forked from tinygrad/tinygrad
Compare commits
1
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
 geohot
|
ccd753e1aa |
+20
-41
@@ -30,6 +30,8 @@ jobs:
|
||||
key: llvm-speed
|
||||
deps: testing_minimal
|
||||
llvm: 'true'
|
||||
- name: External Benchmark Schedule
|
||||
run: python3 test/external/external_benchmark_schedule.py
|
||||
- name: Speed Test
|
||||
run: CPU=1 CPU_LLVM=1 python3 test/speed/external_test_speed_v_torch.py
|
||||
- name: Speed Test (BEAM=2)
|
||||
@@ -46,7 +48,7 @@ jobs:
|
||||
uses: ./.github/actions/setup-tinygrad
|
||||
with:
|
||||
deps: docs
|
||||
pydeps: "capstone torch"
|
||||
pydeps: "capstone"
|
||||
- name: Build wheel and show size
|
||||
run: |
|
||||
pip install build
|
||||
@@ -77,8 +79,6 @@ jobs:
|
||||
run: |
|
||||
python docs/abstractions2.py
|
||||
python docs/abstractions3.py
|
||||
- name: Test README
|
||||
run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py && python README.py
|
||||
- name: Test Quickstart
|
||||
run: awk '/```python/{flag=1;next}/```/{flag=0}flag' docs/quickstart.md > quickstart.py && python quickstart.py
|
||||
- name: Test DEBUG
|
||||
@@ -261,6 +261,8 @@ jobs:
|
||||
key: unittest-12
|
||||
pydeps: "pillow"
|
||||
deps: testing_unit
|
||||
- name: Test README
|
||||
run: awk '/```python/{flag=1;next}/```/{flag=0}flag' README.md > README.py && python README.py
|
||||
- name: Run unit tests
|
||||
run: python -m pytest -n=auto test/unit/ --durations=20
|
||||
- name: Run targetted tests on NULL backend
|
||||
@@ -272,8 +274,6 @@ jobs:
|
||||
# run: NULL=1 python3 examples/llama.py --gen 1 --size 7B --shard 4 --prompt "Hello." --count 3 --temperature 0 --timing
|
||||
- name: Run GC tests
|
||||
run: python test/external/external_uop_gc.py
|
||||
- name: External Benchmark Schedule
|
||||
run: python3 test/external/external_benchmark_schedule.py
|
||||
- name: Run process replay tests
|
||||
uses: ./.github/actions/process-replay
|
||||
- name: Regen dataset on test_tiny
|
||||
@@ -310,7 +310,7 @@ jobs:
|
||||
run: python test/external/fuzz_shape_ops.py
|
||||
|
||||
testopenclimage:
|
||||
name: CL IMAGE Tests
|
||||
name: 'CL IMAGE Tests'
|
||||
runs-on: ubuntu-22.04
|
||||
timeout-minutes: 10
|
||||
steps:
|
||||
@@ -330,7 +330,7 @@ jobs:
|
||||
uses: ./.github/actions/process-replay
|
||||
|
||||
testgpumisc:
|
||||
name: CL Misc tests
|
||||
name: 'CL Misc tests'
|
||||
runs-on: ubuntu-22.04
|
||||
timeout-minutes: 10
|
||||
steps:
|
||||
@@ -355,7 +355,7 @@ jobs:
|
||||
path: /tmp/sops.gz
|
||||
|
||||
testopenpilot:
|
||||
name: openpilot Compile Tests
|
||||
name: 'openpilot Compile Tests'
|
||||
runs-on: ubuntu-22.04
|
||||
timeout-minutes: 15
|
||||
steps:
|
||||
@@ -387,7 +387,7 @@ jobs:
|
||||
# ****** ONNX Tests ******
|
