COMPMID-3735 Remove OpenCL padding: CLSoftmaxLayerKernel
- Renamed SELECT_DATA_TYPE to SELECT_VEC_DATA_TYPE to reflect its usage with vectors. SELECT_DATA_TYPE(dt) will now return the primitive data type
- Changed the interface of VEC_OFFS and V_OFFS in order to receive the primitive data type as a parameter rather than its vector form
- Performed a general cleanup of the kernels, such as creating macro for sum and max reduces, remove reduntant macros, defines, variables, calculations, etc...
- Using VEC_SIZE and VEC_SIZE_LEFTOVER in every kernel in order to allow computation for smaller shapes without adding paddings
- Removed the actual padding from the kernel and adjusting its calculations accordingly. Added asserts for padding removal checks. Removed invalid Validate tests.
Change-Id: If5ccbd5d34e255d38c7f6bfe8740e2b80b28e264
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4277
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/cl_kernels/softmax_layer_quantized.cl b/src/core/CL/cl_kernels/softmax_layer_quantized.cl
index 22b8df8..b7a6e00 100644
--- a/src/core/CL/cl_kernels/softmax_layer_quantized.cl
+++ b/src/core/CL/cl_kernels/softmax_layer_quantized.cl
@@ -23,67 +23,107 @@
*/
#include "helpers_asymm.h"
-#define MAX_OP(x, y, type, size) max((x), (y))
-#define ADD_OP(x, y, type, size) ((x) + (y))
-#define SUB_OP(x, y, type, size) ((x) - (y))
+#if defined(DATA_TYPE) && defined(MIN_VALUE) && defined(VECTOR_SIZE) && defined(VECTOR_SIZE_LEFTOVER) && defined(DIFF_MIN)
+
+#define VEC_BASE VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
+#define VEC_INT VEC_DATA_TYPE(int, VECTOR_SIZE)
+
+/** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel.
+ *
+ * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE, e.g. -DDATA_TYPE=uchar
+ * @note The zero value for the given data type must be given as a preprocessor argument using -DMIN_VALUE, e.g. -DMIN_VALUE=-128
+ * @note Vector size should be given as a preprocessor argument using -DVECTOR_SIZE=size. e.g. -DVECTOR_SIZE=16
+ * @note Leftover vector size has to be passed at compile time using -DVECTOR_SIZE_LEFTOVER. e.g. -DVECTOR_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VECTOR_SIZE
+ * @note Quantized beta can be optionally passed at compile time using -DINPUT_BETA_MULTIPLIER and -DINPUT_BETA_LEFT_SHIFT (if undefined, assume beta equals 1.0)
+ * @note Additional quantization data must be passed at compile time using -DSCALED_DIFF_INT_BITS and -DEXP_ACCUMULATION_INT_BITS.
+ * @note -DDIFF_MIN must be passed at compile time. It is threshold difference between maximum value of input data and current processed value, it defines whether the value will be taken into account or not.
+ * @note In case the input's data type is QASYMM8_SIGNED, -DQASYMM8_SIGNED must be passed.
+ *
+ * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: S32
+ * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[in] sum_ptr Pointer to the sum values tensor slice. Supported data types: same as @p src_ptr
+ * @param[in] sum_stride_x Stride of the sum values tensor in X dimension (in bytes)
+ * @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] sum_stride_y Stride of the sum values tensor in Y dimension (in bytes)
+ * @param[in] sum_step_y sum_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] sum_stride_z Stride of the sum values tensor in Z dimension (in bytes)
+ * @param[in] sum_step_z sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
+ * @param[out] dst_ptr Pointer to the destination tensor slice. Supported data types: QASYMM8/QASYMM8_SIGNED
+ * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void softmax_layer_norm_quantized(
+ TENSOR3D_DECLARATION(src),
+ TENSOR3D_DECLARATION(sum),
+ TENSOR3D_DECLARATION(dst))
+{
+ const int x_offs = max((int)(get_global_id(0) * VECTOR_SIZE - (VECTOR_SIZE - VECTOR_SIZE_LEFTOVER) % VECTOR_SIZE), 0);
+
+ __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x_offs * sizeof(int) + get_global_id(1) * src_stride_y + get_global_id(2) * src_stride_z;
+ __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * dst_stride_y + get_global_id(2) * dst_stride_z;
+
+ Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(sum);
+
+ // Load max value of 1D logits vector (row)
+ int sum_val = *((__global int *)offset(&sum, 0, get_global_id(1)));
+
+ // It will be better to calculate this in prev layer and pass here as parameter
+ uint sum_val_u = convert_uint(sum_val);
+ int headroom_plus_one = clz(sum_val_u);
