|
1583 | 1583 | }
|
1584 | 1584 |
|
1585 | 1585 | BENCHMARK_GEMM(f32_gemm_minmax_ukernel_5x64__asm_amd64_avx512f_broadcast)
|
| 1586 | + |
| 1587 | + static void f32_gemm_minmax_ukernel_1x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1588 | + GEMMBenchmark(state, |
| 1589 | + xnn_f32_gemm_minmax_ukernel_1x16c2__asm_amd64_avx512f_broadcast, |
| 1590 | + xnn_init_f32_minmax_scalar_params, |
| 1591 | + xnn_pack_f32_gemm_goi_w, |
| 1592 | + /*mr=*/1, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1593 | + benchmark::utils::CheckAVX512F); |
| 1594 | + } |
| 1595 | + |
| 1596 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_1x16c2__asm_amd64_avx512f_broadcast) |
| 1597 | + |
| 1598 | + static void f32_gemm_minmax_ukernel_2x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1599 | + GEMMBenchmark(state, |
| 1600 | + xnn_f32_gemm_minmax_ukernel_2x16c2__asm_amd64_avx512f_broadcast, |
| 1601 | + xnn_init_f32_minmax_scalar_params, |
| 1602 | + xnn_pack_f32_gemm_goi_w, |
| 1603 | + /*mr=*/2, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1604 | + benchmark::utils::CheckAVX512F); |
| 1605 | + } |
| 1606 | + |
| 1607 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_2x16c2__asm_amd64_avx512f_broadcast) |
| 1608 | + |
| 1609 | + static void f32_gemm_minmax_ukernel_3x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1610 | + GEMMBenchmark(state, |
| 1611 | + xnn_f32_gemm_minmax_ukernel_3x16c2__asm_amd64_avx512f_broadcast, |
| 1612 | + xnn_init_f32_minmax_scalar_params, |
| 1613 | + xnn_pack_f32_gemm_goi_w, |
| 1614 | + /*mr=*/3, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1615 | + benchmark::utils::CheckAVX512F); |
| 1616 | + } |
| 1617 | + |
| 1618 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_3x16c2__asm_amd64_avx512f_broadcast) |
| 1619 | + |
| 1620 | + static void f32_gemm_minmax_ukernel_4x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1621 | + GEMMBenchmark(state, |
| 1622 | + xnn_f32_gemm_minmax_ukernel_4x16c2__asm_amd64_avx512f_broadcast, |
| 1623 | + xnn_init_f32_minmax_scalar_params, |
| 1624 | + xnn_pack_f32_gemm_goi_w, |
| 1625 | + /*mr=*/4, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1626 | + benchmark::utils::CheckAVX512F); |
| 1627 | + } |
| 1628 | + |
| 1629 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_4x16c2__asm_amd64_avx512f_broadcast) |
| 1630 | + |
| 1631 | + static void f32_gemm_minmax_ukernel_5x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1632 | + GEMMBenchmark(state, |
| 1633 | + xnn_f32_gemm_minmax_ukernel_5x16c2__asm_amd64_avx512f_broadcast, |
| 1634 | + xnn_init_f32_minmax_scalar_params, |
| 1635 | + xnn_pack_f32_gemm_goi_w, |
| 1636 | + /*mr=*/5, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1637 | + benchmark::utils::CheckAVX512F); |
| 1638 | + } |
| 1639 | + |
| 1640 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_5x16c2__asm_amd64_avx512f_broadcast) |
| 1641 | + |
| 1642 | + static void f32_gemm_minmax_ukernel_6x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1643 | + GEMMBenchmark(state, |
| 1644 | + xnn_f32_gemm_minmax_ukernel_6x16c2__asm_amd64_avx512f_broadcast, |
| 1645 | + xnn_init_f32_minmax_scalar_params, |
| 1646 | + xnn_pack_f32_gemm_goi_w, |
| 1647 | + /*mr=*/6, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1648 | + benchmark::utils::CheckAVX512F); |
| 1649 | + } |
| 1650 | + |
| 1651 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_6x16c2__asm_amd64_avx512f_broadcast) |
| 1652 | + |
| 1653 | + static void f32_gemm_minmax_ukernel_7x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1654 | + GEMMBenchmark(state, |
| 1655 | + xnn_f32_gemm_minmax_ukernel_7x16c2__asm_amd64_avx512f_broadcast, |
| 1656 | + xnn_init_f32_minmax_scalar_params, |
| 1657 | + xnn_pack_f32_gemm_goi_w, |
| 1658 | + /*mr=*/7, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1659 | + benchmark::utils::CheckAVX512F); |
| 1660 | + } |
| 1661 | + |
| 1662 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_7x16c2__asm_amd64_avx512f_broadcast) |
| 1663 | + |
| 1664 | + static void f32_gemm_minmax_ukernel_8x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1665 | + GEMMBenchmark(state, |
| 1666 | + xnn_f32_gemm_minmax_ukernel_8x16c2__asm_amd64_avx512f_broadcast, |
| 1667 | + xnn_init_f32_minmax_scalar_params, |
| 1668 | + xnn_pack_f32_gemm_goi_w, |
| 1669 | + /*mr=*/8, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1670 | + benchmark::utils::CheckAVX512F); |
| 1671 | + } |
| 1672 | + |
| 1673 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_8x16c2__asm_amd64_avx512f_broadcast) |
