SWDEV-277559 - Remove AMDIL
The rest of AMDIL support will be removed along with orca backend. Change-Id: I0462501e7147dc4b99870fd02034d0a4a0496e55
This commit is contained in:
Jason Tang
Jason Tang
+74
-85
@@ -113,79 +113,76 @@ std::pair<const Isa*, const Isa*> Isa::supportedIsas() {
|
||||
// Supported Version Features Mem
|
||||
// SIMD Channel LDS LDS
|
||||
// SIMD/ SIMD Instr Bank Size/ Mem
|
||||
// Target ID HSAIL ID AMD IL ID ROC PAL GSL Maj/Min/Stp SRAMECC XNACK CU Width Width Width CU Banks
|
||||
{"gfx600", "Tahiti", "Tahiti", false, false, true, 6, 0, 0, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx601", "Pitcairn", "Pitcairn", false, false, true, 6, 0, 1, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Capeverde
|
||||
{"gfx602", "Oland", "Oland", false, false, true, 6, 0, 2, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Hainan
|
||||
{"gfx700", "Kaveri", "Kalindi", true, false, true, 7, 0, 0, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Spectre, Spooky, Kalindi
|
||||
{"gfx701", "Hawaii", "Hawaii", true, false, true, 7, 0, 1, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Actually Hawaiipro
|
||||
{"gfx702", "gfx702", nullptr, true, false, true, 7, 0, 2, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Actually Hawaii (can execute Hawiipro code)
|
||||
{"gfx703", nullptr, nullptr, false, false, true, 7, 0, 3, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Mullins
|
||||
{"gfx704", "Bonaire", "Bonaire", false, false, true, 7, 0, 4, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx705", "Mullins", "Mullins", false, false, true, 7, 0, 5, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Actually Godavari
|
||||
{"gfx801", nullptr, nullptr, true, true, true, 8, 0, 1, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx801:xnack-", "Carrizo", "Carrizo", true, true, true, 8, 0, 1, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx801:xnack+", nullptr, nullptr, true, true, true, 8, 0, 1, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx802", "Tonga", "Tonga", true, true, true, 8, 0, 2, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Iceland
|
||||
{"gfx803", "Fiji", "Fiji", true, true, true, 8, 0, 3, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Ellesmere/Polaris10, Baffin/Polaris11, Polaris12, Polaris22/VegaM
|
||||
{"gfx805", nullptr, nullptr, true, true, true, 8, 0, 5, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Tongapro
|
||||
{"gfx810", nullptr, nullptr, true, true, true, 8, 1, 0, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx810:xnack-", "Stoney", "Stoney", true, true, true, 8, 1, 0, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx810:xnack+", nullptr, nullptr, true, true, true, 8, 1, 0, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx900", "gfx901", nullptr, true, true, false, 9, 0, 0, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Greenland
|
||||
{"gfx900:xnack-", "gfx900", nullptr, true, true, false, 9, 0, 0, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx900:xnack+", "gfx901", nullptr, true, true, false, 9, 0, 0, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx902", "gfx903", nullptr, true, true, false, 9, 0, 2, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Raven
|
||||
{"gfx902:xnack-", "gfx902", nullptr, true, true, false, 9, 0, 2, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx902:xnack+", "gfx903", nullptr, true, true, false, 9, 0, 2, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx904", "gfx905", nullptr, true, true, false, 9, 0, 4, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Vega12
|
||||
{"gfx904:xnack-", "gfx904", nullptr, true, true, false, 9, 0, 4, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx904:xnack+", "gfx905", nullptr, true, true, false, 9, 0, 4, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906", "gfx907", nullptr, true, true, false, 9, 0, 6, ANY, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Vega20
|
||||
{"gfx906:sramecc-", "gfx907", nullptr, true, true, false, 9, 0, 6, OFF, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc+", nullptr, nullptr, true, true, false, 9, 0, 6, ON, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:xnack-", "gfx906", nullptr, true, true, false, 9, 0, 6, ANY, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:xnack+", "gfx907", nullptr, true, true, false, 9, 0, 6, ANY, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc-:xnack-", "gfx906", nullptr, true, true, false, 9, 0, 6, OFF, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc-:xnack+", "gfx907", nullptr, true, true, false, 9, 0, 6, OFF, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc+:xnack-", nullptr, nullptr, true, true, false, 9, 0, 6, ON, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc+:xnack+", nullptr, nullptr, true, true, false, 9, 0, 6, ON, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908", nullptr, nullptr, true, false, false, 9, 0, 8, ANY, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc-", nullptr, nullptr, true, false, false, 9, 0, 8, OFF, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc+", nullptr, nullptr, true, false, false, 9, 0, 8, ON, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:xnack-", nullptr, nullptr, true, false, false, 9, 0, 8, ANY, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:xnack-", nullptr, nullptr, true, false, false, 9, 0, 8, ANY, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc-:xnack-", nullptr, nullptr, true, false, false, 9, 0, 8, OFF, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc-:xnack+", nullptr, nullptr, true, false, false, 9, 0, 8, OFF, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc+:xnack-", nullptr, nullptr, true, false, false, 9, 0, 8, ON, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc+:xnack+", nullptr, nullptr, true, false, false, 9, 0, 8, ON, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx909", nullptr, nullptr, false, false, false, 9, 0, 9, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Raven2 (can execute Raven code)
|
||||
{"gfx909:xnack-", nullptr, nullptr, false, false, false, 9, 0, 9, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx909:xnack+", nullptr, nullptr, false, false, false, 9, 0, 9, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a", nullptr, nullptr, true, false, false, 9, 0, 10, ANY, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc-", nullptr, nullptr, true, false, false, 9, 0, 10, OFF, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc+", nullptr, nullptr, true, false, false, 9, 0, 10, ON, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:xnack-", nullptr, nullptr, true, false, false, 9, 0, 10, ANY, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:xnack-", nullptr, nullptr, true, false, false, 9, 0, 10, ANY, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc-:xnack-", nullptr, nullptr, true, false, false, 9, 0, 10, OFF, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc-:xnack+", nullptr, nullptr, true, false, false, 9, 0, 10, OFF, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc+:xnack-", nullptr, nullptr, true, false, false, 9, 0, 10, ON, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc+:xnack+", nullptr, nullptr, true, false, false, 9, 0, 10, ON, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90c", nullptr, nullptr, true, true, false, 9, 0, 12, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Renoir
