Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2023 Arm Limited. |
| 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 | */ |
Mohammed Suhail Munshi | a2bb80e | 2023-06-19 14:57:57 +0100 | [diff] [blame^] | 24 | #include "activation_float_helpers.h" |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 25 | #include "helpers.h" |
| 26 | #include "tile_helpers.h" |
| 27 | |
| 28 | #if defined(MAT_MUL_NATIVE_QUANTIZED_NT_NT) |
| 29 | /** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS non-transposed, RHS non-transposed - buffer only |
| 30 | * |
| 31 | * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it |
| 32 | * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension |
| 33 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=uchar) |
| 34 | * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). |
| 35 | * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 36 | * @note The fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output with the relu and bounded relu operations. |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 37 | * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) |
| 38 | * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_QUANTIZED_NT_NT) |
| 39 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 40 | * - M0 > 0 |
| 41 | * - N0 = 1, 2, 3, 4, 8, 16 |
| 42 | * - K0 = 1, 2, 3, 4, 8, 16 |
| 43 | * @note Values > 8 for M0 are not expected to be efficient |
| 44 | * |
| 45 | * @param[in] lhs_ptr Pointer to the lhs matrix. Supported data types: QASYMM8_SIGNED/QASYMM8 |
| 46 | * @param[in] lhs_stride_y Stride of the lhs matrix in Y (2nd) dimension (in bytes) |
| 47 | * @param[in] lhs_stride_z Stride of the lhs tensor in Z (3rd) dimension (in bytes) |
| 48 | * @param[in] lhs_w The width of the lhs tensor |
| 49 | * @param[in] lhs_h The height of the lhs tensor |
| 50 | * @param[in] lhs_n Number of the matrices (buffers) in the batch |
| 51 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix |
| 52 | * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr |
| 53 | * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) |
| 54 | * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) |
| 55 | * @param[in] rhs_w The width of the rhs tensor |
| 56 | * @param[in] rhs_h The height of the rhs tensor |
| 57 | * @param[in] rhs_n Number of the matrices (buffers) in the batch |
| 58 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix |
| 59 | * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr |
| 60 | * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) |
| 61 | * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) |
| 62 | * @param[in] dst_w The width of the dst tensor |
| 63 | * @param[in] dst_h The height of the dst tensor |
| 64 | * @param[in] dst_n Number of the matrices (buffers) in the batch |
| 65 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the dst matrix |
| 66 | */ |
| 67 | __kernel void mat_mul_native_quantized_nt_nt( |
| 68 | TENSOR3D_T(lhs, BUFFER), |
| 69 | TENSOR3D_T(rhs, BUFFER), |
| 70 | TENSOR3D_T(dst, BUFFER)) |
| 71 | { |
| 72 | const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); |
| 73 | const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0); |
| 74 | const uint z = GET_SPATIAL_IDX(2, 1, 0); |
| 75 | |
| 76 | // Compute LHS/RHS/DST matrix address |
| 77 | lhs_offset_first_element_in_bytes += y * lhs_stride_y + z * lhs_stride_z; |
| 78 | rhs_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + z * rhs_stride_z; |
| 79 | dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; |
| 80 | |
| 81 | // Initialize the accumulators |
| 82 | TILE(int, M0, N0, acc); |
| 83 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 84 | { |
| 85 | acc[i].v = K * ((int)LHS_OFFSET) * ((int)RHS_OFFSET); |
| 86 | }) |
| 87 | |
| 88 | TILE(int, 1, N0, b_sum); |
| 89 | b_sum[0].v = 0; |
| 90 | |
| 91 | TILE(int, 1, M0, a_sum); |
| 92 | a_sum[0].v = 0; |
| 93 | |
| 94 | int k; |
