Revert "Use gdev_io_memcpy instead of using memcpy directly"

This reverts commit dc5b1ea.
shinpei0208 · Nov 19, 2013 · 0f37b44 · 0f37b44
1 parent 2bd96d7
commit 0f37b44
Show file tree

Hide file tree

Showing 16 changed files with 42 additions and 83 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -8,8 +8,6 @@ PROJECT(Gdev)
 SET(serial "1.0.0")
 SET(soserial "1")
 
-INCLUDE_DIRECTORIES(util)
-
 FIND_PACKAGE(Doxygen)
 include(UseDoxygen OPTIONAL)
 

diff --git a/common/gdev_api.c b/common/gdev_api.c
@@ -28,7 +28,6 @@
 
 #include "gdev_api.h"
 #include "gdev_device.h"
-#include "gdev_io_memcpy.h"
 #include "gdev_sched.h"
 
 #define __max(x, y) (x) > (y) ? (x) : (y)
@@ -85,7 +84,7 @@ static void __free_dma(gdev_mem_t **dma_mem, int p_count)
  */
 static int __f_memcpy(void *dst, const void *src, uint32_t size)
 {
-	gdev_io_memcpy(dst, src, size);
+	memcpy(dst, src, size);
 	return 0;
 }
 

diff --git a/common/gdev_system.h b/common/gdev_system.h
@@ -34,7 +34,6 @@
 #else
 #include "gdev_lib.h"
 #endif
-#include "gdev_io_memcpy.h"
 
 struct gdev_device; /* prototype declaration */
 
@@ -122,9 +121,9 @@ void gdev_next_compute(struct gdev_device *gdev);
 #define MB()
 #endif
 /* should never used */
-#define COPY_FROM_USER(dst, src, size) gdev_io_memcpy(dst, src, size)
+#define COPY_FROM_USER(dst, src, size) memcpy(dst, src, size) 
 /* should never used */
-#define COPY_TO_USER(dst, src, size) gdev_io_memcpy(dst, src, size)
+#define COPY_TO_USER(dst, src, size) memcpy(dst, src, size)
 #define IOREAD32(addr) *(uint32_t *)(addr)
 #define IOWRITE32(val, addr) *(uint32_t *)(addr) = val
 #endif

diff --git a/cuda/configure b/cuda/configure
@@ -30,11 +30,9 @@ if [ $target = 'libucuda' ] ; then
 	cp -rf ../driver/* .
 	cp -rf ../runtime/ .
 	cp -f ../cudump/* .
-    cp -f ../../util/* .
 elif [ $target = 'kcuda' ] ; then
 	cp -f ../kcuda/* .
 	cp -rf ../driver/* .
-    cp -f ../../util/* .
 	# we don't support Runtime API in kernel
 	runtime='0'
 else

diff --git a/cuda/driver/execution.c b/cuda/driver/execution.c
@@ -28,7 +28,6 @@
 #include "gdev_api.h"
 #include "gdev_autogen.h"
 #include "gdev_cuda.h"
-#include "gdev_io_memcpy.h"
 
 CUresult cuFuncGetAttribute
 (int *pi, CUfunction_attribute attrib, CUfunction hfunc) 
@@ -499,8 +498,8 @@ CUresult cuParamSetv
 
 	k = &func->kernel;
 	f = &func->raw_func;
-	gdev_io_memcpy(&k->param_buf[(f->param_base + offset) / 4], ptr, numbytes);
-
+	memcpy(&k->param_buf[(f->param_base + offset) / 4], ptr, numbytes);
+	
 	return CUDA_SUCCESS;
 }
 
