From 29c39ade39d2d495a608447fed81e8037a80c0cf Mon Sep 17 00:00:00 2001 From: foreman Date: Tue, 4 Sep 2018 16:57:45 -0400 Subject: [PATCH] P4 to Git Change 1601516 by asalmanp@asalmanp-ocl-stg on 2018/09/04 16:37:05 SWDEV-132899 - [OCL][GFX10.1] Adding gfx1010 support and use Pal::GfxIpLevel::GfxIp10_1/Pal::AsicRevision::Navi10 for gfx1010 and Pal::GfxIpLevel::GfxIp10/Pal::AsicRevision::Navi10Lite for Ariel (gfx1000) ReviewRequestURL = http://ocltc.amd.com/reviews/r/15740/ Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudefs.hpp#152 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuprogram.cpp#240 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palcounters.cpp#17 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldefs.hpp#38 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#108 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#54 edit [ROCm/clr commit: 42f5788b2e37d7fafbf51d54070542e8d40ec7a1] --- projects/clr/rocclr/runtime/device/gpu/gpudefs.hpp | 10 ++++------ .../clr/rocclr/runtime/device/gpu/gpuprogram.cpp | 13 ++++--------- .../clr/rocclr/runtime/device/pal/palcounters.cpp | 1 + projects/clr/rocclr/runtime/device/pal/paldefs.hpp | 7 +++++-- .../clr/rocclr/runtime/device/pal/paldevice.cpp | 10 ++++++++++ .../clr/rocclr/runtime/device/pal/palsettings.cpp | 3 +++ 6 files changed, 27 insertions(+), 17 deletions(-) diff --git a/projects/clr/rocclr/runtime/device/gpu/gpudefs.hpp b/projects/clr/rocclr/runtime/device/gpu/gpudefs.hpp index 2317c0302a..72b2701cc0 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpudefs.hpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpudefs.hpp @@ -163,8 +163,8 @@ static const AMDDeviceInfo DeviceInfo[] = { IF(IS_BRAHMA, "", "gfx904"), 4, 16, 1, 256, 64 * Ki, 32, 900 }, /* 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 }, - /* CAL_TARGET_NAVI10 */{ ED_ATI_CAL_MACHINE_NAVI10_ISA, IF(IS_BRAHMA, "", "gfx1000"), - IF(IS_BRAHMA, "", "gfx1000"), 4, 16, 1, 256, 64 * Ki, 32, 1000 }, + /* CAL_TARGET_NAVI10 */{ ED_ATI_CAL_MACHINE_NAVI10_ISA, IF(IS_BRAHMA, "", "gfx1010"), + IF(IS_BRAHMA, "", "gfx1010"), 4, 16, 1, 256, 64 * Ki, 32, 1010 }, }; @@ -184,8 +184,7 @@ enum gfx_handle { gfx905 = 905, gfx906 = 906, gfx907 = 907, - gfx1000 = 1000, - gfx1001 = 1001 + gfx1010 = 1010 }; static const char* Gfx700 = "AMD:AMDGPU:7:0:0"; @@ -202,8 +201,7 @@ static const char* Gfx904 = "AMD:AMDGPU:9:0:4"; static const char* Gfx905 = "AMD:AMDGPU:9:0:5"; static const char* Gfx906 = "AMD:AMDGPU:9:0:6"; static const char* Gfx907 = "AMD:AMDGPU:9:0:7"; -static const char* Gfx1000 = "AMD:AMDGPU:10:0:0"; -static const char* Gfx1001 = "AMD:AMDGPU:10:0:1"; +static const char* Gfx1010 = "AMD:AMDGPU:10:1:0"; // Supported OpenCL versions enum OclVersion { OpenCL10, OpenCL11, OpenCL12, OpenCL20, OpenCL21 }; diff --git a/projects/clr/rocclr/runtime/device/gpu/gpuprogram.cpp b/projects/clr/rocclr/runtime/device/gpu/gpuprogram.cpp index 59ea5d034d..4f7410a0b0 100644 --- a/projects/clr/rocclr/runtime/device/gpu/gpuprogram.cpp +++ b/projects/clr/rocclr/runtime/device/gpu/gpuprogram.cpp @@ -2226,12 +2226,8 @@ hsa_isa_t ORCAHSALoaderContext::IsaFromName(const char* name) { isa.handle = gfx907; return isa; } - if (!strcmp(Gfx1000, name)) { - isa.handle = gfx1000; - return isa; - } - if (!strcmp(Gfx1001, name)) { - isa.handle = gfx1001; + if (!strcmp(Gfx1010, name)) { + isa.handle = gfx1010; return isa; } @@ -2290,11 +2286,10 @@ bool ORCAHSALoaderContext::IsaSupportedByAgent(hsa_agent_t agent, hsa_isa_t isa) assert(0); return false; } - case gfx1000: - case gfx1001: + case gfx1010: switch (program_->dev().hwInfo()->machine_) { case ED_ATI_CAL_MACHINE_NAVI10_ISA: - return isa.handle == gfx1000 || isa.handle == gfx1001; + return isa.handle == gfx1010; default: assert(0); return