| 95 | for(k = 0; k <= K - K0; k += K0) |
| 96 | { |
| 97 | TILE(DATA_TYPE, M0, K0, a); |
| 98 | TILE(DATA_TYPE, N0, K0, b); |
| 99 | |
| 100 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 101 | { |
| 102 | a[i].v = 0; |
| 103 | }) |
| 104 | |
| 105 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 106 | { |
| 107 | b[i].v = 0; |
| 108 | }) |
| 109 | |
| 110 | // Load tile from the lhs tensor |
| 111 | T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 112 | |
| 113 | // Load tile from the rhs tensor in a transposed fashion |
| 114 | // in order to use T_MMUL_NT_T macro because only this macro |
| 115 | // can utilize dot product instruction for Int8/UInt8 by |
| 116 | // directly multiplying the rows of Lhs and Rhs tensors. |
| 117 | T_LOAD_TRANSPOSED(DATA_TYPE, K0, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 118 | |
| 119 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, K0, NT, T, a, b, acc); |
| 120 | |
| 121 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 122 | { |
| 123 | LOOP_UNROLLING(int, j, 0, 1, K0, |
| 124 | { |
| 125 | a_sum[0].s[i] += (int)a[i].s[j]; |
| 126 | }) |
| 127 | }) |
| 128 | |
| 129 | LOOP_UNROLLING(int, i, 0, 1, K0, |
| 130 | { |
| 131 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 132 | { |
| 133 | b_sum[0].s[j] += (int)b[j].s[i]; |
| 134 | }) |
| 135 | }) |
| 136 | |
| 137 | lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE); |
| 138 | rhs_offset_first_element_in_bytes += K0 * rhs_stride_y; |
| 139 | } |
| 140 | |
| 141 | #if((K % K0) != 0) |
| 142 | /* Leftover Loop */ |
| 143 | for(; k < K; ++k) |
| 144 | { |
| 145 | TILE(DATA_TYPE, M0, 1, a); |
| 146 | TILE(DATA_TYPE, N0, 1, b); |
| 147 | |
| 148 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 149 | { |
| 150 | a[i].v = 0; |
| 151 | }) |
| 152 | |
| 153 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 154 | { |
| 155 | b[i].v = 0; |
| 156 | }) |
| 157 | |
| 158 | // Load tile from the lhs tensor |
| 159 | T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 160 | |
| 161 | // Load tile from the rhs tensor in a transposed fashion. |
| 162 | // See the main loop for more explanation |
| 163 | T_LOAD_TRANSPOSED(DATA_TYPE, 1, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 164 | |
| 165 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, 1, NT, T, a, b, acc); |
| 166 | |
| 167 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 168 | { |
| 169 | LOOP_UNROLLING(int, j, 0, 1, 1, |
| 170 | { |
| 171 | a_sum[0].s[i] += (int)a[i].s[j]; |
| 172 | }) |
| 173 | }) |
| 174 | |
| 175 | LOOP_UNROLLING(int, i, 0, 1, 1, |
| 176 | { |
| 177 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 178 | { |
| 179 | b_sum[0].s[j] += (int)b[j].s[i]; |
| 180 | }) |
| 181 | }) |
| 182 | |
| 183 | lhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE); |
| 184 | rhs_offset_first_element_in_bytes += 1 * rhs_stride_y; |
| 185 | } |
| 186 | #endif // ((K % K0) != 0) |
| 187 | |
| 188 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 189 | { |
| 190 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 191 | { |
Mohammed Suhail Munshi | a2bb80e | 2023-06-19 14:57:57 +0100 | [diff] [blame^] | 192 | acc[i].s[j] -= ((int)RHS_OFFSET) * a_sum[0].s[i] + ((int)(LHS_OFFSET)) * b_sum[0].s[j]; |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 193 | }) |
| 194 | }) |
| 195 | |
| 196 | const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0; |
| 197 | const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0; |
| 198 | |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 199 | T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc); |
| 200 | |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 201 | // Quantize the tile |
| 202 | TILE(DATA_TYPE, M0, N0, accq); |
| 203 | T_QUANTIZE8_ASYMMETRIC(int, DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, acc, accq); |
| 204 | |
| 205 | TILE(int, M0, 1, indirect_buffer); |
| 206 | LOOP_UNROLLING(int, _i, 0, 1, M0, |
