AI/rocm-systems
2
0
派生 0
代码 工单 合并请求 工作流 软件包 项目 发布 百科 活动

Move Images code to hsa-runtime folder

浏览代码 
Change-Id: I53c1845d985ac3e9708d952865009c0021f3bb4f


[ROCm/ROCR-Runtime commit: 7e3db20826]
这个提交包含在：
Sean Keely
2020-04-30 00:18:36 -05:00
父节点 bd5ef0eff8
当前提交 1fc7f2dec7
共修改 70 个文件，包含 73713 行新增和 47 行删除
下载 Patch 文件 下载 Diff 文件 展开所有文件 折叠所有文件
projects/rocr-runtime/runtime/hsa-runtime/image/image_manager_nv.cpp
可执行文件
+746
查看文件
@@ -0,0 +1,746 @@
#define NOMINMAX
#include "image_manager_nv.h"
#include <assert.h>
#include <algorithm>
#include <climits>
#include "inc/hsa_ext_amd.h"
#include "addrlib/src/core/addrlib.h"
#include "image_runtime.h"
#include "resource.h"
#include "resource_nv.h"
#include "util.h"
#include "device_info.h"
namespace amd {
//-----------------------------------------------------------------------------
// Workaround switch to combined format/type codes and missing gfx10
// specific look up table. Only covers types used in image_lut_kv.cpp.
//-----------------------------------------------------------------------------
struct formatconverstion_t {
FMT fmt;
type type;
FORMAT format;
};
// Format/Type to combined format code table.
// Sorted and indexed to allow fast searches.
static const formatconverstion_t FormatLUT[] = {
{FMT_1_5_5_5, TYPE_UNORM, CFMT_1_5_5_5_UNORM},
{FMT_10_10_10_2, TYPE_UNORM, CFMT_10_10_10_2_UNORM},
{FMT_10_10_10_2, TYPE_SNORM, CFMT_10_10_10_2_SNORM},
{FMT_10_10_10_2, TYPE_UINT, CFMT_10_10_10_2_UINT},
{FMT_10_10_10_2, TYPE_SINT, CFMT_10_10_10_2_SINT},
{FMT_16, TYPE_UNORM, CFMT_16_UNORM},
{FMT_16, TYPE_SNORM, CFMT_16_SNORM},
{FMT_16, TYPE_UINT, CFMT_16_UINT},
{FMT_16, TYPE_SINT, CFMT_16_SINT},
{FMT_16, TYPE_FLOAT, CFMT_16_FLOAT},
{FMT_16_16, TYPE_UNORM, CFMT_16_16_UNORM},
{FMT_16_16, TYPE_SNORM, CFMT_16_16_SNORM},
{FMT_16_16, TYPE_UINT, CFMT_16_16_UINT},
{FMT_16_16, TYPE_SINT, CFMT_16_16_SINT},
{FMT_16_16, TYPE_FLOAT, CFMT_16_16_FLOAT},
{FMT_16_16_16_16, TYPE_UNORM, CFMT_16_16_16_16_UNORM},
{FMT_16_16_16_16, TYPE_SNORM, CFMT_16_16_16_16_SNORM},
{FMT_16_16_16_16, TYPE_UINT, CFMT_16_16_16_16_UINT},
{FMT_16_16_16_16, TYPE_SINT, CFMT_16_16_16_16_SINT},
{FMT_16_16_16_16, TYPE_FLOAT, CFMT_16_16_16_16_FLOAT},
{FMT_2_10_10_10, TYPE_UNORM, CFMT_2_10_10_10_UNORM},
{FMT_2_10_10_10, TYPE_SNORM, CFMT_2_10_10_10_SNORM},
{FMT_2_10_10_10, TYPE_UINT, CFMT_2_10_10_10_UINT},
{FMT_2_10_10_10, TYPE_SINT, CFMT_2_10_10_10_SINT},
{FMT_24_8, TYPE_UNORM, CFMT_24_8_UNORM},
{FMT_24_8, TYPE_UINT, CFMT_24_8_UINT},
{FMT_32, TYPE_UINT, CFMT_32_UINT},
{FMT_32, TYPE_SINT, CFMT_32_SINT},
{FMT_32, TYPE_FLOAT, CFMT_32_FLOAT},
{FMT_32_32, TYPE_UINT, CFMT_32_32_UINT},
{FMT_32_32, TYPE_SINT, CFMT_32_32_SINT},
{FMT_32_32, TYPE_FLOAT, CFMT_32_32_FLOAT},