||||
|
||||
testonnxcpu:
|
||||
name: ONNX (CPU) Tests
|
||||
name: 'ONNX (CPU) Tests'
|
||||
runs-on: ubuntu-22.04
|
||||
timeout-minutes: 20
|
||||
|
||||
@@ -415,7 +415,7 @@ jobs:
|
||||
uses: ./.github/actions/process-replay
|
||||
|
||||
testopencl:
|
||||
name: ONNX (GPU)+Optimization Tests
|
||||
name: 'ONNX (GPU)+Optimization Tests'
|
||||
runs-on: ubuntu-22.04
|
||||
timeout-minutes: 20
|
||||
steps:
|
||||
@@ -503,8 +503,8 @@ jobs:
|
||||
|
||||
# ****** Feature Tests ******
|
||||
|
||||
testrangeifycpu:
|
||||
name: Linux (rangeify) CPU
|
||||
testrangeify:
|
||||
name: Linux (rangeify)
|
||||
runs-on: ubuntu-24.04
|
||||
timeout-minutes: 15
|
||||
steps:
|
||||
@@ -519,44 +519,23 @@ jobs:
|
||||
llvm: "true"
|
||||
- name: Test CPU=1 RANGEIFY=1
|
||||
# TODO: add more passing tests here
|
||||
# test_symbolic_arange_sym_step is passing now
|
||||
# test_threefry_doesnt_use_long is because there's a contig after the long now
|
||||
# test_embedding issue with jit
|
||||
# test_load_state_dict_sharded_model_dict_same_axis issue with multi
|
||||
# test_instancenorm_3d is very slow
|
||||
run: |
|
||||
CPU=1 CPU_LLVM=0 RANGEIFY=1 python3 -m pytest -n auto --durations 20 \
|
||||
-k "not test_threefry_doesnt_use_long and not test_embedding and not test_load_state_dict_sharded_model_dict_same_axis and not test_instancenorm_3d" \
|
||||
-k "not test_symbolic_arange_sym_step and not test_threefry_doesnt_use_long" \
|
||||
test/test_tiny.py test/test_rangeify.py test/test_ops.py test/test_tensor_variable.py \
|
||||
test/test_outerworld_range.py test/test_sample.py test/test_randomness.py test/test_nn.py
|
||||
- name: Test const folding
|
||||
run: CPU=1 RANGEIFY=1 python3 -m pytest -n auto --durations 20 test/test_const_folding.py -k "not test_cast_padded and not TestReduceOpsConstFolding and not TestMultiConstFolding"
|
||||
test/test_outerworld_range.py test/test_sample.py test/test_randomness.py
|
||||
- name: Test multitensor
|
||||
run: CPU=1 RANGEIFY=1 python3 test/test_multitensor.py TestMultiTensor.test_matmul_shard_1_1 TestMultiTensor.test_simple_add_W
|
||||
run: RANGEIFY=1 PYTHONPATH="." python3 test/test_multitensor.py TestMultiTensor.test_matmul_shard_1_1 TestMultiTensor.test_simple_add_W
|
||||
- name: Test GPU=1 RANGEIFY=1
|
||||
run: GPU=1 RANGEIFY=1 pytest -n auto test/test_ops.py
|
||||
- name: Test CPU=1 RANGEIFY=2
|
||||
run: CPU=1 CPU_LLVM=0 RANGEIFY=2 python3 -m pytest -n auto test/test_tiny.py test/test_rangeify.py test/test_ops.py --durations 20
|
||||
# slow (and still wrong on beautiful_mnist)
|
||||
#- name: Test LLVM RANGEIFY=1 (slow tests)
|
||||
#- name: Test LLVM=1 RANGEIFY=1 (slow tests)
|
||||
# run: CPU=1 CPU_LLVM=1 RANGEIFY=1 python3 -m pytest -n auto test/models/test_mnist.py --durations 20
|
