From: Luo Xionghu <xionghu....@intel.com>
Signed-off-by: Luo Xionghu <xionghu....@intel.com>
---
kernels/compiler_subgroup_image_block_read.cl | 115 ++++++++++++++
src/cl_extensions.h | 1 +
utests/CMakeLists.txt | 1 +
utests/compiler_subgroup_media_block_read.cpp | 212 ++++++++++++++++++++++++++
utests/utest_helper.cpp | 20 +++
utests/utest_helper.hpp | 3 +
6 files changed, 352 insertions(+)
create mode 100644 utests/compiler_subgroup_media_block_read.cpp
diff --git a/kernels/compiler_subgroup_image_block_read.cl
b/kernels/compiler_subgroup_image_block_read.cl
index fa079b7..596265f 100644
--- a/kernels/compiler_subgroup_image_block_read.cl
+++ b/kernels/compiler_subgroup_image_block_read.cl
@@ -62,3 +62,118 @@ __kernel void
compiler_subgroup_image_block_read_us8(image2d_t src, global ushor
dst[id] = tmp;
}
#endif
+#ifdef MEDIA_BLOCK_IO
+__kernel void compiler_subgroup_media_block_read_ui1(image2d_t src, global
uint *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(uint) +
sizeof(uint) * get_group_id(0) * get_local_size(0),yid);
+ uint tmp = intel_sub_group_media_block_read_ui(coord, 16, 1, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_ui2(image2d_t src, global
uint2 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(uint) +
sizeof(uint) * get_group_id(0) * get_local_size(0),yid*2);
+ uint2 tmp = intel_sub_group_media_block_read_ui2(coord, 16, 2, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_ui4(image2d_t src, global
uint4 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(uint) +
sizeof(uint) * get_group_id(0) * get_local_size(0),yid*4);
+ uint4 tmp = intel_sub_group_media_block_read_ui4(coord, 16, 4, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_ui8(image2d_t src, global
uint8 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(uint) +
sizeof(uint) * get_group_id(0) * get_local_size(0),yid*8);
+ uint8 tmp = intel_sub_group_media_block_read_ui8(coord, 16, 8, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_us1(image2d_t src, global
ushort *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(ushort) +
sizeof(ushort) * get_group_id(0) * get_local_size(0),yid);
+ ushort tmp = intel_sub_group_media_block_read_us(coord, 16, 1, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_us2(image2d_t src, global
ushort2 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(ushort) +
sizeof(ushort) * get_group_id(0) * get_local_size(0),yid*2);
+ ushort2 tmp = intel_sub_group_media_block_read_us2(coord, 16, 2, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_us4(image2d_t src, global
ushort4 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(ushort) +
sizeof(ushort) * get_group_id(0) * get_local_size(0),yid*4);
+ ushort4 tmp = intel_sub_group_media_block_read_us4(coord, 16, 4, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_us8(image2d_t src, global
ushort8 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(ushort) +
sizeof(ushort) * get_group_id(0) * get_local_size(0),yid*8);
+ ushort8 tmp = intel_sub_group_media_block_read_us8(coord, 16, 8, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void __attribute__((intel_reqd_sub_group_size(8)))
+compiler_subgroup_media_block_read_us16(image2d_t src, global ushort16 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(ushort) +
sizeof(ushort) * get_group_id(0) * get_local_size(0),yid*16);
+ ushort16 tmp = intel_sub_group_media_block_read_us16(coord, 8, 16, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_uc1(image2d_t src, global
uchar *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(char) +
sizeof(char) * get_group_id(0) * get_local_size(0),yid);
+ uchar tmp = intel_sub_group_media_block_read_uc(coord, 16, 1, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_uc2(image2d_t src, global
uchar2 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(char) +
sizeof(char) * get_group_id(0) * get_local_size(0),yid*2);
+ uchar2 tmp = intel_sub_group_media_block_read_uc2(coord, 16, 2, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_uc4(image2d_t src, global