{FMT_32_32_32, TYPE_UINT, CFMT_32_32_32_UINT},
{FMT_32_32_32, TYPE_SINT, CFMT_32_32_32_SINT},
{FMT_32_32_32, TYPE_FLOAT, CFMT_32_32_32_FLOAT},
{FMT_32_32_32_32, TYPE_UINT, CFMT_32_32_32_32_UINT},
{FMT_32_32_32_32, TYPE_SINT, CFMT_32_32_32_32_SINT},
{FMT_32_32_32_32, TYPE_FLOAT, CFMT_32_32_32_32_FLOAT},
{FMT_5_5_5_1, TYPE_UNORM, CFMT_5_5_5_1_UNORM},
{FMT_5_6_5, TYPE_UNORM, CFMT_5_6_5_UNORM},
{FMT_8, TYPE_UNORM, CFMT_8_UNORM},
{FMT_8, TYPE_SNORM, CFMT_8_SNORM},
{FMT_8, TYPE_UINT, CFMT_8_UINT},
{FMT_8, TYPE_SINT, CFMT_8_SINT},
{FMT_8, TYPE_SRGB, CFMT_8_SRGB},
{FMT_8_24, TYPE_UNORM, CFMT_8_24_UNORM},
{FMT_8_24, TYPE_UINT, CFMT_8_24_UINT},
{FMT_8_8, TYPE_UNORM, CFMT_8_8_UNORM},
{FMT_8_8, TYPE_SNORM, CFMT_8_8_SNORM},
{FMT_8_8, TYPE_UINT, CFMT_8_8_UINT},
{FMT_8_8, TYPE_SINT, CFMT_8_8_SINT},
{FMT_8_8, TYPE_SRGB, CFMT_8_8_SRGB},
{FMT_8_8_8_8, TYPE_UNORM, CFMT_8_8_8_8_UNORM},
{FMT_8_8_8_8, TYPE_SNORM, CFMT_8_8_8_8_SNORM},
{FMT_8_8_8_8, TYPE_UINT, CFMT_8_8_8_8_UINT},
{FMT_8_8_8_8, TYPE_SINT, CFMT_8_8_8_8_SINT},
{FMT_8_8_8_8, TYPE_SRGB, CFMT_8_8_8_8_SRGB}
};
static const int FormatLUTSize = sizeof(FormatLUT)/sizeof(formatconverstion_t);
//Index in FormatLUT to start search, indexed by FMT enum.
static const int FormatEntryPoint[] = {
57,
40,
5,
47,
26,
10,
57,
57,
1,
20,
52,
29,
15,
32,
35,
57,
39,
0,
38,
57,
45,
24
};
static FORMAT GetCombinedFormat(uint8_t fmt, uint8_t type) {
assert(fmt < sizeof(FormatEntryPoint)/sizeof(int) && "FMT out of range.");
int start = FormatEntryPoint[fmt];
int stop = std::min(start + 6, FormatLUTSize); // Only 6 types are used in image_kv_lut.cpp
for(int i=start; i<stop; i++) {
if((FormatLUT[i].fmt == fmt) && (FormatLUT[i].type == type))
return FormatLUT[i].format;
}
return CFMT_INVALID;
};
//-----------------------------------------------------------------------------
// End workaround
//-----------------------------------------------------------------------------
ImageManagerNv::ImageManagerNv() : ImageManagerKv() {}
ImageManagerNv::~ImageManagerNv() {}
// TODO(cfreehil) remove from class, make it a utility function
hsa_status_t ImageManagerNv::CalculateImageSizeAndAlignment(
hsa_agent_t component, const hsa_ext_image_descriptor_t& desc,
hsa_ext_image_data_layout_t image_data_layout,
size_t image_data_row_pitch,
size_t image_data_slice_pitch,
hsa_ext_image_data_info_t& image_info) const {
ADDR2_COMPUTE_SURFACE_INFO_OUTPUT out = {0};
hsa_profile_t profile;
hsa_status_t status = hsa_agent_get_info(component,
HSA_AGENT_INFO_PROFILE, &profile);
Image::TileMode tileMode = Image::TileMode::LINEAR;
if (image_data_layout == HSA_EXT_IMAGE_DATA_LAYOUT_OPAQUE) {
tileMode = (profile == HSA_PROFILE_BASE &&
desc.geometry != HSA_EXT_IMAGE_GEOMETRY_1DB)?
