Merge pull request #34 from NumPower/feat/update_package_1

GD Image loading and save fix, GPU allocation improvement and GPU slicing fixes

Author: henrique-borba

config.w32

@@ -1,73 +1,113 @@
-// Comments in this file start with '//'.
-// Remove where necessary.
+ARG_ENABLE("ndarray",
+  "whether to enable ndarray support",
+  "Enable ndarray support",
+  "no")
 
-ARG_ENABLE('ndarray', 'whether to enable ndarray support', 'no')
+ARG_WITH("cuda", "for CUDA support",
+  "Include CUDA support", "no", "no")
 
-if PHP_NDARRAY != "no"; then
-  // Check for CUBLAS library
-  PHP_CHECK_LIBRARY('cublas', 'cublasDgemm', '
-    ARG_ENABLE("cublas", "Enable CUBLAS support", "no")
-    AC_MSG_RESULT(["CUBLAS detected"])
-    CFLAGS = CFLAGS + " -lcublas -lcudart "
-  ', '
-    AC_MSG_RESULT(["Wrong CUBLAS version or library not found."])
-  ')
+if test x%PHP_CUDA% != xno; then
+    CHECK_LIB("cublas", "cublasDgemm", ,
+      [
+        AC_DEFINE("HAVE_CUBLAS", 1)
+        PHP_ADD_LIBRARY("cublas", , "NDARRAY_SHARED_LIBADD")
+        AC_MSG_RESULT("CUBLAS detected")
+        PHP_ADD_MAKEFILE_FRAGMENT("$abs_srcdir/Makefile.frag", "$abs_builddir")
+        CFLAGS+=" -lcublas -lcudart"
+        AC_CHECK_HEADER("immintrin.h",
+          [
+            AC_DEFINE("HAVE_AVX2", 1)
+            AC_MSG_RESULT("AVX2/SSE detected")
+            CXX+=" -mavx2 -march=native"
+          ],
+          [
+            AC_DEFINE("HAVE_AVX2", 0)
+            AC_MSG_RESULT("AVX2/SSE not found")
+          ],
+          []
+        )
+      ],
+      [
+        AC_MSG_RESULT("wrong cublas version or library not found.")
+        AC_CHECK_HEADER("immintrin.h",
+          [
+            AC_DEFINE("HAVE_AVX2", 1)
+            AC_MSG_RESULT("AVX2/SSE detected")
+            CFLAGS+=" -mavx2 -march=native"
+          ],
+          [
+            AC_DEFINE("HAVE_AVX2", 0)
+            AC_MSG_RESULT("AVX2/SSE not found")
+          ],
+          []
+        )
+      ]
+    )
+else
+    AC_CHECK_HEADER("immintrin.h",
+      [
+        AC_DEFINE("HAVE_AVX2", 1)
+        AC_MSG_RESULT("AVX2/SSE detected")
+        CFLAGS+=" -mavx2 -march=native"
+      ],
+      [
+        AC_DEFINE("HAVE_AVX2", 0)
+        AC_MSG_RESULT("AVX2/SSE not found")
+      ],
+      []
+    )
+fi
 
-  // Check for AVX2 support
-  PHP_CHECK_HEADER('immintrin.h', '
-    AC_DEFINE("HAVE_AVX2", 1, ["Have AVX2/SSE support"])
-    AC_MSG_RESULT(["AVX2/SSE detected"])
-    CFLAGS = CFLAGS + " -mavx2 "
-  ', '
-    AC_DEFINE("HAVE_AVX2", 0, ["Have AVX2/SSE support"])
-    AC_MSG_RESULT(["AVX2/SSE not found"])
-  ')
+if test x%PHP_GD% != xno; then
+    AC_DEFINE("HAVE_GD", 1)
+    AC_MSG_RESULT("GD detected")
+    PHP_ADD_EXTENSION_DEP("ndarray", "gd", true)
+endif
 