||||
|
||||
testrangeifycl:
|
||||
name: Linux (rangeify) CL
|
||||
runs-on: ubuntu-24.04
|
||||
timeout-minutes: 15
|
||||
steps:
|
||||
- name: Checkout Code
|
||||
uses: actions/checkout@v4
|
||||
- name: Setup Environment
|
||||
uses: ./.github/actions/setup-tinygrad
|
||||
with:
|
||||
key: rangeify-minimal-llvm
|
||||
deps: testing_minimal
|
||||
opencl: 'true'
|
||||
llvm: "true"
|
||||
- name: Test CL=1 RANGEIFY=1
|
||||
run: CL=1 RANGEIFY=1 pytest -n auto test/test_ops.py --durations 20
|
||||
- name: Test Fuse
|
||||
run: CL=1 RANGEIFY=2 python3 -m pytest --durations 20 test/test_softmax_fusion.py -k "not test_auto_softmax"
|
||||
|
||||
testdevectorize:
|
||||
name: Linux (devectorize)
|
||||
runs-on: ubuntu-24.04
|
||||
@@ -677,7 +656,7 @@ jobs:
|
||||
run: TRANSCENDENTAL=2 python -m pytest -n=auto test/test_ops.py::TestOps::test_sin test/test_ops.py::TestOps::test_cos test/test_ops.py::TestOps::test_tan test/test_ops.py::TestOps::test_exp test/test_ops.py::TestOps::test_log --durations=20
|
||||
- name: Run TestOps.test_add with SQTT
|
||||
run: |
|
||||
VIZ=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
|
||||
PROFILE=1 SQTT=1 DEBUG=5 python3 test/test_ops.py TestOps.test_add
|
||||
extra/sqtt/rgptool.py create "/tmp/profile.pkl.$USER" -o /tmp/gpu0.rgp
|
||||
- name: Run process replay tests
|
||||
uses: ./.github/actions/process-replay
|
||||
|
||||
@@ -41,6 +41,8 @@ BEAM | [#] | number of beams in kernel beam search
|
||||
DEFAULT_FLOAT | [HALF, ...]| specify the default float dtype (FLOAT32, HALF, BFLOAT16, FLOAT64, ...), default to FLOAT32
|
||||
IMAGE | [1-2] | enable 2d specific optimizations
|
||||
FLOAT16 | [1] | use float16 for images instead of float32
|
||||
PTX | [1] | enable the specialized [PTX](https://docs.nvidia.com/cuda/parallel-thread-execution/) assembler for Nvidia GPUs. If not set, defaults to generic CUDA codegen backend.
|
||||
PROFILE | [1] | enable profiling. This feature is supported in NV, AMD, QCOM and METAL backends.
|
||||
VISIBLE_DEVICES | [list[int]]| restricts the NV/AMD devices that are available. The format is a comma-separated list of identifiers (indexing starts with 0).
|
||||
JIT | [0-2] | 0=disabled, 1=[jit enabled](quickstart.md#jit) (default), 2=jit enabled, but graphs are disabled
|
||||
VIZ | [1] | 0=disabled, 1=[viz enabled](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/viz)
|
||||
|
||||
+11
-18
@@ -2,17 +2,17 @@
|
||||
|
||||
tinygrad supports various runtimes, enabling your code to scale across a wide range of devices. The default runtime can be automatically selected based on the available hardware, or you can force a specific runtime to be default using environment variables (e.g., `CPU=1`).
