AM5728: memcpy from dmabuf is slow

yes. cmem buf alloc code is copy from dual-camera-demo.

attached is the patch to dmabuftest for argb texture.

root@am57xx-evm:~# ./dmabuftest --kmscube --connector 35 -d /dev/video1 -c 1280x720@NV12

...

CMEM buffer pointer is 0xb2147000
PVR:(Error): [ 8660-> 8661] < gbm_pvr_bo_create_common():217|ERROR> condition [ bo_attrs->opt.import_dmabuf.stride != 0] failed [0, ]
ERROR:create_texture:1232: gbm_bo_import failed

diff -ur omapdrmtest/Makefile omapdrmtest.fbo/Makefile
--- omapdrmtest/Makefile	2019-03-25 20:38:48.951233612 +0800
+++ omapdrmtest.fbo/Makefile	2019-03-25 20:14:28.368984221 +0800
@@ -353,7 +353,7 @@
 INSTALL_SCRIPT = ${INSTALL}
 INSTALL_STRIP_PROGRAM = $(install_sh) -c -s
 LD = /media/y/7b0dae3b-6eb2-40a5-814a-23b19833a4a1/ti-processor-sdk-linux-am57xx-evm-05.02.00.10/linux-devkit/sysroots/x86_64-arago-linux/usr/arm-linux-gnueabihf/bin/ld
-LDFLAGS = -lm -lavcodec -lpthread
+LDFLAGS = -lticmem -lm -lavcodec -lpthread
 LIBOBJS = 
 LIBS = 
 LIBTOOL = $(SHELL) $(top_builddir)/libtool
diff -ur omapdrmtest/util/display-kmscube.c omapdrmtest.fbo/util/display-kmscube.c
--- omapdrmtest/util/display-kmscube.c	2019-03-25 20:37:45.285014307 +0800
+++ omapdrmtest.fbo/util/display-kmscube.c	2019-03-25 21:09:28.823936500 +0800
@@ -77,6 +77,53 @@
 #include <pthread.h>
 
 #include "esUtil.h"
+#include <ti/cmem.h>
+#include <linux/dma-buf.h>
+#include <sys/ioctl.h>
+
+#define CMEM_BLOCKID CMEM_CMABLOCKID
+
+CMEM_AllocParams cmem_alloc_params = {
+	CMEM_HEAP,	/* type */
+	CMEM_CACHED,	/* flags */
+	1		/* alignment */
+};
+
+void init_cmem()
+{
+	CMEM_init();
+}
+
+int alloc_cmem_buffer(unsigned int size, unsigned int align, void **cmem_buf)
+{
+	cmem_alloc_params.alignment = align;
+
+	*cmem_buf = CMEM_alloc2(CMEM_BLOCKID, size,
+		&cmem_alloc_params);
+
+	if(*cmem_buf == NULL){
+		printf("CMEM allocation failed");
+		return -1;
+	}
+	printf("CMEM buffer pointer is 0x%x\n",(uint32_t) *cmem_buf );
+	return CMEM_export_dmabuf(*cmem_buf);
+}
+
+void free_cmem_buffer(void *cmem_buffer)
+{
+	CMEM_free(cmem_buffer, &cmem_alloc_params);
+}
+
+int dma_buf_do_cache_operation(int dma_buf_fd, uint32_t cache_operation) 
+{
+	int ret;
+	struct dma_buf_sync sync;
+	sync.flags = cache_operation;
+
+	ret = ioctl(dma_buf_fd, DMA_BUF_IOCTL_SYNC, &sync);
+
+	return ret;
+}
 
 #define MAX_FACES 6
 
@@ -268,11 +315,15 @@
 	return 0;
 }
 
+static struct buffer * alloc_buffer(struct display *disp, uint32_t fourcc, uint32_t w, uint32_t h);
+static struct buffer *fbo_buffer = NULL;
+static int fbo_texId;
 static int init_gl(struct display_kmscube *disp_kmsc)
 {
 	EGLint major, minor, n;
 	GLuint vertex_shader, fragment_shader;
 	GLint ret;
+	struct buffer_kmscube *buf_kmsc;
 
 	static const EGLint context_attribs[] = {
 		EGL_CONTEXT_CLIENT_VERSION, 2,
@@ -397,6 +448,13 @@
 		return -1;
 	}
 
