CLInstanceNormalizationLayer NHWC optimisation
* Make changes to split the workload into two kernels. One kernel precomputes
mean and variance and the second kernel just loads these precomputed values.
* The new approach runs %30 faster than the original code for NHWC workloads
like 32x192x256.
* Resolves MLCE-337
Change-Id: I8356fcefa2d131ab4dcb32268ce7142421d073e4
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5355
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
diff --git a/src/core/CL/cl_kernels/instance_normalization.cl b/src/core/CL/cl_kernels/instance_normalization.cl
index 480d9cd..d2507d9 100644
--- a/src/core/CL/cl_kernels/instance_normalization.cl
+++ b/src/core/CL/cl_kernels/instance_normalization.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -23,6 +23,118 @@
*/
#include "helpers.h"
+#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z)
+/** This function computes the mean and variance of each plane of the input tensor and provides it as output.
+ *
+ * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @attention Data type should be passed using the -DDATA_TYPE=data_type compile flag, e.g. -DDATA_TYPE=float
+ * @attention Dimensions X, Y, and Z should be given as a preprocessor argument with -DDIM_X=value, -DDIM_Y=value, -DDIM_Z=value. e.g. -DDIM_X=6, -DDIM_Y=2, -DDIM_Z=7
+ *
+ * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32
+ * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)
+ * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)
+ * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)
+ * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] input_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] input_step_w input_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor
+ * @param[out] output_ptr (Optional) Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in] output_stride_x (Optional) Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] output_stride_y (Optional) Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] output_stride_z (Optional) Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
+ */
+__kernel void compute_mean_var(
+ TENSOR4D_DECLARATION(input),
+ TENSOR3D_DECLARATION(output))
+{
+ Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
+ Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
+
+#if defined(NHWC)
+ const int ch = get_global_id(0); // Current channel
+ const int batch = get_global_id(1); // Current batch
+ const int elements_plane = DIM_Y * DIM_Z;
+ float part_sum = 0.f;
+ float part_sum_sq = 0.f;
+ const int in_offset = input_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
+ for(int i = 0; i < (DIM_Y * DIM_Z); ++i)
+ {
+ const float data = *((__global DATA_TYPE *)(input_ptr + in_offset + i * input_stride_y));
+ part_sum += data;
+ part_sum_sq += data * data;
+ }
+ float mean = (part_sum / elements_plane);
+ float var = (part_sum_sq / elements_plane) - (mean * mean);
+ __global DATA_TYPE *output_address0 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
+ *output_address0 = mean;
+ __global DATA_TYPE *output_address1 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
+ *output_address1 = var;
+#else // !defined(NHWC)
+ const int ch = get_global_id(2) % DIM_Z; // Current channel
+ const int batch = get_global_id(2) / DIM_Z; // Current batch
+ const int elements_plane = DIM_X * DIM_Y;
+
+ VEC_DATA_TYPE(float, VEC_SIZE)
+ part_sum = 0.f;
+ VEC_DATA_TYPE(float, VEC_SIZE)
+ part_sum_sq = 0.f;
+ // Calculate partial sum
+ for(int y = 0; y < DIM_Y; ++y)
+ {
+ int x = 0;
+ for(; x <= (DIM_X - VEC_SIZE); x += VEC_SIZE)
+ {
+ // Load data
+ VEC_DATA_TYPE(float, VEC_SIZE)
+ data = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)), VEC_DATA_TYPE(float, VEC_SIZE));
+ part_sum += data;
+ part_sum_sq += data * data;
+ }
+ // Left-overs loop
+ for(; x < DIM_X; ++x)
+ {
+ float data = (float)(*((__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)));
+ part_sum.s0 += data;
+ part_sum_sq.s0 += data * data;
+ }
+ }
+ // Perform reduction
+#if VEC_SIZE > 8
+ part_sum.s01234567 += part_sum.s89abcdef;
+ part_sum_sq.s01234567 += part_sum_sq.s89abcdef;
+#endif // VEC_SIZE > 8
+#if VEC_SIZE > 4
+ part_sum.s0123 += part_sum.s4567;
+ part_sum_sq.s0123 += part_sum_sq.s4567;
+#endif // VEC_SIZE > 4
+#if VEC_SIZE > 2
+ part_sum.s01 += part_sum.s23;
+ part_sum_sq.s01 += part_sum_sq.s23;
+#endif // VEC_SIZE > 2
+ part_sum.s0 += part_sum.s1;
+ part_sum_sq.s0 += part_sum_sq.s1;
+
+ float sum = (float)part_sum.s0;
+ float sum_sq = (float)part_sum_sq.s0;
+
+ const float mean = (sum / elements_plane);
+ const float var = (sum_sq / elements_plane) - (mean * mean);
+
+ __global DATA_TYPE *output_address0 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
+ *output_address0 = mean;
+ __global DATA_TYPE *output_address1 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
+ *output_address1 = var;
+
+#endif // defined(NHWC)
+}
+#endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z) */
+
#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(INTERNAL_DATA_TYPE) && defined(GAMMA) && defined(BETA) && defined(EPSILON) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z)
/** This function normalizes the input 2D tensor across the first dimension with respect to mean and standard deviation of the same dimension.
