COMPMID-431 Port OpenCL pooling layer to use fixed point
Change-Id: I6a73cd6582097aaefa83588aad789bdefdc74406
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79967
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl
index 7527b1c..119879a 100644
--- a/src/core/CL/cl_kernels/activation_layer.cl
+++ b/src/core/CL/cl_kernels/activation_layer.cl
@@ -34,7 +34,7 @@
#define SUB_OP(a, b) SUB_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE)
#define MUL_OP(a, b) MUL_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define MLA_OP(a, b, c) MLA_SAT_OP_EXPAND((a), (b), (c), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
-#define DIV_OP(a, b) DIV_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+#define DIV_OP(a, b) DIV_SAT_OP_VEC_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define EXP_OP(a) EXP_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define LOG_OP(a) LOG_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define SQRT_OP(a) DIV_OP(CONST_ONE, INVSQRT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION))
diff --git a/src/core/CL/cl_kernels/fill_border.cl b/src/core/CL/cl_kernels/fill_border.cl
index f511613..2db8c67 100644
--- a/src/core/CL/cl_kernels/fill_border.cl
+++ b/src/core/CL/cl_kernels/fill_border.cl
@@ -24,6 +24,12 @@
#include "fixed_point.h"
#include "helpers.h"
+#if defined(FIXED_POINT_POSITION)
+
+#include "fixed_point.h"
+
+#endif /* FIXED_POINT_POSITION */
+
/** Fill N pixel of the padding edge of a single channel image by replicating the closest valid pixel.
*
* @attention The DATA_TYPE needs to be passed at the compile time.
diff --git a/src/core/CL/cl_kernels/fixed_point.h b/src/core/CL/cl_kernels/fixed_point.h
index d35a46f..478a414 100644
--- a/src/core/CL/cl_kernels/fixed_point.h
+++ b/src/core/CL/cl_kernels/fixed_point.h
@@ -290,7 +290,7 @@
#define MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position) mlal_sat_##type##x##size((a), (b), (c), (position))
#define MLAL_SAT_OP_EXPAND(a, b, c, type, size, position) MLAL_SAT_OP_EXPAND_STR(a, b, c, type, size, position)
-/** Saturate division of two fixed point numbers
+/** Saturate division of two fixed point vectors
*
* @param[in] stype the actual scalar data type.
* @param[in] type the actual data type.
@@ -298,22 +298,27 @@
*
* @return The result of the fixed point division. The result is saturated in case of overflow
*/
-#define DIVQ_SAT_IMPL(stype, type, itype) \
- inline type div_sat_##type(type VopA, type VopB, int fixed_point_position) \
- { \
- itype conv_a = CONVERT((VopA), itype); \
- itype denominator = CONVERT((VopB), itype); \
- itype numerator = conv_a << (itype)(fixed_point_position); \
- itype res = select(numerator / denominator, select((itype)stype##_MAX, (itype)stype##_MIN, conv_a < (itype)0), denominator == (itype)0); \
- return CONVERT_SAT((res), type); \
+#define DIVQ_SAT_IMPL(stype, type, itype) \
+ inline type div_sat_##type(type VopA, type VopB, int fixed_point_position) \
+ { \
+ itype conv_a = CONVERT((VopA), itype); \
+ itype denominator = CONVERT((VopB), itype); \
+ itype numerator = conv_a << (itype)(fixed_point_position); \
+ itype res = select((itype)(numerator / denominator), select((itype)stype##_MAX, (itype)stype##_MIN, (itype)(conv_a < (itype)0)), (itype)(denominator == (itype)0)); \
+ return CONVERT_SAT((res), type); \
}
DIVQ_SAT_IMPL(qs8, qs8x16, qs16x16)
DIVQ_SAT_IMPL(qs16, qs16x8, qs32x8)
DIVQ_SAT_IMPL(qs16, qs16x16, qs32x16)
+DIVQ_SAT_IMPL(qs8, qs8, qs16)
+DIVQ_SAT_IMPL(qs16, qs16, qs32)
-#define DIV_SAT_OP_EXPAND_STR(a, b, type, size, position) div_sat_##type##x##size((a), (b), (position))
-#define DIV_SAT_OP_EXPAND(a, b, type, size, position) DIV_SAT_OP_EXPAND_STR(a, b, type, size, position)