+ int num_bits_over_unit = EXP_ACCUMULATION_INT_BITS - headroom_plus_one;
+ int shifted_sum_minus_one_1 = convert_int((sum_val_u << headroom_plus_one) - (1u << 31));
+ VEC_INT shifted_sum_minus_one = shifted_sum_minus_one_1;
+ VEC_INT shifted_scale = ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1(shifted_sum_minus_one, VECTOR_SIZE);
+
+ // It was already calculated in prev layer, should be stored into tmp output and reused
+ VEC_INT data_diff = VLOAD(VECTOR_SIZE)(0, (__global int *)src_addr);
+ VEC_INT data_diff_mult = data_diff;
+#if defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT)
+ if(INPUT_BETA_MULTIPLIER > 1)
+ {
+ data_diff_mult = ASYMM_MULT(data_diff * (1 << INPUT_BETA_LEFT_SHIFT), INPUT_BETA_MULTIPLIER, VECTOR_SIZE);
+ }
+#endif /* defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT) */
+
+ VEC_INT data = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, VECTOR_SIZE);
+ data = ASYMM_MULT(shifted_scale, data, VECTOR_SIZE);
+ data = ASYMM_ROUNDING_DIVIDE_BY_POW2(data, num_bits_over_unit + 31 - 8, VECTOR_SIZE);
+#ifdef QASYMM8_SIGNED
+ data += (VEC_INT)(MIN_VALUE);
+#endif /* QASYMM8_SIGNED */
+ data = select(MIN_VALUE, data, data_diff >= (VEC_INT)(DIFF_MIN));
+ VEC_BASE data0 = CONVERT_SAT(data, VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE));
+
+ STORE_VECTOR_SELECT(data, DATA_TYPE, dst_addr, VECTOR_SIZE, VECTOR_SIZE_LEFTOVER, VECTOR_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
+}
+
+#if defined(SRC_WIDTH) && defined(LOG_VECTOR_SIZE)
/* Number of workitems in dimension 0. */
#if !defined(GRID_SIZE)
#define GRID_SIZE 1
#endif /* !defined(GRID_SIZE) */
-#if VECTOR_SIZE == 2
-__constant uint2 idx__ = (uint2)(0, 1);
-#define asymm_mult(a, b) ASYMM_MULT(a, b, 2)
-#define asymm_exp_on_negative_values(a, k_integer_bits) ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, 2)
-#define asymm_rescale(value, src_integer_bits, dst_integer_bits) ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, 2)
-
-#elif VECTOR_SIZE == 4
-__constant uint4 idx__ = (uint4)(0, 1, 2, 3);
-#define asymm_mult(a, b) ASYMM_MULT(a, b, 4)
-#define asymm_exp_on_negative_values(a, k_integer_bits) ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, 4)
-#define asymm_rescale(value, src_integer_bits, dst_integer_bits) ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, 4)
-
-#elif VECTOR_SIZE == 8
-__constant uint8 idx__ = (uint8)(0, 1, 2, 3, 4, 5, 6, 7);
-#define asymm_mult(a, b) ASYMM_MULT(a, b, 8)
-#define asymm_exp_on_negative_values(a, k_integer_bits) ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, 8)
-#define asymm_rescale(value, src_integer_bits, dst_integer_bits) ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, 8)
-
-#else /* VECTOR_SIZE DEFAULT */
-#define VECTOR_SIZE 16
-#define LOG_VECTOR_SIZE 4
-__constant uint16 idx__ = (uint16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
-#define asymm_mult(a, b) ASYMM_MULT(a, b, 16)
-#define asymm_exp_on_negative_values(a, k_integer_bits) ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, 16)
-#define asymm_rescale(value, src_integer_bits, dst_integer_bits) ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, 16)
-
-#endif /* VECTOR_SIZE END */
-
-#define VEC_UCHAR VEC_DATA_TYPE(uchar, VECTOR_SIZE)
#define VEC_UINT VEC_DATA_TYPE(uint, VECTOR_SIZE)
-#define VEC_INT VEC_DATA_TYPE(int, VECTOR_SIZE)
-#define VEC_BASE VEC_DATA_TYPE(DATA_TYPE, VECTOR_SIZE)
-#if defined(DIFF_MIN)
-
-VEC_INT mult_by_quantized_multiplier_serial(VEC_INT data)
+VEC_INT mult_by_quantized_multiplier(VEC_INT data)
{
#if defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT)
if(INPUT_BETA_MULTIPLIER > 1)
{
- return asymm_mult(data * (1 << INPUT_BETA_LEFT_SHIFT), INPUT_BETA_MULTIPLIER);
- }
-#endif /* defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT) */
- return data;
-}
-
-int4 mult_by_quantized_multiplier_parallel(int4 data)
-{
-#if defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT)
- if(INPUT_BETA_MULTIPLIER > 1)
- {
- return ASYMM_MULT(data * (1 << INPUT_BETA_LEFT_SHIFT), INPUT_BETA_MULTIPLIER, 4);
+ return ASYMM_MULT(data * (1 << INPUT_BETA_LEFT_SHIFT), INPUT_BETA_MULTIPLIER, VECTOR_SIZE);
}
#endif /* defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT) */
return data;
@@ -92,9 +132,15 @@
/** Shifts the values of the input tensor by the max calculated in softmax_layer_max kernel,
* then gets the exponent of each element as sums all elements across each row.
*
- * @note In case the input is not multiple of 16 -DNON_MULTIPLE_OF_VECTOR_SIZE must be passed.
+ * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE, e.g. -DDATA_TYPE=uchar
+ * @note The zero value for the given data type must be given as a preprocessor argument using -DMIN_VALUE, e.g. -DMIN_VALUE=-128
+ * @note Vector size should be given as a preprocessor argument using -DVECTOR_SIZE=size. e.g. -DVECTOR_SIZE=16
+ * @note Leftover vector size has to be passed at compile time using -DVECTOR_SIZE_LEFTOVER. e.g. -DVECTOR_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VECTOR_SIZE
+ * @note In case the input is not multiple of VECTOR_SIZE -DNON_MULTIPLE_OF_VECTOR_SIZE must be passed.
* @note Quantized beta can be optionally passed at compile time using -DINPUT_BETA_MULTIPLIER and -DINPUT_BETA_LEFT_SHIFT (if undefined, assume beta equals 1.0)
+ * @note Additional quantization data must be passed at compile time using -DSCALED_DIFF_INT_BITS and -DEXP_ACCUMULATION_INT_BITS.
* @note -DDIFF_MIN must be passed at compile time. It is threshold difference between maximum value of input data and current processed value, it defines whether the value will be taken into account or not.
+ * @note In case the input's data type is QASYMM8_SIGNED, -DQASYMM8_SIGNED must be passed.
*
* @param[in] src_ptr Pointer to the source tensor slice. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
@@ -128,111 +174,89 @@
* @param[in] sum_stride_z Stride of the sum values tensor in Z dimension (in bytes)
* @param[in] sum_step_z sum_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
- * @param[in] width Input image width
*/
__kernel void softmax_layer_max_shift_exp_sum_quantized_serial(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(maxo),
TENSOR3D_DECLARATION(dst),
- TENSOR3D_DECLARATION(sum),
- uint width)
+ TENSOR3D_DECLARATION(sum))
{
- Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
- Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+ __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + get_global_id(1) * src_stride_y + get_global_id(2) * src_stride_z;
+ __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * dst_stride_y + get_global_id(2) * dst_stride_z;
+
Image maxo = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(maxo);
Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(sum);
VEC_BASE max_val_vec = (VEC_BASE)(MIN_VALUE);
// Calculate max of row
- const uint width4 = width >> LOG_VECTOR_SIZE;
- for(uint i = 0; i < width4; i++)
- {
- VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)offset(&src, i << LOG_VECTOR_SIZE, 0));
- max_val_vec = MAX_OP(data, max_val_vec, DATA_TYPE, 16);
- }
-
#ifdef NON_MULTIPLE_OF_VECTOR_SIZE
- // Handle non multiple of 16
VEC_BASE vec_min_val = (VEC_BASE)(MIN_VALUE);
- VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)offset(&src, width4 << LOG_VECTOR_SIZE, 0));
- VEC_UCHAR widx = CONVERT(((VEC_UINT)(width4 << LOG_VECTOR_SIZE) + idx__) < width, VEC_UCHAR);
- max_val_vec = MAX_OP(max_val_vec, select(vec_min_val, data, widx), DATA_TYPE, 16);
+ VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)src_addr);
+ VEC_INT widx = (VEC_INT)VECTOR_SIZE_LEFTOVER > VEC_OFFS(int, VECTOR_SIZE);
+ max_val_vec = max(max_val_vec, select(vec_min_val, data, CONVERT(widx, VEC_BASE)));
#endif /* NON_MULTIPLE_OF_VECTOR_SIZE */
- // Perform max reduction
-#if VECTOR_SIZE == 16
- max_val_vec.s01234567 = MAX_OP(max_val_vec.s01234567, max_val_vec.s89ABCDEF, DATA_TYPE, 8);
-#endif /* VECTOR SIZE 16 END */
-#if VECTOR_SIZE >= 8
- max_val_vec.s0123 = MAX_OP(max_val_vec.s0123, max_val_vec.s4567, DATA_TYPE, 4);
-#endif /* VECTOR SIZE 8 END */
-#if VECTOR_SIZE >= 4
- max_val_vec.s01 = MAX_OP(max_val_vec.s01, max_val_vec.s23, DATA_TYPE, 2);
-#endif /* VECTOR SIZE 4 END */
- max_val_vec.s0 = MAX_OP(max_val_vec.s0, max_val_vec.s1, DATA_TYPE, 1);
+ for(uint i = VECTOR_SIZE_LEFTOVER; i < SRC_WIDTH; i += VECTOR_SIZE)
+ {
+ VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr + i * sizeof(DATA_TYPE)));
+ max_val_vec = max(data, max_val_vec);
+ }
- // Store result
- *((__global DATA_TYPE *)maxo.ptr) = max_val_vec.s0;
+ // Perform max reduction
+ DATA_TYPE max_local = MAX_REDUCE(max_val_vec, VECTOR_SIZE);
+ *((__global DATA_TYPE *)maxo.ptr) = max_local;
// Second part