| 207 | { |
| 208 | indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond)); |
| 209 | }); |
| 210 | |
| 211 | T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, accq, indirect_buffer); |
| 212 | } |
| 213 | #endif // defined(MAT_MUL_NATIVE_QUANTIZED_NT_NT) |
| 214 | |
Jakub Sujak | 5e99a3e | 2023-04-18 08:33:56 +0100 | [diff] [blame] | 215 | #if defined(MAT_MUL_NATIVE_QUANTIZED_NT_T) |
| 216 | /** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS non-transposed, RHS transposed - buffer only |
| 217 | * |
| 218 | * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it |
| 219 | * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension |
| 220 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=uchar) |
| 221 | * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). |
| 222 | * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 223 | * @note The fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output bounded activation functions. |
Jakub Sujak | 5e99a3e | 2023-04-18 08:33:56 +0100 | [diff] [blame] | 224 | * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) |
| 225 | * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_QUANTIZED_NT_T) |
| 226 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 227 | * - M0 > 0 |
| 228 | * - N0 = 1, 2, 3, 4, 8, 16 |
| 229 | * - K0 = 1, 2, 3, 4, 8, 16 |
| 230 | * @note Values > 8 for M0, N0, K0 are not expected to be efficient |
| 231 | * |
| 232 | * @param[in] lhs_ptr Pointer to the lhs matrix. Supported data types: QASYMM8/QASYMM8_SIGNED |
| 233 | * @param[in] lhs_stride_y Stride of the lhs matrix in Y (2nd) dimension (in bytes) |
| 234 | * @param[in] lhs_stride_z Stride of the lhs tensor in Z (3rd) dimension (in bytes) |
| 235 | * @param[in] lhs_w The width of the lhs tensor |
| 236 | * @param[in] lhs_h The height of the lhs tensor |
| 237 | * @param[in] lhs_n Number of the matrices (buffers) in the batch |
| 238 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix |
| 239 | * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr |
| 240 | * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) |
| 241 | * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) |
| 242 | * @param[in] rhs_w The width of the rhs tensor |
| 243 | * @param[in] rhs_h The height of the rhs tensor |
| 244 | * @param[in] rhs_n Number of the matrices (buffers) in the batch |
| 245 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix |
| 246 | * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr |
| 247 | * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) |
| 248 | * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) |
| 249 | * @param[in] dst_w The width of the dst tensor |
| 250 | * @param[in] dst_h The height of the dst tensor |
| 251 | * @param[in] dst_n Number of the matrices (buffers) in the batch |
| 252 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the dst matrix |
| 253 | */ |
| 254 | __kernel void mat_mul_native_quantized_nt_t( |
| 255 | TENSOR3D_T(lhs, BUFFER), |
| 256 | TENSOR3D_T(rhs, BUFFER), |
| 257 | TENSOR3D_T(dst, BUFFER)) |
| 258 | { |
| 259 | const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); |
| 260 | const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0); |
| 261 | const uint z = GET_SPATIAL_IDX(2, 1, 0); |
| 262 | |
| 263 | // Compute LHS/RHS/DST matrix address |
| 264 | lhs_offset_first_element_in_bytes += y * lhs_stride_y + z * lhs_stride_z; |
| 265 | rhs_offset_first_element_in_bytes += x * rhs_stride_y + z * rhs_stride_z; |
| 266 | dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; |
| 267 | |
| 268 | // Initialize the accumulators |
| 269 | TILE(int, M0, N0, acc); |
| 270 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 271 | { |
| 272 | acc[i].v = K * ((int)LHS_OFFSET) * ((int)RHS_OFFSET); |
| 273 | }) |
| 274 | |
| 275 | TILE(int, 1, M0, a_sum); |
| 276 | a_sum[0].v = 0; |
| 277 | |
| 278 | TILE(int, 1, N0, b_sum); |
| 279 | b_sum[0].v = 0; |
| 280 | |