|
||||
|
||||
| Runtime | Description | Compiler Options | Requirements |
|
||||
|---------|-------------|------------------|--------------|
|
||||
| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | nvrtc (default)<br>PTX (`NV_PTX=1`) | Ampere/Ada/Blackwell series GPUs.<br>You can select an interface via `NV_IFACE=(NVK\|PCI)`. See [NV interfaces](#nv-interfaces) for details. |
|
||||
| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | LLVM (`AMD_LLVM=1`)<br>HIP/COMGR (`AMD_HIP=1`) | RDNA2 or newer GPUs.<br>You can select an interface via `AMD_IFACE=(KFD\|PCI\|USB)`. See [AMD interfaces](#amd-interfaces) for details. |
|
||||
| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
|
||||
| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for `bfloat` support |
|
||||
| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default)<br> PTX (`CUDA_PTX=1`) | NVIDIA GPU with CUDA support |
|
||||
| [CL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | - | OpenCL 2.0 compatible device |
|
||||
| [CPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang or llvm compiler | Clang JIT (default)<br>LLVM IR (`CPU_LLVM=1`) | `clang` compiler in system `PATH` |
|
||||
| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | - | Dawn library installed and discoverable. Binaries: [pydawn v0.3.0](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0) |
|
||||
|
||||
| Runtime | Description | Requirements |
|
||||
|---------|-------------|--------------|
|
||||
| [NV](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_nv.py) | Provides acceleration for NVIDIA GPUs | Ampere/Ada series GPUs |
|
||||
| [AMD](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_amd.py) | Provides acceleration for AMD GPUs | RDNA2/RDNA3/RDNA4 series GPUs. You can select one of the interfaces for communication by setting `AMD_IFACE=(KFD|PCI)`. See [AMD interfaces](#amd-interfaces) for more details. |
|
||||
| [QCOM](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_qcom.py) | Provides acceleration for QCOM GPUs | 6xx series GPUs |
|
||||
| [METAL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_metal.py) | Utilizes Metal for acceleration on Apple devices | M1+ Macs; Metal 3.0+ for `bfloat` support |
|
||||
| [CUDA](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cuda.py) | Utilizes CUDA for acceleration on NVIDIA GPUs | NVIDIA GPU with CUDA support |
|
||||
| [OpenCL](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cl.py) | Accelerates computations using OpenCL on GPUs | OpenCL 2.0 compatible device |
|
||||
| [CPU (C Code)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_cpu.py) | Runs on CPU using the clang compiler | `clang` compiler in system `PATH` |
|
||||
| [LLVM (LLVM IR)](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_llvm.py) | Runs on CPU using the LLVM compiler infrastructure | llvm libraries installed and findable |
|
||||
| [WEBGPU](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/ops_webgpu.py) | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | Dawn library installed and findable. Download binaries [here](https://github.com/wpmed92/pydawn/releases/tag/v0.3.0). |
|
||||
|
||||
## Interoperability
|
||||
|
||||
@@ -70,12 +70,5 @@ AMD backend supports several interfaces for communicating with devices:
|
||||
|
||||
* `KFD`: uses the amdgpu driver
|
||||
* `PCI`: uses the [AM driver](developer/am.md)
|
||||
* `USB`: USB3 interafce for asm24xx chips.
|
||||
|
||||
You can force an interface by setting `AMD_IFACE` to one of these values. In the case of `AMD_IFACE=PCI`, this may unbind your GPU from the amdgpu driver.
|
||||
|
||||
## NV Interfaces
|
||||
NV backend supports several interfaces for communicating with devices:
|
||||
|
||||
* `NVK`: uses the nvidia driver
|
||||
* `PCI`: uses the [NV driver](https://github.com/tinygrad/tinygrad/tree/master/tinygrad/runtime/support/nv/nvdev.py)
|
||||
|
||||
@@ -177,14 +177,13 @@ def ioctl(fd, request, argp):
|
||||
for i in range(s.numcmds):
|
||||
cmd = get_struct(s.cmdlist+ctypes.sizeof(msm_kgsl.struct_kgsl_command_object)*i, msm_kgsl.struct_kgsl_command_object)
|
||||
if IOCTL > 0: print(f"cmd {i}:", format_struct(cmd))
|
||||
#parse_cmd_buf(get_mem(cmd.gpuaddr, cmd.size))
|
||||
parse_cmd_buf(get_mem(cmd.gpuaddr, cmd.size))
|
||||
for i in range(s.numobjs):
|
||||
obj = get_struct(s.objlist+s.objsize*i, msm_kgsl.struct_kgsl_command_object)
|
||||
if IOCTL > 0:
|
||||
print(f"obj {i}:", format_struct(obj))
|
||||
#print(format_struct(msm_kgsl.struct_kgsl_cmdbatch_profiling_buffer.from_buffer_copy(get_mem(obj.gpuaddr, obj.size))))
|
||||
print(format_struct(msm_kgsl.struct_kgsl_cmdbatch_profiling_buffer.from_buffer_copy(get_mem(obj.gpuaddr, obj.size))))
|
||||
#hexdump(get_mem(obj.gpuaddr, obj.size))
|
||||
pass
|
||||
else:
|
||||
#print(f"ioctl({fd=}, (dir:{idir}, size:0x{size:3X}, type:{itype:d}, nr:0x{nr:2X}), {argp=:X}) = {ret=}")
|
||||
pass
|
||||
|
||||
@@ -4,7 +4,7 @@
|
||||
|
||||
Only supported on 7900XTX, requires either AM (`rmmod amdgpu`) or disabling power gating on AMD (`ppfeaturemask=0xffff3fff`, don't forget to rebuild initramfs)
|
||||
|
||||
SQTT is implemented on top of normal tinygrad profiling, `VIZ=1 SQTT=1` to get profile pickle with sqtt data embedded in it.
