AI/rocm-systems
2
0
Форкнуть 0
Код Задачи Запросы на слияние Действия Пакеты Проекты Релизы Вики Активность

Update addrlib.

Просмотр исходного кода 
Corrects type errors and adds gfx1030 support.

Change-Id: I7fe7b5ded68b1e26fb899676712ad019089a45f3
Этот коммит содержится в:
Sean Keely
2020-06-17 14:09:23 -05:00
родитель 0dae86f686
Коммит 4081c564ee
6 изменённых файлов: 96 добавлений и 124 удалений
Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы
runtime/hsa-runtime/image/addrlib/inc/addrinterface.h
+9 -3
Просмотреть файл
@@ -3349,17 +3349,23 @@ typedef struct _ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT
UINT_32 mipId; ///< mipmap level id
ADDR2_META_FLAGS dccKeyFlags; ///< DCC flags
ADDR2_SURFACE_FLAGS colorFlags; ///< Color surface flags
AddrResourceType resourceType; ///< Color surface type
AddrSwizzleMode swizzleMode; ///< Color surface swizzle mode
UINT_32 bpp; ///< Color surface bits per pixel
UINT_32 unalignedWidth; ///< Color surface original width (of mip0)
UINT_32 unalignedHeight; ///< Color surface original height (of mip0)
UINT_32 numSlices; ///< Color surface original slices (of mip0)
UINT_32 numMipLevels; ///< Color surface mipmap levels
UINT_32 numFrags; ///< Color surface fragment number
UINT_32 pipeXor; ///< pipe Xor setting
UINT_32 pitch; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::pitch
UINT_32 height; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::height
UINT_32 compressBlkWidth; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::compressBlkWidth
UINT_32 compressBlkHeight; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::compressBlkHeight
UINT_32 compressBlkDepth; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::compressBlkDepth
UINT_32 metaBlkWidth; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::metaBlkWidth
UINT_32 metaBlkHeight; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::metaBlkHeight
UINT_32 metaBlkDepth; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::metaBlkDepth
UINT_32 dccRamSliceSize; ///< ADDR2_COMPUTE_DCC_INFO_OUTPUT::dccRamSliceSize
} ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT;
/**
runtime/hsa-runtime/image/addrlib/src/amdgpu_asic_addr.h
+2
Просмотреть файл
@@ -97,6 +97,7 @@
#define AMDGPU_NAVI10_RANGE 0x01, 0x0A
#define AMDGPU_NAVI12_RANGE 0x0A, 0x14
#define AMDGPU_NAVI14_RANGE 0x14, 0x28
#define AMDGPU_SIENNA_RANGE 0x28, 0x32
#define AMDGPU_EXPAND_FIX(x) x
#define AMDGPU_RANGE_HELPER(val, min, max) ((val >= min) && (val < max))
@@ -144,5 +145,6 @@
#define ASICREV_IS_NAVI10_P(r) ASICREV_IS(r, NAVI10)
#define ASICREV_IS_NAVI12(r) ASICREV_IS(r, NAVI12)
#define ASICREV_IS_NAVI14(r) ASICREV_IS(r, NAVI14)
#define ASICREV_IS_SIENNA_M(r) ASICREV_IS(r, SIENNA)
#endif // _AMDGPU_ASIC_ADDR_H
runtime/hsa-runtime/image/addrlib/src/chip/gfx9/gfx9_gb_reg.h
+1 -1
Просмотреть файл
@@ -34,7 +34,7 @@
*
*/
