Merge branch 'master' of https://github.com/scemama/simple_gpu

Author: scemama

.gitignore

@@ -59,3 +59,6 @@ Thumbs.db
 
 # Generated documentation
 /docs/
+
+# CodeQL artifacts
+_codeql_detected_source_root

INSTALL

@@ -41,6 +41,12 @@ sudo make install
 - **NVIDIA cuBLAS library** (included with CUDA Toolkit)
 - **CUDA-capable GPU**
 
+### Optional Dependencies (for AMD GPU support)
+
+- **AMD ROCm platform** (>= 4.0 recommended)
+- **hipBLAS library** (included with ROCm)
+- **ROCm-capable GPU**
+
 ## Installation on Different Systems
 
 ### Ubuntu/Debian
@@ -57,6 +63,12 @@ For NVIDIA GPU support:
 # See: https://developer.nvidia.com/cuda-downloads
 ```
 
+For AMD GPU support:
+```bash
+# Install ROCm from AMD's repository
+# See: https://rocmdocs.amd.com/en/latest/Installation_Guide/Installation-Guide.html
+```
+
 For building from Git:
 ```bash
 sudo apt-get install autoconf automake libtool
@@ -97,6 +109,13 @@ To build only the CPU version (even if CUDA is available):
 ./configure --disable-nvidia
 ```
 
+### Disabling AMD GPU Library
+
+To build only the CPU version (even if ROCm is available):
+```bash
+./configure --disable-amd
+```
+
 ### Specifying CUDA Installation Directory
 
 If CUDA is installed in a non-standard location:
@@ -109,6 +128,18 @@ Example:
 ./configure --with-cuda=/opt/cuda-11.8
 ```
 
+### Specifying ROCm Installation Directory
+
+If ROCm is installed in a non-standard location:
+```bash
+./configure --with-rocm=/path/to/rocm
+```
+
+Example:
+```bash
+./configure --with-rocm=/opt/rocm
+```
+
 ### Specifying BLAS Library
 
 To use a specific BLAS library:
@@ -140,6 +171,12 @@ To install in a custom location (default is /usr/local):
 You can combine multiple options:
 ```bash
 ./configure --prefix=$HOME/local --with-cuda=/opt/cuda --with-blas="-lopenblas"
+
+# Or for AMD GPU
+./configure --prefix=$HOME/local --with-rocm=/opt/rocm --with-blas="-lopenblas"
+
+# Or build both GPU libraries
+./configure --with-cuda=/opt/cuda --with-rocm=/opt/rocm
 ```
 
 ## Build and Installation Steps
@@ -155,6 +192,7 @@ The configure script will:
 - Detect available compilers
 - Check for required libraries
 - Auto-detect CUDA/cuBLAS if available
+- Auto-detect ROCm/hipBLAS if available
 - Display a configuration summary
 
 Example output:
@@ -174,6 +212,12 @@ Example output:
   CUDA_CFLAGS:     -I/usr/local/cuda/include
   CUDA_LDFLAGS:    -L/usr/local/cuda/lib64
   CUDA_LIBS:       -lcudart -lcublas
+
+  AMD GPU:         enabled
+  ROCM_PATH:       /opt/rocm
+  ROCM_CFLAGS:     -I/opt/rocm/include
+  ROCM_LDFLAGS:    -L/opt/rocm/lib
+  ROCM_LIBS:       -lamdhip64 -lhipblas
 ========================================
 ```
 
@@ -187,6 +231,7 @@ make
 This will create:
 - `src/.libs/libsimple_gpu-cpu.so` - CPU-only library
 - `src/.libs/libsimple_gpu-nvidia.so` - NVIDIA GPU library (if CUDA was detected)
+- `src/.libs/libsimple_gpu-amd.so` - AMD GPU library (if ROCm was detected)
 
 ### 3. Test
 
@@ -241,6 +286,13 @@ Explicitly disable NVIDIA support:
 ./configure --disable-nvidia
 ```
 
+### Configure detects ROCm but you don't want it
+
+Explicitly disable AMD support:
+```bash
+./configure --disable-amd
+```
+
 ### Libraries not found at runtime
 
 After installation, you may need to update your library path:
@@ -315,7 +367,8 @@ If you encounter issues:
 1. Check that all prerequisites are installed
 2. Review the configuration summary output
 3. Check `config.log` for detailed error messages
-4. Ensure CUDA libraries are in your library path (for GPU version)
-5. Run tests with `make check` to verify the build
+4. Ensure CUDA libraries are in your library path (for NVIDIA GPU version)
+5. Ensure ROCm libraries are in your library path (for AMD GPU version)
+6. Run tests with `make check` to verify the build
 
 For additional help, please file an issue on the GitHub repository.