|
||||
SQTT is implemented on top of normal tinygrad PROFILE=1, `PROFILE=1 SQTT=1` to get profile pickle with sqtt data embedded in it.
|
||||
|
||||
`SQTT_BUFFER_SIZE=X` to change size of SQTT buffer (per shader engine, 6 SEs on 7900xtx) in megabytes, default 256.
|
||||
|
||||
|
||||
+2
-3
@@ -229,8 +229,7 @@ class TestNN(unittest.TestCase):
|
||||
torch_z = torch_layer(torch_x)
|
||||
torch_z.sum().backward()
|
||||
|
||||
# TODO: why is torch numbers all 0?
|
||||
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-4, rtol=5e-6)
|
||||
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-6, rtol=5e-6)
|
||||
|
||||
def test_layernorm(self):
|
||||
N, C, H, W = 20, 5, 10, 10
|
||||
@@ -333,7 +332,7 @@ class TestNN(unittest.TestCase):
|
||||
|
||||
np.testing.assert_allclose(z.numpy(), torch_z.detach().numpy(), atol=5e-6, rtol=5e-6)
|
||||
np.testing.assert_allclose(x.grad.numpy(), torch_x.grad.detach().numpy(), atol=1e-3, rtol=1e-3)
|
||||
np.testing.assert_allclose(layer.weight.grad.numpy(), torch_layer.weight.grad.detach().numpy(), atol=3e-3, rtol=1e-3)
|
||||
np.testing.assert_allclose(layer.weight.grad.numpy(), torch_layer.weight.grad.detach().numpy(), atol=2e-3, rtol=1e-3)
|
||||
np.testing.assert_allclose(layer.bias.grad.numpy(), torch_layer.bias.grad.detach().numpy(), atol=1e-3, rtol=1e-3)
|
||||
|
||||
def test_rmsnorm(self):
|
||||
|
||||
@@ -17,7 +17,7 @@ def helper_collect_profile(*devs):
|
||||
cpu_events.clear()
|
||||
|
||||
profile_list = []
|
||||
with Context(VIZ=1):
|
||||
with Context(PROFILE=1):
|
||||
yield profile_list
|
||||
for dev in devs: dev.synchronize()
|
||||
for dev in devs: dev._at_profile_finalize()
|
||||
|
||||
@@ -408,7 +408,7 @@ class TestVizProfiler(unittest.TestCase):
|
||||
get_profile(prof)
|
||||
|
||||
def test_python_marker(self):
|
||||
with Context(VIZ=1):
|
||||
with Context(PROFILE=1):
|
||||
a = Tensor.empty(1, device="NULL")
|
||||
b = Tensor.empty(1, device="NULL")
|
||||
(a+b).realize()
|
||||
|
||||
+3
-2
@@ -354,8 +354,9 @@ if PROFILE:
|
||||
|
||||
with open(fn:=temp("profile.pkl", append_user=True), "wb") as f: pickle.dump(cpu_events+Compiled.profile_events+Buffer.profile_events, f)
