Sheri Zhang | b18252d | 2020-04-07 11:04:57 +0100 | [diff] [blame] | 1 | /* |
Adnan AlSinan | 7075fe2 | 2021-07-05 13:12:52 +0100 | [diff] [blame^] | 2 | * Copyright (c) 2020-2021 Arm Limited. |
Sheri Zhang | b18252d | 2020-04-07 11:04:57 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "helpers_asymm.h" |
| 25 | |
| 26 | #if VEC_SIZE == 2 |
| 27 | #define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 2) |
| 28 | #define PERFORM_REDUCTION_IMPL(type) \ |
| 29 | inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 2) sum) \ |
| 30 | { \ |
| 31 | sum.s0 += sum.s1; \ |
| 32 | return sum.s0; \ |
| 33 | } |
| 34 | #elif VEC_SIZE == 4 |
| 35 | #define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 4) |
| 36 | #define PERFORM_REDUCTION_IMPL(type) \ |
| 37 | inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 4) sum) \ |
| 38 | { \ |
| 39 | sum.s01 += sum.s23; \ |
| 40 | sum.s0 += sum.s1; \ |
| 41 | return sum.s0; \ |
| 42 | } |
| 43 | #elif VEC_SIZE == 8 |
| 44 | #define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 8) |
| 45 | #define PERFORM_REDUCTION_IMPL(type) \ |
| 46 | inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 8) sum) \ |
| 47 | { \ |
| 48 | sum.s0123 += sum.s4567; \ |
| 49 | sum.s01 += sum.s23; \ |
| 50 | sum.s0 += sum.s1; \ |
| 51 | return sum.s0; \ |
| 52 | } |
| 53 | #else /* VEC_SIZE DEFAULT */ |
| 54 | #define VEC_SIZE 16 |
| 55 | #define multiply_by_quantized_multiplier(input, qmul, shift) MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, 16) |
| 56 | #define PERFORM_REDUCTION_IMPL(type) \ |
| 57 | inline VEC_DATA_TYPE(type, 1) perform_reduction_##type(VEC_DATA_TYPE(type, 16) sum) \ |
| 58 | { \ |
| 59 | sum.s01234567 += sum.s89abcdef; \ |
| 60 | sum.s0123 += sum.s4567; \ |
| 61 | sum.s01 += sum.s23; \ |
| 62 | sum.s0 += sum.s1; \ |
| 63 | return sum.s0; \ |
| 64 | } |
| 65 | #endif /* VEC_SIZE END */ |
| 66 | |
| 67 | #define PERFORM_REDUCTION_STR(input, type) perform_reduction_##type(input) |
| 68 | #define PERFORM_REDUCTION(input, type) PERFORM_REDUCTION_STR(input, type) |
| 69 | |
| 70 | PERFORM_REDUCTION_IMPL(int) |
| 71 | PERFORM_REDUCTION_IMPL(long) |
| 72 | |
| 73 | /** Compute quantized multiplier and shift for the inverse square root of input. |
| 74 | * Using 3-bit fixed point and 5 iteration of Newton-Raphson method. |
| 75 | * |
| 76 | * @param[in] in Input to use |
| 77 | * @param[in] reverse_shift -1 to reverse the shift direction |
| 78 | * |
| 79 | * @return: |
| 80 | * .s0 Quantized multiplier for inverse square root |
| 81 | * .s1 Shift for inverse square root |
| 82 | * |
| 83 | */ |
| 84 | inline int2 get_invsqrt_quantized_multiplier_exp(int in, int reverse_shift) |
| 85 | { |
| 86 | int2 stddev_inv; |
| 87 | int stddev_inv_multiplier = INT_MAX; |
| 88 | int stddev_inv_shift = 0; |
| 89 | int input = in; |
| 90 | if(input <= 1) |
| 91 | { |
| 92 | stddev_inv.s0 = stddev_inv_multiplier; |
| 93 | stddev_inv.s1 = stddev_inv_shift; |
| 94 | return stddev_inv; |
| 95 | } |
| 96 | |
| 97 | stddev_inv_shift = 11; |
| 98 | while(input >= (1 << 29)) |
| 99 | { |
| 100 | input /= 4; |
| 101 | ++stddev_inv_shift; |
| 102 | } |
| 103 | |
| 104 | const unsigned int max_left_shift_bits = clz(input) - 1; |
| 105 | const unsigned int max_left_shift_bits_pairs = max_left_shift_bits / 2; |
| 106 | const unsigned int left_shift_bit_pairs = max_left_shift_bits_pairs - 1; |
| 107 | stddev_inv_shift -= left_shift_bit_pairs; |
| 108 | input <<= 2 * left_shift_bit_pairs; |
| 109 | |
| 110 | typedef int FixedPointRawType; |
| 111 | const unsigned int fixedpoint_position = 3; |
| 112 | const unsigned int fixedpoint_int_position = sizeof(FixedPointRawType) * 8 - 1 - fixedpoint_position; |
| 113 | typedef FixedPointRawType FixedPoint3; |
| 114 | typedef FixedPointRawType FixedPoint0; |
| 115 | |
