Lines Matching +full:- +full:a
1 // Copyright (c) 2017-2026, Lawrence Livermore National Security, LLC and other CEED contributors.
2 // All Rights Reserved. See the top-level LICENSE and NOTICE files for details.
4 // SPDX-License-Identifier: BSD-2-Clause
19 // c += a * b
21 #define fmadd(c, a, b) (c) = _mm256_fmadd_pd((a), (b), (c)) argument
23 #define fmadd(c, a, b) (c) += _mm256_mul_pd((a), (b)) argument
31 // c += a * b
33 #define fmadd(c, a, b) (c) = _mm_fmadd_ps((a), (b), (c)) argument
35 #define fmadd(c, a, b) (c) += _mm_mul_ps((a), (b)) argument
39 //------------------------------------------------------------------------------
41 //------------------------------------------------------------------------------
42 static inline int CeedTensorContract_Avx_Blocked(CeedTensorContract contract, CeedInt A, CeedInt B,… in CeedTensorContract_Avx_Blocked() argument
52 for (CeedInt a = 0; a < A; a++) { in CeedTensorContract_Avx_Blocked() local
58 … for (CeedInt cc = 0; cc < CC / 4; cc++) vv[jj][cc] = loadu(&v[(a * J + j + jj) * C + c + cc * 4]); in CeedTensorContract_Avx_Blocked()
64 fmadd(vv[jj][cc], tqv, loadu(&u[(a * B + b) * C + c + cc * 4])); in CeedTensorContract_Avx_Blocked()
69 … for (CeedInt cc = 0; cc < CC / 4; cc++) storeu(&v[(a * J + j + jj) * C + c + cc * 4], vv[jj][cc]); in CeedTensorContract_Avx_Blocked()
80 for (CeedInt jj = 0; jj < J - j; jj++) { in CeedTensorContract_Avx_Blocked()
81 … for (CeedInt cc = 0; cc < CC / 4; cc++) vv[jj][cc] = loadu(&v[(a * J + j + jj) * C + c + cc * 4]); in CeedTensorContract_Avx_Blocked()
84 for (CeedInt jj = 0; jj < J - j; jj++) { // doesn't unroll in CeedTensorContract_Avx_Blocked()
88 fmadd(vv[jj][cc], tqv, loadu(&u[(a * B + b) * C + c + cc * 4])); in CeedTensorContract_Avx_Blocked()
92 for (CeedInt jj = 0; jj < J - j; jj++) { in CeedTensorContract_Avx_Blocked()
93 … for (CeedInt cc = 0; cc < CC / 4; cc++) storeu(&v[(a * J + j + jj) * C + c + cc * 4], vv[jj][cc]); in CeedTensorContract_Avx_Blocked()
101 //------------------------------------------------------------------------------
103 //------------------------------------------------------------------------------
104 static inline int CeedTensorContract_Avx_Remainder(CeedTensorContract contract, CeedInt A, CeedInt … in CeedTensorContract_Avx_Remainder() argument
114 const CeedInt J_break = J % JJ ? (J / JJ) * JJ : (J / JJ - 1) * JJ; in CeedTensorContract_Avx_Remainder()
116 for (CeedInt a = 0; a < A; a++) { in CeedTensorContract_Avx_Remainder() local
123 for (CeedInt jj = 0; jj < JJ; jj++) vv[jj] = loadu(&v[(a * J + j + jj) * C + c]); in CeedTensorContract_Avx_Remainder()
127 if (C - c == 1) tqu = set(0.0, 0.0, 0.0, u[(a * B + b) * C + c + 0]); in CeedTensorContract_Avx_Remainder()
128 … else if (C - c == 2) tqu = set(0.0, 0.0, u[(a * B + b) * C + c + 1], u[(a * B + b) * C + c + 0]); in CeedTensorContract_Avx_Remainder()
129 …else if (C - c == 3) tqu = set(0.0, u[(a * B + b) * C + c + 2], u[(a * B + b) * C + c + 1], u[(a *… in CeedTensorContract_Avx_Remainder()
130 else tqu = loadu(&u[(a * B + b) * C + c]); in CeedTensorContract_Avx_Remainder()
135 for (CeedInt jj = 0; jj < JJ; jj++) storeu(&v[(a * J + j + jj) * C + c], vv[jj]); in CeedTensorContract_Avx_Remainder()
143 … for (CeedInt c = (C / CC) * CC; c < C; c++) v[(a * J + j) * C + c] += tq * u[(a * B + b) * C + c]; in CeedTensorContract_Avx_Remainder()
150 //------------------------------------------------------------------------------
152 //------------------------------------------------------------------------------
153 static inline int CeedTensorContract_Avx_Single(CeedTensorContract contract, CeedInt A, CeedInt B, … in CeedTensorContract_Avx_Single() argument
164 for (CeedInt a = 0; a < (A / AA) * AA; a += AA) { in CeedTensorContract_Avx_Single() local
169 for (CeedInt jj = 0; jj < JJ / 4; jj++) vv[aa][jj] = loadu(&v[(a + aa) * J + j + jj * 4]); in CeedTensorContract_Avx_Single()