|
||||
|
||||
from tinygrad.uop.ops import launch_viz
|
||||
launch_viz("PROFILE", fn)
|
||||
if not getenv("SQTT", 0):
|
||||
from tinygrad.uop.ops import launch_viz
|
||||
launch_viz(PROFILE, fn)
|
||||
|
||||
if __name__ == "__main__":
|
||||
from tinygrad import Tensor, Device
|
||||
|
||||
+2
-6
@@ -135,7 +135,7 @@ USE_TC, TC_SELECT, TC_OPT, AMX = ContextVar("TC", 1), ContextVar("TC_SELECT", -1
|
||||
TRANSCENDENTAL, NOLOCALS = ContextVar("TRANSCENDENTAL", 1), ContextVar("NOLOCALS", 0)
|
||||
FUSE_ARANGE, FUSE_CONV_BW = ContextVar("FUSE_ARANGE", 1), ContextVar("FUSE_CONV_BW", 0)
|
||||
SPLIT_REDUCEOP, NO_MEMORY_PLANNER, RING = ContextVar("SPLIT_REDUCEOP", 1), ContextVar("NO_MEMORY_PLANNER", 0), ContextVar("RING", 1)
|
||||
PICKLE_BUFFERS, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("LRU", 1)
|
||||
PICKLE_BUFFERS, PROFILE, LRU = ContextVar("PICKLE_BUFFERS", 1), ContextVar("PROFILE", getenv("VIZ")), ContextVar("LRU", 1)
|
||||
CACHELEVEL, IGNORE_BEAM_CACHE, DEVECTORIZE = ContextVar("CACHELEVEL", 2), ContextVar("IGNORE_BEAM_CACHE", 0), ContextVar("DEVECTORIZE", 1)
|
||||
DISABLE_COMPILER_CACHE, BLOCK_REORDER = ContextVar("DISABLE_COMPILER_CACHE", 0), ContextVar("BLOCK_REORDER", 1)
|
||||
DONT_REALIZE_EXPAND, DONT_GROUP_REDUCES = ContextVar("DONT_REALIZE_EXPAND", 0), ContextVar("DONT_GROUP_REDUCES", 0)
|
||||
@@ -146,7 +146,6 @@ RANGEIFY, FUSE_ATTENTION = ContextVar("RANGEIFY", 0), ContextVar("FUSE_ATTENTION
|
||||
EMULATE = ContextVar("EMULATE", "")
|
||||
CPU_COUNT = ContextVar("CPU_COUNT", max(1, (os.cpu_count() or 1) // (4 if ARCH_X86 else 2))) # take 1/2 of the cores, accounting HT
|
||||
CPU_LLVM, AMD_LLVM = ContextVar("CPU_LLVM", 0), ContextVar("AMD_LLVM", 1)
|
||||
VIZ = PROFILE = ContextVar("VIZ", 0)
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Metadata:
|
||||
@@ -325,10 +324,7 @@ def cpu_objdump(lib, objdump_tool='objdump'):
|
||||
print(subprocess.check_output([objdump_tool, '-d', f.name]).decode('utf-8'))
|
||||
|
||||
def capstone_flatdump(lib: bytes):
|
||||
try: import capstone
|
||||
except ImportError:
|
||||
print("Disassembler Error: Capstone not installed.")