| 1674 | + |
| 1675 | + static void f32_gemm_minmax_ukernel_9x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1676 | + GEMMBenchmark(state, |
| 1677 | + xnn_f32_gemm_minmax_ukernel_9x16c2__asm_amd64_avx512f_broadcast, |
| 1678 | + xnn_init_f32_minmax_scalar_params, |
| 1679 | + xnn_pack_f32_gemm_goi_w, |
| 1680 | + /*mr=*/9, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1681 | + benchmark::utils::CheckAVX512F); |
| 1682 | + } |
| 1683 | + |
| 1684 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_9x16c2__asm_amd64_avx512f_broadcast) |
| 1685 | + |
| 1686 | + static void f32_gemm_minmax_ukernel_10x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1687 | + GEMMBenchmark(state, |
| 1688 | + xnn_f32_gemm_minmax_ukernel_10x16c2__asm_amd64_avx512f_broadcast, |
| 1689 | + xnn_init_f32_minmax_scalar_params, |
| 1690 | + xnn_pack_f32_gemm_goi_w, |
| 1691 | + /*mr=*/10, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1692 | + benchmark::utils::CheckAVX512F); |
| 1693 | + } |
| 1694 | + |
| 1695 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_10x16c2__asm_amd64_avx512f_broadcast) |
| 1696 | + |
| 1697 | + static void f32_gemm_minmax_ukernel_11x16c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1698 | + GEMMBenchmark(state, |
| 1699 | + xnn_f32_gemm_minmax_ukernel_11x16c2__asm_amd64_avx512f_broadcast, |
| 1700 | + xnn_init_f32_minmax_scalar_params, |
| 1701 | + xnn_pack_f32_gemm_goi_w, |
| 1702 | + /*mr=*/11, /*nr=*/16, /*kr=*/2, /*sr=*/1, |
| 1703 | + benchmark::utils::CheckAVX512F); |
| 1704 | + } |
| 1705 | + |
| 1706 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_11x16c2__asm_amd64_avx512f_broadcast) |
| 1707 | + |
| 1708 | + static void f32_gemm_minmax_ukernel_1x32c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1709 | + GEMMBenchmark(state, |
| 1710 | + xnn_f32_gemm_minmax_ukernel_1x32c2__asm_amd64_avx512f_broadcast, |
| 1711 | + xnn_init_f32_minmax_scalar_params, |
| 1712 | + xnn_pack_f32_gemm_goi_w, |
| 1713 | + /*mr=*/1, /*nr=*/32, /*kr=*/2, /*sr=*/1, |
| 1714 | + benchmark::utils::CheckAVX512F); |
| 1715 | + } |
| 1716 | + |
| 1717 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_1x32c2__asm_amd64_avx512f_broadcast) |
| 1718 | + |
| 1719 | + static void f32_gemm_minmax_ukernel_2x32c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1720 | + GEMMBenchmark(state, |
| 1721 | + xnn_f32_gemm_minmax_ukernel_2x32c2__asm_amd64_avx512f_broadcast, |
| 1722 | + xnn_init_f32_minmax_scalar_params, |
| 1723 | + xnn_pack_f32_gemm_goi_w, |
| 1724 | + /*mr=*/2, /*nr=*/32, /*kr=*/2, /*sr=*/1, |
| 1725 | + benchmark::utils::CheckAVX512F); |
| 1726 | + } |
| 1727 | + |
| 1728 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_2x32c2__asm_amd64_avx512f_broadcast) |
| 1729 | + |
| 1730 | + static void f32_gemm_minmax_ukernel_3x32c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1731 | + GEMMBenchmark(state, |
| 1732 | + xnn_f32_gemm_minmax_ukernel_3x32c2__asm_amd64_avx512f_broadcast, |
| 1733 | + xnn_init_f32_minmax_scalar_params, |
| 1734 | + xnn_pack_f32_gemm_goi_w, |
| 1735 | + /*mr=*/3, /*nr=*/32, /*kr=*/2, /*sr=*/1, |
| 1736 | + benchmark::utils::CheckAVX512F); |
| 1737 | + } |
| 1738 | + |
| 1739 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_3x32c2__asm_amd64_avx512f_broadcast) |
| 1740 | + |
| 1741 | + static void f32_gemm_minmax_ukernel_4x32c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1742 | + GEMMBenchmark(state, |
| 1743 | + xnn_f32_gemm_minmax_ukernel_4x32c2__asm_amd64_avx512f_broadcast, |
| 1744 | + xnn_init_f32_minmax_scalar_params, |
| 1745 | + xnn_pack_f32_gemm_goi_w, |
| 1746 | + /*mr=*/4, /*nr=*/32, /*kr=*/2, /*sr=*/1, |
| 1747 | + benchmark::utils::CheckAVX512F); |
| 1748 | + } |
| 1749 | + |
| 1750 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_4x32c2__asm_amd64_avx512f_broadcast) |
| 1751 | + |
| 1752 | + static void f32_gemm_minmax_ukernel_5x32c2__asm_amd64_avx512f_broadcast(benchmark::State& state, const char* net) { |
| 1753 | + GEMMBenchmark(state, |
| 1754 | + xnn_f32_gemm_minmax_ukernel_5x32c2__asm_amd64_avx512f_broadcast, |
| 1755 | + xnn_init_f32_minmax_scalar_params, |
| 1756 | + xnn_pack_f32_gemm_goi_w, |
| 1757 | + /*mr=*/5, /*nr=*/32, /*kr=*/2, /*sr=*/1, |
| 1758 | + benchmark::utils::CheckAVX512F); |
| 1759 | + } |
| 1760 | + |
| 1761 | + BENCHMARK_GEMM(f32_gemm_minmax_ukernel_5x32c2__asm_amd64_avx512f_broadcast) |
1586 | 1762 | #endif // XNN_ENABLE_AVX512F && XNN_ARCH_X86_64 && XNN_ENABLE_ASSEMBLY
|
1587 | 1763 |
|
1588 | 1764 |
|
|
0 commit comments