WIP: redesign the declarative layout APIs

.gitlab-ci.yml

@@ -4,11 +4,19 @@ stages:
 make:generic:
   stage: build
   script:
+    - git clone https://xgitlab.cels.anl.gov/argo/excit.git
+    - cd excit
     - ./autogen.sh
     - mkdir build
     - ./configure --prefix=`pwd`/build
     - make
     - make install
+    - cd ..
+    - ./autogen.sh
+    - mkdir build
+    - PKG_CONFIG_PATH=excit/build/lib/pkgconfig ./configure --prefix=`pwd`/build
+    - make
+    - make install
     - make check
   artifacts:
     when: on_failure
@@ -22,9 +30,17 @@ make:knl:
   stage: build
   script:
     - source /opt/intel/compilers_and_libraries/linux/bin/compilervars.sh intel64
+    - git clone https://xgitlab.cels.anl.gov/argo/excit.git
+    - cd excit
+    - ./autogen.sh
+    - mkdir build
+    - ./configure --prefix=`pwd`/build
+    - make
+    - make install
+    - cd ..
     - ./autogen.sh
     - mkdir build
-    - CC=icc CFLAGS="-mkl -xhost" ./configure --prefix=`pwd`/build --enable-benchmarks
+    - CC=icc CFLAGS="-mkl -xhost" PKG_CONFIG_PATH=excit/build/lib/pkgconfig ./configure --prefix=`pwd`/build --enable-benchmarks
     - make -j64
     - make install
     - make check

configure.ac

@@ -52,6 +52,9 @@ AM_CONDITIONAL([ADD_BENCHMARKS],[test "x$benchmarks" = xtrue])
 AC_CHECK_HEADERS(numa.h)
 AC_CHECK_LIB(numa, move_pages)
 
+# excit iterators
+PKG_CHECK_MODULES([EXCIT],[libexcit])
+
 # internal jemalloc
 ac_configure_args="$ac_configure_args \
 			'--with-jemalloc-prefix=jemk_aml_' \

src/Makefile.am

@@ -1,4 +1,4 @@
-AM_CPPFLAGS = -I$(top_srcdir)/jemalloc/include
+AM_CPPFLAGS = -I$(top_srcdir)/jemalloc/include @EXCIT_CFLAGS@
 lib_LTLIBRARIES = libaml.la
 
 ARENA_JEMALLOC_CSOURCES = arena_jemalloc.c
@@ -10,21 +10,33 @@ AREA_LINUX_CSOURCES = area_linux.c \
 
 AREA_POSIX_CSOURCES = area_posix.c
 
+LAYOUT_CSOURCES = layout.c \
+		  layout_dense.c \
+		  layout_pad.c \
+		  layout_reshape.c
+
 TILING_CSOURCES = tiling.c \
 		  tiling_1d.c \
 		  tiling_2d.c
 
+TILING_ND_CSOURCES = tiling_nd.c \
+		     tiling_nd_resize.c \
+		     tiling_nd_pad.c \
+		     tiling_nd_collapse.c
+
 BINDING_CSOURCES = binding.c \
 		   binding_single.c \
 		   binding_interleave.c
 
 DMA_CSOURCES = dma.c \
 	       dma_linux_par.c \
-	       dma_linux_seq.c
+	       dma_linux_seq.c \
+	       dma_layout.c
 
 SCRATCH_CSOURCES = scratch.c \
 		   scratch_seq.c \
-		   scratch_par.c
+		   scratch_par.c \
+		   scratch_double.c
 
 UTILS_CSOURCES = vector.c
 
@@ -34,12 +46,24 @@ LIBCSOURCES = aml.c area.c arena.c \
 	      $(AREA_LINUX_CSOURCES) \
 	      $(AREA_POSIX_CSOURCES) \
 	      $(TILING_CSOURCES) \
+	      $(TILING_ND_CSOURCES) \
 	      $(BINDING_CSOURCES) \
 	      $(DMA_CSOURCES) \
-	      $(SCRATCH_CSOURCES)
+	      $(SCRATCH_CSOURCES) \
+	      $(LAYOUT_CSOURCES) \
+	      copy.c
 