|
||||
return
|
||||
import capstone
|
||||
match platform.machine():
|
||||
case 'x86_64' | 'AMD64': cs = capstone.Cs(capstone.CS_ARCH_X86, capstone.CS_MODE_64)
|
||||
case 'aarch64' | 'arm64': cs = capstone.Cs(capstone.CS_ARCH_ARM64, capstone.CS_MODE_ARM)
|
||||
|
||||
@@ -2,10 +2,9 @@ from __future__ import annotations
|
||||
from typing import cast
|
||||
import ctypes, functools, hashlib
|
||||
from tinygrad.runtime.autogen import opencl as cl
|
||||
from tinygrad.helpers import init_c_var, to_char_p_p, from_mv, OSX, DEBUG, mv_address, suppress_finalizing, getenv
|
||||
from tinygrad.helpers import init_c_var, to_char_p_p, from_mv, OSX, DEBUG, mv_address, suppress_finalizing
|
||||
from tinygrad.renderer.cstyle import OpenCLRenderer, IntelRenderer
|
||||
from tinygrad.device import BufferSpec, LRUAllocator, Compiled, Compiler, CompileError
|
||||
if getenv("IOCTL"): import extra.qcom_gpu_driver.opencl_ioctl # noqa: F401 # pylint: disable=unused-import
|
||||
|
||||
# see test/external/external_osx_profiling.py to determine this ratio. it's in like GPU clocks or something
|
||||
OSX_TIMING_RATIO = (125/3) if OSX else 1.0
|
||||
|
||||
+7
-5
@@ -7,7 +7,7 @@ from tinygrad.uop import Ops, GroupOp
|
||||
from tinygrad.uop.mathtraits import MathTrait
|
||||
from tinygrad.dtype import ConstType, ImageDType, dtypes, DType, truncate, PtrDType, least_upper_dtype, Invalid, InvalidType
|
||||
from tinygrad.helpers import ContextVar, all_int, prod, getenv, all_same, Context, partition, temp, unwrap, T, argfix, Metadata, flatten, TRACEMETA
|
||||
from tinygrad.helpers import PICKLE_BUFFERS, PROFILE, dedup, cdiv, cmod, diskcache_put, to_function_name, cpu_profile, TracingKey, RANGEIFY, VIZ
|
||||
from tinygrad.helpers import PICKLE_BUFFERS, PROFILE, dedup, cdiv, cmod, diskcache_put, to_function_name, cpu_profile, TracingKey, RANGEIFY
|
||||
if TYPE_CHECKING:
|
||||
from tinygrad.shape.shapetracker import ShapeTracker
|
||||
from tinygrad.device import Buffer, MultiBuffer
|
||||
@@ -835,6 +835,7 @@ def track_uop(u:UOp):
|
||||
|
||||
# *** tracking pattern matcher ***
|
||||
|
||||
VIZ = ContextVar("VIZ", 0)
|
||||
TRACK_MATCH_STATS = ContextVar("TRACK_MATCH_STATS", 2 if VIZ else 0)
|
||||
match_stats:dict[UPat, list[int|float]] = dict()
|
||||
|
||||
@@ -937,7 +938,7 @@ if TRACK_MATCH_STATS or PROFILE:
|
||||
with open(fn:=temp("rewrites.pkl", append_user=True), "wb") as f:
|
||||
print(f"rewrote {len(tracked_ctxs)} graphs and matched {sum(len(r.matches) for x in tracked_ctxs for r in x)} times, saved to {fn}")
|
||||
pickle.dump([(tracked_keys, tracked_ctxs, uop_fields)], f)
|
||||
if VIZ: return launch_viz("VIZ", temp("rewrites.pkl", append_user=True))
|
||||
if VIZ: launch_viz(VIZ, temp("rewrites.pkl", append_user=True))
|
||||
if getenv("PRINT_MATCH_STATS", TRACK_MATCH_STATS.value):
|
||||
ret = [0,0,0.0,0.0]
|
||||
for k,v in sorted(list(match_stats.items()), key=lambda x: x[1][2]+x[1][3]):
|
||||
@@ -947,10 +948,11 @@ if TRACK_MATCH_STATS or PROFILE:
|
||||
print(f"{ret[0]:6d} / {ret[1]:7d} -- {ret[3]*1000.:9.2f} / {(ret[2]+ret[3])*1000.:9.2f} ms -- TOTAL")
|
||||
print(f"{len(match_stats)} rules, {sum(v[0] > 0 for v in match_stats.values())} matched once")
|
||||
|
||||
def launch_viz(env_str:str, data:str):
|
||||
os.environ[env_str] = "0"
|
||||
def launch_viz(var:ContextVar, data:str):
|
||||
os.environ[(env_str:=var.key)] = "0"
|
||||
os.environ[f"{env_str}_DATA"] = data
|
||||
if not int(os.getenv("VIZ", "0")) and not int(os.getenv("PROFILE", "0")) and not int(os.getenv("SQTT", "0")):
|
||||
os.environ[f"{env_str}_VALUE"] = str(var.value)
|
||||
if not int(os.getenv("VIZ", "0")) and not int(os.getenv("PROFILE", "0")):
|
||||
args = ['--kernels', getenv("VIZ_DATA", "")] if getenv("VIZ_DATA", "") else []
|
||||
args += ['--profile', getenv("PROFILE_DATA", "")] if getenv("PROFILE_DATA", "") else []
|
||||
os.execv(sys.executable, [sys.executable] + [os.path.join(os.path.dirname(__file__), "../", "viz", "serve.py")] + args)
|
||||
|
||||
+4
-3
@@ -6,18 +6,19 @@ most uses of DEBUG >= 3
|
||||
tiny-tools
|
||||
|
||||
and a viewer for:
|
||||
SAVE_SCHEDULE=1
|
||||
TRACK_MATCH_STATS=2
|
||||
ProfileEvents
|
||||
PROFILE=1
|
||||
|
||||
to use:
