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

SWDEV-263539 - More Target ID fix in GSL path

Просмотр исходного кода 
On behalf of Tony Tye. The bootleg passes CQE Baffin/Ellesmere test.

Change-Id: I4c21d21b3aaba360682ef15b8a4dda239f8af276


[ROCm/clr commit: d4316141b7]
Этот коммит содержится в:
Jason Tang
2020-12-14 15:40:43 -05:00
коммит произвёл Jason Tang
родитель f7cf40fc02
Коммит 6778b6e4e2
7 изменённых файлов: 145 добавлений и 126 удалений
Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы
projects/clr/rocclr/device/device.hpp
+1 -1
Просмотреть файл
@@ -1276,7 +1276,7 @@ class Device : public RuntimeObject {
typedef std::list<CommandQueue*> CommandQueues;
struct BlitProgram : public amd::HeapObject {
Program* program_; //!< GPU program obejct
Program* program_; //!< GPU program object
Context* context_; //!< A dummy context
BlitProgram(Context* context) : program_(NULL), context_(context) {}
projects/clr/rocclr/device/gpu/gpublit.hpp
+1 -1
Просмотреть файл
@@ -405,7 +405,7 @@ class KernelBlitManager : public DmaBlitManager {
//! Disable operator=
KernelBlitManager& operator=(const KernelBlitManager&);
amd::Program* program_; //!< GPU program obejct
amd::Program* program_; //!< GPU program object
amd::Kernel* kernels_[BlitTotal]; //!< GPU kernels for blit
amd::Memory* constantBuffer_; //!< An internal CB for blits
amd::Memory* xferBuffers_[MaxXferBuffers]; //!< Transfer buffers for images
projects/clr/rocclr/device/gpu/gpudefs.hpp
+84 -64
Просмотреть файл
@@ -83,6 +83,30 @@ static constexpr uint DeviceQueueMaskSize = 32;
//! Defines all supported ASIC families
enum AsicFamilies { Family7xx, Family8xx, FamilyTotal };
// FIXME: Change to use 2 digit for major/minor/stepping.
enum gfx_handle {
gfx600 = 600,
gfx601 = 601,
gfx602 = 602,
gfx700 = 700,
gfx701 = 701,
gfx702 = 702,
gfx704 = 704,
gfx705 = 705,
gfx800 = 800,
gfx801 = 801,
gfx802 = 802,
gfx803 = 803,
gfx810 = 810,
gfx900 = 900,
gfx902 = 902,
gfx904 = 904,
gfx906 = 906,
gfx909 = 909,
gfx90c = 9012
};
// FIXME: Does this need updating?
struct AMDDeviceInfo {
uint machine_; //!< Machine target ID
const char* targetName_; //!< Target name
@@ -94,9 +118,11 @@ struct AMDDeviceInfo {
uint localMemSizePerCU_; //!< Local memory size per CU
uint localMemBanks_; //!< Number of banks of local memory
uint gfxipVersion_; //!< The core engine GFXIP version
uint gfxipMajor_; //!< The core engine GFXIP Major version
uint gfxipMinor_; //!< The core engine GFXIP Minor version
uint gfxipStepping_; //!< The core engine GFXIP Stepping version