Image::TileMode::TILED : Image::TileMode::LINEAR;
}
if (GetAddrlibSurfaceInfoNv(component, desc, tileMode,
image_data_row_pitch, image_data_slice_pitch, out) ==
(uint32_t)(-1)) {
return HSA_STATUS_ERROR;
}
size_t rowPitch = (out.bpp >> 3) * out.pitch;
size_t slicePitch = rowPitch * out.height;
if (desc.geometry != HSA_EXT_IMAGE_GEOMETRY_1DB &&
image_data_layout == HSA_EXT_IMAGE_DATA_LAYOUT_LINEAR &&
((image_data_row_pitch && (rowPitch != image_data_row_pitch)) ||
(image_data_slice_pitch && (slicePitch != image_data_slice_pitch)))) {
return static_cast<hsa_status_t>(
HSA_EXT_STATUS_ERROR_IMAGE_PITCH_UNSUPPORTED);
}
image_info.size = out.surfSize;
assert(image_info.size != 0);
image_info.alignment = out.baseAlign;
assert(image_info.alignment != 0);
return HSA_STATUS_SUCCESS;
}
bool ImageManagerNv::IsLocalMemory(const void* address) const {
return true;
}
hsa_status_t ImageManagerNv::PopulateImageSrd(Image& image,
const metadata_amd_t* descriptor) const {
const metadata_amd_nv_t* desc =
reinterpret_cast<const metadata_amd_nv_t*>(descriptor);
bool atc_access = true;
const void* image_data_addr = image.data;
ImageProperty image_prop =
image_lut_.MapFormat(image.desc.format, image.desc.geometry);
if ((image_prop.cap == HSA_EXT_IMAGE_CAPABILITY_NOT_SUPPORTED) ||
(image_prop.element_size == 0))
return (hsa_status_t)HSA_EXT_STATUS_ERROR_IMAGE_FORMAT_UNSUPPORTED;
const Swizzle swizzle =
image_lut_.MapSwizzle(image.desc.format.channel_order);
if (IsLocalMemory(image.data)) {
atc_access = false;
image_data_addr = reinterpret_cast<const void*>(
reinterpret_cast<uintptr_t>(image.data) - local_memory_base_address_);
}
image.srd[0] = desc->word0.u32All;
image.srd[1] = desc->word1.u32All;
image.srd[2] = desc->word2.u32All;
image.srd[3] = desc->word3.u32All;
image.srd[4] = desc->word4.u32All;
image.srd[5] = desc->word5.u32All;
image.srd[6] = desc->word6.u32All;
image.srd[7] = desc->word7.u32All;
if (image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_1DB) {
SQ_BUF_RSRC_WORD0 word0;
SQ_BUF_RSRC_WORD1 word1;
SQ_BUF_RSRC_WORD2 word2;
SQ_BUF_RSRC_WORD3 word3;
word0.val = 0;
word0.f.BASE_ADDRESS = ext_image::PtrLow32(image_data_addr);
word1.val = image.srd[1];
word1.f.BASE_ADDRESS_HI = ext_image::PtrHigh32(image_data_addr);
word1.f.STRIDE = image_prop.element_size;
word3.val = image.srd[3];
word3.f.DST_SEL_X = swizzle.x;
word3.f.DST_SEL_Y = swizzle.y;
word3.f.DST_SEL_Z = swizzle.z;
word3.f.DST_SEL_W = swizzle.w;
word3.f.FORMAT = GetCombinedFormat(image_prop.data_format, image_prop.data_type);
word3.f.INDEX_STRIDE = image_prop.element_size;
image.srd[0] = word0.val;
image.srd[1] = word1.val;
image.srd[3] = word3.val;
} else {
uint32_t hwPixelSize = image_lut_.GetPixelSize(image_prop.data_format, image_prop.data_type);
if (image_prop.element_size != hwPixelSize) {
return (hsa_status_t)HSA_EXT_STATUS_ERROR_IMAGE_FORMAT_UNSUPPORTED;
}
reinterpret_cast<SQ_IMG_RSRC_WORD0*>(&image.srd[0])->bits.BASE_ADDRESS =
ext_image::PtrLow40Shift8(image_data_addr);
reinterpret_cast<SQ_IMG_RSRC_WORD1*>(&image.srd[1])->bits.BASE_ADDRESS_HI
= ext_image::PtrHigh64Shift40(image_data_addr);
reinterpret_cast<SQ_IMG_RSRC_WORD1*>(&image.srd[1])->bits.FORMAT = GetCombinedFormat(image_prop.data_format, image_prop.data_type);
reinterpret_cast<SQ_IMG_RSRC_WORD3*>(&image.srd[3])->bits.DST_SEL_X =
swizzle.x;
reinterpret_cast<SQ_IMG_RSRC_WORD3*>(&image.srd[3])->bits.DST_SEL_Y =
swizzle.y;
reinterpret_cast<SQ_IMG_RSRC_WORD3*>(&image.srd[3])->bits.DST_SEL_Z =
swizzle.z;
reinterpret_cast<SQ_IMG_RSRC_WORD3*>(&image.srd[3])->bits.DST_SEL_W =
swizzle.w;
if (image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_1DA ||
image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_1D) {
reinterpret_cast<SQ_IMG_RSRC_WORD3*>(&image.srd[3])->bits.TYPE =
image_lut_.MapGeometry(image.desc.geometry);
}
// Imported metadata holds the offset to metadata, add the image base address.