|
||||
{"gfx90c:xnack-", "gfx90c", nullptr, true, true, false, 9, 0, 12, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90c:xnack+", nullptr, nullptr, true, true, false, 9, 0, 12, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1010", "gfx1010", nullptr, true, true, false, 10, 1, 0, NONE, ANY, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1010:xnack-", "gfx1010", nullptr, true, true, false, 10, 1, 0, NONE, OFF, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1010:xnack+", nullptr, nullptr, true, true, false, 10, 1, 0, NONE, ON, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1011", "gfx1011", nullptr, true, true, false, 10, 1, 1, NONE, ANY, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1011:xnack-", "gfx1011", nullptr, true, true, false, 10, 1, 1, NONE, OFF, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1011:xnack+", nullptr, nullptr, true, true, false, 10, 1, 1, NONE, ON, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1012", "gfx1012", nullptr, true, true, false, 10, 1, 2, NONE, ANY, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1012:xnack-", "gfx1012", nullptr, true, true, false, 10, 1, 2, NONE, OFF, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1012:xnack+", nullptr, nullptr, true, true, false, 10, 1, 2, NONE, ON, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1030", "gfx1030", nullptr, true, true, false, 10, 3, 0, NONE, NONE, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1031", "gfx1031", nullptr, true, true, false, 10, 3, 1, NONE, NONE, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1032", "gfx1032", nullptr, true, true, false, 10, 3, 2, NONE, NONE, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
// Target ID HSAIL ID ROC PAL GSL Maj/Min/Stp SRAMECC XNACK CU Width Width Width CU Banks
|
||||
{"gfx700", "Kaveri", true, false, true, 7, 0, 0, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Spectre, Spooky, Kalindi
|
||||
{"gfx701", "Hawaii", true, false, true, 7, 0, 1, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Actually Hawaiipro
|
||||
{"gfx702", "gfx702", true, false, true, 7, 0, 2, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Actually Hawaii (can execute Hawiipro code)
|
||||
{"gfx703", nullptr, false, false, true, 7, 0, 3, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Mullins
|
||||
{"gfx704", "Bonaire", false, false, true, 7, 0, 4, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx705", "Mullins", false, false, true, 7, 0, 5, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Actually Godavari
|
||||
{"gfx801", nullptr, true, true, true, 8, 0, 1, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx801:xnack-", "Carrizo", true, true, true, 8, 0, 1, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx801:xnack+", nullptr, true, true, true, 8, 0, 1, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx802", "Tonga", true, true, true, 8, 0, 2, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Iceland
|
||||
{"gfx803", "Fiji", true, true, true, 8, 0, 3, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Also Ellesmere/Polaris10, Baffin/Polaris11, Polaris12, Polaris22/VegaM
|
||||
{"gfx805", nullptr, true, true, true, 8, 0, 5, NONE, NONE, 4, 16, 1, 256, 64 * Ki, 32}, // Tongapro
|
||||
{"gfx810", nullptr, true, true, true, 8, 1, 0, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx810:xnack-", "Stoney", true, true, true, 8, 1, 0, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx810:xnack+", nullptr, true, true, true, 8, 1, 0, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx900", "gfx901", true, true, false, 9, 0, 0, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Greenland
|
||||
{"gfx900:xnack-", "gfx900", true, true, false, 9, 0, 0, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx900:xnack+", "gfx901", true, true, false, 9, 0, 0, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx902", "gfx903", true, true, false, 9, 0, 2, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Raven
|
||||
{"gfx902:xnack-", "gfx902", true, true, false, 9, 0, 2, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx902:xnack+", "gfx903", true, true, false, 9, 0, 2, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx904", "gfx905", true, true, false, 9, 0, 4, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Vega12
|
||||
{"gfx904:xnack-", "gfx904", true, true, false, 9, 0, 4, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx904:xnack+", "gfx905", true, true, false, 9, 0, 4, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906", "gfx907", true, true, false, 9, 0, 6, ANY, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Vega20
|
||||
{"gfx906:sramecc-", "gfx907", true, true, false, 9, 0, 6, OFF, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc+", nullptr, true, true, false, 9, 0, 6, ON, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:xnack-", "gfx906", true, true, false, 9, 0, 6, ANY, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:xnack+", "gfx907", true, true, false, 9, 0, 6, ANY, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc-:xnack-", "gfx906", true, true, false, 9, 0, 6, OFF, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc-:xnack+", "gfx907", true, true, false, 9, 0, 6, OFF, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc+:xnack-", nullptr, true, true, false, 9, 0, 6, ON, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx906:sramecc+:xnack+", nullptr, true, true, false, 9, 0, 6, ON, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908", nullptr, true, false, false, 9, 0, 8, ANY, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc-", nullptr, true, false, false, 9, 0, 8, OFF, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc+", nullptr, true, false, false, 9, 0, 8, ON, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:xnack-", nullptr, true, false, false, 9, 0, 8, ANY, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:xnack-", nullptr, true, false, false, 9, 0, 8, ANY, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc-:xnack-", nullptr, true, false, false, 9, 0, 8, OFF, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc-:xnack+", nullptr, true, false, false, 9, 0, 8, OFF, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc+:xnack-", nullptr, true, false, false, 9, 0, 8, ON, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx908:sramecc+:xnack+", nullptr, true, false, false, 9, 0, 8, ON, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx909", nullptr, false, false, false, 9, 0, 9, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Raven2 (can execute Raven code)