// Load max value of 1D logits vector (row)
- int max_val = convert_int(*((__global DATA_TYPE *)offset(&maxo, 0, 0)));
+ int max_val = convert_int(max_local);
// Set sum vector, Q(EXP_ACCUMULATION_INT_BITS)
VEC_INT sum1D = 0;
+#ifdef NON_MULTIPLE_OF_VECTOR_SIZE
+ VEC_INT data_fp = CONVERT(data, VEC_INT);
+ VEC_INT data_diff = data_fp - max_val;
+ VEC_INT data_diff_mult = mult_by_quantized_multiplier(data_diff);
+ data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, VECTOR_SIZE);
+ data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, VECTOR_SIZE);
+ VSTORE_PARTIAL(VECTOR_SIZE, VECTOR_SIZE_LEFTOVER)
+ (data_diff, 0, (__global int *)dst_addr);
+ data_fp = select(0, data_fp, data_diff >= (VEC_INT)(DIFF_MIN));
+ sum1D += select(0, data_fp, widx);
+#endif /* NON_MULTIPLE_OF_VECTOR_SIZE */
+
// Shift values, exp and sum
- for(uint i = 0; i < width4; i++)
+ for(uint i = VECTOR_SIZE_LEFTOVER; i < SRC_WIDTH; i += VECTOR_SIZE)
{
- VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)offset(&src, i << LOG_VECTOR_SIZE, 0));
+ VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr + i * sizeof(DATA_TYPE)));
VEC_INT data_fp = CONVERT(data, VEC_INT);
VEC_INT data_diff = data_fp - max_val;
- VEC_INT data_diff_mult = mult_by_quantized_multiplier_serial(data_diff);
- data_fp = asymm_exp_on_negative_values(data_diff_mult, SCALED_DIFF_INT_BITS);
- data_fp = asymm_rescale(data_fp, 0, EXP_ACCUMULATION_INT_BITS);
+ VEC_INT data_diff_mult = mult_by_quantized_multiplier(data_diff);
+ data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, VECTOR_SIZE);
+ data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, VECTOR_SIZE);
VSTORE(VECTOR_SIZE)
- (data_diff, 0, (__global int *)offset(&dst, i << LOG_VECTOR_SIZE, 0));
+ (data_diff, 0, (__global int *)(dst_addr + i * sizeof(int)));
sum1D = sum1D + select(0, data_fp, data_diff >= (VEC_INT)(DIFF_MIN));
}
-#ifdef NON_MULTIPLE_OF_VECTOR_SIZE
- // Handle non multiple of 16
- data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)offset(&src, width4 << LOG_VECTOR_SIZE, 0));
- VEC_INT data_fp = CONVERT(data, VEC_INT);
- VEC_INT data_diff = data_fp - max_val;
- VEC_INT data_diff_mult = mult_by_quantized_multiplier_serial(data_diff);
- data_fp = asymm_exp_on_negative_values(data_diff_mult, SCALED_DIFF_INT_BITS);
- data_fp = asymm_rescale(data_fp, 0, EXP_ACCUMULATION_INT_BITS);
- VEC_INT widx_ = CONVERT(((VEC_UINT)(width4 << LOG_VECTOR_SIZE) + idx__) < width, VEC_INT);
- VSTORE(VECTOR_SIZE)
- (data_diff, 0, (__global int *)offset(&dst, width4 << LOG_VECTOR_SIZE, 0));
- data_fp = select(0, data_fp, data_diff >= (VEC_INT)(DIFF_MIN));
- sum1D = sum1D + select(0, data_fp, widx_);
-#endif /* NON_MULTIPLE_OF_VECTOR_SIZE */
-
// Perform sum reduction
-#if VECTOR_SIZE == 16
- sum1D.s01234567 = ADD_OP(sum1D.s01234567, sum1D.s89ABCDEF, DATA_TYPE, 8);
-#endif /* VECTOR SIZE 16 END */
-#if VECTOR_SIZE >= 8
- sum1D.s0123 = ADD_OP(sum1D.s0123, sum1D.s4567, DATA_TYPE, 4);
-#endif /* VECTOR SIZE 8 END */
-#if VECTOR_SIZE >= 4
- sum1D.s01 = ADD_OP(sum1D.s01, sum1D.s23, DATA_TYPE, 2);
-#endif /* VECTOR SIZE 4 END */
- sum1D.s0 = ADD_OP(sum1D.s0, sum1D.s1, DATA_TYPE, 1);
-
- // Calculate and store result
- *((__global int *)sum.ptr) = sum1D.s0;
+ *((__global int *)sum.ptr) = SUM_REDUCE(sum1D, VECTOR_SIZE);
}
/** Identifies the maximum value across the 1st dimension and shifts the values of the input tensor by this maximum value,
* then gets the exponent of each element as sums all elements across each row.
*
- * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ * @note Datatype must be given as a preprocessor argument using -DDATA_TYPE, e.g. -DDATA_TYPE=uchar
+ * @note The zero value for the given data type must be given as a preprocessor argument using -DMIN_VALUE, e.g. -DMIN_VALUE=-128
+ * @note Vector size should be given as a preprocessor argument using -DVECTOR_SIZE=size. e.g. -DVECTOR_SIZE=16
+ * @note Leftover vector size has to be passed at compile time using -DVECTOR_SIZE_LEFTOVER. e.g. -DVECTOR_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VECTOR_SIZE
* @note In case the input is not a multiple of VECTOR_SIZE (2,4,8,16) -DNON_MULTIPLE_OF_VECTOR_SIZE must be passed.