177 fmadd(vv[aa][jj], tqv, set1(u[(a + aa) * B + b])); in CeedTensorContract_Avx_Single()
182 for (CeedInt jj = 0; jj < JJ / 4; jj++) storeu(&v[(a + aa) * J + j + jj * 4], vv[aa][jj]); in CeedTensorContract_Avx_Single()
187 const CeedInt a = (A / AA) * AA; in CeedTensorContract_Avx_Single() local
192 for (CeedInt aa = 0; aa < A - a; aa++) { in CeedTensorContract_Avx_Single()
193 for (CeedInt jj = 0; jj < JJ / 4; jj++) vv[aa][jj] = loadu(&v[(a + aa) * J + j + jj * 4]); in CeedTensorContract_Avx_Single()
200 for (CeedInt aa = 0; aa < A - a; aa++) { // unroll in CeedTensorContract_Avx_Single()
201 fmadd(vv[aa][jj], tqv, set1(u[(a + aa) * B + b])); in CeedTensorContract_Avx_Single()
205 for (CeedInt aa = 0; aa < A - a; aa++) { in CeedTensorContract_Avx_Single()
206 for (CeedInt jj = 0; jj < JJ / 4; jj++) storeu(&v[(a + aa) * J + j + jj * 4], vv[aa][jj]); in CeedTensorContract_Avx_Single()
210 const CeedInt A_break = A % AA ? (A / AA) * AA : (A / AA - 1) * AA; in CeedTensorContract_Avx_Single()
215 for (CeedInt a = 0; a < A_break; a += AA) { in CeedTensorContract_Avx_Single() local
218 for (CeedInt aa = 0; aa < AA; aa++) vv[aa] = loadu(&v[(a + aa) * J + j]); in CeedTensorContract_Avx_Single()
222 if (J - j == 1) { in CeedTensorContract_Avx_Single()
224 } else if (J - j == 2) { in CeedTensorContract_Avx_Single()
226 } else if (J - 3 == j) { in CeedTensorContract_Avx_Single()
234 fmadd(vv[aa], tqv, set1(u[(a + aa) * B + b])); in CeedTensorContract_Avx_Single()
237 for (CeedInt aa = 0; aa < AA; aa++) storeu(&v[(a + aa) * J + j], vv[aa]); in CeedTensorContract_Avx_Single()
245 for (CeedInt a = A_break; a < A; a++) v[a * J + j] += tq * u[a * B + b]; in CeedTensorContract_Avx_Single() local
251 //------------------------------------------------------------------------------
252 // Tensor Contract - Common Sizes
253 //------------------------------------------------------------------------------
254 static int CeedTensorContract_Avx_Blocked_4_8(CeedTensorContract contract, CeedInt A, CeedInt B, Ce… in CeedTensorContract_Avx_Blocked_4_8() argument
256 return CeedTensorContract_Avx_Blocked(contract, A, B, C, J, t, t_mode, add, u, v, 4, 8); in CeedTensorContract_Avx_Blocked_4_8()
258 static int CeedTensorContract_Avx_Remainder_8_8(CeedTensorContract contract, CeedInt A, CeedInt B, … in CeedTensorContract_Avx_Remainder_8_8() argument
260 return CeedTensorContract_Avx_Remainder(contract, A, B, C, J, t, t_mode, add, u, v, 8, 8); in CeedTensorContract_Avx_Remainder_8_8()
262 static int CeedTensorContract_Avx_Single_4_8(CeedTensorContract contract, CeedInt A, CeedInt B, Cee… in CeedTensorContract_Avx_Single_4_8() argument
264 return CeedTensorContract_Avx_Single(contract, A, B, C, J, t, t_mode, add, u, v, 4, 8); in CeedTensorContract_Avx_Single_4_8()
267 //------------------------------------------------------------------------------
269 //------------------------------------------------------------------------------
270 static int CeedTensorContractApply_Avx(CeedTensorContract contract, CeedInt A, CeedInt B, CeedInt C… in CeedTensorContractApply_Avx() argument
275 for (CeedInt q = 0; q < A * J * C; q++) v[q] = (CeedScalar)0.0; in CeedTensorContractApply_Avx()
280 CeedTensorContract_Avx_Single_4_8(contract, A, B, C, J, t, t_mode, true, u, v); in CeedTensorContractApply_Avx()
283 …if (C >= blk_size) CeedTensorContract_Avx_Blocked_4_8(contract, A, B, C, J, t, t_mode, true, u, v); in CeedTensorContractApply_Avx()
285 …if (C % blk_size) CeedTensorContract_Avx_Remainder_8_8(contract, A, B, C, J, t, t_mode, true, u, v… in CeedTensorContractApply_Avx()
290 //------------------------------------------------------------------------------
292 //------------------------------------------------------------------------------
298 //------------------------------------------------------------------------------