};
static constexpr AMDDeviceInfo DeviceInfo[] = {
// Machine targetName machineTarget
/* CAL_TARGET_600 */ {ED_ATI_CAL_MACHINE_R600_ISA, "", "", 0, 0, 0, 0, 0, 0, 0},
@@ -108,108 +134,102 @@ static constexpr AMDDeviceInfo DeviceInfo[] = {
/* CAL_TARGET_710 */ {ED_ATI_CAL_MACHINE_R710_ISA, "", "", 0, 0, 0, 0, 0, 0, 0},
/* CAL_TARGET_730 */ {ED_ATI_CAL_MACHINE_R730_ISA, "", "", 0, 0, 0, 0, 0, 0, 0},
/* CAL_TARGET_CYPRESS */ {ED_ATI_CAL_MACHINE_CYPRESS_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32,
400},
400, 4, 0, 0},
/* CAL_TARGET_JUNIPER */ {ED_ATI_CAL_MACHINE_JUNIPER_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32,
400},
400, 4, 0, 0},
/* CAL_TARGET_REDWOOD */ {ED_ATI_CAL_MACHINE_REDWOOD_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 16,
400},
/* CAL_TARGET_CEDAR */ {ED_ATI_CAL_MACHINE_CEDAR_ISA, "", "", 1, 8, 5, 256, 32 * Ki, 16, 400},
/* CAL_TARGET_SUMO */ {ED_ATI_CAL_MACHINE_SUMO_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 16, 400},
400, 4, 0, 0},
/* CAL_TARGET_CEDAR */ {ED_ATI_CAL_MACHINE_CEDAR_ISA, "", "", 1, 8, 5, 256, 32 * Ki, 16, 400, 4, 0, 0},
/* CAL_TARGET_SUMO */ {ED_ATI_CAL_MACHINE_SUMO_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 16, 400, 4, 0, 0},
/* CAL_TARGET_SUPERSUMO*/ {ED_ATI_CAL_MACHINE_SUPERSUMO_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 16,
400},
400, 4, 0, 0},
/* CAL_TARGET_WRESTLER*/ {ED_ATI_CAL_MACHINE_WRESTLER_ISA, "", "", 1, 8, 5, 256, 32 * Ki, 16,
400},
400, 4, 0, 0},
/* CAL_TARGET_CAYMAN */ {ED_ATI_CAL_MACHINE_CAYMAN_ISA, "", "", 1, 16, 4, 256, 32 * Ki, 32,
500},
/* CAL_TARGET_KAUAI */ {ED_ATI_CAL_MACHINE_KAUAI_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32, 400},
/* CAL_TARGET_BARTS */ {ED_ATI_CAL_MACHINE_BARTS_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32, 400},
/* CAL_TARGET_TURKS */ {ED_ATI_CAL_MACHINE_TURKS_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32, 400},
500, 5, 0, 0},
/* CAL_TARGET_KAUAI */ {ED_ATI_CAL_MACHINE_KAUAI_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32, 400, 4, 0, 0},
/* CAL_TARGET_BARTS */ {ED_ATI_CAL_MACHINE_BARTS_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32, 400, 4, 0, 0},
/* CAL_TARGET_TURKS */ {ED_ATI_CAL_MACHINE_TURKS_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32, 400, 4, 0, 0},
/* CAL_TARGET_CAICOS */ {ED_ATI_CAL_MACHINE_CAICOS_ISA, "", "", 1, 16, 5, 256, 32 * Ki, 32,
400},
400, 4, 0, 0},
/* CAL_TARGET_TAHITI */ {ED_ATI_CAL_MACHINE_TAHITI_ISA, "Tahiti", "tahiti", 4, 16, 1, 256,
64 * Ki, 32, 600},
64 * Ki, 32, gfx600, 6, 0, 0},