+ * @note Quantized beta can be optionally passed at compile time using -DINPUT_BETA_MULTIPLIER and -DINPUT_BETA_LEFT_SHIFT (if undefined, assume beta equals 1.0)
+ * @note Additional quantization data must be passed at compile time using -DSCALED_DIFF_INT_BITS and -DEXP_ACCUMULATION_INT_BITS.
+ * @note -DDIFF_MIN must be passed at compile time. It is threshold difference between maximum value of input data and current processed value, it defines whether the value will be taken into account or not.
+ * @note In case the input's data type is QASYMM8_SIGNED, -DQASYMM8_SIGNED must be passed.
*
* @param[in] src_ptr Pointer to the source tensor slice. Supported data types: F16/F32
* @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
@@ -266,72 +290,59 @@
* @param[in] sum_stride_z Stride of the sum values tensor in Z dimension (in bytes)
* @param[in] sum_step_z sum_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
- * @param[in] width Input image width
*/
__kernel void softmax_layer_max_shift_exp_sum_quantized_parallel(
TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(maxo),
TENSOR3D_DECLARATION(dst),
- TENSOR3D_DECLARATION(sum),
- uint width)
+ TENSOR3D_DECLARATION(sum))
{
- Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
- Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+ const uint lid = get_local_id(0);
+ const uint x_offs = (VECTOR_SIZE_LEFTOVER + lid * VECTOR_SIZE);
+
+ __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x_offs * sizeof(DATA_TYPE) + get_global_id(1) * src_stride_y + get_global_id(2) * src_stride_z;
+ __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x_offs * sizeof(int) + get_global_id(1) * dst_stride_y + get_global_id(2) * dst_stride_z;
+
Image maxo = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(maxo);
Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(sum);
- const uint4 idx4 = (uint4)(0, 1, 2, 3);
- const uint lid = get_local_id(0);
-
// Define one temporary vector per work-item.
- __local int4 tmp_local[GRID_SIZE];
+ __local VEC_INT tmp_local[GRID_SIZE];
__local DATA_TYPE max_local;
- VEC_DATA_TYPE(DATA_TYPE, 4)
- vec_min_val = (VEC_DATA_TYPE(DATA_TYPE, 4))(MIN_VALUE);
- VEC_DATA_TYPE(DATA_TYPE, 4)
- max_val_vec = vec_min_val;
+ VEC_BASE vec_min_val = (VEC_BASE)(MIN_VALUE);
+ VEC_BASE max_val_vec = vec_min_val;
- // Number of elements per work-item.
- const uint row = width / GRID_SIZE;
// Number of iterations per work-item.
- const uint width_ = row >> 2;
+ const uint width = (SRC_WIDTH / GRID_SIZE) >> LOG_VECTOR_SIZE;
// Calculate max of row
uint i = 0;
- for(; i < width_; i++)
+ for(; i < width; ++i)
{
- VEC_DATA_TYPE(DATA_TYPE, 4)
- data_max = vload4(0, (__global DATA_TYPE *)offset(&src, i * GRID_SIZE * 4, 0));
- max_val_vec = MAX_OP(data_max, max_val_vec, DATA_TYPE, 4);
+ VEC_BASE data_max = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr + (i * GRID_SIZE * VECTOR_SIZE) * sizeof(DATA_TYPE)));
+ max_val_vec = max(data_max, max_val_vec);
}
#ifdef NON_MULTIPLE_OF_GRID_SIZE
// How many work-items needed to complete the computation.
//TODO: Optimize this calculation (avoid %).
- int boundary_workitems = (width % (GRID_SIZE * 4)) / 4;
+ int boundary_workitems = (SRC_WIDTH % (GRID_SIZE * VECTOR_SIZE)) / VECTOR_SIZE;
if(lid < boundary_workitems)
{
- VEC_DATA_TYPE(DATA_TYPE, 4)
- data_max = vload4(0, (__global DATA_TYPE *)offset(&src, i * GRID_SIZE * 4, 0));
- max_val_vec = MAX_OP(data_max, max_val_vec, DATA_TYPE, 4);
+ VEC_BASE data_max = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr + (i * GRID_SIZE * VECTOR_SIZE) * sizeof(DATA_TYPE)));
+ max_val_vec = max(data_max, max_val_vec);
}
#ifdef NON_MULTIPLE_OF_VECTOR_SIZE
- if(boundary_workitems == 0)
- {
- boundary_workitems = GRID_SIZE;
- i--;
- }
- if(lid == (boundary_workitems - 1))
+ VEC_INT widx;
+ if(lid == 0)
{
// Handle non multiple of 4
- VEC_DATA_TYPE(DATA_TYPE, 4)
- data_max = vload4(0, (__global DATA_TYPE *)offset(&src, (GRID_SIZE * i * 4) + 4, 0));
- VEC_DATA_TYPE(DATA_TYPE, 4)
- widx = CONVERT((((uint4)(GRID_SIZE * i * 4) + boundary_workitems * 4 + idx4) < width), VEC_DATA_TYPE(DATA_TYPE, 4));