-  // Check for CBLAS library
-  PHP_CHECK_LIBRARY('cblas', 'cblas_sdot', '
-    ARG_ENABLE("cblas", "Enable CBLAS support", "no")
-    AC_MSG_RESULT(["CBLAS detected"])
-    CFLAGS = CFLAGS + " -lcblas "
-  ', '
-    PHP_CHECK_LIBRARY('openblas', 'cblas_sdot', '
-      ARG_ENABLE("openblas", "Enable OpenBLAS support", "no")
-      AC_MSG_RESULT(["OpenBLAS detected"])
-      AC_DEFINE("HAVE_CBLAS", 1, [""])
-      CFLAGS = CFLAGS + " -lopenblas -lpthread "
-    ', '
-      AC_MSG_ERROR(["Wrong OpenBLAS/BLAS version or library not found."])
-    ', '
-      -lopenblas
-    ')
-  ', '
-    -lcblas
-  ')
+CHECK_LIB("cblas", "cblas_sdot",
+  [
+    AC_DEFINE("HAVE_CBLAS", 1)
+    PHP_ADD_LIBRARY("cblas", , "NDARRAY_SHARED_LIBADD")
+    AC_MSG_RESULT("CBlas detected")
+    CFLAGS+=" -lcblas"
+  ],
+  [
+    CHECK_LIB("openblas", "cblas_sdot",
+      [
+        PHP_ADD_LIBRARY("openblas", , "NDARRAY_SHARED_LIBADD")
+        AC_MSG_RESULT("OpenBLAS detected")
+        AC_DEFINE("HAVE_CBLAS", 1)
+        CFLAGS+=" -lopenblas -lpthread"
+      ],
+      [
+        AC_MSG_ERROR("wrong openblas/blas version or library not found.")
+      ],
+      [
+        "-lopenblas"
+      ]
+    )
+  ],
+  [
+    "-lcblas"
+  ]
+)
 
-  // Check for LAPACKE library
-  PHP_CHECK_LIBRARY('lapack', 'dgesvd_', '
-    ARG_ENABLE("lapacke", "Enable LAPACKE support", "no")
-    AC_MSG_RESULT(["LAPACKE detected"])
-    CFLAGS = CFLAGS + " -llapack -llapacke "
-  ', '
-    AC_MSG_ERROR(["Wrong LAPACKE version or library not found."])
-  ')
+CHECK_LIB("lapacke", "LAPACKE_sgesdd",
+  [
+    AC_DEFINE("HAVE_LAPACKE", 1)
+    PHP_ADD_LIBRARY("lapack", , "NDARRAY_SHARED_LIBADD")
+    AC_MSG_RESULT("LAPACKE detected")
+    CFLAGS+=" -llapack -llapacke"
+  ],
+  [
+    AC_MSG_ERROR("wrong LAPACKE version or library not found. Try `apt install liblapacke-dev`")
+  ]
+)
 
-  // Add your extension's source files
-  PHP_ADD_EXTENSION('ndarray', '
-    numpower.c \
-    src/initializers.c \
-    src/ndmath/double_math.c \
-    src/ndarray.c \
-    src/debug.c \
-    src/buffer.c \
-    src/logic.c \
-    src/gpu_alloc.c \
-    src/ndmath/linalg.c \
-    src/manipulation.c \
-    src/iterators.c \
-    src/indexing.c \
-    src/ndmath/arithmetics.c \
-    src/ndmath/statistics.c \
-    src/types.c
-  ')
+if test x%PHP_NDARRAY% != xno; then
+  AC_DEFINE("HAVE_NDARRAY", 1, "Have ndarray support")
+  PHP_NEW_EXTENSION("ndarray",
+      "numpower.c src/initializers.c src/ndmath/double_math.c src/ndarray.c src/debug.c src/buffer.c src/logic.c src/gpu_alloc.c src/ndmath/linalg.c src/manipulation.c src/iterators.c src/indexing.c src/ndmath/arithmetics.c src/ndmath/statistics.c src/types.c",
+      %ext_shared%
+  )
 endif