+	fbo_buffer = alloc_buffer(&disp_kmsc->base, FOURCC_STR("AR24"), 640, 720);
+	buf_kmsc = to_buffer_kmscube(fbo_buffer);
+	create_texture(disp_kmsc, fbo_buffer);
+	fbo_texId = buf_kmsc->texture_name;
+	fprintf(stderr,"fbo_buffer = %x, fbo_texId = %x fbo_addr = %x\n",fbo_buffer,fbo_texId,fbo_buffer->cmem_buf);
+	
+
 	vertex_shader = glCreateShader(GL_VERTEX_SHADER);
 
 	glShaderSource(vertex_shader, 1, &vertex_shader_source, NULL);
@@ -804,14 +862,29 @@
 
 	switch(fourcc) {
 	case FOURCC('A','R','2','4'):
-		buf->nbo = 1;
+#if 0 
 		buf->bo[0] = alloc_bo(disp, 32, buf->width, buf->height,
 				&bo_handles[0], &buf->pitches[0]);
+                buf->fd[0] = omap_bo_dmabuf(buf->bo[0]);
+#else
+		//Allocate buffer from CMEM and get the buffer descriptor
+		buf->fd[0]  = alloc_cmem_buffer(buf->width*buf->height*32, 1, &buf->cmem_buf);
+		if(buf->fd[0] < 0){
+			free_cmem_buffer(buf->cmem_buf);
+			printf(" Cannot export CMEM buffer\n");
+			return NULL;
+		}
+
+		/* Get the omap bo from the fd allocted using CMEM */
+		buf->bo[0] = omap_bo_from_dmabuf(disp->dev, buf->fd[0]);
+		bo_handles[0] = omap_bo_handle(buf->bo[0]);
+#endif
 		break;
 	case FOURCC('R','G','2','4'):
 		buf->nbo = 1;
 		buf->bo[0] = alloc_bo(disp, 24, buf->width, buf->height,
 				&bo_handles[0], &buf->pitches[0]);
+                buf->fd[0] = omap_bo_dmabuf(buf->bo[0]);
 		break;
 
 	case FOURCC('U','Y','V','Y'):
@@ -819,6 +892,7 @@
 		buf->nbo = 1;
 		buf->bo[0] = alloc_bo(disp, 16, buf->width, buf->height,
 				&bo_handles[0], &buf->pitches[0]);
+                buf->fd[0] = omap_bo_dmabuf(buf->bo[0]);
 		break;
 	case FOURCC('N','V','1','2'):
                 if (disp->multiplanar) {
@@ -844,6 +918,7 @@
 		buf->nbo = 1;
 		buf->bo[0] = alloc_bo(disp, 8, buf->width, (buf->height + buf->height/2),
 				&bo_handles[0], &buf->pitches[0]);
+                buf->fd[0] = omap_bo_dmabuf(buf->bo[0]);
 		break;
 	default:
 		ERROR("invalid format: 0x%08x", fourcc);
@@ -1093,7 +1168,8 @@
 	// TODO: cropping attributes when this will be supported.
 	EGLint attr[20];
 	bool isRGB;
-	int dfd = omap_bo_dmabuf(buf->bo[0]);
+	//int dfd = omap_bo_dmabuf(buf->bo[0]);
+	int dfd = buf->fd[0];
 	struct gbm_import_fd_data gbm_dmabuf = {
 		.fd     = dfd,
 		.width  = buf->width,
@@ -1289,6 +1365,7 @@
 		goto fail;
 	}
 
+	init_cmem();
 	disp->get_buffers = get_buffers;
 	disp->get_vid_buffers = get_vid_buffers;
 	disp->post_buffer = post_buffer;
diff -ur omapdrmtest/util/util.h omapdrmtest.fbo/util/util.h
--- omapdrmtest/util/util.h	2019-03-25 20:35:30.912001817 +0800
+++ omapdrmtest.fbo/util/util.h	2019-03-25 20:25:06.110157762 +0800
@@ -59,6 +59,7 @@
 	bool multiplanar;	/* True when Y and U/V are in separate buffers. */
 	int fd[4];		/* dmabuf */
 	bool noScale;
+	void *cmem_buf;
 };
 

Hello,

From the error, it seems like stride[] is zero. Please check if stride is set.

BR
Margarita

Many .so no source code available,I can't find where to change the stride.

No struct info where the ->stride is ?