- max_val_vec = MAX_OP(max_val_vec, select(vec_min_val, data_max, widx), DATA_TYPE, 4);
+ VEC_BASE data_max = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr - VECTOR_SIZE_LEFTOVER * sizeof(DATA_TYPE)));
+ widx = (VEC_INT)VECTOR_SIZE_LEFTOVER > VEC_OFFS(int, VECTOR_SIZE);
+ max_val_vec = max(max_val_vec, select(vec_min_val, data_max, CONVERT(widx, VEC_BASE)));
}
#endif /* NON_MULTIPLE_OF_VECTOR_SIZE */
#endif /* NON_MULTIPLE_OF_GRID_SIZE */
- tmp_local[lid] = convert_int4(max_val_vec);
+ tmp_local[lid] = CONVERT(max_val_vec, VEC_INT);
barrier(CLK_LOCAL_MEM_FENCE);
@@ -339,7 +350,7 @@
{
if(lid < 128)
{
- tmp_local[lid] = MAX_OP(tmp_local[lid + 128], tmp_local[lid], int, 4);
+ tmp_local[lid] = max(tmp_local[lid + 128], tmp_local[lid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -347,7 +358,7 @@
{
if(lid < 64)
{
- tmp_local[lid] = MAX_OP(tmp_local[lid + 64], tmp_local[lid], int, 4);
+ tmp_local[lid] = max(tmp_local[lid + 64], tmp_local[lid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -355,7 +366,7 @@
{
if(lid < 32)
{
- tmp_local[lid] = MAX_OP(tmp_local[lid + 32], tmp_local[lid], int, 4);
+ tmp_local[lid] = max(tmp_local[lid + 32], tmp_local[lid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -363,7 +374,7 @@
{
if(lid < 16)
{
- tmp_local[lid] = MAX_OP(tmp_local[lid + 16], tmp_local[lid], int, 4);
+ tmp_local[lid] = max(tmp_local[lid + 16], tmp_local[lid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -371,7 +382,7 @@
{
if(lid < 8)
{
- tmp_local[lid] = MAX_OP(tmp_local[lid + 8], tmp_local[lid], int, 4);
+ tmp_local[lid] = max(tmp_local[lid + 8], tmp_local[lid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -379,7 +390,7 @@
{
if(lid < 4)
{
- tmp_local[lid] = MAX_OP(tmp_local[lid + 4], tmp_local[lid], int, 4);
+ tmp_local[lid] = max(tmp_local[lid + 4], tmp_local[lid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -387,72 +398,64 @@
{
if(lid < 2)
{
- tmp_local[lid] = MAX_OP(tmp_local[lid + 2], tmp_local[lid], int, 4);
+ tmp_local[lid] = max(tmp_local[lid + 2], tmp_local[lid]);
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0)
{
- max_val_vec = MAX_OP(CONVERT((tmp_local[lid + 1]), VEC_DATA_TYPE(DATA_TYPE, 4)), CONVERT((tmp_local[lid]), VEC_DATA_TYPE(DATA_TYPE, 4)), DATA_TYPE, 4);
- max_val_vec.s01 = MAX_OP(max_val_vec.s01, max_val_vec.s23, DATA_TYPE, 2);
- max_val_vec.s0 = MAX_OP(max_val_vec.s0, max_val_vec.s1, DATA_TYPE, 1);
- max_local = max_val_vec.s0;
+ max_val_vec = max(CONVERT((tmp_local[lid + 1]), VEC_BASE), CONVERT((tmp_local[lid]), VEC_BASE));
+ max_local = MAX_REDUCE(max_val_vec, VECTOR_SIZE);
}
barrier(CLK_LOCAL_MEM_FENCE);
/* Second section */
// Set sum vector
- int4 sum1D = 0;
- int max_val = convert_int(max_local);
+ VEC_INT sum1D = 0;
+ int max_val = convert_int(max_local);
// Shift values, exp and sum
- for(i = 0; i < width_; i++)
+ for(i = 0; i < width; ++i)
{
- VEC_DATA_TYPE(DATA_TYPE, 4)
- data = vload4(0, (__global DATA_TYPE *)offset(&src, i * GRID_SIZE * 4, 0));
- int4 data_fp = convert_int4(data);
- int4 data_diff = data_fp - max_val;
- int4 data_diff_mult = mult_by_quantized_multiplier_parallel(data_diff);
- data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, 4);
- data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, 4);
- vstore4(data_diff, 0, (__global int *)offset(&dst, i * GRID_SIZE * 4, 0));
- sum1D = sum1D + select(0, data_fp, data_diff >= (int4)(DIFF_MIN));
+ VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr + (i * GRID_SIZE * VECTOR_SIZE) * sizeof(DATA_TYPE)));
+ VEC_INT data_fp = CONVERT(data, VEC_INT);
+ VEC_INT data_diff = data_fp - max_val;
+ VEC_INT data_diff_mult = mult_by_quantized_multiplier(data_diff);
+ data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, VECTOR_SIZE);
+ data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, VECTOR_SIZE);
+ VSTORE(VECTOR_SIZE)
+ (data_diff, 0, (__global int *)(dst_addr + (i * GRID_SIZE * VECTOR_SIZE) * sizeof(int)));
+ sum1D = sum1D + select(0, data_fp, data_diff >= (VEC_INT)(DIFF_MIN));
}
#ifdef NON_MULTIPLE_OF_GRID_SIZE
//TODO: Optimize the calculation (avoid %).