/* CAL_TARGET_PITCAIRN */ {ED_ATI_CAL_MACHINE_PITCAIRN_ISA, "Pitcairn", "pitcairn", 4, 16, 1,
256, 64 * Ki, 32, 600},
256, 64 * Ki, 32, gfx601, 6, 0, 1},
/* CAL_TARGET_CAPEVERDE */ {ED_ATI_CAL_MACHINE_CAPEVERDE_ISA, "Capeverde", "capeverde", 4, 16,
1, 256, 64 * Ki, 32, 600},
1, 256, 64 * Ki, 32, gfx601, 6, 0, 1},
/* CAL_TARGET_DEVASTATOR */ {ED_ATI_CAL_MACHINE_DEVASTATOR_ISA, "", "", 1, 16, 4, 256, 32 * Ki,
32, 500},
32, 500, 5, 0, 0},
/* CAL_TARGET_SCRAPPER */ {ED_ATI_CAL_MACHINE_SCRAPPER_ISA, "", "", 1, 16, 4, 256, 32 * Ki, 32,
500},
500, 5, 0, 0},
/* CAL_TARGET_OLAND */ {ED_ATI_CAL_MACHINE_OLAND_ISA, "Oland", "oland", 4, 16, 1, 256, 64 * Ki,
32, 600},
32, gfx602, 6, 0, 2},
/* CAL_TARGET_BONAIRE */ {ED_ATI_CAL_MACHINE_BONAIRE_ISA, "Bonaire", "bonaire", 4, 16, 1, 256,
64 * Ki, 32, 702},
64 * Ki, 32, gfx704, 7, 0, 4},
/* CAL_TARGET_SPECTRE */ {ED_ATI_CAL_MACHINE_SPECTRE_ISA, "Spectre", "spectre", 4, 16, 1, 256,
64 * Ki, 32, 701},
64 * Ki, 32, gfx700, 7, 0, 0},
/* CAL_TARGET_SPOOKY */ {ED_ATI_CAL_MACHINE_SPOOKY_ISA, "Spooky", "spooky", 4, 16, 1, 256,
64 * Ki, 32, 701},
64 * Ki, 32, gfx700, 7, 0, 0},
/* CAL_TARGET_KALINDI */ {ED_ATI_CAL_MACHINE_KALINDI_ISA, "Kalindi", "kalindi", 4, 16, 1, 256,
64 * Ki, 32, 702},
64 * Ki, 32, gfx700, 7, 0, 0},
/* CAL_TARGET_HAINAN */ {ED_ATI_CAL_MACHINE_HAINAN_ISA, "Hainan", "hainan", 4, 16, 1, 256,
64 * Ki, 32, 600},
64 * Ki, 32, gfx602, 6, 0, 2},
/* CAL_TARGET_HAWAII */ {ED_ATI_CAL_MACHINE_HAWAII_ISA, "Hawaii", "hawaii", 4, 16, 1, 256,
64 * Ki, 32, 702},
64 * Ki, 32, gfx701, 7, 0, 1},
/* CAL_TARGET_ICELAND */ {ED_ATI_CAL_MACHINE_ICELAND_ISA, "Iceland", "iceland", 4, 16, 1, 256,
64 * Ki, 32, 800},
64 * Ki, 32, gfx802, 8, 0, 2},
/* CAL_TARGET_TONGA */ {ED_ATI_CAL_MACHINE_TONGA_ISA, "Tonga", "tonga", 4, 16, 1, 256, 64 * Ki,
32, 800},
32, gfx802, 8, 0, 2},
/* CAL_TARGET_MULLINS */ {ED_ATI_CAL_MACHINE_GODAVARI_ISA, "Mullins", "mullins", 4, 16, 1, 256,
64 * Ki, 32, 702},
64 * Ki, 32, gfx705, 7, 0, 5},
/* CAL_TARGET_FIJI */ {ED_ATI_CAL_MACHINE_FIJI_ISA, "Fiji", "fiji", 4, 16, 1, 256, 64 * Ki, 32,
800},
gfx803, 8, 0, 3},
/* CAL_TARGET_CARRIZO */ {ED_ATI_CAL_MACHINE_CARRIZO_ISA, "Carrizo", "carrizo", 4, 16, 1, 256,
64 * Ki, 32, 800},
64 * Ki, 32, gfx801, 8, 0, 1},
/* CAL_TARGET_ELLESMERE */ {ED_ATI_CAL_MACHINE_ELLESMERE_ISA, "Ellesmere", "ellesmere", 4, 16,
1, 256, 64 * Ki, 32, 800},
1, 256, 64 * Ki, 32, gfx803, 8, 0, 3},
/* CAL_TARGET_BAFFIN */ {ED_ATI_CAL_MACHINE_BAFFIN_ISA, "Baffin", "baffin", 4, 16, 1, 256,
64 * Ki, 32, 800},
64 * Ki, 32, gfx803, 8, 0, 3},
