COMPMID-3239: Implement QSYMM16 LayerNormalizationKernel for CL
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: Ib1577c4a9aa29293a903731b2a4083b0d2243e1e
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2994
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index b707ec8..d6f415b 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -2277,7 +2277,7 @@
* @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)
* @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor
- * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8
+ * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QSYMM16
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x dst_gx_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)
diff --git a/src/core/CL/cl_kernels/helpers_asymm.h b/src/core/CL/cl_kernels/helpers_asymm.h
index 6377dba..37eb246 100644
--- a/src/core/CL/cl_kernels/helpers_asymm.h
+++ b/src/core/CL/cl_kernels/helpers_asymm.h
@@ -145,9 +145,19 @@
b_64 = convert_long##size(b); \
VEC_DATA_TYPE(long, size) \
ab_64 = a_64 * b_64; \
- /* COMPMID-907 */ \
+ /* Revert COMPMID-907 */ \
+ VEC_DATA_TYPE(long, size) \
+ mask1 = 1 << 30; \
+ VEC_DATA_TYPE(long, size) \
+ mask2 = 1 - (1 << 30); \
+ VEC_DATA_TYPE(long, size) \
+ is_positive_or_zero = ab_64 >= 0; \
+ VEC_DATA_TYPE(long, size) \
+ nudge = select(mask2, mask1, is_positive_or_zero); \
+ VEC_DATA_TYPE(long, size) \
+ mask = 1ll << 31; \
VEC_DATA_TYPE(int, size) \
- ab_x2_high32 = convert_int##size(((ab_64 + (1 << 30)) >> 31)); \
+ ab_x2_high32 = convert_int##size((ab_64 + nudge) / mask); \
return select(ab_x2_high32, INT_MAX, overflow); \
}
@@ -397,6 +407,15 @@
#define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b)
#define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) asymm_rescale##size(value, src_integer_bits, dst_integer_bits)
+#define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size) \
+ inline VEC_DATA_TYPE(int, size) multiply_by_quantized_multiplier##size(VEC_DATA_TYPE(int, size) input, int qmul, int shift) \
+ { \
+ const int left_shift = shift > 0 ? shift : 0; \
+ const int right_shift = shift > 0 ? 0 : -shift; \
+ return ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(input * (1 << left_shift), qmul, size), right_shift, size); \
+ }
+#define MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, size) multiply_by_quantized_multiplier##size(input, qmul, shift)
+
QUANTIZE_IMPL(uchar, 1)
QUANTIZE_IMPL(char, 1)
QUANTIZE_IMPL(uint, 1)
@@ -442,16 +461,19 @@
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(8)
ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(16)
+ASYMM_SELECT_USING_MASK_IMPL(1)
ASYMM_SELECT_USING_MASK_IMPL(2)
ASYMM_SELECT_USING_MASK_IMPL(4)
ASYMM_SELECT_USING_MASK_IMPL(8)
ASYMM_SELECT_USING_MASK_IMPL(16)
+ASYMM_MASK_IF_ZERO_IMPL(1)
ASYMM_MASK_IF_ZERO_IMPL(2)
ASYMM_MASK_IF_ZERO_IMPL(4)
ASYMM_MASK_IF_ZERO_IMPL(8)
ASYMM_MASK_IF_ZERO_IMPL(16)
+ASYMM_MASK_IF_NON_ZERO_IMPL(1)
ASYMM_MASK_IF_NON_ZERO_IMPL(2)
ASYMM_MASK_IF_NON_ZERO_IMPL(4)
ASYMM_MASK_IF_NON_ZERO_IMPL(8)
@@ -467,6 +489,7 @@
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(8)
ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(16)
+ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(1)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(2)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(4)
ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(8)
@@ -482,9 +505,16 @@
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(8)
ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(16)
+ASYMM_RESCALE_IMPL(1)
ASYMM_RESCALE_IMPL(2)
ASYMM_RESCALE_IMPL(4)
ASYMM_RESCALE_IMPL(8)
ASYMM_RESCALE_IMPL(16)
+MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(1)
+MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(2)
+MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(4)
+MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(8)
+MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(16)
+
#endif // ARM_COMPUTE_HELPERS_ASYMM_H
diff --git a/src/core/CL/cl_kernels/qlstm_layer_normalization.cl b/src/core/CL/cl_kernels/qlstm_layer_normalization.cl
new file mode 100644
index 0000000..08f0b53
--- /dev/null
+++ b/src/core/CL/cl_kernels/qlstm_layer_normalization.cl
@@ -0,0 +1,260 @@
+/*
+ * Copyright (c) 2020 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "helpers_asymm.h"
+
+#if VEC_SIZE == 2
+#define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 2)
+#define PERFORM_REDUCTION_IMPL(type) \
+ inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 2) sum) \
+ { \
+ sum.s0 += sum.s1; \
+ return sum.s0; \
+ }
+#elif VEC_SIZE == 4
+#define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 4)
+#define PERFORM_REDUCTION_IMPL(type) \
+ inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 4) sum) \
+ { \
+ sum.s01 += sum.s23; \
+ sum.s0 += sum.s1; \
+ return sum.s0; \
+ }
+#elif VEC_SIZE == 8
+#define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 8)
+#define PERFORM_REDUCTION_IMPL(type) \
+ inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 8) sum) \
+ { \
+ sum.s0123 += sum.s4567; \
+ sum.s01 += sum.s23; \
+ sum.s0 += sum.s1; \
+ return sum.s0; \
+ }
+#else /* VEC_SIZE DEFAULT */
+#define VEC_SIZE 16
+#define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 16)
+#define PERFORM_REDUCTION_IMPL(type) \
+ inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 16) sum) \
+ { \
+ sum.s01234567 += sum.s89abcdef; \
+ sum.s0123 += sum.s4567; \
+ sum.s01 += sum.s23; \
+ sum.s0 += sum.s1; \
+ return sum.s0; \
+ }
+#endif /* VEC_SIZE END */
+
+#define PERFORM_REDUCTION_STR(input, type) perform_reduction_##type(input)
+#define PERFORM_REDUCTION(input, type) PERFORM_REDUCTION_STR(input, type)
+
+PERFORM_REDUCTION_IMPL(int)
+PERFORM_REDUCTION_IMPL(long)
+
+/** Compute quantized multiplier and shift for the inverse square root of input.
+ * Using 3-bit fixed point and 5 iteration of Newton-Raphson method.
+ *
+ * @param[in] in Input to use
+ * @param[in] reverse_shift -1 to reverse the shift direction
+ *
+ * @return:
+ * .s0 Quantized multiplier for inverse square root
+ * .s1 Shift for inverse square root
+ *
+ */
+inline int2 get_invsqrt_quantized_multiplier_exp(int in, int reverse_shift)
+{
+ int2 stddev_inv;
+ int stddev_inv_multiplier = INT_MAX;
+ int stddev_inv_shift = 0;
+ int input = in;
+ if(input <= 1)
+ {
+ stddev_inv.s0 = stddev_inv_multiplier;
+ stddev_inv.s1 = stddev_inv_shift;
+ return stddev_inv;
+ }
+
+ stddev_inv_shift = 11;
+ while(input >= (1 << 29))
+ {
+ input /= 4;