uintptr_t meta = uintptr_t(((SQ_IMG_RSRC_WORD7*)(&image.srd[7]))->bits.META_DATA_ADDRESS_HI) << 16;
meta |= uintptr_t(((SQ_IMG_RSRC_WORD6*)(&image.srd[6]))->bits.META_DATA_ADDRESS) << 8;
meta += reinterpret_cast<uintptr_t>(image_data_addr);
((SQ_IMG_RSRC_WORD6*)(&image.srd[6]))->bits.META_DATA_ADDRESS = ext_image::PtrLow16Shift8((void*)meta);
((SQ_IMG_RSRC_WORD7*)(&image.srd[7]))->bits.META_DATA_ADDRESS_HI = ext_image::PtrHigh64Shift16((void*)meta);
}
// Looks like this is only used for CPU copies.
image.row_pitch = 0;
image.slice_pitch = 0;
// Used by HSAIL shader ABI
image.srd[8] = image.desc.format.channel_type;
image.srd[9] = image.desc.format.channel_order;
image.srd[10] = static_cast<uint32_t>(image.desc.width);
return HSA_STATUS_SUCCESS;
}
static TEX_BC_SWIZZLE GetBcSwizzle(const Swizzle& swizzle) {
SEL r = (SEL)swizzle.x;
SEL g = (SEL)swizzle.y;
SEL b = (SEL)swizzle.z;
SEL a = (SEL)swizzle.w;
TEX_BC_SWIZZLE bcSwizzle = TEX_BC_Swizzle_XYZW;
if (a == SEL_X) {
// Have to use either TEX_BC_Swizzle_WZYX or TEX_BC_Swizzle_WXYZ
//
// For the pre-defined border color values (white, opaque black,
// transparent black), the only thing that matters is that the alpha
// channel winds up in the correct place (because the RGB channels are
// all the same) so either of these TEX_BC_Swizzle enumerations will
// work. Not sure what happens with border color palettes.
if (b == SEL_Y) {
// ABGR
bcSwizzle = TEX_BC_Swizzle_WZYX;
} else if ((r == SEL_X) && (g == SEL_X) && (b == SEL_X)) {
// RGBA
bcSwizzle = TEX_BC_Swizzle_XYZW;
} else {
// ARGB
bcSwizzle = TEX_BC_Swizzle_WXYZ;
}
} else if (r == SEL_X) {
// Have to use either TEX_BC_Swizzle_XYZW or TEX_BC_Swizzle_XWYZ
if (g == SEL_Y) {
// RGBA
bcSwizzle = TEX_BC_Swizzle_XYZW;
} else if ((g == SEL_X) && (b == SEL_X) && (a == SEL_W)) {
// RGBA
bcSwizzle = TEX_BC_Swizzle_XYZW;
} else {
// RAGB
bcSwizzle = TEX_BC_Swizzle_XWYZ;
}
} else if (g == SEL_X) {
// GRAB, have to use TEX_BC_Swizzle_YXWZ
bcSwizzle = TEX_BC_Swizzle_YXWZ;
} else if (b == SEL_X) {
// BGRA, have to use TEX_BC_Swizzle_ZYXW
bcSwizzle = TEX_BC_Swizzle_ZYXW;
}
return bcSwizzle;
}
hsa_status_t ImageManagerNv::PopulateImageSrd(Image& image) const {
ImageProperty image_prop =
image_lut_.MapFormat(image.desc.format, image.desc.geometry);
assert(image_prop.cap != HSA_EXT_IMAGE_CAPABILITY_NOT_SUPPORTED);
assert(image_prop.element_size != 0);
bool atc_access = true;
const void* image_data_addr = image.data;
if (IsLocalMemory(image.data)) {
atc_access = false;
image_data_addr = reinterpret_cast<const void*>(
reinterpret_cast<uintptr_t>(image.data) - local_memory_base_address_);
}
if (image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_1DB) {
SQ_BUF_RSRC_WORD0 word0;
SQ_BUF_RSRC_WORD1 word1;
SQ_BUF_RSRC_WORD2 word2;
SQ_BUF_RSRC_WORD3 word3;
word0.val = 0;
word0.f.BASE_ADDRESS = ext_image::PtrLow32(image_data_addr);