uchar4 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(char) +
sizeof(char) * get_group_id(0) * get_local_size(0),yid*4);
+ uchar4 tmp = intel_sub_group_media_block_read_uc4(coord, 16, 4, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_uc8(image2d_t src, global
uchar8 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(char) +
sizeof(char) * get_group_id(0) * get_local_size(0),yid*8);
+ uchar8 tmp = intel_sub_group_media_block_read_uc8(coord, 16, 8, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+__kernel void compiler_subgroup_media_block_read_uc16(image2d_t src, global
uchar16 *dst)
+{
+ int id = get_global_id(0);
+ int yid = get_global_id(1);
+ int2 coord = (int2)(get_sub_group_size()*get_sub_group_id()*sizeof(char) +
sizeof(char) * get_group_id(0) * get_local_size(0),yid*16);
+ uchar16 tmp = intel_sub_group_media_block_read_uc16(coord, 16, 16, src);
+ dst[yid * get_global_size(0) + id] = tmp;
+}
+#endif
diff --git a/src/cl_extensions.h b/src/cl_extensions.h
index 55747a7..af0855e 100644
--- a/src/cl_extensions.h
+++ b/src/cl_extensions.h
@@ -30,6 +30,7 @@
DECL_EXT(intel_motion_estimation) \
DECL_EXT(intel_subgroups) \
DECL_EXT(intel_subgroups_short) \
+ DECL_EXT(intel_media_block_io) \
DECL_EXT(intel_planar_yuv)
#define DECL_GL_EXTENSIONS \
diff --git a/utests/CMakeLists.txt b/utests/CMakeLists.txt
index f7a9046..736b21c 100644
--- a/utests/CMakeLists.txt
+++ b/utests/CMakeLists.txt
@@ -178,6 +178,7 @@ set (utests_sources
compiler_subgroup_buffer_block_write.cpp
compiler_subgroup_image_block_read.cpp
compiler_subgroup_image_block_write.cpp
+ compiler_subgroup_media_block_read.cpp
compiler_async_stride_copy.cpp
compiler_insn_selection_min.cpp
compiler_insn_selection_max.cpp
diff --git a/utests/compiler_subgroup_media_block_read.cpp
b/utests/compiler_subgroup_media_block_read.cpp
new file mode 100644
index 0000000..41289a2
--- /dev/null
+++ b/utests/compiler_subgroup_media_block_read.cpp
@@ -0,0 +1,212 @@
+#include <cstdint>
+#include <cstring>
+#include <iostream>
+#include "utest_helper.hpp"
+
+using namespace std;
+
+/* set to 1 for debug, output of input-expected data */
+#define DEBUG_STDOUT 0
+
+/* NDRANGE */
+#define WG_GLOBAL_SIZE_X 64
+#define WG_GLOBAL_SIZE_Y 1
+#define WG_LOCAL_SIZE_X 64
+#define WG_LOCAL_SIZE_Y 1
+/*
+ * Generic compute-expected function for meida block read
+ */
+template<class T>
+static void compute_expected_media(T* input,
+ T* expected,
+ size_t VEC_SIZE)
+{
+ for(uint32_t w = 0; w < WG_GLOBAL_SIZE_X; ++w) {
+ for (uint32_t h = 0; h < VEC_SIZE * WG_GLOBAL_SIZE_Y; ++h) {
+ expected[h + w * VEC_SIZE * WG_GLOBAL_SIZE_Y] = w;
+ }
+ }
+}
+
+/*
+ * Generic input-expected generate function for media block read
+ */
+template<class T>
+static void generate_data_media(T* &input,
+ T* &expected,
+ size_t VEC_SIZE)
+{
+ /* allocate input and expected arrays */
+ int* input_ui = new int[WG_GLOBAL_SIZE_X * WG_GLOBAL_SIZE_Y * VEC_SIZE];
+ input = (T*)input_ui;
+ expected = new T[WG_GLOBAL_SIZE_X * WG_GLOBAL_SIZE_Y * VEC_SIZE];
+
+ /* seed for random inputs */
+ srand (time(NULL));
+
+#if DEBUG_STDOUT
+ cout << endl << "IN: " << endl;
+#endif
+ uint32_t rpitch = sizeof(uint32_t) * WG_GLOBAL_SIZE_X / sizeof(T);
+ /* generate inputs and expected values */
+ for (uint32_t h = 0; h < VEC_SIZE * WG_GLOBAL_SIZE_Y; ++h) {
+ for(uint32_t w = 0; w < WG_GLOBAL_SIZE_X; ++w)
+ {
+ /* initially 0, augment after */
+ input[w + h * rpitch] = w;
+
+#if DEBUG_STDOUT
+ /* output generated input */
+ cout << setw(4) << (uint32_t)input[w + h * rpitch] << ", " ;
+ if((w+ 1) % 8 == 0)
+ cout << endl;
+#endif
+ }
+ }
+ /* expected values */
+ compute_expected_media(input, expected, VEC_SIZE);
+
+#if DEBUG_STDOUT
+ /* output expected input */
+ cout << endl << "EXP: " << endl;
+ for(uint32_t gid = 0; gid < WG_GLOBAL_SIZE_X; gid++)
+ {
+ cout << "(";
+ for(uint32_t vsz = 0; vsz < VEC_SIZE * WG_GLOBAL_SIZE_Y; vsz++)
+ cout << setw(4) << (uint32_t)expected[gid* VEC_SIZE * WG_GLOBAL_SIZE_Y +
vsz] << ", " ;