+ ++stddev_inv_shift;
+ }
+
+ const unsigned int max_left_shift_bits = clz(input) - 1;
+ const unsigned int max_left_shift_bits_pairs = max_left_shift_bits / 2;
+ const unsigned int left_shift_bit_pairs = max_left_shift_bits_pairs - 1;
+ stddev_inv_shift -= left_shift_bit_pairs;
+ input <<= 2 * left_shift_bit_pairs;
+
+ typedef int FixedPointRawType;
+ const unsigned int fixedpoint_position = 3;
+ const unsigned int fixedpoint_int_position = sizeof(FixedPointRawType) * 8 - 1 - fixedpoint_position;
+ typedef FixedPointRawType FixedPoint3;
+ typedef FixedPointRawType FixedPoint0;
+
+ const FixedPoint3 fixedpoint_input = (input >> 1);
+ const FixedPoint3 fixedpoint_half_input = ASYMM_ROUNDING_DIVIDE_BY_POW2(fixedpoint_input, 1, 1);
+ const FixedPoint3 fixedpoint_half_three = (0x1 << fixedpoint_int_position) + (0x1 << (fixedpoint_int_position - 1));
+ FixedPoint3 x = 0x1 << fixedpoint_int_position;
+
+ const int num_iteration = 5;
+ for(int i = 0; i < num_iteration; i++)
+ {
+ int x3 = ASYMM_RESCALE(ASYMM_MULT(ASYMM_MULT(x, x, 1), x, 1), 9, fixedpoint_position, 1);
+ x = ASYMM_RESCALE(ASYMM_MULT(fixedpoint_half_three, x, 1) - ASYMM_MULT(fixedpoint_half_input, x3, 1), 6, fixedpoint_position, 1);
+ }
+ const FixedPoint0 fixedpoint_half_sqrt_2 = 1518500250;
+ x = ASYMM_MULT(fixedpoint_half_sqrt_2, x, 1);
+ stddev_inv_multiplier = x;
+ if(stddev_inv_shift < 0)
+ {
+ stddev_inv_multiplier <<= -stddev_inv_shift;
+ stddev_inv_shift = 0;
+ }
+ stddev_inv_shift *= reverse_shift;
+
+ stddev_inv.s0 = stddev_inv_multiplier;
+ stddev_inv.s1 = stddev_inv_shift;
+ return stddev_inv;
+}
+
+#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(WIDTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
+/** This function implements QLSTM layer normalization.
+ *
+ * @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 compile flag, e.g. -DDATA_TYPE=float
+ * @attention Width of the input tensor should be passed using the -DWIDTH compile flag, e.g. -DWIDTH=16
+ *
+ * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QSYMM16
+ * @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_offset_first_element_in_bytes The offset of the first element in the first source tensor
+ * @param[in] weight_ptr Pointer to the weight tensor. Supported data type: same as @p input_ptr
+ * @param[in] weight_stride_x Stride of the weight tensor in X dimension (in bytes)
+ * @param[in] weight_step_x weight_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weight_offset_first_element_in_bytes The offset of the first element in the weight tensor
+ * @param[in] bias_ptr Pointer to the bias tensor. Supported data type: S32
+ * @param[in] bias_stride_x Stride of the bias tensor in X dimension (in bytes)
+ * @param[in] bias_step_x bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] bias_offset_first_element_in_bytes The offset of the first element in the biases tensor
+ * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+__kernel void qlstm_layer_normalization(
+ IMAGE_DECLARATION(input),
+ VECTOR_DECLARATION(weight),
+ VECTOR_DECLARATION(bias),
+ IMAGE_DECLARATION(output))
+{
+ // Get pixels pointer
+ Image input = CONVERT_TO_IMAGE_STRUCT(input);
+ Vector weight = CONVERT_TO_VECTOR_STRUCT(weight);
+ Vector bias = CONVERT_TO_VECTOR_STRUCT(bias);