|
||||
{"gfx909:xnack-", nullptr, false, false, false, 9, 0, 9, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx909:xnack+", nullptr, false, false, false, 9, 0, 9, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a", nullptr, true, false, false, 9, 0, 10, ANY, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc-", nullptr, true, false, false, 9, 0, 10, OFF, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc+", nullptr, true, false, false, 9, 0, 10, ON, ANY, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:xnack-", nullptr, true, false, false, 9, 0, 10, ANY, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:xnack-", nullptr, true, false, false, 9, 0, 10, ANY, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc-:xnack-", nullptr, true, false, false, 9, 0, 10, OFF, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc-:xnack+", nullptr, true, false, false, 9, 0, 10, OFF, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc+:xnack-", nullptr, true, false, false, 9, 0, 10, ON, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90a:sramecc+:xnack+", nullptr, true, false, false, 9, 0, 10, ON, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90c", nullptr, true, true, false, 9, 0, 12, NONE, ANY, 4, 16, 1, 256, 64 * Ki, 32}, // Also Renoir
|
||||
{"gfx90c:xnack-", "gfx90c", true, true, false, 9, 0, 12, NONE, OFF, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx90c:xnack+", nullptr, true, true, false, 9, 0, 12, NONE, ON, 4, 16, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1010", "gfx1010", true, true, false, 10, 1, 0, NONE, ANY, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1010:xnack-", "gfx1010", true, true, false, 10, 1, 0, NONE, OFF, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1010:xnack+", nullptr, true, true, false, 10, 1, 0, NONE, ON, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1011", "gfx1011", true, true, false, 10, 1, 1, NONE, ANY, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1011:xnack-", "gfx1011", true, true, false, 10, 1, 1, NONE, OFF, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1011:xnack+", nullptr, true, true, false, 10, 1, 1, NONE, ON, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1012", "gfx1012", true, true, false, 10, 1, 2, NONE, ANY, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1012:xnack-", "gfx1012", true, true, false, 10, 1, 2, NONE, OFF, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1012:xnack+", nullptr, true, true, false, 10, 1, 2, NONE, ON, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1030", "gfx1030", true, true, false, 10, 3, 0, NONE, NONE, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1031", "gfx1031", true, true, false, 10, 3, 1, NONE, NONE, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
{"gfx1032", "gfx1032", true, true, false, 10, 3, 2, NONE, NONE, 2, 32, 1, 256, 64 * Ki, 32},
|
||||
};
|
||||
return std::make_pair(std::begin(supportedIsas_), std::end(supportedIsas_));
|
||||
}
|
||||
@@ -795,7 +792,7 @@ std::string ClBinary::getBIFSymbol(unsigned int symbolID) const {
|
||||
return "";
|
||||
}
|
||||
|
||||
void ClBinary::init(amd::option::Options* optionsObj, bool amdilRequired) {
|
||||
void ClBinary::init(amd::option::Options* optionsObj) {
|
||||
// option has higher priority than environment variable.
|
||||
if ((flags_ & BinarySourceMask) != BinaryRemoveSource) {
|
||||
// set to zero
|
||||
@@ -811,14 +808,6 @@ void ClBinary::init(amd::option::Options* optionsObj, bool amdilRequired) {
|
||||
flags_ |= (optionsObj->oVariables->BinLLVMIR ? BinarySaveLlvmir : BinaryNoSaveLlvmir);
|
||||
}
|
||||
|
||||
// If amdilRequired is true, force to save AMDIL (for correctness)
|
||||
if ((flags_ & BinaryAmdilMask) != BinaryRemoveAmdil || amdilRequired) {
|
||||
// set to zero
|
||||
flags_ = (flags_ & (~BinaryAmdilMask));
|
||||
flags_ |=
|
||||
((optionsObj->oVariables->BinAMDIL || amdilRequired) ? BinarySaveAmdil : BinaryNoSaveAmdil);
|
||||
}
|
||||
|
||||
if ((flags_ & BinaryIsaMask) != BinaryRemoveIsa) {
|
||||
// set to zero
|
||||
flags_ = (flags_ & (~BinaryIsaMask));
|
||||
@@ -852,7 +841,7 @@ bool ClBinary::isRecompilable(std::string& llvmBinary, amd::Elf::ElfPlatform the
|
||||
}
|
||||
if ((platform == amd::Elf::COMPLIB_PLATFORM) &&
|
||||
(((thePlatform == amd::Elf::CAL_PLATFORM) &&
|
||||
((elf_target == (uint16_t)EM_AMDIL) || (elf_target == (uint16_t)EM_HSAIL) ||
|
||||
((elf_target == (uint16_t)EM_HSAIL) ||
|
||||
(elf_target == (uint16_t)EM_HSAIL_64))) ||
|
||||
((thePlatform == amd::Elf::CPU_PLATFORM) &&
|
||||
((elf_target == (uint16_t)EM_386) || (elf_target == (uint16_t)EM_X86_64))))) {
|
||||
@@ -992,8 +981,8 @@ void ClBinary::setFlags(int encryptCode) {
|
||||
if (encryptCode != 0) {
|
||||
flags_ =
|
||||
(flags_ &
|
||||
(~(BinarySourceMask | BinaryLlvmirMask | BinaryAmdilMask | BinaryIsaMask | BinaryASMask)));
|
||||
flags_ |= (BinaryRemoveSource | BinaryRemoveLlvmir | BinaryRemoveAmdil | BinarySaveIsa |
|
||||
(~(BinarySourceMask | BinaryLlvmirMask | BinaryIsaMask | BinaryASMask)));
|
||||
flags_ |= (BinaryRemoveSource | BinaryRemoveLlvmir | BinarySaveIsa |
|
||||
BinaryRemoveAS);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -906,7 +906,7 @@ class ClBinary : public amd::HeapObject {
|
||||
//! Destructor
|
||||
virtual ~ClBinary();
|
||||
|
||||
void init(amd::option::Options* optionsObj, bool amdilRequired = false);
|
||||
void init(amd::option::Options* optionsObj);
|
||||
|
||||
/** called only in loading image routines, never storing routines */
|
||||
bool setBinary(const char* theBinary, size_t theBinarySize, bool allocated = false,
|
||||
@@ -986,7 +986,6 @@ class ClBinary : public amd::HeapObject {
|
||||
|
||||
bool saveSOURCE() { return ((flags_ & BinarySourceMask) == BinarySaveSource); }
|
||||
bool saveLLVMIR() { return ((flags_ & BinaryLlvmirMask) == BinarySaveLlvmir); }
|
||||
bool saveAMDIL() { return ((flags_ & BinaryAmdilMask) == BinarySaveAmdil); }
|
||||
bool saveISA() { return ((flags_ & BinaryIsaMask) == BinarySaveIsa); }
|
||||
|
||||
bool saveAS() { return ((flags_ & BinaryASMask) == BinarySaveAS); }
|
||||
@@ -1015,12 +1014,6 @@ class ClBinary : public amd::HeapObject {
|
||||
BinaryNoSaveLlvmir = 0x10,
|
||||
BinaryLlvmirMask = 0x18,
|
||||
|
||||
// AMDIL control
|
||||
BinarySaveAmdil = 0x0, // 0: default
|
||||
BinaryRemoveAmdil = 0x20, // for encrypted binary
|
||||
BinaryNoSaveAmdil = 0x40,
|
||||
BinaryAmdilMask = 0x60,
|
||||
|
||||
// ISA control
|
||||
BinarySaveIsa = 0x0, // 0: default
|
||||
BinaryRemoveIsa = 0x80, // for encrypted binary
|
||||
@@ -1293,11 +1286,6 @@ class Isa {
|
||||
return hsailId_;
|
||||
}
|
||||
|
||||
/// @returns This Isa's name to use with the AMD IL compiler.