+ cout << ")";
+ if((gid + 1) % 8 == 0)
+ cout << endl;
+ cout << endl;
+ }
+#endif
+}
+
+/*
+ * Generic subgroup utest function for media block read
+ */
+template<class T>
+static void subgroup_generic_media(T* input,
+ T* expected,
+ size_t VEC_SIZE,
+ const char* kernel_name)
+{
+ OCL_CALL(cl_kernel_init, "compiler_subgroup_image_block_read.cl",
kernel_name,
+ SOURCE, "-DMEDIA_BLOCK_IO");
+
+ cl_image_format format;
+ cl_image_desc desc;
+
+ memset(&desc, 0x0, sizeof(cl_image_desc));
+ memset(&format, 0x0, sizeof(cl_image_format));
+
+ /* get simd size */
+ globals[0] = WG_GLOBAL_SIZE_X;
+ globals[1] = WG_GLOBAL_SIZE_Y;
+ locals[0] = WG_LOCAL_SIZE_X;
+ locals[1] = WG_LOCAL_SIZE_Y;
+ size_t SIMD_SIZE = 0;
+
OCL_CALL(utestclGetKernelSubGroupInfoKHR,kernel,device,CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE_KHR,sizeof(size_t)*1,locals,sizeof(size_t),&SIMD_SIZE,NULL);
+
+ if(SIMD_SIZE != 16) {
+ printf("skip! this case could only run under simd16 mode!\n");
+ return;
+ }
+
+ size_t buf_sz = VEC_SIZE * WG_GLOBAL_SIZE_X * WG_GLOBAL_SIZE_Y;
+ /* input and expected data */
+ generate_data_media(input, expected, VEC_SIZE);
+
+ /* prepare input for datatype */
+ format.image_channel_order = CL_R;
+ format.image_channel_data_type = CL_UNSIGNED_INT32;
+ desc.image_type = CL_MEM_OBJECT_IMAGE2D;
+ desc.image_width = WG_GLOBAL_SIZE_X;
+ desc.image_height = VEC_SIZE * WG_GLOBAL_SIZE_Y;
+ desc.image_row_pitch = WG_GLOBAL_SIZE_X * sizeof(uint32_t);
+
+ OCL_CREATE_IMAGE(buf[0], CL_MEM_COPY_HOST_PTR, &format, &desc, input);
+ OCL_CREATE_BUFFER(buf[1], 0, buf_sz * sizeof(T), NULL);
+
+ OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
+ OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]);
+
+ /* run the kernel on GPU */
+ OCL_NDRANGE(2);
+
+ /* check if mismatch */
+ OCL_MAP_BUFFER(1);
+ uint32_t mismatches = 0;
+
+ for (uint32_t i = 0; i < buf_sz; i++)
+ if(((T *)buf_data[1])[i] != *(expected + i))
+ {
+ /* found mismatch, increment */
+ mismatches++;
+
+#if DEBUG_STDOUT
+ /* output mismatch */
+ cout << "Err at " << i << ", " <<
+ (uint32_t)((T *)buf_data[1])[i] << " != " << (uint32_t)*(expected + i)
<< endl;
+#endif
+ }
+
+#if DEBUG_STDOUT
+ /* output mismatch count */
+ cout << "mismatches " << mismatches << endl;
+#endif
+
+ OCL_UNMAP_BUFFER(1);
+
+ OCL_ASSERT(mismatches == 0);
+ free(input);
+ free(expected);
+}
+
+#define compiler_subgroup_media_block_read_ui(name, kernel, vec_size) \
+void name(void) \
+{ \
+ if(!cl_check_media_block_io()) \
+ return; \
+ cl_uint *input = NULL; \
+ cl_uint *expected = NULL; \
+ subgroup_generic_media(input, expected, vec_size, #kernel); \
+} \
+MAKE_UTEST_FROM_FUNCTION(name);
+compiler_subgroup_media_block_read_ui(compiler_subgroup_media_block_read_ui1,
compiler_subgroup_media_block_read_ui1, 1)
+compiler_subgroup_media_block_read_ui(compiler_subgroup_media_block_read_ui2,
compiler_subgroup_media_block_read_ui2, 2)
+compiler_subgroup_media_block_read_ui(compiler_subgroup_media_block_read_ui4,
compiler_subgroup_media_block_read_ui4, 4)
+compiler_subgroup_media_block_read_ui(compiler_subgroup_media_block_read_ui8,
compiler_subgroup_media_block_read_ui8, 8)
+
+#define compiler_subgroup_media_block_read_us(name, kernel, vec_size) \
+void name(void) \
+{ \
+ if(!cl_check_media_block_io()) \
+ return; \
+ cl_ushort *input = NULL; \
+ cl_ushort *expected = NULL; \
+ subgroup_generic_media(input, expected, vec_size, #kernel); \
+} \
+MAKE_UTEST_FROM_FUNCTION(name);
+compiler_subgroup_media_block_read_us(compiler_subgroup_media_block_read_us1,
compiler_subgroup_media_block_read_us1, 1)
+compiler_subgroup_media_block_read_us(compiler_subgroup_media_block_read_us2,
compiler_subgroup_media_block_read_us2, 2)
+compiler_subgroup_media_block_read_us(compiler_subgroup_media_block_read_us4,
compiler_subgroup_media_block_read_us4, 4)
+compiler_subgroup_media_block_read_us(compiler_subgroup_media_block_read_us8,
compiler_subgroup_media_block_read_us8, 8)
+//the us16 case could only run under SIMD8 mode.