- boundary_workitems = (width % (GRID_SIZE * 4)) / 4;
+ boundary_workitems = (SRC_WIDTH % (GRID_SIZE * VECTOR_SIZE)) / VECTOR_SIZE;
if(lid < boundary_workitems)
{
- VEC_DATA_TYPE(DATA_TYPE, 4)
- data = vload4(0, (__global DATA_TYPE *)offset(&src, i * GRID_SIZE * 4, 0));
- int4 data_fp = convert_int4(data);
- int4 data_diff = data_fp - max_val;
- int4 data_diff_mult = mult_by_quantized_multiplier_parallel(data_diff);
- data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, 4);
- data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, 4);
- vstore4(data_diff, 0, (__global int *)offset(&dst, i * GRID_SIZE * 4, 0));
- sum1D = sum1D + select(0, data_fp, data_diff >= (int4)(DIFF_MIN));
+ VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr + (i * GRID_SIZE * VECTOR_SIZE) * sizeof(DATA_TYPE)));
+ VEC_INT data_fp = CONVERT(data, VEC_INT);
+ VEC_INT data_diff = data_fp - max_val;
+ VEC_INT data_diff_mult = mult_by_quantized_multiplier(data_diff);
+ data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, VECTOR_SIZE);
+ data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, VECTOR_SIZE);
+ VSTORE(VECTOR_SIZE)
+ (data_diff, 0, (__global int *)(dst_addr + (i * GRID_SIZE * VECTOR_SIZE) * sizeof(int)));
+ sum1D = sum1D + select(0, data_fp, data_diff >= (VEC_INT)(DIFF_MIN));
}
#ifdef NON_MULTIPLE_OF_VECTOR_SIZE
- if(boundary_workitems == 0)
- {
- boundary_workitems = GRID_SIZE;
- i--;
- }
- if(lid == (boundary_workitems - 1))
+ if(lid == 0)
{
// Handle non multiple of vector size ((GRID_SIZE * i * 4) + 4, 0); move 4 float positions ahead, *4 is due to the stride
- VEC_DATA_TYPE(DATA_TYPE, 4)
- data = vload4(0, (__global DATA_TYPE *)offset(&src, i * GRID_SIZE * 4 + 4, 0));
- int4 data_fp = convert_int4(data);
- int4 data_diff = data_fp - max_val;
- int4 data_diff_mult = mult_by_quantized_multiplier_parallel(data_diff);
- data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, 4);
- data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, 4);
- int4 widx = convert_int4(((uint4)(GRID_SIZE * i * 4) + boundary_workitems * 4 + idx4) < width);
- vstore4(data_diff, 0, (__global int *)offset(&dst, i * GRID_SIZE * 4 + 4, 0));
- data_fp = select(MIN_VALUE, data_fp, data_diff >= (int4)(DIFF_MIN));
+ VEC_BASE data = VLOAD(VECTOR_SIZE)(0, (__global DATA_TYPE *)(src_addr - VECTOR_SIZE_LEFTOVER * sizeof(DATA_TYPE)));
+ VEC_INT data_fp = CONVERT(data, VEC_INT);
+ VEC_INT data_diff = data_fp - max_val;
+ VEC_INT data_diff_mult = mult_by_quantized_multiplier(data_diff);
+ data_fp = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, VECTOR_SIZE);
+ data_fp = ASYMM_RESCALE(data_fp, 0, EXP_ACCUMULATION_INT_BITS, VECTOR_SIZE);
+ VSTORE_PARTIAL(VECTOR_SIZE, VECTOR_SIZE_LEFTOVER)
+ (data_diff, 0, (__global int *)(dst_addr - VECTOR_SIZE_LEFTOVER * sizeof(int)));
+ data_fp = select(MIN_VALUE, data_fp, data_diff >= (VEC_INT)(DIFF_MIN));
data_fp = select(0, data_fp, widx);
sum1D = sum1D + data_fp;
}
@@ -466,7 +469,7 @@
{
if(lid < 128)
{
- tmp_local[lid] = ADD_OP(tmp_local[lid + 128], tmp_local[lid], int, 4);
+ tmp_local[lid] += tmp_local[lid + 128];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -474,7 +477,7 @@
{
if(lid < 64)
{
- tmp_local[lid] = ADD_OP(tmp_local[lid + 64], tmp_local[lid], int, 4);
+ tmp_local[lid] += tmp_local[lid + 64];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -482,7 +485,7 @@
{
if(lid < 32)
{
- tmp_local[lid] = ADD_OP(tmp_local[lid + 32], tmp_local[lid], int, 4);
+ tmp_local[lid] += tmp_local[lid + 32];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -490,7 +493,7 @@
{
if(lid < 16)
{
- tmp_local[lid] = ADD_OP(tmp_local[lid + 16], tmp_local[lid], int, 4);
+ tmp_local[lid] += tmp_local[lid + 16];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -498,7 +501,7 @@
{
if(lid < 8)
{
- tmp_local[lid] = ADD_OP(tmp_local[lid + 8], tmp_local[lid], int, 4);
+ tmp_local[lid] += tmp_local[lid + 8];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -506,7 +509,7 @@
{
if(lid < 4)
{
- tmp_local[lid] = ADD_OP(tmp_local[lid + 4], tmp_local[lid], int, 4);
+ tmp_local[lid] += tmp_local[lid + 4];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
@@ -514,88 +517,16 @@
{
if(lid < 2)
{
- tmp_local[lid] = ADD_OP(tmp_local[lid + 2], tmp_local[lid], int, 4);
+ tmp_local[lid] += tmp_local[lid + 2];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0)
{
- sum1D = ADD_OP(tmp_local[lid + 1], tmp_local[lid], int, 4);
- // Perform max reduction
- sum1D.s01 = ADD_OP(sum1D.s01, sum1D.s23, int, 2);
- sum1D.s0 = ADD_OP(sum1D.s0, sum1D.s1, int, 1);
- *((__global int *)sum.ptr) = sum1D.s0;
+ sum1D = (tmp_local[lid + 1] + tmp_local[lid]);
+ // Perform sum reduction
+ *((__global int *)sum.ptr) = SUM_REDUCE(sum1D, VECTOR_SIZE);
}
}
-
-/** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel.