Makefile.am

@@ -1,7 +1,7 @@
 ACLOCAL_AMFLAGS = -I m4
 AM_CPPFLAGS = -I$(srcdir)/include
 
-EXTRA_DIST = README.md LICENSE pkgconfig/simple_gpu-cpu.pc.in pkgconfig/simple_gpu-nvidia.pc.in
+EXTRA_DIST = README.md LICENSE pkgconfig/simple_gpu-cpu.pc.in pkgconfig/simple_gpu-nvidia.pc.in pkgconfig/simple_gpu-amd.pc.in
 
 include_HEADERS = include/simple_gpu.h include/simple_gpu.F90
 
@@ -13,6 +13,10 @@ if BUILD_NVIDIA
 pkgconfig_DATA += pkgconfig/simple_gpu-nvidia.pc
 endif
 
+if BUILD_AMD
+pkgconfig_DATA += pkgconfig/simple_gpu-amd.pc
+endif
+
 # ===== Source library section (from src/Makefile.am) =====
 lib_LTLIBRARIES = libsimple_gpu-cpu.la
 
@@ -33,6 +37,16 @@ libsimple_gpu_nvidia_la_LIBADD = $(CUDA_LIBS)
 libsimple_gpu_nvidia_la_LDFLAGS = -version-info 1:0:0 $(CUDA_LDFLAGS)
 endif
 
+# AMD GPU version library (conditional)
+if BUILD_AMD
+lib_LTLIBRARIES += libsimple_gpu-amd.la
+libsimple_gpu_amd_la_SOURCES = src/gpu_amd.c include/simple_gpu.F90
+libsimple_gpu_amd_la_CFLAGS = -I$(top_srcdir)/include $(ROCM_CFLAGS)
+libsimple_gpu_amd_la_FCFLAGS =
+libsimple_gpu_amd_la_LIBADD = $(ROCM_LIBS)
+libsimple_gpu_amd_la_LDFLAGS = -version-info 1:0:0 $(ROCM_LDFLAGS)
+endif
+
 # ===== Tests section (from tests/Makefile.am) =====
 # CPU tests (always built)
 TESTS = \
@@ -158,6 +172,77 @@ tests_test_sgeam_gpu_LDADD = libsimple_gpu-nvidia.la
 tests_test_sgeam_gpu_LDFLAGS = -ldl
 endif
 
+# AMD GPU tests (only when AMD is available)
+if BUILD_AMD
+TESTS += \
+	tests/test_ddot_amd \
+	tests/test_sdot_amd \
+	tests/test_dgemv_amd \
+	tests/test_sgemv_amd \
+	tests/test_dgemm_amd \
+	tests/test_sgemm_amd \
+	tests/test_dgeam_amd \
+	tests/test_sgeam_amd
+
+check_PROGRAMS += \
+	tests/test_ddot_amd \
+	tests/test_sdot_amd \
+	tests/test_dgemv_amd \
+	tests/test_sgemv_amd \
+	tests/test_dgemm_amd \
+	tests/test_sgemm_amd \
+	tests/test_dgeam_amd \
+	tests/test_sgeam_amd
+
+# DDOT AMD test
+tests_test_ddot_amd_SOURCES = tests/test_ddot.F90
+tests_test_ddot_amd_FCFLAGS = -I$(top_builddir)
+tests_test_ddot_amd_LDADD = libsimple_gpu-amd.la
+tests_test_ddot_amd_LDFLAGS = -ldl
+
+# SDOT AMD test
+tests_test_sdot_amd_SOURCES = tests/test_sdot.F90
+tests_test_sdot_amd_FCFLAGS = -I$(top_builddir)
+tests_test_sdot_amd_LDADD = libsimple_gpu-amd.la
+tests_test_sdot_amd_LDFLAGS = -ldl
+
+# DGEMV AMD test
+tests_test_dgemv_amd_SOURCES = tests/test_dgemv.F90
+tests_test_dgemv_amd_FCFLAGS = -I$(top_builddir)
+tests_test_dgemv_amd_LDADD = libsimple_gpu-amd.la
+tests_test_dgemv_amd_LDFLAGS = -ldl
+
+# SGEMV AMD test
+tests_test_sgemv_amd_SOURCES = tests/test_sgemv.F90
+tests_test_sgemv_amd_FCFLAGS = -I$(top_builddir)
+tests_test_sgemv_amd_LDADD = libsimple_gpu-amd.la
+tests_test_sgemv_amd_LDFLAGS = -ldl
+
+# DGEMM AMD test
+tests_test_dgemm_amd_SOURCES = tests/test_dgemm.F90
+tests_test_dgemm_amd_FCFLAGS = -I$(top_builddir)
+tests_test_dgemm_amd_LDADD = libsimple_gpu-amd.la
+tests_test_dgemm_amd_LDFLAGS = -ldl
+
+# SGEMM AMD test
+tests_test_sgemm_amd_SOURCES = tests/test_sgemm.F90
+tests_test_sgemm_amd_FCFLAGS = -I$(top_builddir)
+tests_test_sgemm_amd_LDADD = libsimple_gpu-amd.la
+tests_test_sgemm_amd_LDFLAGS = -ldl
+
+# DGEAM AMD test
+tests_test_dgeam_amd_SOURCES = tests/test_dgeam.F90
+tests_test_dgeam_amd_FCFLAGS = -I$(top_builddir)
+tests_test_dgeam_amd_LDADD = libsimple_gpu-amd.la
+tests_test_dgeam_amd_LDFLAGS = -ldl
+
+# SGEAM AMD test
+tests_test_sgeam_amd_SOURCES = tests/test_sgeam.F90
+tests_test_sgeam_amd_FCFLAGS = -I$(top_builddir)
+tests_test_sgeam_amd_LDADD = libsimple_gpu-amd.la
+tests_test_sgeam_amd_LDFLAGS = -ldl
+endif
+
 # Clean up Fortran module files
 CLEANFILES = *.mod
 