|
||||
const char *amdIlName() const {
|
||||
return amdIlId_;
|
||||
}
|
||||
|
||||
/// @returns If the ROCm runtime supports the ISA.
|
||||
bool runtimeRocSupported() const {
|
||||
return runtimeRocSupported_;
|
||||
@@ -1397,7 +1385,7 @@ class Isa {
|
||||
|
||||
private:
|
||||
|
||||
constexpr Isa(const char* targetId, const char* hsailId, const char* amdIlId,
|
||||
constexpr Isa(const char* targetId, const char* hsailId,
|
||||
bool runtimeRocSupported, bool runtimePalSupported, bool runtimeGslSupported,
|
||||
uint32_t versionMajor, uint32_t versionMinor, uint32_t versionStepping,
|
||||
Feature sramecc, Feature xnack, uint32_t simdPerCU, uint32_t simdWidth,
|
||||
@@ -1405,7 +1393,6 @@ class Isa {
|
||||
uint32_t localMemSizePerCU, uint32_t localMemBanks)
|
||||
: targetId_(targetId),
|
||||
hsailId_(hsailId),
|
||||
amdIlId_(amdIlId),
|
||||
runtimeRocSupported_(runtimeRocSupported),
|
||||
runtimePalSupported_(runtimePalSupported),
|
||||
runtimeGslSupported_(runtimeGslSupported),
|
||||
@@ -1433,10 +1420,6 @@ class Isa {
|
||||
// unsupported.
|
||||
const char* hsailId_;
|
||||
|
||||
// @brief Isa's AMD IL name. Used for the Compiler Library for AMD IL
|
||||
// compilation using the Shader Compiler. Empty string if unsupported.
|
||||
const char* amdIlId_;
|
||||
|
||||
bool runtimeRocSupported_; //!< ROCm runtime is supported.
|
||||
bool runtimePalSupported_; //!< PAL runtime is supported.
|
||||
bool runtimeGslSupported_; //!< GSL runtime is supported.
|
||||
|
||||
@@ -1435,7 +1435,7 @@ bool Program::initBuild(amd::option::Options* options) {
|
||||
bool tempFile = false;
|
||||
|
||||
// true means hsail required
|
||||
clBinary()->init(options, true);
|
||||
clBinary()->init(options);
|
||||
if (options->isDumpFlagSet(amd::option::DUMP_BIF)) {
|
||||
outFileName = options->getDumpFileName(".bin");
|
||||
} else {
|
||||
@@ -1773,7 +1773,6 @@ std::vector<std::string> Program::ProcessOptions(amd::option::Options* options)
|
||||
|
||||
std::string processorName = device().isa().processorName();
|
||||
const char* hsailName = device().isa().hsailName();
|
||||
const char* amdIlName = device().isa().amdIlName();
|
||||
|
||||
optionsVec.push_back(std::string("-D__") + processorName + "__=1");
|
||||
optionsVec.push_back(std::string("-D__") + processorName + "=1");
|
||||
@@ -1781,11 +1780,6 @@ std::vector<std::string> Program::ProcessOptions(amd::option::Options* options)
|
||||
optionsVec.push_back(std::string("-D__") + hsailName + "__=1");
|
||||
optionsVec.push_back(std::string("-D__") + hsailName + "=1");
|
||||
}
|
||||
if (amdIlName && (strcmp(amdIlName, processorName.c_str()) != 0) &&
|
||||
(!hsailName || strcmp(amdIlName, hsailName) != 0)) {
|
||||
optionsVec.push_back(std::string("-D__") + amdIlName + "__=1");
|
||||
optionsVec.push_back(std::string("-D__") + amdIlName + "=1");
|
||||
}
|
||||
|
||||
// Set options for the standard device specific options
|
||||
// All our devices support these options now
|
||||
@@ -1977,7 +1971,7 @@ bool Program::initClBinary(const char* binaryIn, size_t size, amd::Os::FileDesc
|
||||
aclBinaryOptions binOpts = {0};
|
||||
binOpts.struct_size = sizeof(binOpts);
|
||||
binOpts.elfclass =
|
||||
(info().arch_id == aclX64 || info().arch_id == aclAMDIL64 || info().arch_id == aclHSAIL64)
|
||||
(info().arch_id == aclX64 || info().arch_id == aclHSAIL64)
|
||||
? ELFCLASS64
|
||||
: ELFCLASS32;
|
||||
binOpts.bitness = ELFDATA2LSB;
|
||||
|
||||
@@ -66,7 +66,7 @@ bool ClBinary::loadKernels(NullProgram& program, bool* hasRecompiled) {
|
||||
if (platform == amd::Elf::COMPLIB_PLATFORM) {
|
||||
// BIF 3.0
|
||||
uint32_t flag;
|
||||
aclTargetInfo tgtInfo = aclGetTargetInfo("amdil", dev().isa().amdIlName(), NULL);
|
||||
aclTargetInfo tgtInfo = aclGetTargetInfo("amdil", nullptr, NULL);
|
||||
if (!elfIn()->getFlags(flag)) {
|
||||
LogError("The OCL binary image loading failed: incorrect format");
|
||||
return false;
|
||||
@@ -241,7 +241,7 @@ bool ClBinary::loadKernels(NullProgram& program, bool* hasRecompiled) {
|
||||
} else if (!elfsymbol->IsKernel) {
|
||||
// Not a kernel. Add its metadata to the OCL binary in case recompilation happens
|
||||
// and the new binary is needed.