diff --git a/cuda/driver/gdev_cuda.c b/cuda/driver/gdev_cuda.c
@@ -26,7 +26,6 @@
 
 #include "cuda.h"
 #include "gdev_cuda.h"
-#include "gdev_io_memcpy.h"
 #ifdef __KERNEL__
 #include <linux/errno.h>
 #else
@@ -964,7 +963,7 @@ CUresult gdev_cuda_construct_kernels
 		/* the following c[] setup is NVIDIA's nvcc-specific. */
 		k->cmem_count = GDEV_NVIDIA_CONST_SEGMENT_MAX_COUNT;
 		/* c0[] is a parameter list. */
-		gdev_io_memcpy(k->param_buf, f->cmem[0].buf, f->param_base);
+		memcpy(k->param_buf, f->cmem[0].buf, f->param_base);
 		k->cmem[0].size = gdev_cuda_align_cmem_size(f->param_size + cmem_size_align);
 		k->cmem[0].offset = 0;
 		for (i = 1; i < GDEV_NVIDIA_CONST_SEGMENT_MAX_COUNT; i++) {
@@ -1135,7 +1134,7 @@ CUresult gdev_cuda_memcpy_code(struct CUmod_st *mod, void *buf)
 	for (i = 0; i < GDEV_NVIDIA_CONST_SEGMENT_MAX_COUNT; i++) {
 		if (mod->cmem[i].buf) {
 			offset = mod->cmem[i].addr - addr;
-			gdev_io_memcpy(buf + offset, mod->cmem[i].buf, mod->cmem[i].raw_size);
+			memcpy(buf + offset, mod->cmem[i].buf, mod->cmem[i].raw_size);
 		}
 	}
 
@@ -1144,12 +1143,12 @@ CUresult gdev_cuda_memcpy_code(struct CUmod_st *mod, void *buf)
 		f = &func->raw_func;
 		if (f->code_buf) {
 			offset = k->code_addr - addr;
-			gdev_io_memcpy(buf + offset, f->code_buf, f->code_size);
+			memcpy(buf + offset, f->code_buf, f->code_size);
 		}
 		for (i = 0; i < GDEV_NVIDIA_CONST_SEGMENT_MAX_COUNT; i++) {
 			if (f->cmem[i].buf) {
 				offset = k->cmem[i].addr - addr;
-				gdev_io_memcpy(buf + offset, f->cmem[i].buf, f->cmem[i].size);
+				memcpy(buf + offset, f->cmem[i].buf, f->cmem[i].size);
 			}
 		}
 	}

diff --git a/lib/configure b/lib/configure
@@ -45,12 +45,10 @@ done
 if [ $target = 'kernel' ] ; then
 	# all files are driver-independent 
 	cp -f $topdir/$common/*.h .
-    cp -f $topdir/util/* .
 	cp -f ../kernel/* .
 elif [ $target = 'user' ] ; then
 	# copy driver-independent files
 	cp -f $topdir/$common/* .
-    cp -f $topdir/util/* .
 	cp -f ../user/gdev/* .
 	# detect which driver to be used
 	sh ./autogen.sh

diff --git a/lib/kernel/gdev_lib.h b/lib/kernel/gdev_lib.h
@@ -32,8 +32,6 @@
 #include <stdlib.h> /* malloc/free, etc. */
 #include <string.h> /* memcpy, etc. */
 
-#include "gdev_io_memcpy.h"
-
 #define GDEV_PRINT(fmt, arg...) fprintf(stderr, "[gdev] " fmt, ##arg)
 #define GDEV_DPRINT(fmt, arg...)					\
 	if (GDEV_DEBUG_PRINT)							\
@@ -44,7 +42,7 @@
 #define FREE(x) free(x)
 #define SCHED_YIELD() sched_yield()
 #define MB() //mb()
-#define COPY_FROM_USER(dst, src, size) gdev_io_memcpy(dst, src, size)
-#define COPY_TO_USER(dst, src, size) gdev_io_memcpy(dst, src, size)
+#define COPY_FROM_USER(dst, src, size) memcpy(dst, src, size)
+#define COPY_TO_USER(dst, src, size) memcpy(dst, src, size)
 
 #endif
diff --git a/lib/user/nouveau/nouveau_gdev.c b/lib/user/nouveau/nouveau_gdev.c
@@ -24,7 +24,6 @@
 