/* CAL_TARGET_GREENLAND */ {ED_ATI_CAL_MACHINE_GREENLAND_ISA, IF(IS_BRAHMA, "", "gfx900"),
IF(IS_BRAHMA, "", "gfx900"), 4, 16, 1, 256, 64 * Ki, 32, 900},
IF(IS_BRAHMA, "", "gfx900"), 4, 16, 1, 256, 64 * Ki, 32, gfx900, 9, 0, 0},
/* CAL_TARGET_STONEY */ {ED_ATI_CAL_MACHINE_STONEY_ISA, "Stoney", "stoney", 4, 16, 1, 256,
64 * Ki, 32, 800},
/* CAL_TARGET_LEXA */ {ED_ATI_CAL_MACHINE_LEXA_ISA, "gfx804", "gfx804", 4, 16, 1, 256, 64 * Ki,
32, 800},
64 * Ki, 32, gfx810, 8, 1, 0},
/* CAL_TARGET_LEXA */ {ED_ATI_CAL_MACHINE_LEXA_ISA, "gfx803", "gfx803", 4, 16, 1, 256, 64 * Ki,
32, gfx803, 8, 0, 3},
/* CAL_TARGET_RAVEN */ {ED_ATI_CAL_MACHINE_RAVEN_ISA, IF(IS_BRAHMA, "", "gfx902"),
IF(IS_BRAHMA, "", "gfx902"), 4, 16, 1, 256, 64 * Ki, 32, 900},
/* CAL_TARGET_RAVEN2 */ {ED_ATI_CAL_MACHINE_RAVEN2_ISA, IF(IS_BRAHMA, "", "gfx902"),
IF(IS_BRAHMA, "", "gfx902"), 4, 16, 1, 256, 64 * Ki, 32, 900},
/* CAL_TARGET_RENOIR */{ ED_ATI_CAL_MACHINE_RENOIR_ISA, IF(IS_BRAHMA, "", "gfx902"),
IF(IS_BRAHMA, "", "gfx902"), 4, 16, 1, 256, 64 * Ki, 32, 900 },
/* CAL_TARGET_POLARIS22 */ {ED_ATI_CAL_MACHINE_POLARIS22_ISA, IF(IS_BRAHMA, "", "gfx804"),
IF(IS_BRAHMA, "", "gfx804"), 4, 16, 1, 256, 64 * Ki, 32, 800},
IF(IS_BRAHMA, "", "gfx902"), 4, 16, 1, 256, 64 * Ki, 32, gfx902, 9, 0, 2},
/* CAL_TARGET_RAVEN2 */ {ED_ATI_CAL_MACHINE_RAVEN2_ISA, IF(IS_BRAHMA, "", "gfx909"),
IF(IS_BRAHMA, "", "gfx902"), 4, 16, 1, 256, 64 * Ki, 32, gfx909, 9, 0, 9},
/* CAL_TARGET_RENOIR */{ ED_ATI_CAL_MACHINE_RENOIR_ISA, IF(IS_BRAHMA, "", "gfx90c"),
IF(IS_BRAHMA, "", "gfx90c"), 4, 16, 1, 256, 64 * Ki, 32, gfx90c, 9, 0, 12},
/* CAL_TARGET_POLARIS22 */ {ED_ATI_CAL_MACHINE_POLARIS22_ISA, IF(IS_BRAHMA, "", "gfx803"),
IF(IS_BRAHMA, "", "gfx803"), 4, 16, 1, 256, 64 * Ki, 32, gfx803, 8, 0, 3},
/* CAL_TARGET_VEGA12 */{ ED_ATI_CAL_MACHINE_VEGA12_ISA, IF(IS_BRAHMA, "", "gfx904"),
IF(IS_BRAHMA, "", "gfx904"), 4, 16, 1, 256, 64 * Ki, 32, 900 },
IF(IS_BRAHMA, "", "gfx904"), 4, 16, 1, 256, 64 * Ki, 32, gfx904, 9, 0, 4},
/* CAL_TARGET_VEGA20 */{ ED_ATI_CAL_MACHINE_VEGA20_ISA, IF(IS_BRAHMA, "", "gfx906"),
IF(IS_BRAHMA, "", "gfx906"), 4, 16, 1, 256, 64 * Ki, 32, 900 },
};
enum gfx_handle {
gfx700 = 700,
gfx701 = 701,
gfx702 = 702,
gfx800 = 800,
gfx801 = 801,
gfx804 = 804,
gfx810 = 810,
gfx900 = 900,
gfx902 = 902,
gfx903 = 903,
gfx904 = 904,
gfx906 = 906
IF(IS_BRAHMA, "", "gfx906"), 4, 16, 1, 256, 64 * Ki, 32, gfx906, 9, 0, 6 },
};
// FIXME: These need updating to new Target ID format. Or is all this code nw
// obsolete and should be deleted? How is XNACK and SRAMECC settings supported?