+//compiler_subgroup_media_block_read_us(compiler_subgroup_media_block_read_us16,
compiler_subgroup_media_block_read_us16, 16)
+
+#define compiler_subgroup_media_block_read_uc(name, kernel, vec_size) \
+void name(void) \
+{ \
+ if(!cl_check_media_block_io()) \
+ return; \
+ cl_uchar *input = NULL; \
+ cl_uchar *expected = NULL; \
+ subgroup_generic_media(input, expected, vec_size, #kernel); \
+} \
+MAKE_UTEST_FROM_FUNCTION(name);
+compiler_subgroup_media_block_read_uc(compiler_subgroup_media_block_read_uc1,
compiler_subgroup_media_block_read_uc1, 1)
+compiler_subgroup_media_block_read_uc(compiler_subgroup_media_block_read_uc2,
compiler_subgroup_media_block_read_uc2, 2)
+compiler_subgroup_media_block_read_uc(compiler_subgroup_media_block_read_uc4,
compiler_subgroup_media_block_read_uc4, 4)
+compiler_subgroup_media_block_read_uc(compiler_subgroup_media_block_read_uc8,
compiler_subgroup_media_block_read_uc8, 8)
+compiler_subgroup_media_block_read_uc(compiler_subgroup_media_block_read_uc16,
compiler_subgroup_media_block_read_uc16, 16)
diff --git a/utests/utest_helper.cpp b/utests/utest_helper.cpp
index 7052a14..f4487c1 100644
--- a/utests/utest_helper.cpp
+++ b/utests/utest_helper.cpp
@@ -939,6 +939,26 @@ int cl_check_subgroups_short(void)
return 1;
}
+int cl_check_media_block_io(void)
+{
+ if (!cl_check_subgroups())
+ return 0;
+ std::string extStr;
+ size_t param_value_size;
+ OCL_CALL(clGetDeviceInfo, device, CL_DEVICE_EXTENSIONS, 0, 0,
¶m_value_size);
+ std::vector<char> param_value(param_value_size);
+ OCL_CALL(clGetDeviceInfo, device, CL_DEVICE_EXTENSIONS, param_value_size,
+ param_value.empty() ? NULL : ¶m_value.front(),
¶m_value_size);
+ if (!param_value.empty())
+ extStr = std::string(¶m_value.front(), param_value_size-1);
+
+ if (std::strstr(extStr.c_str(), "cl_intel_media_block_io") == NULL) {
+ printf("No cl_intel_media_block_io, Skip!");
+ return 0;
+ }
+ return 1;
+}
+
int cl_check_ocl20(bool or_beignet)
{
size_t param_value_size;
diff --git a/utests/utest_helper.hpp b/utests/utest_helper.hpp
index e2a6a88..5dc381e 100644
--- a/utests/utest_helper.hpp
+++ b/utests/utest_helper.hpp
@@ -303,6 +303,9 @@ extern int cl_check_beignet(void);
/* Check is intel subgroups enabled. */
extern int cl_check_subgroups(void);
+/* Check is intel_media_block_io enabled. */
+extern int cl_check_media_block_io(void);
+
typedef cl_int(clGetKernelSubGroupInfoKHR_cb)(cl_kernel, cl_device_id,
cl_kernel_sub_group_info, size_t,
const void *, size_t, void *,
--
2.5.0
_______________________________________________
Beignet mailing list
Beignet@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/beignet