|
||||
1. Run tinygrad with VIZ=1 (this saves the pkls and launches the server (new process please!))
|
||||
1. Run tinygrad with VIZ=1 and/or PROFILE=1 (this saves the pkls and launches the server (new process please!))
|
||||
2. That's it!
|
||||
|
||||
This should be able to:
|
||||
1. See all schedules (VIZ=1)
|
||||
2. See all graphs and how they were rewritten (VIZ=1)
|
||||
3. See generated code (VIZ=1)
|
||||
4. See profile (click on 'profiler')
|
||||
4. See profile (PROFILE=1)
|
||||
|
||||
bunch of dev rules:
|
||||
* everything must be responsive to keyboard smashing! lag should never happen
|
||||
|
||||
@@ -102,10 +102,10 @@
|
||||
fill: none;
|
||||
stroke-width: 1.4px;
|
||||
}
|
||||
g.node.highlight rect, .edgePath.highlight, g.port circle {
|
||||
.highlight rect, .edgePath.highlight, g.port circle {
|
||||
stroke: #89C9A2;
|
||||
}
|
||||
g.highlight.child rect, .edgePath.highlight.child {
|
||||
.highlight.child rect, .edgePath.highlight.child {
|
||||
stroke: #C888B0;
|
||||
}
|
||||
#edge-labels g.port.highlight {
|
||||
|
||||
@@ -721,7 +721,7 @@ appendResizer(document.querySelector(".metadata-parent"), { minWidth: 20, maxWid
|
||||
|
||||
// **** keyboard shortcuts
|
||||
|
||||
document.addEventListener("keydown", (event) => {
|
||||
document.addEventListener("keydown", async function(event) {
|
||||
const { currentCtx, currentStep, currentRewrite, expandSteps } = state;
|
||||
// up and down change the step or context from the list
|
||||
const changeStep = expandSteps && ctxs[currentCtx].steps?.length;
|
||||
|
||||
@@ -295,7 +295,7 @@ class TCPServerWithReuse(socketserver.TCPServer): allow_reuse_address = True
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument('--kernels', type=pathlib.Path, help='Path to kernels', default=pathlib.Path(temp("rewrites.pkl", append_user=True)))
|
||||
parser.add_argument('--profile', type=pathlib.Path, help='Path to profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
|
||||
parser.add_argument('--profile', type=pathlib.Path, help='Path profile', default=pathlib.Path(temp("profile.pkl", append_user=True)))
|
||||
args = parser.parse_args()
|
||||
|
||||
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
|
||||
@@ -308,7 +308,7 @@ if __name__ == "__main__":
|
||||
|
||||
ctxs = get_metadata(load_pickle(args.kernels))
|
||||
|
||||
profile_ret = get_profile(load_pickle(args.profile))
|
||||
profile_ret = get_profile(profile:=load_pickle(args.profile))
|
||||
|
||||
server = TCPServerWithReuse(('', PORT), Handler)
|
||||
reloader_thread = threading.Thread(target=reloader)
|
||||
|
||||
Reference in New Issue
Block a user