word1.val = 0;
word1.f.BASE_ADDRESS_HI = ext_image::PtrHigh32(image_data_addr);
word1.f.STRIDE = image_prop.element_size;
word1.f.SWIZZLE_ENABLE = false;
word1.f.CACHE_SWIZZLE = false;
word2.f.NUM_RECORDS = image.desc.width * image_prop.element_size;
const Swizzle swizzle =
image_lut_.MapSwizzle(image.desc.format.channel_order);
word3.val = 0;
word3.f.RESOURCE_LEVEL = 1;
word3.f.DST_SEL_X = swizzle.x;
word3.f.DST_SEL_Y = swizzle.y;
word3.f.DST_SEL_Z = swizzle.z;
word3.f.DST_SEL_W = swizzle.w;
word3.f.FORMAT = GetCombinedFormat(image_prop.data_format, image_prop.data_type);
word3.f.INDEX_STRIDE = image_prop.element_size;
word3.f.TYPE = image_lut_.MapGeometry(image.desc.geometry);
image.srd[0] = word0.val;
image.srd[1] = word1.val;
image.srd[2] = word2.val;
image.srd[3] = word3.val;
image.row_pitch = image.desc.width * image_prop.element_size;
image.slice_pitch = image.row_pitch;
} else {
SQ_IMG_RSRC_WORD0 word0;
SQ_IMG_RSRC_WORD1 word1;
SQ_IMG_RSRC_WORD2 word2;
SQ_IMG_RSRC_WORD3 word3;
SQ_IMG_RSRC_WORD4 word4;
SQ_IMG_RSRC_WORD5 word5;
SQ_IMG_RSRC_WORD5 word6;
SQ_IMG_RSRC_WORD5 word7;
ADDR2_COMPUTE_SURFACE_INFO_OUTPUT out = {0};
uint32_t swizzleMode = GetAddrlibSurfaceInfoNv(
image.component, image.desc, image.tile_mode,
image.row_pitch, image.slice_pitch, out);
if (swizzleMode == (uint32_t)(-1)) {
return HSA_STATUS_ERROR;
}
assert((out.bpp / 8) == image_prop.element_size);
const size_t row_pitch_size = out.pitch * image_prop.element_size;
word0.f.BASE_ADDRESS = ext_image::PtrLow40Shift8(image_data_addr);
word1.val = 0;
word1.f.BASE_ADDRESS_HI = ext_image::PtrHigh64Shift40(image_data_addr);
word1.f.MIN_LOD = 0;
word1.f.FORMAT = GetCombinedFormat(image_prop.data_format, image_prop.data_type);
// Only take the lowest 2 bits of (image.desc.width - 1)
word1.f.WIDTH = ext_image::BitSelect<0, 1>(image.desc.width - 1);
word2.val = 0;
// Take the high 12 bits of (image.desc.width - 1)
word2.f.WIDTH_HI = ext_image::BitSelect<2, 13>(image.desc.width - 1);
word2.f.HEIGHT = image.desc.height ? image.desc.height - 1 : 0;
word2.f.RESOURCE_LEVEL = 1;
const Swizzle swizzle =
image_lut_.MapSwizzle(image.desc.format.channel_order);
word3.val = 0;
word3.f.DST_SEL_X = swizzle.x;
word3.f.DST_SEL_Y = swizzle.y;
word3.f.DST_SEL_Z = swizzle.z;
word3.f.DST_SEL_W = swizzle.w;
word3.f.SW_MODE = swizzleMode;
word3.f.BC_SWIZZLE = GetBcSwizzle(swizzle);
word3.f.TYPE = image_lut_.MapGeometry(image.desc.geometry);
const bool image_array =
(image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_1DA ||
image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_2DA ||
image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_2DADEPTH);
const bool image_3d = (image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_3D);
word4.val = 0;
word4.f.DEPTH =
(image_array) // Doesn't hurt but isn't array_size already >0?