numpower.c

@@ -474,20 +474,24 @@ PHP_METHOD(NDArray, setDevice) {
 #endif
 }
 
-ZEND_BEGIN_ARG_INFO(arginfo_reshape, 1)
+// @todo Indices conversion lose precision, we must convert it directly to a integer vector in C
+//       without relying on ZVAL_TO_NDARRAY. We must apply the same for all other cases where a
+//       PHP array of longs is converted to NDArray before being converted to a C integer.
+ZEND_BEGIN_ARG_INFO(arginfo_reshape, 2)
+ZEND_ARG_INFO(0, a)
 ZEND_ARG_INFO(0, shape_zval)
 ZEND_END_ARG_INFO();
 PHP_METHOD(NDArray, reshape) {
     int *new_shape;
     zval *shape_zval;
-    zval *current = getThis();
+    zval *a;
     NDArray *rtn;
-    ZEND_PARSE_PARAMETERS_START(1, 1)
-    Z_PARAM_ZVAL(shape_zval)
+    ZEND_PARSE_PARAMETERS_START(2, 2)
+        Z_PARAM_ZVAL(a)
+        Z_PARAM_ZVAL(shape_zval)
     ZEND_PARSE_PARAMETERS_END();
-    NDArray* target = ZVAL_TO_NDARRAY(current);
+    NDArray* target = ZVAL_TO_NDARRAY(a);
     NDArray* shape = ZVAL_TO_NDARRAY(shape_zval);
-
     new_shape = NDArray_ToIntVector(shape);
 
     rtn = NDArray_Reshape(target, new_shape, NDArray_NUMELEMENTS(shape));
@@ -501,6 +505,7 @@ PHP_METHOD(NDArray, reshape) {
     if (Z_TYPE_P(shape_zval) == IS_ARRAY) {
         NDArray_FREE(shape);
     }
+    CHECK_INPUT_AND_FREE(a, target);
     RETURN_NDARRAY(rtn, return_value);
 }
 
@@ -3947,7 +3952,7 @@ static const zend_function_entry class_NDArray_methods[] = {
     ZEND_ME(NDArray, max, arginfo_ndarray_max, ZEND_ACC_PUBLIC | ZEND_ACC_STATIC)
 
     // MANIPULATION
-    ZEND_ME(NDArray, reshape, arginfo_reshape, ZEND_ACC_PUBLIC)
+    ZEND_ME(NDArray, reshape, arginfo_reshape, ZEND_ACC_PUBLIC | ZEND_ACC_STATIC)
     ZEND_ME(NDArray, toArray, arginfo_toArray, ZEND_ACC_PUBLIC)
     ZEND_ME(NDArray, toImage, arginfo_toImage, ZEND_ACC_PUBLIC)
     ZEND_ME(NDArray, copy, arginfo_ndarray_copy, ZEND_ACC_PUBLIC | ZEND_ACC_STATIC)

src/buffer.c

@@ -62,7 +62,6 @@ void buffer_free() {
  */
 void buffer_ndarray_free(int uuid) {
     if (MAIN_MEM_STACK.buffer != NULL) {
-        // @todo investigate double free problem
         if (MAIN_MEM_STACK.lastFreed == -1) {
             MAIN_MEM_STACK.lastFreed = uuid;
         }

src/gpu_alloc.c

@@ -1,4 +1,5 @@
 #include "../config.h"
+#include <Zend/zend.h>
 
 #ifdef HAVE_CUBLAS
 #include "gpu_alloc.h"
@@ -9,27 +10,26 @@
 void
 NDArray_VMALLOC(void** target, unsigned int size) {
     MAIN_MEM_STACK.totalGPUAllocated++;
-    cudaMalloc(target, size);
-    cudaDeviceSynchronize();
+    cublasStatus_t stat = cudaMalloc(target, size);
+    if (stat != cudaSuccess) {
+        zend_throw_error(NULL, "device memory allocation failed");
+    }
 }
 