- *
- * @note Quantized beta can be optionally passed at compile time using -DINPUT_BETA_MULTIPLIER and -DINPUT_BETA_LEFT_SHIFT (if undefined, assume beta equals 1.0)
- * @note -DDIFF_MIN must be passed at compile time. It is threshold difference between maximum value of input data and current processed value, it defines whether the value will be taken into account or not.
- *
- * @param[in] src_ptr Pointer to the source tensor slice. Supported data types: S32
- * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
- * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[in] sum_ptr Pointer to the sum values tensor slice. Supported data types: same as @p src_ptr
- * @param[in] sum_stride_x Stride of the sum values tensor in X dimension (in bytes)
- * @param[in] sum_step_x sum_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] sum_stride_y Stride of the sum values tensor in Y dimension (in bytes)
- * @param[in] sum_step_y sum_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] sum_stride_z Stride of the sum values tensor in Z dimension (in bytes)
- * @param[in] sum_step_z sum_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in] sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
- * @param[out] dst_ptr Pointer to the destination tensor slice. Supported data types: QASYMM8/QASYMM8_SIGNED
- * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- */
-__kernel void softmax_layer_norm_quantized(
- TENSOR3D_DECLARATION(src),
- TENSOR3D_DECLARATION(sum),
- TENSOR3D_DECLARATION(dst))
-{
- Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
- Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
- Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(sum);
-
- // Load max value of 1D logits vector (row)
- int sum_val = *((__global int *)offset(&sum, 0, get_global_id(1)));
-
- // It will be better to calculate this in prev layer and pass here as parameter
- uint sum_val_u = convert_uint(sum_val);
- int headroom_plus_one = clz(sum_val_u);
- int num_bits_over_unit = EXP_ACCUMULATION_INT_BITS - headroom_plus_one;
- int shifted_sum_minus_one_1 = convert_int((sum_val_u << headroom_plus_one) - (1u << 31));
- int16 shifted_sum_minus_one = shifted_sum_minus_one_1;
- int16 shifted_scale = ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1(shifted_sum_minus_one, 16);
-
- // It was already calculated in prev layer, should be stored into tmp output and reused
- int16 data_diff = vload16(0, (__global int *)offset(&src, 0, 0));
- int16 data_diff_mult = data_diff;
-#if defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT)
- if(INPUT_BETA_MULTIPLIER > 1)
- {
- data_diff_mult = ASYMM_MULT(data_diff * (1 << INPUT_BETA_LEFT_SHIFT), INPUT_BETA_MULTIPLIER, 16);
- }
-#endif /* defined(INPUT_BETA_MULTIPLIER) && defined(INPUT_BETA_LEFT_SHIFT) */
-
- int16 data = ASYMM_EXP_ON_NEGATIVE_VALUES(data_diff_mult, SCALED_DIFF_INT_BITS, 16);
- data = ASYMM_MULT(shifted_scale, data, 16);
- data = ASYMM_ROUNDING_DIVIDE_BY_POW2(data, num_bits_over_unit + 31 - 8, 16);
-#ifdef QASYMM8_SIGNED
- data = ADD_OP(data, (int16)(MIN_VALUE), int, 16);
-#endif /* QASYMM8_SIGNED */
- data = select(MIN_VALUE, data, data_diff >= (int16)(DIFF_MIN));
- vstore16(CONVERT_SAT(data, VEC_DATA_TYPE(DATA_TYPE, 16)), 0, (__global DATA_TYPE *)offset(&dst, 0, 0));
-}
-
-#endif /* defined(DIFF_MIN) */
+#endif // #if defined(SRC_WIDTH) && defined(LOG_VECTOR_SIZE)
+#endif /* defined(DATA_TYPE) && defined(DIFF_MIN) && defined(VECTOR_SIZE) && defined(VECTOR_SIZE_LEFTOVER) && defined(MIN_VALUE) */