| 281 | int k; |
| 282 | for(k = 0; k <= K - K0; k += K0) |
| 283 | { |
| 284 | TILE(DATA_TYPE, M0, K0, a); |
| 285 | TILE(DATA_TYPE, N0, K0, b); |
| 286 | |
| 287 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 288 | { |
| 289 | a[i].v = 0; |
| 290 | }) |
| 291 | |
| 292 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 293 | { |
| 294 | b[i].v = 0; |
| 295 | }) |
| 296 | |
| 297 | // Load tile from lhs/rhs tensors |
| 298 | T_LOAD(DATA_TYPE, M0, K0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 299 | T_LOAD(DATA_TYPE, N0, K0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 300 | |
| 301 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, K0, NT, T, a, b, acc); |
| 302 | |
| 303 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 304 | { |
| 305 | LOOP_UNROLLING(int, j, 0, 1, K0, |
| 306 | { |
| 307 | a_sum[0].s[i] += (int)a[i].s[j]; |
| 308 | }) |
| 309 | }) |
| 310 | |
| 311 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 312 | { |
| 313 | LOOP_UNROLLING(int, j, 0, 1, K0, |
| 314 | { |
| 315 | b_sum[0].s[i] += (int)b[i].s[j]; |
| 316 | }) |
| 317 | }) |
| 318 | |
| 319 | lhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE); |
| 320 | rhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE); |
| 321 | } |
| 322 | |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 323 | #if((K % K0) != 0) |
Jakub Sujak | 5e99a3e | 2023-04-18 08:33:56 +0100 | [diff] [blame] | 324 | // Leftover loop |
| 325 | for(; k < K; ++k) |
| 326 | { |
| 327 | TILE(DATA_TYPE, M0, 1, a); |
| 328 | TILE(DATA_TYPE, N0, 1, b); |
| 329 | |
| 330 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 331 | { |
| 332 | a[i].v = 0; |
| 333 | }) |
| 334 | |
| 335 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 336 | { |
| 337 | b[i].v = 0; |
| 338 | }) |
| 339 | |
| 340 | // Load tile from lhs/rhs tensors |
| 341 | T_LOAD(DATA_TYPE, M0, 1, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 342 | T_LOAD(DATA_TYPE, N0, 1, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 343 | |
| 344 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, 1, NT, T, a, b, acc); |
| 345 | |
| 346 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 347 | { |
| 348 | LOOP_UNROLLING(int, j, 0, 1, 1, |
| 349 | { |
| 350 | a_sum[0].s[i] += (int)a[i].s[j]; |
| 351 | }) |
| 352 | }) |
| 353 | |
| 354 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 355 | { |
| 356 | LOOP_UNROLLING(int, j, 0, 1, 1, |
| 357 | { |
| 358 | b_sum[0].s[i] += (int)b[i].s[j]; |
| 359 | }) |
| 360 | }) |
| 361 | |
| 362 | lhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE); |
| 363 | rhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE); |
| 364 | } |
| 365 | #endif // ((K % K0) != 0) |
| 366 | |
| 367 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 368 | { |
| 369 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 370 | { |
Mohammed Suhail Munshi | a2bb80e | 2023-06-19 14:57:57 +0100 | [diff] [blame^] | 371 | acc[i].s[j] -= ((int)(RHS_OFFSET)) * a_sum[0].s[i] + ((int)(LHS_OFFSET)) * b_sum[0].s[j]; |
Jakub Sujak | 5e99a3e | 2023-04-18 08:33:56 +0100 | [diff] [blame] | 372 | }) |
| 373 | }) |
| 374 | |
| 375 | const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0; |
| 376 | const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0; |
| 377 | |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 378 | T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc); |
| 379 | |
Jakub Sujak | 5e99a3e | 2023-04-18 08:33:56 +0100 | [diff] [blame] | 380 | // Quantize the tile |
| 381 | TILE(DATA_TYPE, M0, N0, accq); |
| 382 | T_QUANTIZE8_ASYMMETRIC(int, DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, acc, accq); |
| 383 | |
| 384 | TILE(int, M0, 1, indirect_buffer); |
| 385 | LOOP_UNROLLING(int, _i, 0, 1, M0, |
| 386 | { |
| 387 | indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond)); |
| 388 | }); |
| 389 | |
| 390 | T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, accq, indirect_buffer); |