| 116 | const FixedPoint3 fixedpoint_input = (input >> 1); |
| 117 | const FixedPoint3 fixedpoint_half_input = ASYMM_ROUNDING_DIVIDE_BY_POW2(fixedpoint_input, 1, 1); |
| 118 | const FixedPoint3 fixedpoint_half_three = (0x1 << fixedpoint_int_position) + (0x1 << (fixedpoint_int_position - 1)); |
| 119 | FixedPoint3 x = 0x1 << fixedpoint_int_position; |
| 120 | |
| 121 | const int num_iteration = 5; |
| 122 | for(int i = 0; i < num_iteration; i++) |
| 123 | { |
| 124 | int x3 = ASYMM_RESCALE(ASYMM_MULT(ASYMM_MULT(x, x, 1), x, 1), 9, fixedpoint_position, 1); |
| 125 | x = ASYMM_RESCALE(ASYMM_MULT(fixedpoint_half_three, x, 1) - ASYMM_MULT(fixedpoint_half_input, x3, 1), 6, fixedpoint_position, 1); |
| 126 | } |
| 127 | const FixedPoint0 fixedpoint_half_sqrt_2 = 1518500250; |
| 128 | x = ASYMM_MULT(fixedpoint_half_sqrt_2, x, 1); |
| 129 | stddev_inv_multiplier = x; |
| 130 | if(stddev_inv_shift < 0) |
| 131 | { |
| 132 | stddev_inv_multiplier <<= -stddev_inv_shift; |
| 133 | stddev_inv_shift = 0; |
| 134 | } |
| 135 | stddev_inv_shift *= reverse_shift; |
| 136 | |
| 137 | stddev_inv.s0 = stddev_inv_multiplier; |
| 138 | stddev_inv.s1 = stddev_inv_shift; |
| 139 | return stddev_inv; |
| 140 | } |
| 141 | |
| 142 | #if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(WIDTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) |
| 143 | /** This function implements QLSTM layer normalization. |
| 144 | * |
| 145 | * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 |
| 146 | * @attention Data type should be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float |
| 147 | * @attention Width of the input tensor should be passed using the -DWIDTH compile flag, e.g. -DWIDTH=16 |
| 148 | * |
| 149 | * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QSYMM16 |
| 150 | * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 151 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 152 | * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 153 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 154 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 155 | * @param[in] weight_ptr Pointer to the weight tensor. Supported data type: same as @p input_ptr |
| 156 | * @param[in] weight_stride_x Stride of the weight tensor in X dimension (in bytes) |
| 157 | * @param[in] weight_step_x weight_stride_x * number of elements along X processed per workitem(in bytes) |
| 158 | * @param[in] weight_offset_first_element_in_bytes The offset of the first element in the weight tensor |
| 159 | * @param[in] bias_ptr Pointer to the bias tensor. Supported data type: S32 |
| 160 | * @param[in] bias_stride_x Stride of the bias tensor in X dimension (in bytes) |
| 161 | * @param[in] bias_step_x bias_stride_x * number of elements along X processed per workitem(in bytes) |
| 162 | * @param[in] bias_offset_first_element_in_bytes The offset of the first element in the biases tensor |
| 163 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 164 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 165 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 166 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 167 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 168 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 169 | */ |
| 170 | __kernel void qlstm_layer_normalization( |
| 171 | IMAGE_DECLARATION(input), |
| 172 | VECTOR_DECLARATION(weight), |
| 173 | VECTOR_DECLARATION(bias), |
| 174 | IMAGE_DECLARATION(output)) |
| 175 | { |
| 176 | // Get pixels pointer |
| 177 | Image input = CONVERT_TO_IMAGE_STRUCT(input); |
| 178 | Vector weight = CONVERT_TO_VECTOR_STRUCT(weight); |
| 179 | Vector bias = CONVERT_TO_VECTOR_STRUCT(bias); |
| 180 | Image output = CONVERT_TO_IMAGE_STRUCT(output); |
| 181 | |
| 182 | VEC_DATA_TYPE(int, VEC_SIZE) |
| 183 | sum = 0; |
| 184 | VEC_DATA_TYPE(long, VEC_SIZE) |