-LIBHSOURCES = aml.h
+LIBHSOURCES = aml.h \
+	      aml-layout.h \
+	      aml-layout-dense.h \
+	      aml-layout-pad.h \
+	      aml-layout-reshape.h \
+	      aml-tiling.h \
+	      aml-tiling-resize.h \
+	      aml-tiling-pad.h \
+	      aml-tiling-collapse.h \
+	      aml-copy.h
 
 libaml_la_SOURCES = $(LIBCSOURCES) $(LIBHSOURCES)
-libaml_la_LIBADD = -L$(top_srcdir)/jemalloc/lib/ -ljemalloc-aml
+libaml_la_LIBADD = -L$(top_srcdir)/jemalloc/lib/ -ljemalloc-aml @EXCIT_LIBS@
 include_HEADERS = $(LIBHSOURCES)

src/aml-copy.h

@@ -0,0 +1,192 @@
+#ifndef AML_COPY_H
+#define AML_COPY_H 1
+
+ /*******************************************************************************
+ * Hypervolume copy and transpose functions.
+ ******************************************************************************/
+
+/*
+ * Copies a (sub-)hypervolume to another (sub-)hypervolume.
+ * "d": number of dimensions.
+ * "dst": pointer to the destination hypervolume.
+ * "dst_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the destination hypervolume.
+ * "src": pointer to the source hypervolume.
+ * "src_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the source hypervolume.
+ * "elem_number": pointer to d values representing the number of elements
+ *                in each dimension of the (sub-)hypervolume to copy.
+ * "elem_size": size of memory elements.
+ * Returns 0 if successful; an error code otherwise.
+ */
+int aml_copy_nd(size_t d, void *dst, const size_t *dst_pitch,
+		const void *src, const size_t *src_pitch,
+		const size_t *elem_number, const size_t elem_size);
+/*
+ * Copies a (sub-)hypervolume to another (sub-)hypervolume while transposing.
+ * Reverse of aml_copy_rtnd.
+ * Example a[3][4][5] -> b[5][3][4] (C notation).
+ * "d": number of dimensions.
+ * "dst": pointer to the destination hypervolume.
+ * "dst_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the destination hypervolume.
+ * "src": pointer to the source hypervolume.
+ * "src_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the source hypervolume.
+ * "elem_number": pointer to d values representing the number of elements
+ *                in each dimension of the (sub-)hypervolume to copy.
+ * "elem_size": size of memory elements in the src hypervolume order.
+ * Returns 0 if successful; an error code otherwise.
+ */
+int aml_copy_tnd(size_t d, void *dst, const size_t *dst_pitch,
+		 const void *src, const size_t *src_pitch,
+		 const size_t *elem_number, const size_t elem_size);
+/*
+ * Copies a (sub-)hypervolume to another (sub-)hypervolume while transposing.
+ * Reverse of aml_copy_tnd.
+ * Example a[3][4][5] -> b[4][5][3] (C notation).
+ * "d": number of dimensions.
+ * "dst": pointer to the destination hypervolume.
+ * "dst_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the destination hypervolume.
+ * "src": pointer to the source hypervolume.
+ * "src_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the source hypervolume.
+ * "elem_number": pointer to d values representing the number of elements
+ *                in each dimension of the (sub-)hypervolume to copy.
+ * "elem_size": size of memory elements in the src hypervolume order.
+ * Returns 0 if successful; an error code otherwise.
+ */
+int aml_copy_rtnd(size_t d, void *dst, const size_t *dst_pitch,
+		  const void *src, const size_t *src_pitch,
+		  const size_t *elem_number, const size_t elem_size);
+
+/*
+ * Copies a (sub-)hypervolume to another (sub-)hypervolume while shuffling
+ * dimensions. Example a[4][2][3][5] -> b[5][4][3][2] (C notation).
+ * "d": number of dimensions.
+ * "target_dims": array of d dimension index representing the mapping
+ *                between the source dimensions and the target dimensions.
+ *                Example [3, 1, 0, 2]
+ * "dst": pointer to the destination hypervolume.
+ * "dst_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the destination hypervolume.
+ * "src": pointer to the source hypervolume.
+ * "src_pitch": pointer to d-1 pitch values representing the pitch
+ *              in each dimension of the source hypervolume.
+ * "elem_number": pointer to d values representing the number of elements
+ *                in each dimension of the (sub-)hypervolume to copy.
+ * "elem_size": size of memory elements in the src hypervolume order.
+ * Returns 0 if successful; an error code otherwise.
+ */
+int aml_copy_shnd(size_t d, const size_t *target_dims, void *dst,
+		  const size_t *dst_pitch, const void *src,
+		  const size_t *src_pitch, const size_t *elem_number,
+		  const size_t elem_size);
+/*
+ * Strided version of aml_copy_nd.