static constexpr const char* Gfx600 = "amdgcn-amd-amdhsa--gfx600";
static constexpr const char* Gfx601 = "amdgcn-amd-amdhsa--gfx601";
static constexpr const char* Gfx602 = "amdgcn-amd-amdhsa--gfx602";
static constexpr const char* Gfx700 = "amdgcn-amd-amdhsa--gfx700";
static constexpr const char* Gfx701 = "amdgcn-amd-amdhsa--gfx701";
static constexpr const char* Gfx800 = "amdgcn-amd-amdhsa--gfx801:xnack-";
static constexpr const char* Gfx702 = "amdgcn-amd-amdhsa--gfx702";
static constexpr const char* Gfx704 = "amdgcn-amd-amdhsa--gfx704";
static constexpr const char* Gfx705 = "amdgcn-amd-amdhsa--gfx705";
static constexpr const char* Gfx801 = "amdgcn-amd-amdhsa--gfx801:xnack+";
static constexpr const char* Gfx804 = "amdgcn-amd-amdhsa--gfx803";
static constexpr const char* Gfx802 = "amdgcn-amd-amdhsa--gfx802";
static constexpr const char* Gfx803 = "amdgcn-amd-amdhsa--gfx803";
static constexpr const char* Gfx810 = "amdgcn-amd-amdhsa--gfx810:xnack+";
static constexpr const char* Gfx900 = "amdgcn-amd-amdhsa--gfx900:xnack-";
static constexpr const char* Gfx902 = "amdgcn-amd-amdhsa--gfx902:xnack-";
static constexpr const char* Gfx903 = "amdgcn-amd-amdhsa--gfx903:xnack+"; // NOTE: gfx903 is gfx902 with xnack
static constexpr const char* Gfx902 = "amdgcn-amd-amdhsa--gfx902:xnack+";
static constexpr const char* Gfx904 = "amdgcn-amd-amdhsa--gfx904:xnack-";
static constexpr const char* Gfx906 = "amdgcn-amd-amdhsa--gfx906:xnack-";
static constexpr const char* Gfx909 = "amdgcn-amd-amdhsa--gfx909:xnack+";
static constexpr const char* Gfx90c = "amdgcn-amd-amdhsa--gfx90c:xnack+";
// Supported OpenCL versions
enum OclVersion { OpenCL10, OpenCL11, OpenCL12, OpenCL20, OpenCL21 };
projects/clr/rocclr/device/gpu/gpudevice.cpp
+3 -3
Просмотреть файл
@@ -594,9 +594,9 @@ void NullDevice::fillDeviceInfo(const CALdeviceattribs& calAttr, const gslMemInf
info_.globalMemChannelBankWidth_ = hwInfo()->memChannelBankWidth_;
info_.localMemSizePerCU_ = hwInfo()->localMemSizePerCU_;
info_.localMemBanks_ = hwInfo()->localMemBanks_;
info_.gfxipMajor_ = hwInfo()->gfxipVersion_ / 100;
info_.gfxipMinor_ = hwInfo()->gfxipVersion_ / 10 % 10;
info_.gfxipStepping_ = hwInfo()->gfxipVersion_ % 10;
info_.gfxipMajor_ = hwInfo()->gfxipMajor_;
info_.gfxipMinor_ = hwInfo()->gfxipMinor_;
info_.gfxipStepping_ = hwInfo()->gfxipStepping_;
info_.numAsyncQueues_ = numComputeRings;
projects/clr/rocclr/device/gpu/gpuprogram.cpp
+54 -55
Просмотреть файл
@@ -1793,6 +1793,18 @@ bool HSAILProgram::saveBinaryAndSetType(type_t type) {
hsa_isa_t ORCAHSALoaderContext::IsaFromName(const char* name) {
hsa_isa_t isa = {0};
if (!strcmp(Gfx600, name)) {
isa.handle = gfx600;
return isa;
}
if (!strcmp(Gfx601, name)) {
isa.handle = gfx601;
return isa;
}
if (!strcmp(Gfx602, name)) {
isa.handle = gfx602;
return isa;
}
if (!strcmp(Gfx700, name)) {
isa.handle = gfx700;
return isa;
@@ -1801,16 +1813,24 @@ hsa_isa_t ORCAHSALoaderContext::IsaFromName(const char* name) {
isa.handle = gfx701;
return isa;
}
if (!strcmp(Gfx800, name)) {
isa.handle = gfx800;
if (!strcmp(Gfx702, name)) {
isa.handle = gfx702;
return isa;
}
if (!strcmp(Gfx705, name)) {
isa.handle = gfx702;
return isa;
}
if (!strcmp(Gfx801, name)) {
isa.handle = gfx801;
return isa;
}
if (!strcmp(Gfx804, name)) {
isa.handle = gfx804;
if (!strcmp(Gfx802, name)) {
isa.handle = gfx802;
return isa;
}
if (!strcmp(Gfx803, name)) {
isa.handle = gfx803;
return isa;
}
if (!strcmp(Gfx810, name)) {
@@ -1825,10 +1845,6 @@ hsa_isa_t ORCAHSALoaderContext::IsaFromName(const char* name) {
isa.handle = gfx902;
return isa;
}
if (!strcmp(Gfx903, name)) {
isa.handle = gfx903;
return isa;
}
if (!strcmp(Gfx904, name)) {
isa.handle = gfx904;
return isa;
@@ -1837,62 +1853,45 @@ hsa_isa_t ORCAHSALoaderContext::IsaFromName(const char* name) {
isa.handle = gfx906;
return isa;
}
if (!strcmp(Gfx909, name)) {
isa.handle = gfx909;
return isa;
}
if (!strcmp(Gfx90c, name)) {
isa.handle = gfx90c;
return isa;
}
LogPrintfError("Couldn't find: %s!", name);
return isa;
}
bool ORCAHSALoaderContext::IsaSupportedByAgent(hsa_agent_t agent, hsa_isa_t isa) {
switch (program_->dev().hwInfo()->gfxipVersion_) {
default:
LogError("Unsupported gfxip version");
return false;
uint dev_gfxip = program_->dev().hwInfo()->gfxipVersion_;
uint isa_gfxip = isa.handle;
switch (dev_gfxip) {
case gfx700:
return isa.handle == gfx700;
case gfx704:
case gfx801:
case gfx802:
case gfx803:
case gfx810:
case gfx900:
case gfx902:
case gfx904:
case gfx906:
case gfx909:
case gfx90c:
return isa_gfxip == dev_gfxip;
case gfx701:
case gfx702:
// gfx701 only differs from gfx702 by faster fp operations and can be loaded on either device.