+#define DIV_SAT_OP_EXPAND_STR(a, b, type, position) div_sat_##type((a), (b), (position))
+#define DIV_SAT_OP_EXPAND(a, b, type, position) DIV_SAT_OP_EXPAND_STR(a, b, type, position)
+
+#define DIV_SAT_OP_VEC_EXPAND_STR(a, b, type, size, position) div_sat_##type##x##size((a), (b), (position))
+#define DIV_SAT_OP_VEC_EXPAND(a, b, type, size, position) DIV_SAT_OP_VEC_EXPAND_STR(a, b, type, size, position)
/** Saturate exponential of a fixed point vector
*
@@ -372,7 +377,7 @@
type B = -(type)(0x56AE >> (15 - fixed_point_position)); /* -0.6771900 */ \
type C = (type)(0x2933 >> (15 - fixed_point_position)); /* 0.3218538 */ \
type D = -(type)(0x0AA7 >> (15 - fixed_point_position)); /* -0.0832229 */ \
- type inter_a = select(VopA, DIV_SAT_OP_EXPAND(const_one, VopA, stype, size, fixed_point_position), VopA < const_one); \
+ type inter_a = select(VopA, DIV_SAT_OP_VEC_EXPAND(const_one, VopA, stype, size, fixed_point_position), VopA < const_one); \
type shift_val = (type)(15 - stype##_SHIFT) - clz(inter_a >> (type)fixed_point_position); \
inter_a = inter_a >> shift_val; \
inter_a = sub_sat(inter_a, const_one); \
@@ -444,7 +449,7 @@
type exp2x = EXP_OP_EXPAND(MUL_SAT_OP_EXPAND(const_two, VopA, stype, size, fixed_point_position), stype, size, fixed_point_position); \
type num = SUB_SAT_OP_EXPAND(exp2x, const_one, stype, size); \
type den = ADD_SAT_OP_EXPAND(exp2x, const_one, stype, size); \
- return DIV_SAT_OP_EXPAND(num, den, stype, size, fixed_point_position); \
+ return DIV_SAT_OP_VEC_EXPAND(num, den, stype, size, fixed_point_position); \
}
TANHQ_IMPL(qs8, qs8x16, 16)
diff --git a/src/core/CL/cl_kernels/normalization_layer.cl b/src/core/CL/cl_kernels/normalization_layer.cl
index 598b734..e2a5c40 100644
--- a/src/core/CL/cl_kernels/normalization_layer.cl
+++ b/src/core/CL/cl_kernels/normalization_layer.cl
@@ -28,7 +28,7 @@
#include "fixed_point.h"
#define MUL_OP(x, y) MUL_SAT_OP_EXPAND((x), (y), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define ADD_OP(x, y) ADD_SAT_OP_EXPAND((x), (y), DATA_TYPE, VEC_SIZE)
-#define DIV_OP(x, y) DIV_SAT_OP_EXPAND((x), (y), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
+#define DIV_OP(x, y) DIV_SAT_OP_VEC_EXPAND((x), (y), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define EXP_OP(x) EXP_OP_EXPAND((x), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define LOG_OP(x) LOG_OP_EXPAND((x), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
#define POW_OP(x, y) EXP_OP(MUL_OP(LOG_OP((x)), (y)))
diff --git a/src/core/CL/cl_kernels/pooling_layer.cl b/src/core/CL/cl_kernels/pooling_layer.cl
index 06989aa..18ad4a6 100644
--- a/src/core/CL/cl_kernels/pooling_layer.cl
+++ b/src/core/CL/cl_kernels/pooling_layer.cl
@@ -23,12 +23,31 @@
*/
#include "helpers.h"
-#ifdef POOL_AVG
+#ifdef FIXED_POINT_POSITION
+
+#include "fixed_point.h"
+
+#if defined(POOL_AVG)
+#define POOL_OP(x, y) add_sat(x, y)
+#else /* POOL_AVG */
+#define POOL_OP(x, y) (max((x), (y)))
+#endif /* POOL_AVG */
+
+#define DIV_OP1(x, y) DIV_SAT_OP_EXPAND((x), y, DATA_TYPE, FIXED_POINT_POSITION)
+#define DIV_OP(x, y) DIV_OP1(x, y << FIXED_POINT_POSITION)
+
+#else /* FIXED_POINT_POSITION */
+
+#if defined(POOL_AVG)
#define POOL_OP(x, y) ((x) + (y))
#else /* POOL_AVG */
#define POOL_OP(x, y) (fmax((x), (y)))
#endif /* POOL_AVG */
+#define DIV_OP(x, y) (x * (1.f / y))
+
+#endif /* FIXED_POINT_POSITION */
+
#if STRIDE_X == 1
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE1(res, input, output)
#elif STRIDE_X == 2 /* STRIDE_X == 1 */
@@ -37,9 +56,6 @@
#define POOLING3x3(res, input, output) POOLING3x3_STRIDE3(res, input, output)
#endif /* STRIDE_X == 3 */
-#define CONVERT_OP(data_type) convert_##data_type##4
-#define CONVERT_VECTOR4(data_type) CONVERT_OP(data_type)
-
#define POOLING3x3_STRIDE1(res, input, output) \