|
||||
if (saveAMDIL() && (elfsymbol->SymInfo[NDX_METADATA].size > 0)) {
|
||||
if (false && (elfsymbol->SymInfo[NDX_METADATA].size > 0)) {
|
||||
std::string fmetadata = "__OpenCL_";
|
||||
fmetadata.append(it.first);
|
||||
fmetadata.append("_fmetadata");
|
||||
@@ -371,14 +371,14 @@ bool ClBinary::loadKernels(NullProgram& program, bool* hasRecompiled) {
|
||||
bool ClBinary::storeKernel(const std::string& name, const NullKernel* nullKernel,
|
||||
Kernel::InitData* initData, const std::string& metadata,
|
||||
const std::string& ilSource) {
|
||||
if (!saveISA() && !saveAMDIL()) {
|
||||
if (!saveISA()) {
|
||||
return true;
|
||||
}
|
||||
|
||||
// should we save kernel metadata only under saveAMDIL()?
|
||||
bool kernelMetaStored = false;
|
||||
|
||||
if (saveAMDIL() && (ilSource.size() > 0)) {
|
||||
if (false && (ilSource.size() > 0)) {
|
||||
// Save IL (this is the per-kernel IL)
|
||||
std::string ilName = "__OpenCL_" + name + "_amdil";
|
||||
if (!elfOut()->addSymbol(amd::Elf::ILTEXT, ilName.c_str(), ilSource.data(),
|
||||
|
||||
+27
-103
@@ -77,11 +77,6 @@ struct CalDevice {
|
||||
static constexpr CalDevice supportedCalDevices[] = {
|
||||
// Prefer - NullDevice -
|
||||
// GFX Version GSL Machine CAL Name CAL Target PAL double OCL200
|
||||
{6, 0, 0, ED_ATI_CAL_MACHINE_TAHITI_ISA, "Tahiti", CAL_TARGET_TAHITI, false, true, false},
|
||||
{6, 0, 1, ED_ATI_CAL_MACHINE_PITCAIRN_ISA, "Pitcairn", CAL_TARGET_PITCAIRN, false, true, false},
|
||||
{6, 0, 1, ED_ATI_CAL_MACHINE_CAPEVERDE_ISA, "Capeverde", CAL_TARGET_CAPEVERDE, false, true, false},
|
||||
{6, 0, 2, ED_ATI_CAL_MACHINE_OLAND_ISA, "Oland", CAL_TARGET_OLAND, false, true, false},
|
||||
{6, 0, 2, ED_ATI_CAL_MACHINE_HAINAN_ISA, "Hainan", CAL_TARGET_HAINAN, false, true, false},
|
||||
{7, 0, 0, ED_ATI_CAL_MACHINE_KALINDI_ISA, "Kalindi", CAL_TARGET_KALINDI, false, true, true },
|
||||
{7, 0, 0, ED_ATI_CAL_MACHINE_SPECTRE_ISA, "Spectre", CAL_TARGET_SPECTRE, false, true, true },
|
||||
{7, 0, 0, ED_ATI_CAL_MACHINE_SPOOKY_ISA, "Spooky", CAL_TARGET_SPOOKY, false, true, true },
|
||||
@@ -270,45 +265,19 @@ bool NullDevice::create(const char* calName, const amd::Isa& isa, CALtarget targ
|
||||
// Fill the device info structure
|
||||
fillDeviceInfo(calAttr, memInfo, 4096, 1, 0);
|
||||
|
||||
if (NULL == compiler_) {
|
||||
#if !defined(ATI_OS_LINUX)
|
||||
char CompilerLibrary[220] = "";
|
||||
strcat_s(CompilerLibrary, "amdocl12cl" LP64_SWITCH("", "64") ".dll");
|
||||
#endif
|
||||
const char* library = getenv("COMPILER_LIBRARY");
|
||||
// Runtime doesn't know what local size could be on the real board
|
||||
info_.maxGlobalVariableSize_ = static_cast<size_t>(512 * Mi);
|
||||
|
||||
if (NULL == hsaCompiler_) {
|
||||
const char* library = getenv("HSA_COMPILER_LIBRARY");
|
||||
aclCompilerOptions opts = {
|
||||
sizeof(aclCompilerOptions_0_8),
|
||||
#if defined(ATI_OS_LINUX)
|
||||
library ? library : LINUX_ONLY("lib") "amdocl12cl" LP64_SWITCH(LINUX_SWITCH("32", ""), "64")
|
||||
LINUX_SWITCH(".so", ".dll"),
|
||||
#else
|
||||
library ? library : CompilerLibrary,
|
||||
#endif
|
||||
NULL,
|
||||
NULL,
|
||||
NULL,
|
||||
NULL,
|
||||
NULL,
|
||||
AMD_OCL_SC_LIB
|
||||
};
|
||||
compiler_ = aclCompilerInit(&opts, NULL);
|
||||
}
|
||||
|
||||
if (settings().hsail_ || (settings().oclVersion_ >= OpenCL20)) {
|
||||
// Runtime doesn't know what local size could be on the real board
|
||||
info_.maxGlobalVariableSize_ = static_cast<size_t>(512 * Mi);
|
||||
|
||||
if (NULL == hsaCompiler_) {
|
||||
const char* library = getenv("HSA_COMPILER_LIBRARY");
|
||||
aclCompilerOptions opts = {
|
||||
sizeof(aclCompilerOptions_0_8), library, NULL, NULL, NULL, NULL, NULL, AMD_OCL_SC_LIB};
|
||||
// Initialize the compiler handle
|
||||
acl_error error;
|
||||
hsaCompiler_ = aclCompilerInit(&opts, &error);
|
||||
if (error != ACL_SUCCESS) {
|
||||
LogPrintfError("Error initializing the compiler for offline CAL device %s", isa.targetId());
|
||||
return false;
|
||||
}
|
||||
sizeof(aclCompilerOptions_0_8), library, NULL, NULL, NULL, NULL, NULL, AMD_OCL_SC_LIB};
|
||||
// Initialize the compiler handle
|
||||
acl_error error;
|
||||
hsaCompiler_ = aclCompilerInit(&opts, &error);
|
||||
if (error != ACL_SUCCESS) {
|
||||
LogPrintfError("Error initializing the compiler for offline CAL device %s", isa.targetId());
