Merge pull request #22 from SeisSol/davschneller/half

Cleanup, Green Sparsity, RISC-V V, and AVX10 Support (Version 0.3.0)

.github/workflows/codegen.yml

@@ -38,140 +38,39 @@ jobs:
 
           echo "Success!"
 
-  pspamm-codegen-avx:
-    name: pspamm-codegen-avx
+  pspamm-codegen:
+    name: pspamm-codegen
     runs-on: ubuntu-24.04
     needs: install-pspamm
-    steps:
-      - name: apt-get
-        run: |
-          set -euo pipefail
-          sudo apt-get update
-          sudo apt-get install g++ qemu-user-static
-
-      - name: setup-python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.11'
-
-      - name: checkout-pspamm
-        uses: actions/checkout@v4
-
-      - name: pip-pspamm
-        run: |
-          pip install .
-
-      - name: pspamm-tests-generate
-        run: |
-          cd tests/
-          python unit_tests_hsw.py
-      
-      - name: pspamm-tests-compile
-        run: |
-          cd tests/
-          g++ -static -mavx512f build/hsw_testsuite.cpp -o build/hsw-test
-      
-      - name: pspamm-tests-run
-        run: |
-          cd tests/
-          qemu-x86_64-static -cpu Haswell build/hsw-test
-
-  pspamm-codegen-avx512-no-run:
-    name: pspamm-codegen-avx512-no-run
-    runs-on: ubuntu-24.04
-    needs: install-pspamm
-    steps:
-      - name: apt-get
-        run: |
-          set -euo pipefail
-          sudo apt-get update
-          sudo apt-get install g++ qemu-user-static
-
-      - name: setup-python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.11'
-
-      - name: checkout-pspamm
-        uses: actions/checkout@v4
-
-      - name: pip-pspamm
-        run: |
-          pip install .
-
-      - name: pspamm-tests-generate
-        run: |
-          cd tests/
-          python unit_tests_knl.py
-      
-      - name: pspamm-tests-compile
-        run: |
-          cd tests/
-          g++ -static -mavx512f build/knl_testsuite.cpp -o build/knl-test
-      
-      # disabled, since qemu doesn't support AVX512F (yet) with of Ubuntu 24.04
-      # - name: pspamm-tests-run
-      #   run: |
-      #     cd tests/
-      #     qemu-x86_64-static -cpu Skylake-Server build/knl-test
-
-  pspamm-codegen-aarch64:
-    name: pspamm-codegen-aarch64
-    runs-on: ubuntu-24.04
-    needs: install-pspamm
-    steps:
-      - name: apt-get
-        run: |
-          set -euo pipefail
-          sudo apt-get update
-          sudo apt-get install g++-aarch64-linux-gnu qemu-user-static
-
-      - name: setup-python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.11'
-
-      - name: checkout-pspamm
-        uses: actions/checkout@v4
-
-      - name: pip-pspamm
-        run: |
-          pip install .
-
-      - name: pspamm-tests-generate
-        run: |
-          cd tests/
-          python unit_tests_arm.py
-      
-      - name: pspamm-tests-compile
-        run: |
-          cd tests/
-          aarch64-linux-gnu-g++ -static -march=armv8.2-a build/arm_testsuite.cpp -o build/arm-test
-      
-      - name: pspamm-tests-run
-        run: |
-          cd tests/
-          qemu-aarch64-static -cpu max build/arm-test
-
-  pspamm-codegen-armsve:
-    name: pspamm-codegen-armsve
-    runs-on: ubuntu-24.04
-    needs: install-pspamm
-    # include vector lengths for SVE manually (for now)
+    # include all vector lengths manually for now
+    # not supported:
+    # * RVV >= 1024
+    # * ARM/SVE which is not a power of 2
     strategy:
+      fail-fast: false
       matrix:
-        vectorlen:
-          - 128
-          - 256
-          - 512
-          - 1024
-          - 2048
+        arch:
+          - hsw128
+          - hsw256
+          - knl128
+          - knl256
+          - knl512
+          - arm128
+          - arm_sve128
+          - arm_sve256
+          - arm_sve512
+          - arm_sve1024
+          - arm_sve2048
+          - rvv128
+          - rvv256
+          - rvv512
+          - rvv1024
     steps:
       - name: apt-get
         run: |
           set -euo pipefail
           sudo apt-get update
-          sudo apt-get install g++-aarch64-linux-gnu qemu-user-static
+          sudo apt-get install g++-aarch64-linux-gnu g++-riscv64-linux-gnu qemu-user-static
 
       - name: setup-python
         uses: actions/setup-python@v4
@@ -185,17 +84,12 @@ jobs:
         run: |
           pip install .
 