+ */
+int aml_copy_ndstr(size_t d, void *dst, const size_t *dst_pitch,
+		   const size_t *dst_stride, const void *src,
+		   const size_t *src_pitch, const size_t *src_stride,
+		   const size_t *elem_number, const size_t elem_size);
+/*
+ * Strided version of aml_copy_tnd.
+ */
+int aml_copy_tndstr(size_t d, void *dst, const size_t *dst_pitch,
+		    const size_t *dst_stride, const void *src,
+		    const size_t *src_pitch, const size_t *src_stride,
+		    const size_t *elem_number, const size_t elem_size);
+/*
+ * Strided version of aml_copy_rtnd.
+ */
+int aml_copy_rtndstr(size_t d, void *dst, const size_t *dst_pitch,
+		     const size_t *dst_stride, const void *src,
+		     const size_t *src_pitch, const size_t *src_stride,
+		     const size_t *elem_number, const size_t elem_size);
+/*
+ * Strided version of aml_copy_shnd.
+ */
+int aml_copy_shndstr(size_t d, const size_t *target_dims, void *dst,
+		     const size_t *dst_pitch, const size_t *dst_stride,
+		     const void *src, const size_t *src_pitch,
+		     const size_t *src_stride, const size_t *elem_number,
+		     const size_t elem_size);
+/*
+ * Version of aml_copy_nd using cumulative pitch.
+ */
+int aml_copy_nd_c(size_t d, void *dst, const size_t *cumul_dst_pitch,
+		  const void *src, const size_t *cumul_src_pitch,
+		  const size_t *elem_number, const size_t elem_size);
+/*
+ * Version of aml_copy_ndstr using cumulative pitch.
+ */
+int aml_copy_ndstr_c(size_t d, void *dst, const size_t *dst_pitch,
+		     const size_t *cumul_dst_stride, const void *src,
+		     const size_t *src_pitch, const size_t *cumul_src_stride,
+		     const size_t *elem_number, const size_t elem_size);
+/*
+ * Version of aml_copy_nd using cumulative pitch.
+ */
+int aml_copy_tnd_c(size_t d, void *dst, const size_t *cumul_dst_pitch,
+		   const void *src, const size_t *cumul_src_pitch,
+		   const size_t *elem_number, const size_t elem_size);
+/*
+ * Version of aml_copy_nd using cumulative pitch.
+ */
+int aml_copy_rtnd_c(size_t d, void *dst, const size_t *cumul_dst_pitch,
+		    const void *src, const size_t *cumul_src_pitch,
+		    const size_t *elem_number, const size_t elem_size);
+/*
+ * Version of aml_copy_shnd using cumulative pitch.
+ */
+int aml_copy_shnd_c(size_t d, const size_t *target_dims, void *dst,
+		    const size_t *cumul_dst_pitch, const void *src,
+		    const size_t *cumul_src_pitch, const size_t *elem_number,
+		    const size_t elem_size);
+/*
+ * Version of aml_copy_tndstr using cumulative pitch.
+ */
+int aml_copy_tndstr_c(size_t d, void *dst, const size_t *cumul_dst_pitch,
+		      const size_t *dst_stride, const void *src,
+		      const size_t *cumul_src_pitch, const size_t *src_stride,
+		      const size_t *elem_number, const size_t elem_size);
+/*
+ * Version of aml_copy_rtndstr using cumulative pitch.
+ */
+int aml_copy_rtndstr_c(size_t d, void *dst, const size_t *cumul_dst_pitch,
+		       const size_t *dst_stride, const void *src,
+		       const size_t *cumul_src_pitch, const size_t *src_stride,
+		       const size_t *elem_number, const size_t elem_size);
+/*
+ * Version of aml_copy_shndstr using cumulative pitch.
+ */
+int aml_copy_shndstr_c(size_t d, const size_t *target_dims, void *dst,
+		       const size_t *cumul_dst_pitch, const size_t *dst_stride,
+		       const void *src, const size_t *cumul_src_pitch,
+		       const size_t *src_stride, const size_t *elem_number,
+		       const size_t elem_size);
+
+ /*******************************************************************************
+ * Generic building block API: Native version
+ * Native means using AML-internal layouts.
+ ******************************************************************************/
+
+int aml_copy_layout_native(struct aml_layout *dst,
+			   const struct aml_layout *src);
+int aml_copy_layout_transform_native(struct aml_layout *dst,
+				     const struct aml_layout *src,
+				     const size_t *target_dims);
+int aml_copy_layout_generic(struct aml_layout *dst,
+			    const struct aml_layout *src);
+int aml_copy_layout_transform_generic(struct aml_layout *dst,
+				      const struct aml_layout *src,
+				      const size_t *target_dims);
+int aml_copy_layout_transpose_native(struct aml_layout *dst, const struct aml_layout *src);
+int aml_copy_layout_reverse_transpose_native(struct aml_layout *dst,
+					     const struct aml_layout *src);
+int aml_copy_layout_transpose_generic(struct aml_layout *dst, const struct aml_layout *src);
+int aml_copy_layout_reverse_transpose_generic(struct aml_layout *dst,
+					      const struct aml_layout *src);
+
+#endif