 #include "gdev_api.h"
 #include "gdev_device.h"
-#include "gdev_io_memcpy.h"
 #include "gdev_nvidia.h"
 #include "gdev_nvidia_fifo.h"
 #include "gdev_nvidia_nve4.h"
@@ -44,6 +43,16 @@ struct gdev_nouveau_ctx_objects {
 	struct nouveau_object *m2mf;
 };
 
+/* this ensures that SSE is not applied to memcpy. */
+void* __nouveau_io_memcpy(void* s1, const void* s2, size_t n)
+{
+    volatile char* out = (volatile char*)s1;
+    const volatile char* in = (const volatile char*)s2;
+    size_t i;
+    for (i = 0; i < n; ++i) out[i] = in[i];
+    return s1;
+}
+
 void __nouveau_fifo_space(struct gdev_ctx *ctx, uint32_t len)
 {
 	struct nouveau_pushbuf *push = (struct nouveau_pushbuf *)ctx->pctx;
@@ -645,7 +654,7 @@ int gdev_raw_read(struct gdev_mem *mem, void *buf, uint64_t addr, uint32_t size)
 	uint64_t offset = addr - bo->offset;
 
 	if (bo->map) {
-		gdev_io_memcpy(buf, bo->map + offset, size);
+		__nouveau_io_memcpy(buf, bo->map + offset, size);
 		return 0;
 	}
 	else {
@@ -660,7 +669,7 @@ int gdev_raw_write(struct gdev_mem *mem, uint64_t addr, const void *buf, uint32_
 	uint64_t offset = addr - bo->offset;
 
 	if (bo->map) {
-		gdev_io_memcpy(bo->map + offset, buf, size);
+		__nouveau_io_memcpy(bo->map + offset, buf, size);
 		return 0;
 	}
 	else {

diff --git a/lib/user/nvrm/nvrm_gdev.c b/lib/user/nvrm/nvrm_gdev.c
@@ -24,7 +24,6 @@
 
 #include "gdev_api.h"
 #include "gdev_device.h"
-#include "gdev_io_memcpy.h"
 #include "gdev_nvidia.h"
 #include "gdev_nvidia_fifo.h"
 #include "nvrm.h"
@@ -38,6 +37,16 @@
 struct gdev_device *lgdev; /* local gdev_device structure for user-space scheduling */
 static struct nvrm_context *nvrm_ctx = 0;
 
+/* this ensures that SSE is not applied to memcpy. */
+void* __nvrm_io_memcpy(void* s1, const void* s2, size_t n)
+{
+    volatile char* out = (volatile char*)s1;
+    const volatile char* in = (const volatile char*)s2;
+    size_t i;
+    for (i = 0; i < n; ++i) out[i] = in[i];
+    return s1;
+}
+
 int gdev_raw_query(struct gdev_device *gdev, uint32_t type, uint64_t *result)
 {
 #ifndef GDEV_SCHED_DISABLED/* fix this */
@@ -489,7 +498,7 @@ int gdev_raw_read(struct gdev_mem *mem, void *buf, uint64_t addr, uint32_t size)
 	}
 
 	uint64_t offset = addr - mem->addr;
-	gdev_io_memcpy(buf, ptr + offset, size);
+	__nvrm_io_memcpy(buf, ptr + offset, size);
 
 	if (!mem->map) {
 		nvrm_bo_host_unmap(bo);
@@ -508,7 +517,7 @@ int gdev_raw_write(struct gdev_mem *mem, uint64_t addr, const void *buf, uint32_
 	}
 
 	uint64_t offset = addr - mem->addr;
-	gdev_io_memcpy(ptr + offset, buf, size);
+	__nvrm_io_memcpy(ptr + offset, buf, size);
 
 	if (!mem->map) {
 		nvrm_bo_host_unmap(bo);

diff --git a/lib/user/pscnv/libpscnv.c b/lib/user/pscnv/libpscnv.c
@@ -5,7 +5,6 @@
 #include "drm.h"
 #include "libpscnv.h"
 #include "pscnv_drm.h"
-#include "gdev_io_memcpy.h"
 