union GB_ADDR_CONFIG {
union GB_ADDR_CONFIG_gfx9 {
struct {
#if defined(LITTLEENDIAN_CPU)
unsigned int NUM_PIPES : 3;
runtime/hsa-runtime/image/addrlib/src/core/coord.cpp
+2 -2
Просмотреть файл
@@ -83,11 +83,11 @@ BOOL_32 Coordinate::operator<(const Coordinate& b)
}
else
{
if (dim == 's' || b.dim == 'm')
if (dim == DIM_S || b.dim == DIM_M)
{
ret = TRUE;
}
else if (b.dim == 's' || dim == 'm')
else if (b.dim == DIM_S || dim == DIM_M)
{
ret = FALSE;
}
runtime/hsa-runtime/image/addrlib/src/gfx10/gfx10addrlib.cpp
+55 -70
Просмотреть файл
@@ -673,88 +673,67 @@ ADDR_E_RETURNCODE Gfx10Lib::HwlComputeDccAddrFromCoord(
}
else
{
ADDR2_COMPUTE_DCCINFO_INPUT input = {0};
input.size = sizeof(input);
input.dccKeyFlags = pIn->dccKeyFlags;
input.colorFlags = pIn->colorFlags;
input.swizzleMode = pIn->swizzleMode;
input.resourceType = pIn->resourceType;
input.bpp = pIn->bpp;
input.unalignedWidth = Max(pIn->unalignedWidth, 1u);
input.unalignedHeight = Max(pIn->unalignedHeight, 1u);
input.numSlices = Max(pIn->numSlices, 1u);
input.numFrags = Max(pIn->numFrags, 1u);
input.numMipLevels = Max(pIn->numMipLevels, 1u);
const UINT_32 elemLog2 = Log2(pIn->bpp >> 3);
const UINT_32 numPipeLog2 = m_pipesLog2;
const UINT_32 pipeMask = (1 << numPipeLog2) - 1;
UINT_32 index = m_dccBaseIndex + elemLog2;
const UINT_8* patIdxTable;
ADDR2_COMPUTE_DCCINFO_OUTPUT output = {0};
output.size = sizeof(output);
returnCode = ComputeDccInfo(&input, &output);
if (returnCode == ADDR_OK)
if (m_settings.supportRbPlus)
{
const UINT_32 elemLog2 = Log2(pIn->bpp >> 3);
const UINT_32 numPipeLog2 = m_pipesLog2;
const UINT_32 pipeMask = (1 << numPipeLog2) - 1;
UINT_32 index = m_dccBaseIndex + elemLog2;
const UINT_8* patIdxTable;
patIdxTable = DCC_64K_R_X_RBPLUS_PATIDX;
if (m_settings.supportRbPlus)
if (pIn->dccKeyFlags.pipeAligned)
{
patIdxTable = DCC_64K_R_X_RBPLUS_PATIDX;
index += MaxNumOfBpp;
if (pIn->dccKeyFlags.pipeAligned)
if (m_numPkrLog2 < 2)
{
index += MaxNumOfBpp;
if (m_numPkrLog2 < 2)
{
index += m_pipesLog2 * MaxNumOfBpp;
}
else
{
// 4 groups for "m_numPkrLog2 < 2" case
index += 4 * MaxNumOfBpp;
const UINT_32 dccPipePerPkr = 3;
index += (m_numPkrLog2 - 2) * dccPipePerPkr * MaxNumOfBpp +
(m_pipesLog2 - m_numPkrLog2) * MaxNumOfBpp;
}
}
}
else
{
patIdxTable = DCC_64K_R_X_PATIDX;
if (pIn->dccKeyFlags.pipeAligned)
{
index += (numPipeLog2 + UnalignedDccType) * MaxNumOfBpp;
index += m_pipesLog2 * MaxNumOfBpp;
}
else
{
index += Min(numPipeLog2, UnalignedDccType - 1) * MaxNumOfBpp;
// 4 groups for "m_numPkrLog2 < 2" case
index += 4 * MaxNumOfBpp;
const UINT_32 dccPipePerPkr = 3;
index += (m_numPkrLog2 - 2) * dccPipePerPkr * MaxNumOfBpp +
(m_pipesLog2 - m_numPkrLog2) * MaxNumOfBpp;
}
}
const UINT_32 blkSizeLog2 = Log2(output.metaBlkWidth) + Log2(output.metaBlkHeight) + elemLog2 - 8;
const UINT_32 blkMask = (1 << blkSizeLog2) - 1;