 void
 NDArray_VMEMCPY_D2D(char* target, char* dst, unsigned int size) {
     cudaMemcpy(dst, target, size, cudaMemcpyDeviceToDevice);
-    cudaDeviceSynchronize();
 }
 
 void
 NDArray_VMEMCPY_H2D(char* target, char* dst, unsigned int size) {
     cudaMemcpy(dst, target, size, cudaMemcpyHostToDevice);
-    cudaDeviceSynchronize();
 }
 
 void
 NDArray_VFREE(void* target) {
     MAIN_MEM_STACK.totalGPUAllocated--;
     cudaFree(target);
-    cudaDeviceSynchronize();
 }
 
 void

src/logic.c

@@ -18,7 +18,6 @@
 /**
  * Check if all values are not 0
  *
- * @todo Implement non-AVX2 logic
  * @param a
  * @return
  */

src/manipulation.c

@@ -161,7 +161,7 @@ linearize_FLOAT_matrix(float *dst_in,
                     cublasHandle_t handle;
                     cublasCreate(&handle);
                     cublasScopy(handle, columns,
-                                (const float*)((const float*)src + (columns - 1) * column_strides),
+                                (const float*)src,
                                 column_strides, dst, one);
 #endif
                 }
@@ -178,7 +178,7 @@ linearize_FLOAT_matrix(float *dst_in,
 #ifdef HAVE_CUBLAS
                     cublasHandle_t handle;
                     cublasCreate(&handle);
-                    cublasScopy(handle, columns, (const float*)((const char*)src + (columns - 1) * column_strides),
+                    cublasScopy(handle, columns, (const float*)src,
                                 column_strides / sizeof(float), dst, one);
 #endif
                 }
@@ -206,18 +206,17 @@ linearize_FLOAT_matrix(float *dst_in,
 
 NDArray*
 NDArray_Slice(NDArray* array, NDArray** indexes, int num_indices, int return_view) {
+    if (num_indices > NDArray_NDIM(array)) {
+        zend_throw_error(NULL, "too many indices for array");
+        return NULL;
+    }
+
     NDArray *slice, *rtn;
     int slice_ndim = NDArray_NDIM(array);
     int *slice_shape = emalloc(sizeof(int) * slice_ndim);
     int *slice_strides = emalloc(sizeof(int) * slice_ndim);
     int i, offset = 0;
     int start = 0, stop = 0, step = 0;
-
-    if (num_indices > NDArray_NDIM(array)) {
-        zend_throw_error(NULL, "too many indices for array");
-        return NULL;
-    }
-
     if (NDArray_NDIM(array) == 1) {
         int out_ndim = NDArray_NDIM(array);
         if (NDArray_NUMELEMENTS(indexes[0]) >= 1) {
@@ -265,7 +264,6 @@ NDArray_Slice(NDArray* array, NDArray** indexes, int num_indices, int return_vie
         slice_shape[i] = (int)floorf(((float)stop - (float)start) / (float)step);
         offset += start * NDArray_STRIDES(array)[i];
     }
-
     for (; i < slice_ndim; i++) {
         slice_shape[i] = NDArray_SHAPE(array)[i];
     }
@@ -276,12 +274,12 @@ NDArray_Slice(NDArray* array, NDArray** indexes, int num_indices, int return_vie
     if (NDArray_DEVICE(array) == NDARRAY_DEVICE_CPU) {
         rtn_data = emalloc(NDArray_ELSIZE(array) * NDArray_NUMELEMENTS(slice));
     }
+
 #ifdef HAVE_CUBLAS
     if (NDArray_DEVICE(array) == NDARRAY_DEVICE_GPU) {
         NDArray_VMALLOC((void**)&rtn_data, NDArray_ELSIZE(array) * NDArray_NUMELEMENTS(slice));
     }
 #endif
-
     linearize_FLOAT_matrix(rtn_data, NDArray_FDATA(slice), slice);
     slice->data = (char*)rtn_data;
     slice->strides = Generate_Strides(slice_shape, slice_ndim, NDArray_ELSIZE(slice));