return isa.handle == gfx701 || isa.handle == gfx702;
case gfx800:
switch (program_->dev().hwInfo()->machine_) {
case ED_ATI_CAL_MACHINE_ICELAND_ISA:
case ED_ATI_CAL_MACHINE_TONGA_ISA:
return isa.handle == gfx800;
case ED_ATI_CAL_MACHINE_CARRIZO_ISA:
return isa.handle == gfx801;
case ED_ATI_CAL_MACHINE_FIJI_ISA:
case ED_ATI_CAL_MACHINE_ELLESMERE_ISA:
case ED_ATI_CAL_MACHINE_BAFFIN_ISA:
case ED_ATI_CAL_MACHINE_LEXA_ISA:
case ED_ATI_CAL_MACHINE_POLARIS22_ISA:
// gfx800 ISA has only sgrps limited and can be loaded.
// gfx801 ISA has XNACK limitations and can be loaded.
return isa.handle == gfx800 || isa.handle == gfx801 || isa.handle == gfx804;
case ED_ATI_CAL_MACHINE_STONEY_ISA:
return isa.handle == gfx810;
default:
assert(0);
return false;
}
case gfx900:
case gfx902:
case gfx903:
case gfx904:
case gfx906:
switch (program_->dev().hwInfo()->machine_) {
case ED_ATI_CAL_MACHINE_GREENLAND_ISA:
return isa.handle == gfx900;
case ED_ATI_CAL_MACHINE_RAVEN_ISA:
case ED_ATI_CAL_MACHINE_RENOIR_ISA:
return isa.handle == gfx902 || isa.handle == gfx903;
case ED_ATI_CAL_MACHINE_VEGA12_ISA:
return isa.handle == gfx904;
case ED_ATI_CAL_MACHINE_VEGA20_ISA:
return isa.handle == gfx906;
default:
assert(0);
return false;
}
return isa_gfxip == gfx701|| isa_gfxip == gfx702;
case gfx600:
case gfx601:
case gfx602:
default:
LogPrintfError("Unsupported gfxip version gfx%d", dev_gfxip);
return false;
}
}
projects/clr/rocclr/device/pal/palblit.hpp
+1 -1
Просмотреть файл
@@ -417,7 +417,7 @@ class KernelBlitManager : public DmaBlitManager {
//! Disable operator=
KernelBlitManager& operator=(const KernelBlitManager&);
amd::Program* program_; //!< GPU program obejct
amd::Program* program_; //!< GPU program object
amd::Kernel* kernels_[BlitTotal]; //!< GPU kernels for blit
amd::Memory* xferBuffers_[MaxXferBuffers]; //!< Transfer buffers for images
size_t xferBufferSize_; //!< Transfer buffer size
projects/clr/rocclr/device/rocm/rocblit.hpp
+1 -1
Просмотреть файл
@@ -463,7 +463,7 @@ class KernelBlitManager : public DmaBlitManager {
//! Disable operator=
KernelBlitManager& operator=(const KernelBlitManager&);
amd::Program* program_; //!< GPU program obejct
amd::Program* program_; //!< GPU program object
amd::Kernel* kernels_[BlitTotal]; //!< GPU kernels for blit
amd::Memory* constantBuffer_; //!< An internal CB for blits
mutable uint32_t constantBufferOffset_; //!< Current offset in the constant buffer