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -317,7 +286,6 @@ bool NullDevice::create(const char* calName, const amd::Isa& isa, CALtarget targ
|
||||
|
||||
bool NullDevice::isHsailProgram(amd::option::Options* options) {
|
||||
bool isCIPlus = settings().ciPlus_;
|
||||
bool isHSAILcapable = settings().hsail_;
|
||||
bool isBlit = false;
|
||||
bool isSPIRV = false;
|
||||
bool isClang = false;
|
||||
@@ -369,10 +337,10 @@ bool NullDevice::isHsailProgram(amd::option::Options* options) {
|
||||
}
|
||||
isInputOptions = false;
|
||||
}
|
||||
if (isSPIRV || (isBlit && isCIPlus && isHSAILcapable) || isClang || isOCL20) {
|
||||
if (isSPIRV || (isBlit && isCIPlus) || isClang || isOCL20) {
|
||||
return true;
|
||||
}
|
||||
if (isLegacy || !isHSAILcapable || isEDG) {
|
||||
if (isLegacy || isEDG) {
|
||||
return false;
|
||||
}
|
||||
return true;
|
||||
@@ -1048,51 +1016,15 @@ bool Device::create(CALuint ordinal, CALuint numOfDevices) {
|
||||
fillDeviceInfo(getAttribs(), getMemInfo(), static_cast<size_t>(getMaxTextureSize()),
|
||||
engines().numComputeRings(), engines().numComputeRingsRT());
|
||||
|
||||
if (NULL == compiler_) {
|
||||
#if !defined(ATI_OS_LINUX)
|
||||
char CompilerLibrary[220] = "";
|
||||
strcat_s(CompilerLibrary, "amdocl12cl" LP64_SWITCH("", "64") ".dll");
|
||||
#endif
|
||||
|
||||
const char* library = getenv("COMPILER_LIBRARY");
|
||||
if (NULL == hsaCompiler_) {
|
||||
const char* library = getenv("HSA_COMPILER_LIBRARY");
|
||||
aclCompilerOptions opts = {
|
||||
sizeof(aclCompilerOptions_0_8),
|
||||
#if defined(ATI_OS_LINUX)
|
||||
library ? library : LINUX_ONLY("lib") "amdocl12cl" LP64_SWITCH(LINUX_SWITCH("32", ""), "64")
|
||||
LINUX_SWITCH(".so", ".dll"),
|
||||
#else
|
||||
library ? library : CompilerLibrary,
|
||||
#endif
|
||||
NULL,
|
||||
NULL,
|
||||
NULL,
|
||||
NULL,
|
||||
NULL,
|
||||
AMD_OCL_SC_LIB
|
||||
};
|
||||
compiler_ = aclCompilerInit(&opts, NULL);
|
||||
}
|
||||
|
||||
if (settings().hsail_ || (settings().oclVersion_ >= OpenCL20)) {
|
||||
if (NULL == hsaCompiler_) {
|
||||
const char* library = getenv("HSA_COMPILER_LIBRARY");
|
||||
aclCompilerOptions opts = {
|
||||
sizeof(aclCompilerOptions_0_8), library, NULL, NULL, NULL, NULL, NULL, AMD_OCL_SC_LIB};
|
||||
// Initialize the compiler handle
|
||||
acl_error error;
|
||||
hsaCompiler_ = aclCompilerInit(&opts, &error);
|
||||
if (error != ACL_SUCCESS) {
|
||||
LogError("Error initializing the compiler");
|
||||
return false;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
blitProgram_ = new BlitProgram(context_);
|
||||
// Create blit programs
|
||||
if (blitProgram_ == NULL || !blitProgram_->create(this)) {
|
||||
delete blitProgram_;
|
||||
blitProgram_ = NULL;
|
||||
LogError("Couldn't create blit kernels!");
|
||||
sizeof(aclCompilerOptions_0_8), library, NULL, NULL, NULL, NULL, NULL, AMD_OCL_SC_LIB};
|
||||
// Initialize the compiler handle
|
||||
acl_error error;
|
||||
hsaCompiler_ = aclCompilerInit(&opts, &error);
|
||||
if (error != ACL_SUCCESS) {
|
||||
LogError("Error initializing the compiler");
|
||||
return false;
|
||||
}
|
||||
}
|
||||
@@ -1694,20 +1626,12 @@ device::Memory* Device::createMemory(amd::Memory& owner) const {
|
||||
|
||||
bool Device::createSampler(const amd::Sampler& owner, device::Sampler** sampler) const {
|
||||
*sampler = NULL;
|
||||
if (settings().hsail_ || (settings().oclVersion_ >= OpenCL20)) {
|
||||
Sampler* gpuSampler = new Sampler(*this);
|
||||
if ((NULL == gpuSampler) || !gpuSampler->create(owner)) {
|
||||
delete gpuSampler;
|
||||
return false;
|
||||
}
|
||||
*sampler = gpuSampler;
|
||||
} else {
|
||||
//! Currently allocate the base device class for AMDIL path
|
||||
*sampler = new device::Sampler();
|
||||
if (*sampler == nullptr) {
|
||||
return false;
|
||||
}
|
||||
Sampler* gpuSampler = new Sampler(*this);
|
||||
if ((NULL == gpuSampler) || !gpuSampler->create(owner)) {
|
||||
delete gpuSampler;
|
||||
return false;
|
||||
}
|
||||
*sampler = gpuSampler;
|
||||
return true;
|
||||
}
|
||||
|
||||
|
||||
@@ -644,7 +644,7 @@ bool NullKernel::create(const std::string& code, const std::string& metadata,
|
||||
if ((binaryCode == NULL) && (binarySize == 0) && !code.empty()) {
|
||||
acl_error err;
|
||||
aclTargetInfo info = aclGetTargetInfo(nullDev().settings().use64BitPtr_ ? "amdil64" : "amdil",