| 391 | } |
| 392 | #endif // defined(MAT_MUL_NATIVE_QUANTIZED_NT_T) |
| 393 | |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 394 | #if defined(MAT_MUL_NATIVE_QUANTIZED_T_NT) |
| 395 | /** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS non-transposed |
| 396 | * |
| 397 | * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it |
| 398 | * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension |
| 399 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=uchar) |
| 400 | * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). |
| 401 | * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 402 | * @note The fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output with the relu and bounded relu operations. |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 403 | * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) |
| 404 | * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_QUANTIZED_T_NT) |
| 405 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 406 | * - M0 > 0 |
| 407 | * - N0 = 1, 2, 3, 4, 8, 16 |
| 408 | * - K0 = 1, 2, 3, 4, 8, 16 |
| 409 | * @note Values > 8 for M0, N0 and K0 are not expected to be efficient |
| 410 | * |
| 411 | * @param[in] lhs_ptr Pointer to the lhs matrix. Supported data types: QASYMM8/QASYMM8_SIGNED |
| 412 | * @param[in] lhs_stride_y Stride of the lhs matrix in Y (2nd) dimension (in bytes) |
| 413 | * @param[in] lhs_stride_z Stride of the lhs tensor in Z (3rd) dimension (in bytes) |
| 414 | * @param[in] lhs_w The width of the lhs tensor |
| 415 | * @param[in] lhs_h The height of the lhs tensor |
| 416 | * @param[in] lhs_n Number of the matrices (buffers) in the batch |
| 417 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix |
| 418 | * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr |
| 419 | * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) |
| 420 | * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) |
| 421 | * @param[in] rhs_w The width of the rhs tensor |
| 422 | * @param[in] rhs_h The height of the rhs tensor |
| 423 | * @param[in] rhs_n Number of the matrices (buffers) in the batch |
| 424 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix |
| 425 | * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr |
| 426 | * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) |
| 427 | * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) |
| 428 | * @param[in] dst_w The width of the dst tensor |
| 429 | * @param[in] dst_h The height of the dst tensor |
| 430 | * @param[in] dst_n Number of the matrices (buffers) in the batch |
| 431 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the dst matrix |
| 432 | */ |
| 433 | __kernel void mat_mul_native_quantized_t_nt( |
| 434 | TENSOR3D_T(lhs, BUFFER), |
| 435 | TENSOR3D_T(rhs, BUFFER), |
| 436 | TENSOR3D_T(dst, BUFFER)) |
| 437 | { |
| 438 | const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); |
| 439 | const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0); |
| 440 | const uint z = GET_SPATIAL_IDX(2, 1, 0); |
| 441 | |
| 442 | // Compute LHS/RHS/DST matrix address |
| 443 | lhs_offset_first_element_in_bytes += y * sizeof(DATA_TYPE) + z * lhs_stride_z; |
| 444 | rhs_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + z * rhs_stride_z; |
| 445 | dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; |
| 446 | |
| 447 | // Initialize the accumulators |
| 448 | TILE(int, M0, N0, acc); |
| 449 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 450 | { |
| 451 | acc[i].v = K * ((int)LHS_OFFSET) * ((int)RHS_OFFSET); |
| 452 | }) |
| 453 | |
| 454 | TILE(int, 1, N0, b_sum); |
| 455 | b_sum[0].v = 0; |
| 456 | |
| 457 | TILE(int, 1, M0, a_sum); |
| 458 | a_sum[0].v = 0; |
| 459 | |
| 460 | int k; |
| 461 | for(k = 0; k <= K - K0; k += K0) |
| 462 | { |
| 463 | TILE(DATA_TYPE, M0, K0, a); |
| 464 | TILE(DATA_TYPE, N0, K0, b); |
| 465 | |
| 466 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 467 | { |