? std::max(image.desc.array_size, static_cast<size_t>(1)) - 1
: (image_3d) ? image.desc.depth - 1 : out.pitch - 1;
word5.val = 0;
word6.val = 0;
word7.val = 0;
image.srd[0] = word0.val;
image.srd[1] = word1.val;
image.srd[2] = word2.val;
image.srd[3] = word3.val;
image.srd[4] = word4.val;
image.srd[5] = word5.val;
image.srd[6] = word6.val;
image.srd[7] = word7.val;
image.row_pitch = row_pitch_size;
image.slice_pitch = out.sliceSize;
}
image.srd[8] = image.desc.format.channel_type;
image.srd[9] = image.desc.format.channel_order;
image.srd[10] = static_cast<uint32_t>(image.desc.width);
return HSA_STATUS_SUCCESS;
}
hsa_status_t ImageManagerNv::ModifyImageSrd(
Image& image, hsa_ext_image_format_t& new_format) const {
image.desc.format = new_format;
ImageProperty image_prop =
image_lut_.MapFormat(image.desc.format, image.desc.geometry);
assert(image_prop.cap != HSA_EXT_IMAGE_CAPABILITY_NOT_SUPPORTED);
assert(image_prop.element_size != 0);
if (image.desc.geometry == HSA_EXT_IMAGE_GEOMETRY_1DB) {
const Swizzle swizzle =
image_lut_.MapSwizzle(image.desc.format.channel_order);
SQ_BUF_RSRC_WORD3* word3 =
reinterpret_cast<SQ_BUF_RSRC_WORD3*>(&image.srd[3]);
word3->bits.DST_SEL_X = swizzle.x;
word3->bits.DST_SEL_Y = swizzle.y;
word3->bits.DST_SEL_Z = swizzle.z;
word3->bits.DST_SEL_W = swizzle.w;
word3->bits.FORMAT = GetCombinedFormat(image_prop.data_format, image_prop.data_type);
} else {
SQ_IMG_RSRC_WORD1* word1 =
reinterpret_cast<SQ_IMG_RSRC_WORD1*>(&image.srd[1]);
word1->bits.FORMAT = GetCombinedFormat(image_prop.data_format, image_prop.data_type);
const Swizzle swizzle =
image_lut_.MapSwizzle(image.desc.format.channel_order);
SQ_IMG_RSRC_WORD3* word3 =
reinterpret_cast<SQ_IMG_RSRC_WORD3*>(&image.srd[3]);
word3->bits.DST_SEL_X = swizzle.x;
word3->bits.DST_SEL_Y = swizzle.y;
word3->bits.DST_SEL_Z = swizzle.z;
word3->bits.DST_SEL_W = swizzle.w;
}
image.srd[8] = image.desc.format.channel_type;
image.srd[9] = image.desc.format.channel_order;
image.srd[10] = static_cast<uint32_t>(image.desc.width);
return HSA_STATUS_SUCCESS;
}
hsa_status_t ImageManagerNv::PopulateSamplerSrd(amd::Sampler& sampler) const {
const hsa_ext_sampler_descriptor_t sampler_descriptor = sampler.desc;
SQ_IMG_SAMP_WORD0 word0;
SQ_IMG_SAMP_WORD1 word1;
SQ_IMG_SAMP_WORD2 word2;
SQ_IMG_SAMP_WORD3 word3;
word0.u32All = 0;
switch (sampler_descriptor.address_mode) {
case HSA_EXT_SAMPLER_ADDRESSING_MODE_CLAMP_TO_EDGE:
word0.bits.CLAMP_X = static_cast<int>(SQ_TEX_CLAMP_LAST_TEXEL);
break;
case HSA_EXT_SAMPLER_ADDRESSING_MODE_CLAMP_TO_BORDER:
word0.bits.CLAMP_X = static_cast<int>(SQ_TEX_CLAMP_BORDER);
break;
case HSA_EXT_SAMPLER_ADDRESSING_MODE_MIRRORED_REPEAT:
word0.bits.CLAMP_X = static_cast<int>(SQ_TEX_MIRROR);
break;
case HSA_EXT_SAMPLER_ADDRESSING_MODE_UNDEFINED:
case HSA_EXT_SAMPLER_ADDRESSING_MODE_REPEAT:
word0.bits.CLAMP_X = static_cast<int>(SQ_TEX_WRAP);
break;
default:
return HSA_STATUS_ERROR_INVALID_ARGUMENT;
}
word0.bits.CLAMP_Y = word0.bits.CLAMP_X;
word0.bits.CLAMP_Z = word0.bits.CLAMP_X;
word0.bits.FORCE_UNNORMALIZED = (sampler_descriptor.coordinate_mode ==
HSA_EXT_SAMPLER_COORDINATE_MODE_UNNORMALIZED);
word1.u32All = 0;
word1.bits.MAX_LOD = 4095;
word2.u32All = 0;
switch (sampler_descriptor.filter_mode) {
case HSA_EXT_SAMPLER_FILTER_MODE_NEAREST:
word2.bits.XY_MAG_FILTER = static_cast<int>(SQ_TEX_XY_FILTER_POINT);
break;
case HSA_EXT_SAMPLER_FILTER_MODE_LINEAR:
word2.bits.XY_MAG_FILTER = static_cast<int>(SQ_TEX_XY_FILTER_BILINEAR);
break;
default:
return HSA_STATUS_ERROR_INVALID_ARGUMENT;
}
word2.bits.XY_MIN_FILTER = word2.bits.XY_MAG_FILTER;
word2.bits.Z_FILTER = SQ_TEX_Z_FILTER_NONE;
word2.bits.MIP_FILTER = SQ_TEX_MIP_FILTER_NONE;
word3.u32All = 0;
// TODO: check this bit with HSAIL spec.
