[Beignet] [PATCH 2/2] Fix bug of clEnqueueCopyBufferToImage and clEnqueueCopyImageToBuffer.
yan.wang at linux.intel.com
yan.wang at linux.intel.com
Thu May 25 02:57:06 UTC 2017
From: Yan Wang <yan.wang at linux.intel.com>
"imagedim_non_pow_2" cases of basic modudle of confrmance shows
regression after use TILE_Y mode for large image by previous patch.
This bug comes from the non-align16 kernel of clEnqueueCopyBufferToImage
and clEnqueueCopyImageToBuffer.
It will force CL_RGBA/CL_UNORM_INT8/8191x8192 image of conformance test
to CL_R/CL_UNSIGNED_INT8/32764x8192 image for copying.
So it makes width as 8191 x 4 = 32764 and its width will exceed the maximum
width (16 x 1024 = 16384) of GEN surface state structure which only has 14 bits.
So use align4 copy kernel to avoid this bug.
Signed-off-by: Yan Wang <yan.wang at linux.intel.com>
---
src/CMakeLists.txt | 1 +
src/cl_context.h | 2 ++
src/cl_mem.c | 32 ++++++++++++++++++++++
.../cl_internal_copy_buffer_to_image_2d_align4.cl | 18 ++++++++++++
.../cl_internal_copy_image_2d_to_buffer_align4.cl | 18 ++++++++++++
5 files changed, 71 insertions(+)
create mode 100644 src/kernels/cl_internal_copy_buffer_to_image_2d_align4.cl
create mode 100644 src/kernels/cl_internal_copy_image_2d_to_buffer_align4.cl
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 77a1c87..6433566 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -53,6 +53,7 @@ cl_internal_copy_image_2d_array_to_2d_array cl_internal_copy_image_2d_array_to_2
cl_internal_copy_image_2d_array_to_3d cl_internal_copy_image_3d_to_2d_array
cl_internal_copy_image_2d_to_buffer cl_internal_copy_image_2d_to_buffer_align16 cl_internal_copy_image_3d_to_buffer
cl_internal_copy_buffer_to_image_2d cl_internal_copy_buffer_to_image_2d_align16 cl_internal_copy_buffer_to_image_3d
+cl_internal_copy_buffer_to_image_2d_align4 cl_internal_copy_image_2d_to_buffer_align4
cl_internal_fill_buf_align8 cl_internal_fill_buf_align4
cl_internal_fill_buf_align2 cl_internal_fill_buf_unalign
cl_internal_fill_buf_align128 cl_internal_fill_image_1d
diff --git a/src/cl_context.h b/src/cl_context.h
index 8ba499f..75bf895 100644
--- a/src/cl_context.h
+++ b/src/cl_context.h
@@ -62,9 +62,11 @@ enum _cl_internal_ker_type {
CL_ENQUEUE_COPY_IMAGE_3D_TO_2D_ARRAY, //copy image 3d to image 2d array
CL_ENQUEUE_COPY_IMAGE_2D_TO_BUFFER, //copy image 2d to buffer
CL_ENQUEUE_COPY_IMAGE_2D_TO_BUFFER_ALIGN16,
+ CL_ENQUEUE_COPY_IMAGE_2D_TO_BUFFER_ALIGN4,
CL_ENQUEUE_COPY_IMAGE_3D_TO_BUFFER, //copy image 3d tobuffer
CL_ENQUEUE_COPY_BUFFER_TO_IMAGE_2D, //copy buffer to image 2d
CL_ENQUEUE_COPY_BUFFER_TO_IMAGE_2D_ALIGN16,
+ CL_ENQUEUE_COPY_BUFFER_TO_IMAGE_2D_ALIGN4,
CL_ENQUEUE_COPY_BUFFER_TO_IMAGE_3D, //copy buffer to image 3d
CL_ENQUEUE_FILL_BUFFER_UNALIGN, //fill buffer with 1 aligne pattern, pattern size=1
CL_ENQUEUE_FILL_BUFFER_ALIGN2, //fill buffer with 2 aligne pattern, pattern size=2
diff --git a/src/cl_mem.c b/src/cl_mem.c
index 0c49c3d..3b9a3be 100644
--- a/src/cl_mem.c
+++ b/src/cl_mem.c
@@ -2159,6 +2159,7 @@ cl_mem_copy_image_to_buffer(cl_command_queue queue, cl_event event, struct _cl_m
size_t origin0, region0;
size_t kn_dst_offset;
int align16 = 0;
+ int align4 = 0;
size_t align_size = 1;
size_t w_saved;
@@ -2183,6 +2184,13 @@ cl_mem_copy_image_to_buffer(cl_command_queue queue, cl_event event, struct _cl_m
align16 = 1;
align_size = 16;
}
+ else if((image->image_type == CL_MEM_OBJECT_IMAGE2D) && ((image->w * image->bpp) % 4 == 0) &&
+ ((src_origin[0] * bpp) % 4 == 0) && (region0 % 4 == 0) && (dst_offset % 4 == 0)){
+ fmt.image_channel_order = CL_R;
+ fmt.image_channel_data_type = CL_UNSIGNED_INT32;
+ align4 = 1;
+ align_size = 4;
+ }
else{
fmt.image_channel_order = CL_R;
fmt.image_channel_data_type = CL_UNSIGNED_INT8;
@@ -2206,6 +2214,14 @@ cl_mem_copy_image_to_buffer(cl_command_queue queue, cl_event event, struct _cl_m
cl_internal_copy_image_2d_to_buffer_align16_str,