| 468 | a[i].v = 0; |
| 469 | }) |
| 470 | |
| 471 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 472 | { |
| 473 | b[i].v = 0; |
| 474 | }) |
| 475 | |
| 476 | // Load tile from the lhs/rhs tensors in a transposed fashion |
| 477 | // see mat_mul_native_quantized_nt_nt main loop for more explanation |
| 478 | T_LOAD_TRANSPOSED(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 479 | T_LOAD_TRANSPOSED(DATA_TYPE, K0, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 480 | |
| 481 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, K0, NT, T, a, b, acc); |
| 482 | |
| 483 | LOOP_UNROLLING(int, i, 0, 1, K0, |
| 484 | { |
| 485 | LOOP_UNROLLING(int, j, 0, 1, M0, |
| 486 | { |
| 487 | a_sum[0].s[j] += (int)a[j].s[i]; |
| 488 | }) |
| 489 | }) |
| 490 | |
| 491 | LOOP_UNROLLING(int, i, 0, 1, K0, |
| 492 | { |
| 493 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 494 | { |
| 495 | b_sum[0].s[j] += (int)b[j].s[i]; |
| 496 | }) |
| 497 | }) |
| 498 | |
| 499 | lhs_offset_first_element_in_bytes += K0 * lhs_stride_y; |
| 500 | rhs_offset_first_element_in_bytes += K0 * rhs_stride_y; |
| 501 | } |
| 502 | |
| 503 | #if((K % K0) != 0) |
| 504 | /* Leftover Loop */ |
| 505 | for(; k < K; ++k) |
| 506 | { |
| 507 | TILE(DATA_TYPE, M0, 1, a); |
| 508 | TILE(DATA_TYPE, N0, 1, b); |
| 509 | |
| 510 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 511 | { |
| 512 | a[i].v = 0; |
| 513 | }) |
| 514 | |
| 515 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 516 | { |
| 517 | b[i].v = 0; |
| 518 | }) |
| 519 | |
| 520 | // Load tile from the lhs/rhs tensors in a transposed fashion |
| 521 | // see mat_mul_native_quantized_nt_nt main loop for more explanation |
| 522 | T_LOAD_TRANSPOSED(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 523 | T_LOAD_TRANSPOSED(DATA_TYPE, 1, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 524 | |
| 525 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, 1, NT, T, a, b, acc); |
| 526 | |
| 527 | LOOP_UNROLLING(int, i, 0, 1, 1, |
| 528 | { |
| 529 | LOOP_UNROLLING(int, j, 0, 1, M0, |
| 530 | { |
| 531 | a_sum[0].s[j] += (int)a[j].s[i]; |
| 532 | }) |
| 533 | }) |
| 534 | |
| 535 | LOOP_UNROLLING(int, i, 0, 1, 1, |
| 536 | { |
| 537 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 538 | { |
| 539 | b_sum[0].s[j] += (int)b[j].s[i]; |
| 540 | }) |
| 541 | }) |
| 542 | |
| 543 | lhs_offset_first_element_in_bytes += 1 * lhs_stride_y; |
| 544 | rhs_offset_first_element_in_bytes += 1 * rhs_stride_y; |
| 545 | } |
| 546 | #endif // ((K % K0) != 0) |
| 547 | |
| 548 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 549 | { |
| 550 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 551 | { |
Mohammed Suhail Munshi | a2bb80e | 2023-06-19 14:57:57 +0100 | [diff] [blame^] | 552 | acc[i].s[j] -= ((int)(RHS_OFFSET)) * a_sum[0].s[i] + ((int)(LHS_OFFSET)) * b_sum[0].s[j]; |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 553 | }) |
| 554 | }) |
| 555 | |
| 556 | const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0; |
| 557 | const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0; |
| 558 | |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 559 | T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc); |
| 560 | |
Gunes Bayir | 9d0c4de | 2023-04-13 18:22:58 +0100 | [diff] [blame] | 561 | // Quantize the tile |
| 562 | TILE(DATA_TYPE, M0, N0, accq); |
| 563 | T_QUANTIZE8_ASYMMETRIC(int, DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, acc, accq); |
| 564 | |
| 565 | TILE(int, M0, 1, indirect_buffer); |
| 566 | LOOP_UNROLLING(int, _i, 0, 1, M0, |
| 567 | { |
| 568 | indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond)); |
| 569 | }); |
| 570 | |
| 571 | T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, accq, indirect_buffer); |
| 572 | } |
| 573 | #endif // defined(MAT_MUL_NATIVE_QUANTIZED_T_NT) |
Omar Al Khatib | 467daef | 2023-04-13 14:56:23 +0100 | [diff] [blame] | 574 | |