|
||||
nullDev().isa().amdIlName(), &err);
|
||||
nullptr, &err);
|
||||
if (err != ACL_SUCCESS) {
|
||||
LogWarning("aclGetTargetInfo failed");
|
||||
return false;
|
||||
|
||||
@@ -44,7 +44,7 @@ namespace gpu {
|
||||
const aclTargetInfo& NullProgram::info() {
|
||||
acl_error err;
|
||||
info_ = aclGetTargetInfo(gpuNullDevice().settings().use64BitPtr_ ? "amdil64" : "amdil",
|
||||
device().isa().amdIlName(), &err);
|
||||
nullptr, &err);
|
||||
if (err != ACL_SUCCESS) {
|
||||
LogWarning("aclGetTargetInfo failed");
|
||||
}
|
||||
@@ -392,28 +392,6 @@ bool NullProgram::linkImpl(amd::option::Options* options) {
|
||||
buildLog_ += "Internal error: adding a kernel into OpenCL binary failed!\n";
|
||||
return false;
|
||||
}
|
||||
} else {
|
||||
// Non-kernel function, save metadata symbols for recompilation
|
||||
if (clBinary()->saveAMDIL()) {
|
||||
size_t metadataSize = baseFunc->metadata_.end_ - baseFunc->metadata_.begin_;
|
||||
if (metadataSize <= 0) {
|
||||
continue;
|
||||
}
|
||||
std::string metadataStr;
|
||||
// Get the metadata string
|
||||
metadataStr.insert(0, ilProgram_, baseFunc->metadata_.begin_, metadataSize);
|
||||
|
||||
std::stringstream aStream;
|
||||
aStream << "__OpenCL_" << baseFunc->name_ << "_fmetadata";
|
||||
std::string metaName = aStream.str();
|
||||
// Save metadata symbols in .rodata
|
||||
if (!clBinary()->elfOut()->addSymbol(amd::Elf::RODATA, metaName.c_str(),
|
||||
metadataStr.data(), metadataStr.size())) {
|
||||
buildLog_ += "Internal error: addSymbol failed!\n";
|
||||
LogError("AddSymbol failed");
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -755,28 +733,6 @@ bool NullProgram::linkImpl(const std::vector<device::Program*>& inputPrograms,
|
||||
buildLog_ += "Internal error: adding a kernel into OpenCL binary failed!\n";
|
||||
return false;
|
||||
}
|
||||
} else {
|
||||
// Non-kernel function, save metadata symbols for recompilation
|
||||
if (clBinary()->saveAMDIL()) {
|
||||
size_t metadataSize = baseFunc->metadata_.end_ - baseFunc->metadata_.begin_;
|
||||
if (metadataSize <= 0) {
|
||||
continue;
|
||||
}
|
||||
std::string metadataStr;
|
||||
// Get the metadata string
|
||||
metadataStr.insert(0, ilProgram_, baseFunc->metadata_.begin_, metadataSize);
|
||||
|
||||
std::stringstream aStream;
|
||||
aStream << "__OpenCL_" << baseFunc->name_ << "_fmetadata";
|
||||
std::string metaName = aStream.str();
|
||||
// Save metadata symbols in .rodata
|
||||
if (!clBinary()->elfOut()->addSymbol(amd::Elf::RODATA, metaName.c_str(),
|
||||
metadataStr.data(), metadataStr.size())) {
|
||||
buildLog_ += "Internal error: addSymbol failed!\n";
|
||||
LogError("AddSymbol failed");
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -470,11 +470,7 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
|
||||
// Check resource cache first for an appropriate resource
|
||||
gslRef_ = dev().resourceCache().findCalResource(&cal_);
|
||||
if (memType == Scratch) {
|
||||
if ((dev().settings().hsail_) || (dev().settings().oclVersion_ >= OpenCL20)) {
|
||||
desc.minAlignment = 64 * Ki;
|
||||
} else {
|
||||
desc.vaBase = static_cast<mcaddr>(0x100000000ULL);
|
||||
}
|
||||
desc.minAlignment = 64 * Ki;
|
||||
} else if ((gslRef_ != NULL) && (!dev().settings().use64BitPtr_)) {
|
||||
// Make sure runtime didn't pick a resource with > 4GB address
|
||||
if ((cal()->dimension_ == GSL_MOA_BUFFER) &&
|
||||
@@ -947,7 +943,7 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
|
||||
}
|
||||
}
|
||||
|
||||
if ((dev().settings().hsail_ || (dev().settings().oclVersion_ >= OpenCL20)) && !cal()->buffer_) {
|
||||
if (!cal()->buffer_) {
|
||||
hwSrd_ = dev().srds().allocSrdSlot(reinterpret_cast<address*>(&hwState_));
|
||||
if (0 == hwSrd_) {
|
||||
return false;
|
||||
@@ -1055,7 +1051,7 @@ void Resource::free() {
|
||||
}
|
||||
|
||||
// Free SRD for images
|
||||
if ((dev().settings().hsail_ || (dev().settings().oclVersion_ >= OpenCL20)) && !cal()->buffer_) {
|
||||
if (!cal()->buffer_) {
|
||||
dev().srds().freeSrdSlot(hwSrd_);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -136,9 +136,6 @@ Settings::Settings() {
|
||||
numDeviceEvents_ = 1024;
|
||||
numWaitEvents_ = 8;
|
||||
|
||||
// Disable HSAIL by default
|
||||
hsail_ = false;
|
||||
|
||||
// Don't support platform atomics by default.
