| /external/ruy/ruy/ |
| D | pack_arm.h | 59 int packed_zero_point, std::int32_t* sums_ptr,
69 std::int8_t* packed_ptr, std::int32_t* sums_ptr,
76 std::int32_t* sums_ptr, int input_xor);
82 std::int32_t* sums_ptr, int input_xor);
87 std::int32_t* sums_ptr, int input_xor);
93 int packed_stride, std::int32_t* sums_ptr,
102 const std::int32_t* sums_ptr;
115 const std::int32_t* sums_ptr,
124 params->sums_ptr = sums_ptr;
190 std::int32_t* sums_ptr = sums ? sums + block_col : nullptr;
[all …]
|
| D | pack_arm.cc | 38 std::int8_t* packed_ptr, std::int32_t* sums_ptr, in Pack8bitColMajorForNeon() argument
202 [packed_ptr] "+r"(packed_ptr), [sums_ptr] "+r"(sums_ptr) in Pack8bitColMajorForNeon()
238 static_assert(offsetof(Params, sums_ptr) == RUY_OFFSET_SUMS_PTR, ""); in CheckOffsetsInPackParams8bit()
634 std::int32_t* sums_ptr, int input_xor) { in Pack8bitColMajorForNeonA55ish() argument
828 [ packed_ptr ] "+r"(packed_ptr), [ sums_ptr ] "+r"(sums_ptr) in Pack8bitColMajorForNeonA55ish()
844 std::int32_t* sums_ptr, int input_xor) { in Pack8bitColMajorForNeonDotprodA55ish() argument
1075 [src_ptr3] "+r"(src_ptr3), [packed_ptr] "+r"(packed_ptr), [sums_ptr] "+r"(sums_ptr) in Pack8bitColMajorForNeonDotprodA55ish()
1091 std::int32_t* sums_ptr, int input_xor) { in Pack8bitColMajorForNeonDotprod() argument
1573 [packed_ptr] "+r"(packed_ptr), [sums_ptr] "+r"(sums_ptr) in Pack8bitColMajorForNeonDotprod()
1592 int packed_stride, std::int32_t* sums_ptr, in Pack8bitRowMajorForNeonDotprod() argument
[all …]
|
| D | pack_avx512.cc | 105 std::int8_t* packed_ptr, std::int32_t* sums_ptr,
167 if (sums_ptr) {
170 sums_ptr[i] = 0;
194 if (sums_ptr) {
437 if (sums_ptr) {
441 _mm256_loadu_si256(reinterpret_cast<const __m256i*>(sums_ptr));
453 _mm256_storeu_si256(reinterpret_cast<__m256i*>(sums_ptr), sums);
638 std::int32_t* sums_ptr) {
663 sums_ptr ? sums_ptr + kHalfLayoutCols : nullptr;
666 remaining_src_cols, src_rows, packed_ptr, sums_ptr,
[all …]
|
| D | pack_avx2_fma.cc | 69 int src_rows, std::int8_t* packed_ptr, std::int32_t* sums_ptr,
131 if (sums_ptr) {
134 sums_ptr[i] = 0;
156 if (sums_ptr) {
511 if (sums_ptr) {
515 _mm256_loadu_si256(reinterpret_cast<const __m256i*>(sums_ptr));
532 _mm256_storeu_si256(reinterpret_cast<__m256i*>(sums_ptr), sums);
548 std::int8_t* packed_ptr, std::int32_t* sums_ptr) {
568 sums_ptr, trailing_buf);
|
| D | pack_x86.h | 62 std::int8_t* packed_ptr, std::int32_t* sums_ptr);
89 std::int32_t* sums_ptr = sums ? sums + block_col : nullptr;
101 remaining_src_cols, src_matrix.layout.rows, packed_ptr, sums_ptr);
109 std::int8_t* packed_ptr, std::int32_t* sums_ptr);
136 std::int32_t* sums_ptr = sums ? sums + block_col : nullptr;
148 remaining_src_cols, src_matrix.layout.rows, packed_ptr, sums_ptr);
227 std::int8_t* packed_ptr, std::int32_t* sums_ptr);
257 std::int32_t* sums_ptr = sums ? sums + block_col : nullptr;
269 remaining_src_cols, src_matrix.layout.rows, packed_ptr, sums_ptr);
|
| D | pack_avx.cc | 219 std::int32_t* sums_ptr,
281 if (sums_ptr) {
284 sums_ptr[i] = 0;
306 if (sums_ptr) {
653 if (sums_ptr) {
657 _mm256_loadu_si256(reinterpret_cast<const __m256i*>(sums_ptr));
671 _mm256_storeu_si256(reinterpret_cast<__m256i*>(sums_ptr), sums);
690 std::int8_t* packed_ptr, std::int32_t* sums_ptr) {
710 sums_ptr, trailing_buf);
|
| /external/gemmlowp/internal/ |
| D | pack_neon.h | 278 std::int32_t* sums_ptr = dst->sums_of_each_slice() + start_width; in Pack() local
305 int32x4_t sum = vld1q_s32(sums_ptr); in Pack()
308 vst1q_s32(sums_ptr, sum); in Pack()
311 int32x2_t sum = vld1_s32(sums_ptr); in Pack()
315 vst1_s32(sums_ptr, sum); in Pack()
343 std::int32_t* sums_ptr = dst->sums_of_each_slice() + start_width; in Pack() local
366 int32x4_t sum = vld1q_s32(sums_ptr); in Pack()
369 vst1q_s32(sums_ptr, sum); in Pack()
372 int32x2_t sum = vld1_s32(sums_ptr); in Pack()
376 vst1_s32(sums_ptr, sum); in Pack()
|
| D | pack_msa.h | 372 std::int32_t* sums_ptr = dst->sums_of_each_slice() + start_width; in Pack() local
409 v4i32 sum = __builtin_msa_ld_w(const_cast<std::int32_t*>(sums_ptr), 0); in Pack()
415 __builtin_msa_st_w(sum, sums_ptr, 0); in Pack()
418 std::int32_t sum[2] = { sums_ptr[0], sums_ptr[1] }; in Pack()
422 sums_ptr[0] = sum[0]; in Pack()
423 sums_ptr[1] = sum[1]; in Pack()
|