word3.bits.BORDER_COLOR_TYPE = SQ_TEX_BORDER_COLOR_TRANS_BLACK;
sampler.srd[0] = word0.u32All;
sampler.srd[1] = word1.u32All;
sampler.srd[2] = word2.u32All;
sampler.srd[3] = word3.u32All;
return HSA_STATUS_SUCCESS;
}
uint32_t ImageManagerNv::GetAddrlibSurfaceInfoNv(
hsa_agent_t component, const hsa_ext_image_descriptor_t& desc,
Image::TileMode tileMode,
size_t image_data_row_pitch,
size_t image_data_slice_pitch,
ADDR2_COMPUTE_SURFACE_INFO_OUTPUT& out) const {
const ImageProperty image_prop =
GetImageProperty(component, desc.format, desc.geometry);
const AddrFormat addrlib_format = GetAddrlibFormat(image_prop);
const uint32_t width = static_cast<uint32_t>(desc.width);
const uint32_t height = static_cast<uint32_t>(desc.height);
static const size_t kMinNumSlice = 1;
const uint32_t num_slice = static_cast<uint32_t>(
std::max(kMinNumSlice, std::max(desc.array_size, desc.depth)));
ADDR2_COMPUTE_SURFACE_INFO_INPUT in = {0};
in.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_INPUT);
in.format = addrlib_format;
in.bpp = static_cast<unsigned int>(image_prop.element_size) * 8;
in.width = width;
in.height = height;
in.numSlices = num_slice;
switch (desc.geometry) {
case HSA_EXT_IMAGE_GEOMETRY_1D:
case HSA_EXT_IMAGE_GEOMETRY_1DB:
case HSA_EXT_IMAGE_GEOMETRY_1DA:
in.resourceType = ADDR_RSRC_TEX_1D;
break;
case HSA_EXT_IMAGE_GEOMETRY_2D:
case HSA_EXT_IMAGE_GEOMETRY_2DDEPTH:
case HSA_EXT_IMAGE_GEOMETRY_2DA:
case HSA_EXT_IMAGE_GEOMETRY_2DADEPTH:
in.resourceType = ADDR_RSRC_TEX_2D;
break;
case HSA_EXT_IMAGE_GEOMETRY_3D:
in.resourceType = ADDR_RSRC_TEX_3D;
break;
}
in.swizzleMode =
(tileMode == Image::TileMode::LINEAR)? ADDR_SW_LINEAR : ADDR_SW_4KB;
in.flags.texture = 1;
ADDR2_GET_PREFERRED_SURF_SETTING_INPUT prefSettingsInput = { 0 };
ADDR2_GET_PREFERRED_SURF_SETTING_OUTPUT prefSettingsOutput = { 0 };
prefSettingsInput.size = sizeof(prefSettingsInput);
prefSettingsInput.flags = in.flags;
prefSettingsInput.bpp = in.bpp;
prefSettingsInput.format = in.format;
prefSettingsInput.width = in.width;
prefSettingsInput.height = in.height;
prefSettingsInput.numFrags = in.numFrags;
prefSettingsInput.numSamples = in.numSamples;
prefSettingsInput.numMipLevels = in.numMipLevels;
prefSettingsInput.numSlices = in.numSlices;
prefSettingsInput.resourceLoction = ADDR_RSRC_LOC_UNDEF;
prefSettingsInput.resourceType = in.resourceType;
if (tileMode == Image::TileMode::LINEAR) {
// this should force linear.
prefSettingsInput.forbiddenBlock.macroThin4KB = 1;
prefSettingsInput.forbiddenBlock.macroThick4KB = 1;
prefSettingsInput.forbiddenBlock.macroThin64KB = 1;
prefSettingsInput.forbiddenBlock.macroThick64KB = 1;
prefSettingsInput.forbiddenBlock.micro = 1;
prefSettingsInput.forbiddenBlock.var = 1;
} else {
// this should not allow linear.