|
||||
svmAtomics_ = false;
|
||||
|
||||
@@ -229,19 +226,11 @@ bool Settings::create(const CALdeviceattribs& calAttr, bool reportAsOCL12Device,
|
||||
case CAL_TARGET_HAWAII:
|
||||
ciPlus_ = true;
|
||||
sdmaProfiling_ = true;
|
||||
hsail_ = GPU_HSAIL_ENABLE;
|
||||
threadTraceEnable_ = AMD_THREAD_TRACE_ENABLE;
|
||||
// Fall through to SI ...
|
||||
case CAL_TARGET_PITCAIRN:
|
||||
case CAL_TARGET_CAPEVERDE:
|
||||
case CAL_TARGET_OLAND:
|
||||
case CAL_TARGET_HAINAN:
|
||||
reportFMAF_ = false;
|
||||
if (target == CAL_TARGET_HAWAII) {
|
||||
reportFMAF_ = true;
|
||||
}
|
||||
// Fall through ...
|
||||
case CAL_TARGET_TAHITI:
|
||||
// Cache line size is 64 bytes
|
||||
cacheLineSize_ = 64;
|
||||
// L1 cache size is 16KB
|
||||
@@ -282,10 +271,10 @@ bool Settings::create(const CALdeviceattribs& calAttr, bool reportAsOCL12Device,
|
||||
// This needs to be cleaned once 64bit addressing is stable
|
||||
if (oclVersion_ < OpenCL20) {
|
||||
use64BitPtr_ = flagIsDefault(GPU_FORCE_64BIT_PTR)
|
||||
? LP64_SWITCH(false, calAttr.isWorkstation || hsail_)
|
||||
? LP64_SWITCH(false, true)
|
||||
: GPU_FORCE_64BIT_PTR;
|
||||
} else {
|
||||
if (GPU_FORCE_64BIT_PTR || LP64_SWITCH(false, (hsail_ || (oclVersion_ >= OpenCL20)))) {
|
||||
if (GPU_FORCE_64BIT_PTR || LP64_SWITCH(false, true)) {
|
||||
use64BitPtr_ = true;
|
||||
}
|
||||
}
|
||||
|
||||
@@ -73,7 +73,6 @@ class Settings : public device::Settings {
|
||||
uint linearPersistentImage_ : 1; //!< Allocates linear images in persistent
|
||||
uint useSingleScratch_ : 1; //!< Allocates single scratch per device
|
||||
uint sdmaProfiling_ : 1; //!< Enables SDMA profiling
|
||||
uint hsail_ : 1; //!< Enables HSAIL compilation
|
||||
uint svmAtomics_ : 1; //!< SVM device atomics
|
||||
uint svmFineGrainSystem_ : 1; //!< SVM fine grain system support
|
||||
uint useDeviceQueue_ : 1; //!< Submit to separate device queue
|
||||
|
||||
@@ -71,11 +71,6 @@ struct PalDevice {
|
||||
|
||||
static constexpr PalDevice supportedPalDevices[] = {
|
||||
// GFX Version PAL GFX IP Level PAL Name PAL ASIC Revision
|
||||
{6, 0, 0, Pal::GfxIpLevel::GfxIp6, "Tahiti", Pal::AsicRevision::Tahiti},
|
||||
{6, 0, 1, Pal::GfxIpLevel::GfxIp6, "Pitcairn", Pal::AsicRevision::Pitcairn},
|
||||
{6, 0, 1, Pal::GfxIpLevel::GfxIp6, "Capeverde", Pal::AsicRevision::Capeverde},
|
||||
{6, 0, 2, Pal::GfxIpLevel::GfxIp6, "Oland", Pal::AsicRevision::Oland},
|
||||
{6, 0, 2, Pal::GfxIpLevel::GfxIp6, "Hainan", Pal::AsicRevision::Hainan},
|
||||
{7, 0, 0, Pal::GfxIpLevel::GfxIp7, "Kalindi", Pal::AsicRevision::Kalindi},
|
||||
{7, 0, 0, Pal::GfxIpLevel::GfxIp7, "Spectre", Pal::AsicRevision::Spectre},
|
||||
{7, 0, 0, Pal::GfxIpLevel::GfxIp7, "Spooky", Pal::AsicRevision::Spooky},
|
||||
|
||||
@@ -120,9 +120,6 @@ release(bool, GPU_USE_SINGLE_SCRATCH, false, \
|
||||
"Use single scratch buffer per device instead of per HW ring") \
|
||||
release(bool, AMD_OCL_WAIT_COMMAND, false, \
|
||||
"1 = Enable a wait for every submitted command") \
|
||||
/* HSAIL is by default, except Linux 32bit, because of known Catalyst 32bit issue */ \
|
||||
release(bool, GPU_HSAIL_ENABLE, LP64_SWITCH(LINUX_SWITCH(false,true),true), \
|
||||
"Enable HSAIL on dGPU stack (requires CI+ HW)") \
|
||||
release(uint, GPU_PRINT_CHILD_KERNEL, 0, \
|
||||
"Prints the specified number of the child kernels") \
|
||||
release(bool, GPU_USE_DEVICE_QUEUE, false, \
|
||||
|
||||
Reference in New Issue
Block a user