OSX GPUOcelot (#8209)

* add patches

* add osx test in ci

* macos specific uvm, gpfifo mask

* only do that for now

* Revert "add patches"

This reverts commit 80d3112a57.

* use fork for now

* workflow only one worker

* merge osxtests with tests

* Revert "merge osxtests with tests"

This reverts commit 3461c8f46c.

* macos pagesize 16384

---------

Co-authored-by: nimlgen <138685161+nimlgen@users.noreply.github.com>
Co-authored-by: George Hotz <72895+geohot@users.noreply.github.com>

.github/actions/setup-tinygrad/action.yml

@@ -143,14 +143,20 @@ runs:
 
     # **** CUDA ****
 
-    - name: Install packages (cuda)
+    - name: Install cuda packages (Linux)
-      if: inputs.cuda == 'true'
+      if: inputs.cuda == 'true' && runner.os == 'Linux'
       shell: bash
       run: |
         echo 'Acquire::http::Pipeline-Depth "5";' | sudo tee -a /etc/apt/apt.conf.d/99parallel
         sudo apt update -y || true
         sudo apt install -y --no-install-recommends git g++ cmake ninja-build llvm-15-dev zlib1g-dev libglew-dev \
           flex bison libfl-dev libboost-thread-dev libboost-filesystem-dev nvidia-cuda-toolkit-gcc libzstd-dev
+    - name: Install gpuocelot dependencies (MacOS)
+      if: inputs.cuda == 'true' && runner.os == 'macOS'
+      shell: bash
+      run: |
+        brew update
+        brew install cmake ninja llvm@15 zlib glew flex bison boost zstd ncurses
     - name: Cache gpuocelot
       if: inputs.cuda == 'true'
       id: cache-build
@@ -159,7 +165,7 @@ runs:
         cache-name: cache-gpuocelot-build
       with:
         path: ${{ github.workspace }}/gpuocelot/ocelot
-        key: ubuntu22.04-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-0
+        key: ${{ runner.os }}-gpuocelot-b16039dc940dc6bc4ea0a98380495769ff35ed99-rebuild-0
     - name: Clone/compile gpuocelot
       if: inputs.cuda == 'true' && steps.cache-build.outputs.cache-hit != 'true'
       shell: bash
@@ -176,7 +182,7 @@ runs:
       shell: bash
       run: |
         cd ${{ github.workspace }}/gpuocelot/ocelot/build
-        sudo cp libgpuocelot.so /usr/lib/libgpuocelot.so
+        sudo cp libgpuocelot.${{ runner.os == 'macOS' && 'dylib' || 'so' }} /usr/${{ runner.os == 'macOS' && 'local/' || ''}}lib/
 
     # **** WebGPU ****
 

.github/workflows/test.yml

@@ -513,6 +513,7 @@ jobs:
         deps: testing
         python-version: '3.11'
         amd: 'true'
+        cuda: 'true'
     - name: Run real world test
       run: JIT=2 METAL=1 python -m pytest -n=auto test/models/test_real_world.py --durations=20
     - name: Test models (Metal)
@@ -540,6 +541,14 @@ jobs:
         FORWARD_ONLY: 1
       run: |
         python3 -m pytest -n=auto test/test_hcq.py test/test_tiny.py --durations=20
+    - name: Run pytest (ptx)
+      env:
+        MOCKGPU: 1
+        PTX: 1
+        NV: 1
+        FORWARD_ONLY: 1
+      run: |
+        python3 -m pytest -n=auto test/test_hcq.py test/test_tiny.py --durations=20
     - name: Run process replay tests
       uses: ./.github/actions/process-replay
 

test/mockgpu/nv/nvgpu.py

@@ -55,7 +55,7 @@ class GPFIFO:
 
   def _reset_buf_state(self): self.buf, self.buf_ptr = None, 0
   def _set_buf_state(self, gpfifo_entry):
-    ptr = ((gpfifo_entry >> 2) & 0xfffffffff) << 2
+    ptr = ((gpfifo_entry >> 2) & 0x3fffffffff) << 2
     sz = ((gpfifo_entry >> 42) & 0x1fffff) << 2
     self.buf = to_mv(ptr, sz).cast("I")
     self.buf_sz = sz // 4

tinygrad/runtime/ops_nv.py

@@ -7,7 +7,7 @@ from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, H
 from tinygrad.runtime.support.hcq import HWInterface, MOCKGPU
 from tinygrad.ops import sint
 from tinygrad.device import BufferSpec, CPUProgram
-from tinygrad.helpers import getenv, mv_address, init_c_struct_t, to_mv, round_up, data64, data64_le, DEBUG, prod
+from tinygrad.helpers import getenv, mv_address, init_c_struct_t, to_mv, round_up, data64, data64_le, DEBUG, prod, OSX
 from tinygrad.renderer.ptx import PTXRenderer
 from tinygrad.renderer.cstyle import NVRenderer
 from tinygrad.runtime.support.compiler_cuda import CUDACompiler, PTXCompiler, PTX, NVPTXCompiler, NVCompiler
@@ -295,8 +295,8 @@ class NVDevice(HCQCompiled[NVSignal]):
   # TODO: Need a proper allocator for va addresses
   # 0x1000000000 - 0x2000000000, reserved for system/cpu mappings
   # VA space is 48bits.
-  low_uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=0x1000000000, base=0x1000000000, wrap=False)
+  low_uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=0x1000000000, base=0x8000000000 if OSX else 0x1000000000, wrap=False)
-  uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=(1 << 48) - 1, base=0x2000000000, wrap=False)
+  uvm_vaddr_allocator: BumpAllocator = BumpAllocator(size=(1 << 48) - 1, base=low_uvm_vaddr_allocator.base + low_uvm_vaddr_allocator.size, wrap=False)
   host_object_enumerator: int = 0x1000
 
   def _new_gpu_fd(self):
@@ -314,7 +314,7 @@ class NVDevice(HCQCompiled[NVSignal]):
 
   def _gpu_alloc(self, size:int, host=False, uncached=False, cpu_access=False, contiguous=False, map_flags=0, tag="") -> HCQBuffer:
     # Uncached memory is "system". Use huge pages only for gpu memory.
-    page_size = (4 << 10) if uncached or host else ((2 << 20) if size >= (8 << 20) else (4 << 10))
+    page_size = (4 << (12 if OSX else 10)) if uncached or host else ((2 << 20) if size >= (8 << 20) else (4 << (12 if OSX else 10)))
     size = round_up(size, page_size)
     va_addr = self._alloc_gpu_vaddr(size, alignment=page_size, force_low=cpu_access)