| 575 | #if defined(MAT_MUL_NATIVE_QUANTIZED_T_T) |
| 576 | /** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS transposed |
| 577 | * |
| 578 | * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it |
| 579 | * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension |
| 580 | * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=uchar) |
| 581 | * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). |
| 582 | * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 583 | * @note The fused activation function used should be passed with -DACTIVATION_TYPE, -DA_VAL and -DB_VAL are used for min and max output with the relu and bounded relu operations. |
Omar Al Khatib | 467daef | 2023-04-13 14:56:23 +0100 | [diff] [blame] | 584 | * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) |
| 585 | * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_QUANTIZED_T_T) |
| 586 | * @note Only the following configurations of M0, N0 and K0 are currently supported: |
| 587 | * - M0 = 1, 2, 3, 4, 8, 16 |
| 588 | * - N0 = 1, 2, 3, 4, 8, 16 |
| 589 | * - K0 = 1, 2, 3, 4, 8, 16 |
| 590 | * @note Values > 8 for M0, N0 and K0 are not expected to be efficient |
| 591 | * |
| 592 | * @param[in] lhs_ptr Pointer to the lhs matrix. Supported data types: QASYMM8/QASYMM8_SIGNED |
| 593 | * @param[in] lhs_stride_y Stride of the lhs matrix in Y (2nd) dimension (in bytes) |
| 594 | * @param[in] lhs_stride_z Stride of the lhs tensor in Z (3rd) dimension (in bytes) |
| 595 | * @param[in] lhs_w The width of the lhs tensor |
| 596 | * @param[in] lhs_h The height of the lhs tensor |
| 597 | * @param[in] lhs_n Number of the matrices (buffers) in the batch |
| 598 | * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix |
| 599 | * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr |
| 600 | * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) |
| 601 | * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) |
| 602 | * @param[in] rhs_w The width of the rhs tensor |
| 603 | * @param[in] rhs_h The height of the rhs tensor |
| 604 | * @param[in] rhs_n Number of the matrices (buffers) in the batch |
| 605 | * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix |
| 606 | * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr |
| 607 | * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) |
| 608 | * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) |
| 609 | * @param[in] dst_w The width of the dst tensor |
| 610 | * @param[in] dst_h The height of the dst tensor |
| 611 | * @param[in] dst_n Number of the matrices (buffers) in the batch |
| 612 | * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the dst matrix |
| 613 | */ |
| 614 | __kernel void mat_mul_native_quantized_t_t( |
| 615 | TENSOR3D_T(lhs, BUFFER), |
| 616 | TENSOR3D_T(rhs, BUFFER), |
| 617 | TENSOR3D_T(dst, BUFFER)) |
| 618 | { |
| 619 | const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); |
| 620 | const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0); |
| 621 | const uint z = GET_SPATIAL_IDX(2, 1, 0); |
| 622 | |
| 623 | // Compute LHS/RHS/DST matrix address |
| 624 | lhs_offset_first_element_in_bytes += y * sizeof(DATA_TYPE) + z * lhs_stride_z; |
| 625 | rhs_offset_first_element_in_bytes += x * rhs_stride_y + z * rhs_stride_z; |
| 626 | dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; |
| 627 | |
| 628 | // Initialize the accumulators |
| 629 | TILE(int, M0, N0, acc); |
| 630 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 631 | { |
| 632 | acc[i].v = K * ((int)LHS_OFFSET) * ((int)RHS_OFFSET); |
| 633 | }) |
| 634 | |
| 635 | TILE(int, 1, M0, a_sum); |
| 636 | a_sum[0].v = 0; |
| 637 | |
| 638 | TILE(int, 1, N0, b_sum); |
| 639 | b_sum[0].v = 0; |
| 640 | |
| 641 | int k; |
| 642 | for(k = 0; k <= K - K0; k += K0) |
| 643 | { |
| 644 | TILE(DATA_TYPE, M0, K0, a); |
| 645 | TILE(DATA_TYPE, N0, K0, b); |
| 646 | |
| 647 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 648 | { |
| 649 | a[i].v = 0; |