(size_t)cl_internal_copy_image_2d_to_buffer_align16_str_size, NULL);
}
+ else if(align4){
+ extern char cl_internal_copy_image_2d_to_buffer_align4_str[];
+ extern size_t cl_internal_copy_image_2d_to_buffer_align4_str_size;
+
+ ker = cl_context_get_static_kernel_from_bin(queue->ctx, CL_ENQUEUE_COPY_IMAGE_2D_TO_BUFFER_ALIGN4,
+ cl_internal_copy_image_2d_to_buffer_align4_str,
+ (size_t)cl_internal_copy_image_2d_to_buffer_align4_str_size, NULL);
+ }
else{
extern char cl_internal_copy_image_2d_to_buffer_str[];
extern size_t cl_internal_copy_image_2d_to_buffer_str_size;
@@ -2263,6 +2279,7 @@ cl_mem_copy_buffer_to_image(cl_command_queue queue, cl_event event, cl_mem buffe
size_t origin0, region0;
size_t kn_src_offset;
int align16 = 0;
+ int align4 = 0;
size_t align_size = 1;
size_t w_saved = 0;
@@ -2287,6 +2304,13 @@ cl_mem_copy_buffer_to_image(cl_command_queue queue, cl_event event, cl_mem buffe
align16 = 1;
align_size = 16;
}
+ else if((image->image_type == CL_MEM_OBJECT_IMAGE2D) && ((image->w * image->bpp) % 4 == 0) &&
+ ((dst_origin[0] * bpp) % 4 == 0) && (region0 % 4 == 0) && (src_offset % 4 == 0)){
+ fmt.image_channel_order = CL_R;
+ fmt.image_channel_data_type = CL_UNSIGNED_INT32;
+ align4 = 1;
+ align_size = 4;
+ }
else{
fmt.image_channel_order = CL_R;
fmt.image_channel_data_type = CL_UNSIGNED_INT8;
@@ -2310,6 +2334,14 @@ cl_mem_copy_buffer_to_image(cl_command_queue queue, cl_event event, cl_mem buffe
cl_internal_copy_buffer_to_image_2d_align16_str,
(size_t)cl_internal_copy_buffer_to_image_2d_align16_str_size, NULL);
}
+ else if(align4){
+ extern char cl_internal_copy_buffer_to_image_2d_align4_str[];
+ extern size_t cl_internal_copy_buffer_to_image_2d_align4_str_size;
+
+ ker = cl_context_get_static_kernel_from_bin(queue->ctx, CL_ENQUEUE_COPY_BUFFER_TO_IMAGE_2D_ALIGN4,
+ cl_internal_copy_buffer_to_image_2d_align4_str,
+ (size_t)cl_internal_copy_buffer_to_image_2d_align4_str_size, NULL);
+ }
else{
extern char cl_internal_copy_buffer_to_image_2d_str[];
extern size_t cl_internal_copy_buffer_to_image_2d_str_size;
diff --git a/src/kernels/cl_internal_copy_buffer_to_image_2d_align4.cl b/src/kernels/cl_internal_copy_buffer_to_image_2d_align4.cl
new file mode 100644
index 0000000..79a3d8c
--- /dev/null
+++ b/src/kernels/cl_internal_copy_buffer_to_image_2d_align4.cl
@@ -0,0 +1,18 @@
+kernel void __cl_copy_buffer_to_image_2d_align4(__write_only image2d_t image, global uint* buffer,
+ unsigned int region0, unsigned int region1, unsigned int region2,
+ unsigned int dst_origin0, unsigned int dst_origin1, unsigned int dst_origin2,
+ unsigned int src_offset)
+{
+ int i = get_global_id(0);
+ int j = get_global_id(1);
+ uint4 color = (uint4)(0);
+ int2 dst_coord;
+ if((i >= region0) || (j>= region1))
+ return;
+ dst_coord.x = dst_origin0 + i;
+ dst_coord.y = dst_origin1 + j;
+ src_offset += j * region0 + i;
+ color.x = buffer[src_offset];
+ write_imageui(image, dst_coord, color.x);
+}
+
diff --git a/src/kernels/cl_internal_copy_image_2d_to_buffer_align4.cl b/src/kernels/cl_internal_copy_image_2d_to_buffer_align4.cl
new file mode 100644
index 0000000..dc76e02
--- /dev/null
+++ b/src/kernels/cl_internal_copy_image_2d_to_buffer_align4.cl
@@ -0,0 +1,18 @@
+kernel void __cl_copy_image_2d_to_buffer_align4( __read_only image2d_t image, global uint* buffer,
+ unsigned int region0, unsigned int region1, unsigned int region2,
+ unsigned int src_origin0, unsigned int src_origin1, unsigned int src_origin2,
+ unsigned int dst_offset)
+{
+
+ int i = get_global_id(0);
+ int j = get_global_id(1);
+ if((i >= region0) || (j>= region1))
+ return;
+ uint4 color;
+ const sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
+ int2 src_coord;
+ src_coord.x = src_origin0 + i;
+ src_coord.y = src_origin1 + j;
+ color = read_imageui(image, sampler, src_coord);
+ *(buffer + dst_offset + region0*j + i) = color.x;
+}
--
2.7.4
More information about the Beignet
mailing list