prefSettingsInput.forbiddenBlock.linear = 1;
// Debug setting, simplifies buffer alignment.
prefSettingsInput.forbiddenBlock.macroThin64KB = 1;
prefSettingsInput.forbiddenBlock.macroThick64KB = 1;
}
// but don't ever allow the 256b swizzle modes
//prefSettingsInput.forbiddenBlock.micro = 1;
// and don't allow variable-size block modes
//prefSettingsInput.forbiddenBlock.var = 1;
if (ADDR_OK != Addr2GetPreferredSurfaceSetting(addr_lib_,
&prefSettingsInput, &prefSettingsOutput)) {
return (uint32_t)(-1);
}
in.swizzleMode = prefSettingsOutput.swizzleMode;
out.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_OUTPUT);
if (ADDR_OK != Addr2ComputeSurfaceInfo(addr_lib_, &in, &out)) {
return (uint32_t)(-1);
}
if (out.surfSize == 0) {
return (uint32_t)(-1);
}
return in.swizzleMode;
}
hsa_status_t ImageManagerNv::FillImage(const Image& image, const void* pattern,
const hsa_ext_image_region_t& region) {
if (BlitQueueInit().queue_ == NULL) {
return HSA_STATUS_ERROR_OUT_OF_RESOURCES;
}
Image* image_view = const_cast<Image*>(&image);
SQ_BUF_RSRC_WORD3* word3_buff = NULL;
SQ_IMG_RSRC_WORD3* word3_image = NULL;
uint32_t dst_sel_w_original = 0;
if (image_view->desc.format.channel_type ==
HSA_EXT_IMAGE_CHANNEL_TYPE_UNORM_SHORT_101010) {
// Force GPU to ignore the last two bits (alpha bits).
if (image_view->desc.geometry == HSA_EXT_IMAGE_GEOMETRY_1DB) {
word3_buff = reinterpret_cast<SQ_BUF_RSRC_WORD3*>(&image_view->srd[3]);
dst_sel_w_original = word3_buff->bits.DST_SEL_W;
word3_buff->bits.DST_SEL_W = SEL_0;
} else {
word3_image = reinterpret_cast<SQ_IMG_RSRC_WORD3*>(&image_view->srd[3]);
dst_sel_w_original = word3_image->bits.DST_SEL_W;
word3_image->bits.DST_SEL_W = SEL_0;
}
}
SQ_IMG_RSRC_WORD1* word1 = NULL;
uint32_t num_format_original = 0;
const void* new_pattern = pattern;
float fill_value[4] = {0};
switch (image_view->desc.format.channel_order) {
case HSA_EXT_IMAGE_CHANNEL_ORDER_SRGBA:
case HSA_EXT_IMAGE_CHANNEL_ORDER_SRGB:
case HSA_EXT_IMAGE_CHANNEL_ORDER_SRGBX:
case HSA_EXT_IMAGE_CHANNEL_ORDER_SBGRA: {
// KV and CZ don't have write support for SRGBA image, so convert pattern
// to standard form and treat the image as RGBA image.
const float* pattern_f = reinterpret_cast<const float*>(pattern);
fill_value[0] = LinearToStandardRGB(pattern_f[0]);
fill_value[1] = LinearToStandardRGB(pattern_f[1]);
fill_value[2] = LinearToStandardRGB(pattern_f[2]);
fill_value[3] = pattern_f[3];
new_pattern = fill_value;
ImageProperty image_prop = image_lut_.MapFormat(image.desc.format, image.desc.geometry);
word1 = reinterpret_cast<SQ_IMG_RSRC_WORD1*>(&image_view->srd[1]);
num_format_original = word1->bits.FORMAT;
word1->bits.FORMAT = GetCombinedFormat(image_prop.data_format, TYPE_UNORM);
} break;
default:
break;
}
hsa_status_t status =
ext_image::ImageRuntime::instance()->blit_kernel().FillImage(
blit_queue_, blit_code_catalog_, *image_view, new_pattern, region);
// Revert back original configuration.
if (word3_buff != NULL) {
word3_buff->bits.DST_SEL_W = dst_sel_w_original;
}
if (word3_image != NULL) {
word3_image->bits.DST_SEL_W = dst_sel_w_original;
}
if (word1 != NULL) {
word1->bits.FORMAT = num_format_original;
}
return status;
}
} // namespace amd