 int drmIoctl(int fd, unsigned long request, void *arg)
 {
@@ -93,7 +92,7 @@ int pscnv_gem_new(int fd, uint32_t cookie, uint32_t flags, uint32_t tile_flags,
 	req.tile_flags = tile_flags;
 	req.size = size;
 	if (user)
-		gdev_io_memcpy(req.user, user, sizeof(req.user));
+		memcpy(req.user, user, sizeof(req.user));
 	ret = drmCommandWriteRead(fd, DRM_PSCNV_GEM_NEW, &req, sizeof(req));
 	if (ret)
 		return ret;
@@ -122,7 +121,7 @@ int pscnv_gem_info(int fd, uint32_t handle, uint32_t *cookie, uint32_t *flags, u
 	if (map_handle)
 		*map_handle = req.map_handle;
 	if (user)
-		gdev_io_memcpy(user, req.user, sizeof(req.user));
+		memcpy(user, req.user, sizeof(req.user));
 	return 0;
 }
 

diff --git a/lib/user/pscnv/pscnv_gdev.c b/lib/user/pscnv/pscnv_gdev.c
@@ -24,7 +24,6 @@
 
 #include "gdev_api.h"
 #include "gdev_device.h"
-#include "gdev_io_memcpy.h"
 #include "gdev_nvidia.h"
 #include "gdev_nvidia_fifo.h"
 #include "libpscnv.h"
@@ -375,7 +374,7 @@ int gdev_raw_read(struct gdev_mem *mem, void *buf, uint64_t addr, uint32_t size)
 	uint64_t offset = addr - bo->vm_base;
 
 	if (bo->map) {
-		gdev_io_memcpy(buf, bo->map + offset, size);
+		memcpy(buf, bo->map + offset, size);
 		return 0;
 	}
 	else
@@ -391,7 +390,7 @@ int gdev_raw_write(struct gdev_mem *mem, uint64_t addr, const void *buf, uint32_
 	uint64_t offset = addr - bo->vm_base;
 
 	if (bo->map) {
-		gdev_io_memcpy(bo->map + offset, buf, size);
+		memcpy(bo->map + offset, buf, size);
 		return 0;
 	}
 	else

diff --git a/lib/user/sched/gdev_api.c b/lib/user/sched/gdev_api.c
@@ -28,7 +28,6 @@
 
 #include "gdev_api.h"
 #include "gdev_device.h"
-#include "gdev_io_memcpy.h"
 #include "gdev_sched.h"
 
 #define __max(x, y) (x) > (y) ? (x) : (y)
@@ -85,7 +84,7 @@ static void __free_dma(gdev_mem_t **dma_mem, int p_count)
  */
 static int __f_memcpy(void *dst, const void *src, uint32_t size)
 {
-	gdev_io_memcpy(dst, src, size);
+	memcpy(dst, src, size);
 	return 0;
 }
 

diff --git a/mod/CMakeLists.txt b/mod/CMakeLists.txt
@@ -4,14 +4,13 @@ SET(soserial "1")
 include(UseDoxygen OPTIONAL)
 
 file(GLOB common_src "../common/*")
-file(GLOB util_src "../util/*")
 file(GLOB gmod_src "./gdev/*")
 
 set(target gdev_module)
 
-set(mod_src ${common_src} ${util_src} ${gmod_src})
+set(mod_src ${common_src} ${gmod_src})
 
-include_directories(../common/ ../util/ /usr/local/gdev/include ./gdev)
+include_directories(../common/ /usr/local/gdev/include ./gdev)
 
 SET( MODULE_NAME
     gdev

diff --git a/mod/configure b/mod/configure
@@ -23,7 +23,6 @@ done
 # copy driver-independent files
 cp -fr ../$gdev/* ./
 cp -fr ../../$common/* ./
-cp -fr ../../util/* ./
 
 # detect which driver to be used
 sh ./autogen.sh
diff --git a/util/gdev_io_memcpy.h b/util/gdev_io_memcpy.h