl_kernels/pixelwise_mul_int.cl b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
index b0bd338..92a7e6f 100644
--- a/src/core/CL/cl_kernels/pixelwise_mul_int.cl
+++ b/src/core/CL/cl_kernels/pixelwise_mul_int.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -36,6 +36,10 @@
#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
#if defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(ACC_DATA_TYPE) && defined(DATA_TYPE_OUT)
+
+#define VEC_ACC_TYPE VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE_OUT)
+#define VEC_OUT_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
+
/** Performs a pixelwise multiplication with integer scale of integer inputs.
*
* @attention The inputs and output data types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT:
@@ -75,27 +79,29 @@
TENSOR3D_DECLARATION(out),
const uint scale)
{
- // Get pixels pointer
- Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
- Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
- Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+ size_t x = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
+ size_t y = get_global_id(1);
+ size_t z = get_global_id(2);
+
+ __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + x * in1_stride_x + y * in1_stride_y + z * in1_stride_z;
+ __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + x * in2_stride_x + y * in2_stride_y + z * in2_stride_z;
+ __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + x * out_stride_x + y * out_stride_y + z * out_stride_z;
// Load data
- VEC_DATA_TYPE(ACC_DATA_TYPE, 16)
- in1_data = CONVERT(vload16(0, (__global DATA_TYPE_IN1 *)in1.ptr), VEC_DATA_TYPE(ACC_DATA_TYPE, 16));
- VEC_DATA_TYPE(ACC_DATA_TYPE, 16)
- in2_data = CONVERT(vload16(0, (__global DATA_TYPE_IN2 *)in2.ptr), VEC_DATA_TYPE(ACC_DATA_TYPE, 16));
+ VEC_ACC_TYPE in1_data = CONVERT((VEC_DATA_TYPE(DATA_TYPE_IN1, VEC_SIZE_OUT))VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE_IN1 *)in1_addr), VEC_ACC_TYPE);
+ VEC_ACC_TYPE in2_data = CONVERT((VEC_DATA_TYPE(DATA_TYPE_IN2, VEC_SIZE_OUT))VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE_IN2 *)in2_addr), VEC_ACC_TYPE);
// Perform multiplication and store result
- vstore16(MUL_OP(in1_data, in2_data, scale, DATA_TYPE_OUT, 16), 0, (__global DATA_TYPE_OUT *)out.ptr);
+ VEC_OUT_TYPE out_data0 = MUL_OP(in1_data, in2_data, scale, DATA_TYPE_OUT, VEC_SIZE_OUT);
+ STORE_VECTOR_SELECT(out_data, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
}
#endif /* defined(DATA_TYPE_IN1) && defined(DATA_TYPE_IN2) && defined(ACC_DATA_TYPE) && defined(DATA_TYPE_OUT) */
-#if defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) && defined(VEC_SIZE)
+#if defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) && defined(VEC_SIZE_OUT)
-#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
-#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
-#define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)
+#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE_OUT)
+#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE_OUT)
+#define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)
/** Performs a pixelwise multiplication with float scale of quantized inputs.
*
@@ -141,14 +147,17 @@
TENSOR3D_DECLARATION(out),
const float scale)
{
- // Get pixels pointer
- Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1);
- Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2);
- Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out);
+ size_t x = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
+ size_t y = get_global_id(1);
+ size_t z = get_global_id(2);
+
+ __global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + x * in1_stride_x + y * in1_stride_y + z * in1_stride_z;
+ __global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + x * in2_stride_x + y * in2_stride_y + z * in2_stride_z;
+ __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + x * out_stride_x + y * out_stride_y + z * out_stride_z;
// Load data
- VEC_INT in_a = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_OUT *)in1.ptr), VEC_INT);
- VEC_INT in_b = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_OUT *)in2.ptr), VEC_INT);
+ VEC_INT in_a = CONVERT((VEC_TYPE)(VLOAD(VEC_SIZE_OUT)(0, (__global DATA_TYPE_OUT *)in1_addr)), VEC_INT);
+ VEC_INT in_b = CONVERT((VEC_TYPE)(VLOAD(VEC_SIZE_OUT)(0, (__global DATA_TYPE_OUT *)in2_addr)), VEC_INT);
// Dequantize
#if defined(OFFSET_IN1)
@@ -165,10 +174,9 @@
#else // defined(OFFSET_OUT)
const VEC_FLOAT qresf32 = (in1f32 * in2f32 * scale) / ((VEC_FLOAT)(float)SCALE_OUT);
#endif // defined(OFFSET_OUT)
- const VEC_TYPE res = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
+ const VEC_TYPE res0 = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
// Store result
- VSTORE(VEC_SIZE)
- (res, 0, (__global DATA_TYPE_OUT *)out.ptr);
+ STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
}
-#endif /* defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) && defined(VEC_SIZE) */
+#endif /* defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE_OUT) && defined(VEC_SIZE_OUT) */