| 650 | }) |
| 651 | |
| 652 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 653 | { |
| 654 | b[i].v = 0; |
| 655 | }) |
| 656 | |
| 657 | // Load tile from the lhs tensor in a transposed fashion |
| 658 | // see mat_mul_native_quantized_nt_nt main loop for more explanation |
| 659 | T_LOAD_TRANSPOSED(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 660 | |
| 661 | // Load tile from the rhs tensor |
| 662 | T_LOAD(DATA_TYPE, N0, K0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 663 | |
| 664 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, K0, NT, T, a, b, acc); |
| 665 | |
| 666 | LOOP_UNROLLING(int, i, 0, 1, K0, |
| 667 | { |
| 668 | LOOP_UNROLLING(int, j, 0, 1, M0, |
| 669 | { |
| 670 | a_sum[0].s[j] += (int)a[j].s[i]; |
| 671 | }) |
| 672 | }) |
| 673 | |
| 674 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 675 | { |
| 676 | LOOP_UNROLLING(int, j, 0, 1, K0, |
| 677 | { |
| 678 | b_sum[0].s[i] += (int)b[i].s[j]; |
| 679 | }) |
| 680 | }) |
| 681 | |
| 682 | lhs_offset_first_element_in_bytes += K0 * lhs_stride_y; |
| 683 | rhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE); |
| 684 | } |
| 685 | |
| 686 | #if((K % K0) != 0) |
| 687 | /* Leftover Loop */ |
| 688 | for(; k < K; ++k) |
| 689 | { |
| 690 | TILE(DATA_TYPE, M0, 1, a); |
| 691 | TILE(DATA_TYPE, N0, 1, b); |
| 692 | |
| 693 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 694 | { |
| 695 | a[i].v = 0; |
| 696 | }) |
| 697 | |
| 698 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 699 | { |
| 700 | b[i].v = 0; |
| 701 | }) |
| 702 | |
| 703 | // Load tile from the lhs tensor in a transposed fashion |
| 704 | // see mat_mul_native_quantized_nt_nt main loop for more explanation |
| 705 | T_LOAD_TRANSPOSED(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); |
| 706 | |
| 707 | // Load tile from the rhs tensor |
| 708 | T_LOAD(DATA_TYPE, N0, 1, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); |
| 709 | |
| 710 | T_MMUL(DATA_TYPE, DATA_TYPE, int, M0, N0, 1, NT, T, a, b, acc); |
| 711 | |
| 712 | LOOP_UNROLLING(int, i, 0, 1, 1, |
| 713 | { |
| 714 | LOOP_UNROLLING(int, j, 0, 1, M0, |
| 715 | { |
| 716 | a_sum[0].s[j] += (int)a[j].s[i]; |
| 717 | }) |
| 718 | }) |
| 719 | |
| 720 | LOOP_UNROLLING(int, i, 0, 1, N0, |
| 721 | { |
| 722 | LOOP_UNROLLING(int, j, 0, 1, 1, |
| 723 | { |
| 724 | b_sum[0].s[i] += (int)b[i].s[j]; |
| 725 | }) |
| 726 | }) |
| 727 | |
| 728 | lhs_offset_first_element_in_bytes += 1 * lhs_stride_y; |
| 729 | rhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE); |
| 730 | } |
| 731 | #endif // ((K % K0) != 0) |
| 732 | |
| 733 | LOOP_UNROLLING(int, i, 0, 1, M0, |
| 734 | { |
| 735 | LOOP_UNROLLING(int, j, 0, 1, N0, |
| 736 | { |
Mohammed Suhail Munshi | a2bb80e | 2023-06-19 14:57:57 +0100 | [diff] [blame^] | 737 | acc[i].s[j] -= ((int)RHS_OFFSET) * a_sum[0].s[i] + ((int)(LHS_OFFSET)) * b_sum[0].s[j]; |
Omar Al Khatib | 467daef | 2023-04-13 14:56:23 +0100 | [diff] [blame] | 738 | }) |
| 739 | }) |
| 740 | |
| 741 | const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0; |
| 742 | const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0; |
| 743 | |
| 744 | // Quantize the tile |
Mohammed Suhail Munshi | 94abde4 | 2023-05-25 16:48:43 +0100 | [diff] [blame] | 745 | T_ACTIVATION(int, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, acc, acc); |
| 746 | |
Omar Al Khatib | 467daef | 2023-04-13 14:56:23 +0100 | [diff] [blame] | 747 | TILE(DATA_TYPE, M0, N0, accq); |
| 748 | T_QUANTIZE8_ASYMMETRIC(int, DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, acc, accq); |
| 749 | |
| 750 | TILE(int, M0, 1, indirect_buffer); |
| 751 | LOOP_UNROLLING(int, _i, 0, 1, M0, |
| 752 | { |
| 753 | indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond)); |
| 754 | }); |
| 755 | |
| 756 | T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, accq, indirect_buffer); |
| 757 | } |
| 758 | #endif // defined(MAT_MUL_NATIVE_QUANTIZED_T_T) |