const UINT_32 blkOffset = ComputeOffsetFromSwizzlePattern(DCC_64K_R_X_SW_PATTERN[patIdxTable[index]],
blkSizeLog2 + 1, // +1 for nibble offset
pIn->x,
pIn->y,
pIn->slice,
0);
const UINT_32 xb = pIn->x / output.metaBlkWidth;
const UINT_32 yb = pIn->y / output.metaBlkHeight;
const UINT_32 pb = output.pitch / output.metaBlkWidth;
const UINT_32 blkIndex = (yb * pb) + xb;
const UINT_32 pipeXor = ((pIn->pipeXor & pipeMask) << m_pipeInterleaveLog2) & blkMask;
pOut->addr = (static_cast<UINT_64>(output.dccRamSliceSize) * pIn->slice) +
(blkIndex * (1 << blkSizeLog2)) +
((blkOffset >> 1) ^ pipeXor);
}
else
{
patIdxTable = DCC_64K_R_X_PATIDX;
if (pIn->dccKeyFlags.pipeAligned)
{
index += (numPipeLog2 + UnalignedDccType) * MaxNumOfBpp;
}
else
{
index += Min(numPipeLog2, UnalignedDccType - 1) * MaxNumOfBpp;
}
}
const UINT_32 blkSizeLog2 = Log2(pIn->metaBlkWidth) + Log2(pIn->metaBlkHeight) + elemLog2 - 8;
const UINT_32 blkMask = (1 << blkSizeLog2) - 1;
const UINT_32 blkOffset = ComputeOffsetFromSwizzlePattern(DCC_64K_R_X_SW_PATTERN[patIdxTable[index]],
blkSizeLog2 + 1, // +1 for nibble offset
pIn->x,
pIn->y,
pIn->slice,
0);
const UINT_32 xb = pIn->x / pIn->metaBlkWidth;
const UINT_32 yb = pIn->y / pIn->metaBlkHeight;
const UINT_32 pb = pIn->pitch / pIn->metaBlkWidth;
const UINT_32 blkIndex = (yb * pb) + xb;
const UINT_32 pipeXor = ((pIn->pipeXor & pipeMask) << m_pipeInterleaveLog2) & blkMask;
pOut->addr = (static_cast<UINT_64>(pIn->dccRamSliceSize) * pIn->slice) +
(blkIndex * (1 << blkSizeLog2)) +
((blkOffset >> 1) ^ pipeXor);
}
return returnCode;
@@ -943,6 +922,12 @@ ChipFamily Gfx10Lib::HwlConvertChipFamily(
{
case FAMILY_NV:
m_settings.isDcn2 = 1;
if (ASICREV_IS_SIENNA_M(chipRevision))
{
m_settings.supportRbPlus = 1;
m_settings.dccUnsup3DSwDis = 0;
}
break;
default:
ADDR_ASSERT(!"Unknown chip family");
runtime/hsa-runtime/image/addrlib/src/gfx9/gfx9addrlib.cpp
+27 -48
Просмотреть файл
@@ -987,62 +987,41 @@ ADDR_E_RETURNCODE Gfx9Lib::HwlComputeDccAddrFromCoord(
}
else
{
ADDR2_COMPUTE_DCCINFO_INPUT input = {0};
input.size = sizeof(input);
input.dccKeyFlags = pIn->dccKeyFlags;
input.colorFlags = pIn->colorFlags;
input.swizzleMode = pIn->swizzleMode;
input.resourceType = pIn->resourceType;
input.bpp = pIn->bpp;
input.unalignedWidth = Max(pIn->unalignedWidth, 1u);
input.unalignedHeight = Max(pIn->unalignedHeight, 1u);
input.numSlices = Max(pIn->numSlices, 1u);
input.numFrags = Max(pIn->numFrags, 1u);
input.numMipLevels = Max(pIn->numMipLevels, 1u);
UINT_32 elementBytesLog2 = Log2(pIn->bpp >> 3);
UINT_32 numSamplesLog2 = Log2(pIn->numFrags);
UINT_32 metaBlkWidthLog2 = Log2(pIn->metaBlkWidth);
UINT_32 metaBlkHeightLog2 = Log2(pIn->metaBlkHeight);
UINT_32 metaBlkDepthLog2 = Log2(pIn->metaBlkDepth);
UINT_32 compBlkWidthLog2 = Log2(pIn->compressBlkWidth);
UINT_32 compBlkHeightLog2 = Log2(pIn->compressBlkHeight);
UINT_32 compBlkDepthLog2 = Log2(pIn->compressBlkDepth);