| 185 | sum_sq = 0; |
| 186 | // Calculate partial sum |
| 187 | int i = 0; |
| 188 | for(; i <= (WIDTH - VEC_SIZE); i += VEC_SIZE) |
| 189 | { |
| 190 | // Load data |
| 191 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 192 | data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)offset(&input, i, 0)); |
| 193 | |
| 194 | sum += CONVERT(data, VEC_DATA_TYPE(int, VEC_SIZE)); |
| 195 | sum_sq += CONVERT(data, VEC_DATA_TYPE(long, VEC_SIZE)) * CONVERT(data, VEC_DATA_TYPE(long, VEC_SIZE)); |
| 196 | } |
| 197 | // Perform reduction |
| 198 | sum.s0 = PERFORM_REDUCTION(sum, int); |
| 199 | sum_sq.s0 = PERFORM_REDUCTION(sum_sq, long); |
| 200 | |
| 201 | // Left-overs loop |
| 202 | for(; i < WIDTH; ++i) |
| 203 | { |
| 204 | DATA_TYPE data = *((__global DATA_TYPE *)offset(&input, i, 0)); |
| 205 | |
| 206 | sum.s0 += CONVERT(data, int); |
| 207 | sum_sq.s0 += CONVERT(data, long) * CONVERT(data, long); |
| 208 | } |
| 209 | |
| 210 | int temp = 0x100000 / WIDTH; |
| 211 | int mean = (int)(sum.s0 * 1024 / WIDTH); |
| 212 | int var2 = ((sum_sq.s0 * (long)temp) - ((long)mean * (long)mean)) / 0x100000; |
| 213 | int2 stddev_inv = get_invsqrt_quantized_multiplier_exp(var2, -1); |
| 214 | |
| 215 | i = 0; |
| 216 | for(; i <= (WIDTH - VEC_SIZE); i += VEC_SIZE) |
| 217 | { |
| 218 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 219 | data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)offset(&input, i, 0)); |
| 220 | VEC_DATA_TYPE(int, VEC_SIZE) |
| 221 | res = CONVERT(data, VEC_DATA_TYPE(int, VEC_SIZE)) * 1024 - mean; |
| 222 | res = multiply_by_quantized_multiplier(res, stddev_inv.s0, stddev_inv.s1); |
| 223 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 224 | w = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)vector_offset(&weight, i)); |
| 225 | res = res * CONVERT(w, VEC_DATA_TYPE(int, VEC_SIZE)); |
| 226 | res = res + VLOAD(VEC_SIZE)(0, (__global int *)vector_offset(&bias, i)); |
| 227 | // Due to different rounding scheme, we might need to revisit in the future: res = select(res - 512, res + 512, res > 0) / 1024; |
| 228 | res = (res + 512) >> 10; |
| 229 | res = multiply_by_quantized_multiplier(res, OUTPUT_MULTIPLIER, OUTPUT_SHIFT + 12); |
| 230 | #if defined(MIN_BOUND) |
| 231 | res = max(res, (VEC_DATA_TYPE(int, VEC_SIZE))MIN_BOUND); |
| 232 | #endif // defined(MIN_BOUND) |
| 233 | #if defined(MAX_BOUND) |
| 234 | res = min(res, (VEC_DATA_TYPE(int, VEC_SIZE))MAX_BOUND); |
| 235 | #endif // defined(MAX_BOUND) |
| 236 | VSTORE(VEC_SIZE) |
| 237 | (CONVERT(res, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), 0, (__global DATA_TYPE *)offset(&output, i, 0)); |
| 238 | } |
| 239 | for(; i < WIDTH; ++i) |
| 240 | { |
| 241 | DATA_TYPE data = *((__global DATA_TYPE *)offset(&input, i, 0)); |
| 242 | int res = (int)data * 1024 - mean; |
| 243 | res = MULTIPLY_BY_QUANTIZED_MULTIPLIER(res, stddev_inv.s0, stddev_inv.s1, 1); |
| 244 | DATA_TYPE w = *((__global DATA_TYPE *)vector_offset(&weight, i)); |
| 245 | res = res * (int)w; |
| 246 | int b = *((__global int *)vector_offset(&bias, i)); |
| 247 | res = res + b; |
| 248 | // Due to different rounding scheme, we might need to revisit in the future: res = select(res - 512, res + 512, res > 0) / 1024; |
| 249 | res = (res + 512) >> 10; |
| 250 | res = MULTIPLY_BY_QUANTIZED_MULTIPLIER(res, OUTPUT_MULTIPLIER, OUTPUT_SHIFT + 12, 1); |
| 251 | #if defined(MIN_BOUND) |
| 252 | res = max(res, MIN_BOUND); |
| 253 | #endif // defined(MIN_BOUND) |
| 254 | #if defined(MAX_BOUND) |
| 255 | res = min(res, MAX_BOUND); |
| 256 | #endif // defined(MAX_BOUND) |
| 257 | *((__global DATA_TYPE *)offset(&output, i, 0)) = (DATA_TYPE)res; |
| 258 | } |
| 259 | } |
| 260 | #endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(WIDTH) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT) */ |