src/aml-dma-layout.h

@@ -0,0 +1,41 @@
+#ifndef AML_DMA_LAYOUT_H
+#define AML_DMA_LAYOUT_H 1
+
+/*******************************************************************************
+ * Layout aware DMA
+ * DMA using layouts as source and destination.
+ ******************************************************************************/
+
+extern struct aml_dma_ops aml_dma_ops_layout;
+
+struct aml_dma_request_layout {
+	int type;
+	struct aml_layout *dest;
+	struct aml_layout *src;
+};
+
+typedef int (*aml_dma_operator)(struct aml_layout *, struct aml_layout *, void*);
+struct aml_dma_layout {
+	struct aml_vector requests;
+	pthread_mutex_t lock;
+	aml_dma_operator do_work;
+	void *work_arg;
+};
+
+#define AML_DMA_LAYOUT_DECL(name) \
+	struct aml_dma_layout __ ##name## _inner_data; \
+	struct aml_dma name = { \
+		&aml_dma_ops_layout, \
+		(struct aml_dma_data *)&__ ## name ## _inner_data, \
+	};
+
+#define AML_DMA_LAYOUT_ALLOCSIZE \
+	(sizeof(struct aml_dma_layout) + \
+	 sizeof(struct aml_dma))
+
+int aml_dma_layout_create(struct aml_dma **dma, ...);
+int aml_dma_layout_init(struct aml_dma *dma, ...);
+int aml_dma_layout_vinit(struct aml_dma *dma, va_list args);
+int aml_dma_layout_destroy(struct aml_dma *dma);
+
+#endif