geohot

# -*- coding: utf-8 -*-
#
# TARGET arch is: ['-D__HIP_PLATFORM_AMD__', '-I/opt/rocm/include']
# WORD_SIZE is: 8
# POINTER_SIZE is: 8
# LONGDOUBLE_SIZE is: 16
#
import ctypes

geohot

# tiny@tiny9:~/tinygrad$ python3 examples/benchmark_copies.py
# CPU copy 6.18 ms, 16.28 GB/s
# GPU copy 4.38 ms, 23.00 GB/s
# GPU  6x  1.85 ms, 54.54 GB/s

import time
def timeit(fxn):
  tms = []
  for _ in range(10):
    st = time.perf_counter()

geohot

__kernel void matmul(__global float* data0, const __global float* data1, const __global float* data2) {
  int gidx0 = get_group_id(1); /* 512 */
  int gidx1 = get_group_id(0); /* 512 */
  float2 acc0 = (float2)(0.0f,0.0f);
  float2 acc1 = (float2)(0.0f,0.0f);
  for (int ridx0 = 0; ridx0 < 512; ++ridx0) {
    float2 val0 = (float2)(*((__global float2*)(data1+(gidx0*2048)+(ridx0*2))));
    float2 val1 = (float2)(*((__global float2*)(data1+(gidx0*2048)+(ridx0*2)+1024)));
    float2 val2 = (float2)(*((__global float2*)(data2+(gidx1*2)+(ridx0*2048))));
    float2 val3 = (float2)(*((__global float2*)(data2+(gidx1*2)+(ridx0*2048)+1024)));

geohot

***    0 E_64_32_6_6n5                         arg   2 sz [64, 1, 1]         [32, 1, 1]   OPs     33M/   0.00G  mem  3.07 GB tm      3.20us/     0.00ms (10483.20 GFLOPS,  297.02 GB/s)                                                                                                                                                                                                        
***    1 r_128_31_31_3_2_3_2_2_2_8n26          arg   3 sz [31, 31, 128]      [2, 3, 1]    OPs    283M/   0.03G  mem  3.07 GB tm    218.44us/     0.22ms ( 1297.42 GFLOPS,  216.24 GB/s)                                                                                                                                                                                                        
***    2 r_1024_32_16_2_3_4_4_8n6              arg   3 sz [32, 1024, 1]      [2, 16, 1]   OPs    805M/   0.32G  mem  3.07 GB tm     64.68us/     0.29ms (12450.43 GFLOPS, 1426.62 GB/s)                                                                         

geohot

import os
import sys
import time
import torch
import torch.distributed as dist
import torch.multiprocessing as mp

def all_reduce_latency(nbytes, rank):
  buf = torch.randn(nbytes // 4).cuda(rank)

geohot

# given a number in the range [amin, amax] (inclusive)
# what are the min and max of that number after modding it by b?

# c style modulus
def modn(a, b): return -((-a)%b) if a < 0 else a%b

# aka a fast version of
def slow_modrange(amin, amax, b):
  values = [modn(rv, b) for rv in range(amin, amax+1)]
  return min(values), max(values)

geohot

import torch
torch.set_grad_enabled(False)
model = torch.nn.Linear(1, 1, bias=False).cuda()
model.weight[:] = 1.
print(model(torch.Tensor([2349.]).cuda()))

geohot

# instructions to install openpilot on a pixel 3

# enter fastboot with power + volume down
# make sure bootloader is unlocked
# make sure modern version of android platform tools is installed
mkdir pixel
wget https://dl.google.com/dl/android/aosp/blueline-pq3a.190801.002-factory-f3d66c49.zip
unzip blueline-pq3a.190801.002-factory-f3d66c49.zip
cd blueline-pq3a.190801.002/
./flash-all.sh

geohot

%pylab inline
%load_ext autoreload
%autoreload 2

from tools.lib.route import Route
from tools.lib.logreader import LogReader
r,num = Route("ce2fbd370f78ef21|2020-11-27--16-27-28"),10
#r,num = Route("f66032c2b5aa18ac|2020-12-04--09-33-54"),30
alr = []
for n in range(num-1, num+5):

geohot

diff --git a/clang/include/clang/AST/Stmt.h b/clang/include/clang/AST/Stmt.h
index 13f265223..61b0a83c6 100644
--- a/clang/include/clang/AST/Stmt.h
+++ b/clang/include/clang/AST/Stmt.h
@@ -2459,13 +2459,16 @@ class ForStmt : public Stmt {
 public:
   ForStmt(const ASTContext &C, Stmt *Init, Expr *Cond, VarDecl *condVar,
           Expr *Inc, Stmt *Body, SourceLocation FL, SourceLocation LP,
-          SourceLocation RP);
+          SourceLocation RP, bool is_fore_statement=false);