-      - name: pspamm-tests-generate
+      - name: pspamm-tests
         run: |
           cd tests/
-          python unit_tests_arm_sve.py ${{matrix.vectorlen}}
-      
-      - name: pspamm-tests-compile
-        run: |
-          cd tests/
-          aarch64-linux-gnu-g++ -static -march=armv8.2-a+sve -msve-vector-bits=${{matrix.vectorlen}} build/arm_sve${{matrix.vectorlen}}_testsuite.cpp -o build/arm_sve${{matrix.vectorlen}}-test
-      
-      - name: pspamm-tests-run
-        run: |
-          cd tests/
-          qemu-aarch64-static -cpu max,sve${{matrix.vectorlen}}=on,sve-default-vector-length=-1 build/arm_sve${{matrix.vectorlen}}-test
+          LOCALARCH=${{matrix.arch}}
+          if [[ ${LOCALARCH:0:3} == "knl" ]]; then
+            ./runlocal.sh ${{matrix.arch}} norun
+          else
+            ./runlocal.sh ${{matrix.arch}}
+          fi

README.md

@@ -1,22 +1,32 @@
-# Code Generator for Sparse Matrix Multiplication
-Generates inline-Assembly for sparse Matrix Multiplication.
+# PSpaMM
 
-Currently Intel Xeon Phi 'Knights Landing' (AVX512), Haswell/Zen2 (AVX2), and ARM Cortex-A53 (ARMv8) are supported.
+A code generator for small matrix multiplications.
+
+Currently supported:
+
+* x86_64: AVX2, AVX512/AVX10.1
+* ARM/AARCH64: NEON, SVE (128,256,512,1024,2048 bit)
+* RISC-V: V (128,256,512,1024,2048,4096,8192 bit)
 
 ## Installation
 
-PspaMM is a Python package. I.e. you may do
+PspaMM is a Python package. I.e. after cloning, may install it via pip.
 
-```
-pip install .
-```
+Alternatively, you can install it directly by running
 
-to install it.
+```bash
 
-## Usage 
+pip install git+https://github.com/SeisSol/PSpaMM.git
 
 ```
-pspamm-generator M N K LDA LDB LDC ALPHA BETA --arch {arm,arm_sve{128,256,512,1024,2048},knl,hsw} \
-    --mtx_filename MTX_FILE_PATH --output_funcname FUNCTION_NAME --output_filename OUTPUT_NAME
+
+## Usage
+
+```bash
+
+pspamm-generator M N K LDA LDB LDC ALPHA BETA \
+    --arch {arm,arm_sve{128..2048},knl{128..512},hsw{128..256},rvv{128..8192}} \
+    --amtx_filename MTX_FILE_PATH --bmtx_filename MTX_FILE_PATH \
+    --output_funcname FUNCTION_NAME --output_filename OUTPUT_NAME
 
 ```

pspamm/VERSION

@@ -1 +1 @@
-0.2.3
+0.3.0

pspamm/architecture.py

@@ -8,17 +8,5 @@ def init():
     generator = None
     operands = None
 
-
-
-#https://stackoverflow.com/questions/452969/does-python-have-an-equivalent-to-java-class-forname
-
 def get_class( kls ):
     return import_module(kls)
-    parts = kls.split('.')
-    module = ".".join(parts[:-1])
-    m = __import__( module )
-    for comp in parts[1:]:
-        m = getattr(m, comp)
-    return m
-
-

pspamm/codegen/analysis.py

@@ -3,57 +3,12 @@
 
 from typing import List, Set
 
-class Analyzer(Visitor):
-
+class Analyzer:
     def __init__(self, starting_regs: List[Register] = None):
         self.clobbered_registers = set(starting_regs)
-        self.stack = []
-
-    def visitFma(self, stmt: FmaStmt):
-        self.clobbered_registers.add(stmt.add_dest)
-
-    def visitMul(self, stmt: FmaStmt):
-        self.clobbered_registers.add(stmt.dest)
-        self.clobbered_registers.add(stmt.mult_src)
-
-    def visitBcst(self, stmt: FmaStmt):
-        self.clobbered_registers.add(stmt.dest)
-
-    def visitAdd(self, stmt: AddStmt):
-        self.clobbered_registers.add(stmt.dest)
-
-    def visitLabel(self, stmt: LabelStmt):
-        pass
-
-    def visitJump(self, stmt: JumpStmt):
-        pass
-
-    def visitMov(self, stmt: MovStmt):
-        if isinstance(stmt.dest, Register):
-            self.clobbered_registers.add(stmt.dest)
-
-    def visitLea(self, stmt: MovStmt):
-            self.clobbered_registers.add(stmt.dest)
-
-    def visitStore(self, stmt: MovStmt):
-        if isinstance(stmt.dest, Register):
-            self.clobbered_registers.add(stmt.dest)
-
-    def visitLoad(self, stmt: MovStmt):
-        if isinstance(stmt.dest, Register):
-            self.clobbered_registers.add(stmt.dest)
-
-    def visitPrefetch(self, stmt: PrefetchStmt):
-        self.clobbered_registers.add(stmt.dest.base)
-
-    def visitCmp(self, stmt: CmpStmt):
-        pass
-
-    def visitBlock(self, block: Block):
-        self.stack.append(block)
-        for stmt in block.contents:
-            stmt.accept(self)
-        self.stack.pop()
-
-
+        self.input_operands = set()
 
+    def collect(self, block: Block):
+        for instr in block.flatten():
+            self.clobbered_registers |= instr.regs()
+            self.input_operands |= instr.args()