README.md

@@ -10,11 +10,11 @@ A simple and easy-to-use library for GPU computing in Fortran, providing transpa
 
 Simple GPU is a library designed to simplify GPU computing in Fortran applications. It provides:
 
-- **Dual implementation**: CPU-only version using standard BLAS, and GPU-accelerated version using NVIDIA cuBLAS
+- **Dual implementation**: CPU-only version using standard BLAS, and GPU-accelerated version using NVIDIA cuBLAS or AMD hipBLAS
 - **Transparent interface**: Same Fortran API for both CPU and GPU versions
 - **Memory management**: Easy GPU memory allocation and data transfer
 - **BLAS operations**: Common BLAS operations (GEMM, GEMV, DOT, GEAM) for both single and double precision
-- **Stream support**: Asynchronous operations through CUDA streams
+- **Stream support**: Asynchronous operations through CUDA or HIP streams
 
 ## Features
 
@@ -64,6 +64,11 @@ All BLAS operations have variants that accept 64-bit integers for dimensions. Th
 - NVIDIA cuBLAS library
 - CUDA-capable GPU
 
+#### For AMD GPU version (optional):
+- AMD ROCm platform
+- hipBLAS library
+- ROCm-capable GPU
+
 ### Building from Source
 
 1. **Generate configure script** (if building from git):
@@ -81,16 +86,21 @@ All BLAS operations have variants that accept 64-bit integers for dimensions. Th
    **Configuration options**:
    - `--disable-nvidia`: Disable NVIDIA GPU library even if CUDA is available
    - `--with-cuda=DIR`: Specify CUDA installation directory (default: `/usr/local/cuda`)
+   - `--disable-amd`: Disable AMD GPU library even if ROCm is available
+   - `--with-rocm=DIR`: Specify ROCm installation directory (default: auto-detect)
    - `--with-blas=LIB`: Specify BLAS library to use
 
    **Example configurations**:
    ```bash
    # CPU version only
-   ./configure --disable-nvidia
+   ./configure --disable-nvidia --disable-amd
 
    # Specify CUDA location
    ./configure --with-cuda=/opt/cuda
 
+   # Specify ROCm location
+   ./configure --with-rocm=/opt/rocm
+   
    # Use specific BLAS library
    ./configure --with-blas="-lmkl_rt"
    ```
@@ -112,7 +122,7 @@ All BLAS operations have variants that accept 64-bit integers for dimensions. Th
 
 ## Library Versions
 
-Simple GPU provides two shared libraries:
+Simple GPU provides three shared libraries:
 
 1. **libsimple_gpu-cpu.so**: CPU-only version
    - Uses standard BLAS library
@@ -124,7 +134,12 @@ Simple GPU provides two shared libraries:
    - Requires CUDA-capable GPU
    - Provides GPU acceleration for supported operations
 
-Both libraries provide the same Fortran interface, allowing seamless switching between CPU and GPU implementations.
+3. **libsimple_gpu-amd.so**: AMD GPU version (if ROCm is available)
+   - Uses AMD hipBLAS library
+   - Requires ROCm-capable GPU
+   - Provides GPU acceleration for supported operations
+
+All libraries provide the same Fortran interface, allowing seamless switching between CPU and GPU implementations.
 
 ## Usage
 
@@ -324,6 +339,9 @@ gfortran -o myapp myapp.f90 -lsimple_gpu-cpu
 
 # GPU version (NVIDIA)
 gfortran -o myapp myapp.f90 -lsimple_gpu-nvidia
+
+# GPU version (AMD)
+gfortran -o myapp myapp.f90 -lsimple_gpu-amd
 ```
 
 ## Testing
@@ -371,12 +389,17 @@ Specify BLAS library explicitly:
 
 ### Runtime errors with GPU version
 
-1. Ensure NVIDIA drivers are properly installed
-2. Check that CUDA libraries are in your library path:
+1. Ensure NVIDIA drivers are properly installed (for NVIDIA GPUs)
+2. Ensure ROCm drivers are properly installed (for AMD GPUs)
+3. Check that CUDA libraries are in your library path:
    ```bash
    export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
    ```
-3. Verify GPU is accessible with `nvidia-smi`
+4. Check that ROCm libraries are in your library path:
+   ```bash
+   export LD_LIBRARY_PATH=/opt/rocm/lib:$LD_LIBRARY_PATH
+   ```
+5. Verify GPU is accessible with `nvidia-smi` (NVIDIA) or `rocm-smi` (AMD)
 
 ## Contributing