*
@@ -51,105 +163,50 @@
* @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
*/
__kernel void instance_normalization(
- TENSOR4D_DECLARATION(input)
+ TENSOR4D_DECLARATION(input),
+ TENSOR3D_DECLARATION(mean_var)
#ifndef IN_PLACE
,
TENSOR4D_DECLARATION(output)
#endif /* IN_PLACE */
)
{
- Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
+ Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0);
+ Tensor3D mean_var = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(mean_var);
#ifndef IN_PLACE
Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0);
#endif /* IN_PLACE */
- INTERNAL_DATA_TYPE sum = 0.f;
- INTERNAL_DATA_TYPE sum_sq = 0.f;
+#if defined(NHWC)
+ const int ch = get_global_id(0); // Current channel
+ const int batch = get_global_id(2); // Current batch
+#else /* defined(NHWC) */
+ const int ch = get_global_id(2) % DIM_Z; // Current channel
+ const int batch = get_global_id(2) / DIM_Z; // Current batch
+#endif /* defined(NHWC) */
+
+ const __global DATA_TYPE *mean_ptr = (__global DATA_TYPE *)tensor3D_offset(&mean_var, ch, 0, batch);
+ const __global DATA_TYPE *var_ptr = (__global DATA_TYPE *)tensor3D_offset(&mean_var, ch, 1, batch);
+ const INTERNAL_DATA_TYPE mean = (INTERNAL_DATA_TYPE) * mean_ptr;
+ const INTERNAL_DATA_TYPE var = (INTERNAL_DATA_TYPE) * var_ptr;
+ const INTERNAL_DATA_TYPE multip = GAMMA / sqrt(var + EPSILON);
+ const INTERNAL_DATA_TYPE beta = (INTERNAL_DATA_TYPE)BETA;
#if defined(NHWC)
-
- const int ch = get_global_id(0); // Current channel
- const int batch = get_global_id(2); // Current batch
- const int elements_plane = DIM_Y * DIM_Z;
-
- for(int i_w = 0; i_w < DIM_Y; ++i_w)
- {
- for(int i_h = 0; i_h < DIM_Z; ++i_h)
- {
- INTERNAL_DATA_TYPE data = (INTERNAL_DATA_TYPE) * ((__global DATA_TYPE *)tensor4D_offset(&in, ch, i_w, i_h, batch));
- sum += data;
- sum_sq += data * data;
- }
- }
-
-#else // !defined(NHWC)
- const int ch = get_global_id(2) % DIM_Z; // Current channel
- const int batch = get_global_id(2) / DIM_Z; // Current batch
- const int elements_plane = DIM_X * DIM_Y;
-
- VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
- part_sum = 0.f;
- VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
- part_sum_sq = 0.f;
- // Calculate partial sum
- for(int y = 0; y < DIM_Y; ++y)
- {
- int x = 0;
- for(; x <= (DIM_X - VEC_SIZE); x += VEC_SIZE)
- {
- // Load data
- VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
- data = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)), VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE));
- part_sum += data;
- part_sum_sq += data * data;
- }
- // Left-overs loop
- for(; x < DIM_X; ++x)
- {
- INTERNAL_DATA_TYPE data = (INTERNAL_DATA_TYPE)(*((__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)));
- part_sum.s0 += data;
- part_sum_sq.s0 += data * data;
- }
- }
- // Perform reduction
-#if VEC_SIZE > 8
- part_sum.s01234567 += part_sum.s89abcdef;
- part_sum_sq.s01234567 += part_sum_sq.s89abcdef;
-#endif // VEC_SIZE > 8
-#if VEC_SIZE > 4
- part_sum.s0123 += part_sum.s4567;
- part_sum_sq.s0123 += part_sum_sq.s4567;
-#endif // VEC_SIZE > 4
-#if VEC_SIZE > 2
- part_sum.s01 += part_sum.s23;
- part_sum_sq.s01 += part_sum_sq.s23;
-#endif // VEC_SIZE > 2
- part_sum.s0 += part_sum.s1;
- part_sum_sq.s0 += part_sum_sq.s1;
-
- sum = (INTERNAL_DATA_TYPE)part_sum.s0;
- sum_sq = (INTERNAL_DATA_TYPE)part_sum_sq.s0;
-
-#endif // defined(NHWC)
-
- const INTERNAL_DATA_TYPE mean = (sum / elements_plane);
- const INTERNAL_DATA_TYPE var = (sum_sq / elements_plane) - (mean * mean);
- const INTERNAL_DATA_TYPE multip = GAMMA / sqrt(var + EPSILON);
-
-#if defined(NHWC)
-
- for(int i_w = 0; i_w < DIM_Y; ++i_w)
- {
- for(int i_h = 0; i_h < DIM_Z; ++i_h)
- {
- __global DATA_TYPE *input_address = (__global DATA_TYPE *)tensor4D_offset(&in, ch, i_w, i_h, batch);
-#ifdef IN_PLACE
- __global DATA_TYPE *output_address = input_address;
-#else /* !IN_PLACE */
- __global DATA_TYPE *output_address = (__global DATA_TYPE *)tensor4D_offset(&out, ch, i_w, i_h, batch);
+ const int in_offset = input_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
+#ifndef IN_PLACE
+ const int out_offset = output_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
#endif /* IN_PLACE */
- *(output_address) = (*(input_address) - mean) * multip + (INTERNAL_DATA_TYPE)BETA;
- }
+
+ for(int i = 0; i < (DIM_Y * DIM_Z); ++i)
+ {
+ __global DATA_TYPE *input_address = (__global DATA_TYPE *)(input_ptr + in_offset + i * input_stride_y);
+#ifdef IN_PLACE
+ __global DATA_TYPE *output_address = input_address;
+#else /* !IN_PLACE */
+ __global DATA_TYPE *output_address = (__global DATA_TYPE *)(output_ptr + out_offset + i * output_stride_y);
+#endif /* IN_PLACE */
+ *(output_address) = (*(input_address) - mean) * multip + beta;
}
#else // !defined(NHWC)