({ \
VEC_DATA_TYPE(DATA_TYPE, 4) \
@@ -142,30 +158,19 @@
const int start_y = get_global_id(1) * stride_y - pad_y;
const int end_x = min(start_x + pool_size, upper_bound_w);
const int end_y = min(start_y + pool_size, upper_bound_h);
- return 1.f / ((end_y - start_y) * (end_x - start_x));
-}
-
-VEC_DATA_TYPE(DATA_TYPE, 4)
-calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upper_bound_h,
- const int pad_x, const int pad_y, const int stride_x, const int stride_y)
-{
- const int4 start_x = ((int4)get_global_id(0) * 4 + (int4)(0, 1, 2, 3)) * (int4)stride_x - (int4)pad_x;
- const int start_y = get_global_id(1) * stride_y - pad_y;
- const int4 end_x = min(start_x + (int4)pool_size, (int4)upper_bound_w);
- const int end_y = min(start_y + pool_size, upper_bound_h);
- return (VEC_DATA_TYPE(DATA_TYPE, 4))(1.f) / CONVERT_VECTOR4(DATA_TYPE)(((int4)(end_y - start_y)) * (end_x - start_x));
+ return ((end_y - start_y) * (end_x - start_x));
}
/** Performs a pooling function of pool size equal to 2.
*
- * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
+ * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
- * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
+ * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32
* @param[in] input_stride_x Stride of the source image 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 source image in Y dimension (in bytes)
@@ -202,7 +207,7 @@
// Divide by pool region in case of average pooling
#ifdef POOL_AVG
- res *= calculate_avg_scale(2, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+ res = DIV_OP(res, calculate_avg_scale(2, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* POOL_AVG */
// Store result
@@ -211,14 +216,14 @@
/** Performs a pooling function of pool size equal to 3
*
- * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
+ * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
- * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
+ * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32
* @param[in] input_stride_x Stride of the source image 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 source image in Y dimension (in bytes)
@@ -258,17 +263,32 @@
// Divide by pool region in case of average pooling
#ifdef POOL_AVG
- res *= calculate_avg_scale(3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
-#endif //POOL_AVG
+ res = DIV_OP(res, calculate_avg_scale(3, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
+#endif /* POOL_AVG */
// Store result
*(__global DATA_TYPE *)output.ptr = res;
}
-#if defined(POOLING3x3)
+#if defined(POOLING3x3) && !defined(FIXED_POINT_POSITION)
+
+#define CONVERT_OP(data_type) convert_##data_type##4
+#define CONVERT_VECTOR4(data_type) CONVERT_OP(data_type)
+
+VEC_DATA_TYPE(DATA_TYPE, 4)
+calculate_avg_scale4(const int pool_size, const int upper_bound_w, const int upper_bound_h,
+ const int pad_x, const int pad_y, const int stride_x, const int stride_y)
+{
+ const int4 start_x = ((int4)get_global_id(0) * 4 + (int4)(0, 1, 2, 3)) * (int4)stride_x - (int4)pad_x;
+ const int start_y = get_global_id(1) * stride_y - pad_y;
+ const int4 end_x = min(start_x + (int4)pool_size, (int4)upper_bound_w);
+ const int end_y = min(start_y + pool_size, upper_bound_h);
+ return (VEC_DATA_TYPE(DATA_TYPE, 4))(1.f) / CONVERT_VECTOR4(DATA_TYPE)(((int4)(end_y - start_y)) * (end_x - start_x));
+}
+
/** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3
*
- * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
+ * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
@@ -313,18 +333,18 @@
vstore4(res, 0, (__global DATA_TYPE *)output.ptr);
}
-#endif // defined(POOLING3x3)
+#endif // defined(POOLING3x3) && !defined(FIXED_POINT_POSITION)
/** Performs a pooling function of pool size equal to 7.