false; diff --git a/projects/clr/rocclr/runtime/device/pal/palcounters.cpp b/projects/clr/rocclr/runtime/device/pal/palcounters.cpp index 3020a66f04..ded7c82dec 100644 --- a/projects/clr/rocclr/runtime/device/pal/palcounters.cpp +++ b/projects/clr/rocclr/runtime/device/pal/palcounters.cpp @@ -650,6 +650,7 @@ void PerfCounter::convertInfo() { } break; case Pal::GfxIpLevel::GfxIp10: + case Pal::GfxIpLevel::GfxIp10_1: if (info_.blockIndex_ < gfx10BlockIdPal.size()) { auto p = gfx10BlockIdPal[info_.blockIndex_]; info_.blockIndex_ = std::get<0>(p); diff --git a/projects/clr/rocclr/runtime/device/pal/paldefs.hpp b/projects/clr/rocclr/runtime/device/pal/paldefs.hpp index 658eb17418..ac6d5e3ddb 100644 --- a/projects/clr/rocclr/runtime/device/pal/paldefs.hpp +++ b/projects/clr/rocclr/runtime/device/pal/paldefs.hpp @@ -173,8 +173,11 @@ static const AMDDeviceInfo Gfx9PlusSubDeviceInfo[] = { /* Raven2 */{"gfx902", "gfx902", 4, 16, 1, 256, 64 * Ki, 32, 902, false}, /* Raven2 XNACK */{ LIGHTNING_SWITCH("gfx902","gfx903"), LIGHTNING_SWITCH("gfx902","gfx903"), 4, 16, 1, 256, 64 * Ki, 32, LIGHTNING_SWITCH(902, 903), true}, - /* Navi10 */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1000, false}, - /* Navi10 XNACK */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1000, true}, + /* Navi10 */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1010, false}, + /* Navi10 XNACK */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1010, true}, + /* Navi10Lite */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1000, false}, + /* Navi10Lite XNACK */{ "", "", 4, 16, 1, 256, 64 * Ki, 32, 1000, true}, + }; // Supported OpenCL versions diff --git a/projects/clr/rocclr/runtime/device/pal/paldevice.cpp b/projects/clr/rocclr/runtime/device/pal/paldevice.cpp index aaa4987605..95ac784907 100644 --- a/projects/clr/rocclr/runtime/device/pal/paldevice.cpp +++ b/projects/clr/rocclr/runtime/device/pal/paldevice.cpp @@ -162,13 +162,20 @@ bool NullDevice::init() { Pal::GfxIpLevel ipLevel = Pal::GfxIpLevel::_None; uint ipLevelMajor = round(pal::Gfx9PlusSubDeviceInfo[id].gfxipVersion_ / 100); + uint ipLevelMinor = round(pal::Gfx9PlusSubDeviceInfo[id].gfxipVersion_ / 10 % 10); switch (ipLevelMajor) { case 9: ipLevel = Pal::GfxIpLevel::GfxIp9; break; case 10: + switch (ipLevelMinor) { + case 0: ipLevel = Pal::GfxIpLevel::GfxIp10; break; + case 1: + ipLevel = Pal::GfxIpLevel::GfxIp10_1; + break; + } } Pal::AsicRevision revision = Pal::AsicRevision::Unknown; @@ -192,6 +199,9 @@ bool NullDevice::init() { revision = Pal::AsicRevision::Vega20; break; case 1000: + revision = Pal::AsicRevision::Navi10Lite; + break; + case 1010: revision = Pal::AsicRevision::Navi10; break; } diff --git a/projects/clr/rocclr/runtime/device/pal/palsettings.cpp b/projects/clr/rocclr/runtime/device/pal/palsettings.cpp index efd00ce489..65cdc0965b 100644 --- a/projects/clr/rocclr/runtime/device/pal/palsettings.cpp +++ b/projects/clr/rocclr/runtime/device/pal/palsettings.cpp @@ -175,6 +175,7 @@ bool Settings::create(const Pal::DeviceProperties& palProp, switch (palProp.revision) { case Pal::AsicRevision::Unknown: switch (palProp.gfxLevel) { + case Pal::GfxIpLevel::GfxIp10_1: case Pal::GfxIpLevel::GfxIp10: gfx10Plus_ = true; case Pal::GfxIpLevel::GfxIp9: @@ -185,6 +186,7 @@ bool Settings::create(const Pal::DeviceProperties& palProp, return false; } case Pal::AsicRevision::Navi10: + case Pal::AsicRevision::Navi10Lite: gfx10Plus_ = true; // Fall through to AI (gfx9) ... case Pal::AsicRevision::Vega20: @@ -445,6 +447,7 @@ bool Settings::create(const Pal::DeviceProperties& palProp, #if defined(WITH_LIGHTNING_COMPILER) switch (palProp.gfxLevel) { + case Pal::GfxIpLevel::GfxIp10_1: case Pal::GfxIpLevel::GfxIp10: case Pal::GfxIpLevel::GfxIp9: singleFpDenorm_ = true;