zoecarver

# SPDX-FileCopyrightText: (c) 2025 Tenstorrent AI ULC
#
# SPDX-License-Identifier: Apache-2.0

"""
Sweep benchmark: mcast matmul, balanced matmul, balanced matmul+relu vs ttnn.

  mcast:    A+B mcast both on dm_read, grid="auto" (works at any size)
  balanced: A on dm_read, B on dm_write (two-NoC), 1 block/core
  relu:     balanced matmul with fused relu on last K iteration

zoecarver

  tt-lang compute (M=8, N=8, K=8, 21ms)                                                                                                                               
                                          
  // Phase 1: Zero-fill 64 acc tiles (8x8 output block)                                                                                                               
  cb_reserve_back(cb2, 64);                                                                                                                                           
  for (m = 0; m < 8; m++)                    // M=8                                                                                                                   
    for (n = 0; n < 8; n += 4)               // N=8, subblock of 4                                                                                                    
      acquire; fill x4; commit; wait; pack x4; release;                                                                                           

zoecarver

# ===== RMSNorm per-row kernels =====

@ttl.compute()
def rmsnorm_compute_row(x_dfb, sc_dfb, ms_dfb, sq_dfb, red_dfb, acc_dfb,
                        bcast_dfb, rsq_dfb, out_dfb):
    sc = sc_dfb.pop_back_val()
    ms = ms_dfb.pop_back_val()

    # Pass 1: sum of squares
    x0 = x_dfb.pop_back_val()

zoecarver

@ttl.compute()
def compute(x_dfb, sc_dfb, ms_dfb, sq_dfb, red_dfb, acc_dfb, bcast_dfb, rsq_dfb, out_dfb):
    core_x, _ = ttl.node(dims=2)
    sc = sc_dfb.pop_back_val()
    ms = ms_dfb.pop_back_val()

    for local_t in range(tiles_per_core):
        tile_idx = core_x * tiles_per_core + local_t
        if tile_idx < seq_tiles:
            # Pass 1: sum of squares across dim tiles

zoecarver

@ttl.operation(grid="auto")
    def rmsnorm_kernel(x, scaler, mean_scale, out):
        grid_cols, _ = ttl.grid_size(dims=2)
        seq_tiles = x.shape[0] // TILE
        tiles_per_core = -(-seq_tiles // grid_cols)

        x_dfb = ttl.make_dataflow_buffer_like(x, shape=(1, 1), buffer_factor=2)
        sc_dfb = ttl.make_dataflow_buffer_like(scaler, shape=(1, 1), buffer_factor=1)
        ms_dfb = ttl.make_dataflow_buffer_like(mean_scale, shape=(1, 1), buffer_factor=1)
        sq_dfb = ttl.make_dataflow_buffer_like(x, shape=(1, 1), buffer_factor=2)

zoecarver

Input: 

# =============================================================================
# Tensor accessor page print
# =============================================================================


@ttl.kernel(grid=(1, 1))
def tensor_accessor_print(inp, out):
    inp_dfb = ttl.make_dataflow_buffer_like(inp, shape=(1, 1), buffer_factor=2)

zoecarver

"""
Cell-list molecular dynamics on Tenstorrent hardware using TT-Lang.

Full Ewald electrostatics: erfc-damped real-space (cell-list, TT kernel) +
u-series reciprocal-space (separable Gaussian convolution, TT kernel + host).
LJ short-range forces. Periodic boundary conditions.
On-device Verlet integration (f32) with bf16 force kernels.

Validated: 10K atoms, 10K steps, 1.1ms/step (non-rebuild), 12 min total.
"""

zoecarver

# Saturn Privacy Policy

**Last Updated:** January 2025

## Overview

Saturn is a voice assistant app that connects you to Google's Gemini AI and the internet while prioritizing your privacy. This policy explains how we handle your data.

**Note:** Saturn connects to the internet and processes content through third-party AI services, which may not be suitable for children.

zoecarver

*** with-deleted.txt	2021-04-22 11:20:57.000000000 -0700
--- without-deleted.txt	2021-04-22 11:21:38.000000000 -0700
***************
*** 1,42 ****
! /Users/zoecarver/Developer/llvm-source/llvm-project/libcxx/test/std/ranges/range.access/range.prim/size.compile.pass.cpp:199:3: error: call to deleted function call operator in type 'const std::ranges::__size::__fn'
    std::ranges::size(d);
    ^~~~~~~~~~~~~~~~~
- /Users/zoecarver/Developer/llvm-source/build/include/c++/v1/__ranges/size.h:105:10: note: candidate function [with auto:1 = InvalidMinusBeginEnd &] has been explicitly deleted
-   size_t operator()(auto &&) const = delete;
-          ^

zoecarver

import std_config

var SWIFT_LLVM_ODR_SAFE: Int32 { get }
extension swift {
  typealias SmallBitVector = llvm.SmallBitVector
  typealias StringLiteral = llvm.StringLiteral
  typealias StringRef = llvm.StringRef
  typealias Twine = llvm.Twine
  typealias APFloat = llvm.APFloat
  typealias APInt = llvm.APInt