ADDR2_COMPUTE_DCCINFO_OUTPUT output = {0};
output.size = sizeof(output);
MetaEqParams metaEqParams = {pIn->mipId, elementBytesLog2, numSamplesLog2, pIn->dccKeyFlags,
Gfx9DataColor, pIn->swizzleMode, pIn->resourceType,
metaBlkWidthLog2, metaBlkHeightLog2, metaBlkDepthLog2,
compBlkWidthLog2, compBlkHeightLog2, compBlkDepthLog2};
returnCode = ComputeDccInfo(&input, &output);
const CoordEq* pMetaEq = GetMetaEquation(metaEqParams);
if (returnCode == ADDR_OK)
{
UINT_32 elementBytesLog2 = Log2(pIn->bpp >> 3);
UINT_32 numSamplesLog2 = Log2(pIn->numFrags);
UINT_32 metaBlkWidthLog2 = Log2(output.metaBlkWidth);
UINT_32 metaBlkHeightLog2 = Log2(output.metaBlkHeight);
UINT_32 metaBlkDepthLog2 = Log2(output.metaBlkDepth);
UINT_32 compBlkWidthLog2 = Log2(output.compressBlkWidth);
UINT_32 compBlkHeightLog2 = Log2(output.compressBlkHeight);
UINT_32 compBlkDepthLog2 = Log2(output.compressBlkDepth);
UINT_32 xb = pIn->x / pIn->metaBlkWidth;
UINT_32 yb = pIn->y / pIn->metaBlkHeight;
UINT_32 zb = pIn->slice / pIn->metaBlkDepth;
MetaEqParams metaEqParams = {pIn->mipId, elementBytesLog2, numSamplesLog2, pIn->dccKeyFlags,
Gfx9DataColor, pIn->swizzleMode, pIn->resourceType,
metaBlkWidthLog2, metaBlkHeightLog2, metaBlkDepthLog2,
compBlkWidthLog2, compBlkHeightLog2, compBlkDepthLog2};
UINT_32 pitchInBlock = pIn->pitch / pIn->metaBlkWidth;
UINT_32 sliceSizeInBlock = (pIn->height / pIn->metaBlkHeight) * pitchInBlock;
UINT_32 blockIndex = zb * sliceSizeInBlock + yb * pitchInBlock + xb;
const CoordEq* pMetaEq = GetMetaEquation(metaEqParams);
UINT_32 coords[] = { pIn->x, pIn->y, pIn->slice, pIn->sample, blockIndex };
UINT_64 address = pMetaEq->solve(coords);
UINT_32 xb = pIn->x / output.metaBlkWidth;
UINT_32 yb = pIn->y / output.metaBlkHeight;
UINT_32 zb = pIn->slice / output.metaBlkDepth;
pOut->addr = address >> 1;
UINT_32 pitchInBlock = output.pitch / output.metaBlkWidth;
UINT_32 sliceSizeInBlock = (output.height / output.metaBlkHeight) * pitchInBlock;
UINT_32 blockIndex = zb * sliceSizeInBlock + yb * pitchInBlock + xb;
UINT_32 numPipeBits = GetPipeLog2ForMetaAddressing(pIn->dccKeyFlags.pipeAligned,
pIn->swizzleMode);
UINT_32 coords[] = { pIn->x, pIn->y, pIn->slice, pIn->sample, blockIndex };
UINT_64 address = pMetaEq->solve(coords);
UINT_64 pipeXor = static_cast<UINT_64>(pIn->pipeXor & ((1 << numPipeBits) - 1));
pOut->addr = address >> 1;
UINT_32 numPipeBits = GetPipeLog2ForMetaAddressing(pIn->dccKeyFlags.pipeAligned,
pIn->swizzleMode);
UINT_64 pipeXor = static_cast<UINT_64>(pIn->pipeXor & ((1 << numPipeBits) - 1));
pOut->addr ^= (pipeXor << m_pipeInterleaveLog2);
}
pOut->addr ^= (pipeXor << m_pipeInterleaveLog2);
}
return returnCode;
@@ -1067,7 +1046,7 @@ BOOL_32 Gfx9Lib::HwlInitGlobalParams(
if (m_settings.isArcticIsland)
{
GB_ADDR_CONFIG gbAddrConfig;
GB_ADDR_CONFIG_gfx9 gbAddrConfig;
gbAddrConfig.u32All = pCreateIn->regValue.gbAddrConfig;