Fix DSP transcendentals (tinygrad#9542)

Author: quortus

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

@@ -215,7 +215,7 @@ runs:
         wget -qO- https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc
         echo "deb http://apt.llvm.org/$(lsb_release -cs)/ llvm-toolchain-$(lsb_release -cs)-19 main" | sudo tee /etc/apt/sources.list.d/llvm.list
         sudo apt update -y || true
-        sudo apt install -y --no-install-recommends libllvm19
+        sudo apt install -y --no-install-recommends libllvm19 clang-19 lld-19
 
     - name: Install LLVM (macOS)
       if: inputs.llvm == 'true' && runner.os == 'macOS'

.github/workflows/test.yml

@@ -498,6 +498,8 @@ jobs:
         cache-to: type=gha,mode=min
     - name: Run test_tiny on DSP
       run: DEBUG=2 DSP=1 python test/test_tiny.py
+    - name: Test transcendentals
+      run: CC=clang-19 PYTHONPATH="." DEBUG=2 DSP=1 python test/test_transcendental.py TestTranscendentalVectorized
     - name: Test quantize onnx
       run: PYTHONPATH="." DEBUG=2 DSP=1 python3 test/test_quantize_onnx.py
     - name: Test LLVM=1 DEVECTORIZE=0

test/test_transcendental.py

@@ -148,12 +148,16 @@ def test_exp2_vectorized(self):
   def test_log2_vectorized(self):
     for vec_size in [1,2,3,4,5,127,128]: self._test_vectorized_op(Tensor.log2, np.log2, (0.001, 200), vec_size)
 
+  @unittest.skipIf(getenv("DSP"), "requires int division")
   def test_sin_vectorized(self):
     for vec_size in [1,2,3,4,5,127,128]: self._test_vectorized_op(Tensor.sin, np.sin, (-100, 100), vec_size)
 
   def test_pow_vectorized(self):
     # np.pow returns nan for negative values raised to a non-integral power
     for vec_size in [1,2,3,4,5,127,128]: self._test_vectorized_op(Tensor.pow, np.pow, (0.001, 200), vec_size, param_range=(-10, 10))
 
+  def test_sqrt_vectorized(self):
+    for vec_size in [1,2,3,4,5,127,128]: self._test_vectorized_op(Tensor.sqrt, np.sqrt, (0, 100), vec_size)
+
 if __name__ == '__main__':
   unittest.main()

tinygrad/runtime/ops_dsp.py

@@ -41,9 +41,7 @@ class DSPRenderer(ClangRenderer):
   extra_matcher = dsp_pm_late+ClangRenderer.extra_matcher
   string_rewrite = dsp_string+ClangRenderer.string_rewrite
   type_map = { **ClangRenderer.type_map, dtypes.uint64: "unsigned long long", dtypes.int64: "long long" }
-  code_for_op = {**ClangRenderer.code_for_op, Ops.SIN: lambda x,dtype: f"__builtin_sin({x})",
-                 Ops.LOG2: lambda x,dtype: f"__builtin_log2l({x})" if dtype == dtypes.float64 else f"__builtin_log2f({x})",
-                 Ops.EXP2: lambda x,dtype: f"__builtin_exp2l({x})" if dtype == dtypes.float64 else f"__builtin_exp2f({x})"}
+  code_for_op = {k:v for k,v in ClangRenderer.code_for_op.items() if k != Ops.SQRT}
 
   def render_kernel(self, function_name:str, kernel:list[str], bufs:list[tuple[str,tuple[DType,bool]]], uops:list[UOp], prefix=None) -> str:
     ret = super().render_kernel(function_name, kernel, bufs, uops, prefix)