*
- * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32;
+ * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are QS8/QS16/F16/F32;
* @note In case of average pooling the following information must be passed at compile time:
* -DPOOL_AVG must be provided otherwise max pooling will be performed.
* -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
* -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions
* -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension
*
- * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32
+ * @param[in] input_ptr Pointer to the source image. Supported data types: QS8/QS16/F16/F32
* @param[in] input_stride_x Stride of the source image 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 source image in Y dimension (in bytes)
@@ -389,7 +409,7 @@
// Divide by pool region in case of average pooling
#ifdef POOL_AVG
- res *= calculate_avg_scale(7, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+ res = DIV_OP(res, calculate_avg_scale(7, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y));
#endif /* POOL_AVG */
// Store result
diff --git a/src/core/CL/cl_kernels/softmax_layer.cl b/src/core/CL/cl_kernels/softmax_layer.cl
index e895bc1..9b24380 100644
--- a/src/core/CL/cl_kernels/softmax_layer.cl
+++ b/src/core/CL/cl_kernels/softmax_layer.cl
@@ -29,7 +29,7 @@
#define MAX_OP(x, y, type, size) MAX_OP_EXPAND(x, y, type, size)
#define ADD_OP(x, y, type, size) ADD_SAT_OP_EXPAND((x), (y), type, size)
#define SUB_OP(x, y, type, size) SUB_SAT_OP_EXPAND((x), (y), type, size)
-#define DIV_OP(x, y, type, size) DIV_SAT_OP_EXPAND((x), (y), type, size, FIXED_POINT_POSITION)
+#define DIV_OP(x, y, type, size) DIV_SAT_OP_VEC_EXPAND((x), (y), type, size, FIXED_POINT_POSITION)
#define EXP_OP(x, type, size) EXP_OP_EXPAND((x), type, size, FIXED_POINT_POSITION)
#define MIN_VAL_EXPAND(type) type##_MIN
diff --git a/src/core/CL/kernels/CLFillBorderKernel.cpp b/src/core/CL/kernels/CLFillBorderKernel.cpp
index 7667491..6ff1521 100644
--- a/src/core/CL/kernels/CLFillBorderKernel.cpp
+++ b/src/core/CL/kernels/CLFillBorderKernel.cpp
@@ -90,6 +90,10 @@
build_opts.emplace(("-DBORDER_SIZE_BOTTOM=" + support::cpp11::to_string(border_size.bottom)));
build_opts.emplace(("-DBORDER_SIZE_LEFT=" + support::cpp11::to_string(border_size.left)));
build_opts.emplace(("-DBORDER_SIZE_RIGHT=" + support::cpp11::to_string(border_size.right)));
+ if(is_data_type_fixed_point(tensor->info()->data_type()))
+ {
+ build_opts.emplace("-DFIXED_POINT_POSITION");
+ }
// Create kernel
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
index 6b2e881..3ef4725 100644
--- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp
+++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
@@ -67,10 +67,12 @@
static const std::set<int> supported_pool_sizes = { 2, 3, 7 };
ARM_COMPUTE_UNUSED(supported_pool_sizes);
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_ERROR_ON_NULLPTR(output);
ARM_COMPUTE_ERROR_ON(supported_pool_sizes.find(pool_size) == supported_pool_sizes.end());
ARM_COMPUTE_ERROR_ON(pool_pad_x >= pool_size || pool_pad_y >= pool_size);
+ ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
// Check output dimensions
std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(0),
@@ -94,7 +96,7 @@
// Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenCL kernel where
// each thread computes 4 output elements
- const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3);
+ const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3) && !is_data_type_fixed_point(input->info()->data_type());
int num_elements_read_per_iteration = (pool_size == 7) ? 8 : pool_size;
if(is_pool3x3_stride_le3)
@@ -120,6 +122,11 @@
std::set<std::string> build_opts;
build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
build_opts.emplace(("-DPOOL_" + ((PoolingType::MAX == pool_type) ? std::string("MAX") : std::string("AVG"))));
+ if(is_data_type_fixed_point(input->info()->data_type()))
+ {
+ build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
+ }
+
build_opts.emplace(("-DSTRIDE_X=" + support::cpp11::to_string(pool_stride_x)));
if(pool_type == PoolingType::AVG)
{