+ Image output = CONVERT_TO_IMAGE_STRUCT(output);
+
+ VEC_DATA_TYPE(int, VEC_SIZE)
+ sum = 0;
+ VEC_DATA_TYPE(long, VEC_SIZE)
+ sum_sq = 0;
+ // Calculate partial sum
+ int i = 0;
+ for(; i <= (WIDTH - VEC_SIZE); i += VEC_SIZE)
+ {
+ // Load data
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)offset(&input, i, 0));
+
+ sum += CONVERT(data, VEC_DATA_TYPE(int, VEC_SIZE));
+ sum_sq += CONVERT(data, VEC_DATA_TYPE(long, VEC_SIZE)) * CONVERT(data, VEC_DATA_TYPE(long, VEC_SIZE));
+ }
+ // Perform reduction
+ sum.s0 = PERFORM_REDUCTION(sum, int);
+ sum_sq.s0 = PERFORM_REDUCTION(sum_sq, long);
+
+ // Left-overs loop
+ for(; i < WIDTH; ++i)
+ {
+ DATA_TYPE data = *((__global DATA_TYPE *)offset(&input, i, 0));
+
+ sum.s0 += CONVERT(data, int);
+ sum_sq.s0 += CONVERT(data, long) * CONVERT(data, long);
+ }
+
+ int temp = 0x100000 / WIDTH;
+ int mean = (int)(sum.s0 * 1024 / WIDTH);
+ int var2 = ((sum_sq.s0 * (long)temp) - ((long)mean * (long)mean)) / 0x100000;
+ int2 stddev_inv = get_invsqrt_quantized_multiplier_exp(var2, -1);
+
+ i = 0;
+ for(; i <= (WIDTH - VEC_SIZE); i += VEC_SIZE)
+ {
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)offset(&input, i, 0));
+ VEC_DATA_TYPE(int, VEC_SIZE)
+ res = CONVERT(data, VEC_DATA_TYPE(int, VEC_SIZE)) * 1024 - mean;
+ res = multiply_by_quantized_multiplier(res, stddev_inv.s0, stddev_inv.s1);
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ w = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)vector_offset(&weight, i));
+ res = res * CONVERT(w, VEC_DATA_TYPE(int, VEC_SIZE));
+ res = res + VLOAD(VEC_SIZE)(0, (__global int *)vector_offset(&bias, i));
+ // Due to different rounding scheme, we might need to revisit in the future: res = select(res - 512, res + 512, res > 0) / 1024;
+ res = (res + 512) >> 10;
+ res = multiply_by_quantized_multiplier(res, OUTPUT_MULTIPLIER, OUTPUT_SHIFT + 12);
+#if defined(MIN_BOUND)
+ res = max(res, (VEC_DATA_TYPE(int, VEC_SIZE))MIN_BOUND);
+#endif // defined(MIN_BOUND)
+#if defined(MAX_BOUND)
+ res = min(res, (VEC_DATA_TYPE(int, VEC_SIZE))MAX_BOUND);
+#endif // defined(MAX_BOUND)
+ VSTORE(VEC_SIZE)
+ (CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), 0, (__global DATA_TYPE *)offset(&output, i, 0));
+ }
+ for(; i < WIDTH; ++i)
+ {
+ DATA_TYPE data = *((__global DATA_TYPE *)offset(&input, i, 0));
+ int res = (int)data * 1024 - mean;
+ res = MULTIPLY_BY_QUANTIZED_MULTIPLIER(res, stddev_inv.s0, stddev_inv.s1, 1);
+ DATA_TYPE w = *((__global DATA_TYPE *)vector_offset(&weight, i));
+ res = res * (int)w;
+ int b = *((__global int *)vector_offset(&bias, i));
+ res = res + b;
+ // Due to different rounding scheme, we might need to revisit in the future: res = select(res - 512, res + 512, res > 0) / 1024;
+ res = (res + 512) >> 10;
+ res = MULTIPLY_BY_QUANTIZED_MULTIPLIER(res, OUTPUT_MULTIPLIER, OUTPUT_SHIFT + 12, 1);
+#if defined(MIN_BOUND)
+ res = max(res, MIN_BOUND);
+#endif // defined(MIN_BOUND)
+#if defined(MAX_BOUND)
+ res = min(res, MAX_BOUND);
+#endif // defined(MAX_BOUND)
+ *((__global DATA_TYPE *)offset(&output, i, 0)) = (DATA_TYPE)res;
+ }
+}
+#endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(WIDTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) */
\ No newline at end of file