AI/rocm-systems
2
0
Fork 0
Koodi Ongelmat Pull-pyynnöt Actions Paketit Projektit Julkaisut Wiki Toiminta

Correct gfx10.3+ trap handler.

Selaa lähdekoodia 
Change-Id: I77d2b41c8882014a430d741ecd777718a1f61561


[ROCm/ROCR-Runtime commit: 8808ed3177]
This commit is contained in:
Laurent Morichetti
2021-01-20 16:54:09 -08:00
vanhempi 0aa0ebe2ee
commit 3eaae50cc6
2 muutettua tiedostoa jossa 34 lisäystä ja 3 poistoa
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocr-runtime/runtime/hsa-runtime/core/inc/amd_gpu_shaders.h
+24 -1
Näytä tiedosto
@@ -241,6 +241,9 @@ static const unsigned int kCodeTrapHandler9[] = {
// Restore m0
s_mov_b32 m0, ttmp2
.if ((.amdgcn.gfx_generation_number == 10 && .amdgcn.gfx_generation_minor >= 3) || .amdgcn.gfx_generation_number > 10)
s_branch .halt_wave
.else
// If PC is at an s_endpgm instruction then don't halt the wavefront.
s_and_b32 ttmp1, ttmp1, SQ_WAVE_PC_HI_ADDRESS_MASK
s_load_dword ttmp2, [ttmp0, ttmp1]
@@ -257,6 +260,7 @@ static const unsigned int kCodeTrapHandler9[] = {
s_and_b32 ttmp12, ttmp12, ~SQ_WAVE_STATUS_HALT_MASK
mExitTrap
.endif
.excp_raised:
s_bitset1_b32 ttmp11, TTMP11_EXCP_RAISED_BIT
@@ -299,6 +303,9 @@ static const unsigned int kCodeTrapHandler9[] = {
s_bitset1_b32 ttmp11, TTMP11_TRAP_RAISED_BIT
.signal_error:
.if (.amdgcn.gfx_generation_number == 10 && .amdgcn.gfx_generation_minor >= 3)
// This needs to be rewriten for gfx10.3 as scalar stores are not available.
.else
// FIXME: don't trash ttmp4/ttmp5 when exception handling is unified.
s_mov_b32 ttmp4, ttmp3
@@ -348,6 +355,7 @@ static const unsigned int kCodeTrapHandler9[] = {
s_mov_b32 m0, 0x0
s_nop 0
s_sendmsg sendmsg(MSG_INTERRUPT)
.endif
.skip_event_trigger:
// Since we trashed ttmp4/ttmp5, reset the wave_id to 0
@@ -497,7 +505,7 @@ static const unsigned int kCodeFill10[] = {
0xD70F6A02, 0x0002040C, 0xD5286A03, 0x01A90103, 0xBF82FFF5, 0xBF810000,
};
static const unsigned int kCodeTrapHandler10[] = {
static const unsigned int kCodeTrapHandler1010[] = {
0x93eeff6d, 0x00080010, 0xbf850044, 0xb96ef803, 0x876fff6e, 0x00000900,
0xbf850033, 0x876fff6e, 0x00007080, 0xbf840003, 0xbef71d88, 0xbf820001,
0xbef71d87, 0xbeee037e, 0xbeef037f, 0xbefe03ff, 0x80000000, 0xbf90000a,
@@ -523,6 +531,21 @@ static const unsigned int kCodeTrapHandler10[] = {
0xb9eef807, 0x87fe7e7e, 0x87ea6a6a, 0xb9f8f802, 0xbe80226c,
};
static const unsigned int kCodeTrapHandler10[] = {
0x93eeff6d, 0x00080010, 0xbf85002b, 0xb96ef803, 0x876fff6e, 0x00000900,
0xbf85001a, 0x876fff6e, 0x00007080, 0xbf840003, 0xbef71d88, 0xbf820001,
0xbef71d87, 0xbeee037e, 0xbeef037f, 0xbefe03ff, 0x80000000, 0xbf90000a,
0xbf800007, 0xbf0c9f7e, 0xbf84fffd, 0xbeff036f, 0x877eff7e, 0x00000fff,
0xbeef037e, 0xbefe036e, 0xbeef1d96, 0xbeee037c, 0xbefc036f, 0xbf800000,
0xbf900001, 0xbefc036e, 0xbf820019, 0xbef71d88, 0x876fff6e, 0x10000100,
0xbf06ff6f, 0x00000100, 0xbeef03ff, 0x20000000, 0xbf85000f, 0x876fff6e,
0x00000800, 0xbeef03f4, 0xbf85000b, 0xbf82000c, 0xbf09836e, 0xbf85ffdc,
0xbf06826e, 0xbeef03ff, 0x80000000, 0xbf850003, 0x806c846c, 0x826d806d,
0xbf820005, 0xbef71d87, 0xbef00380, 0xbef10380, 0x8878ff78, 0x00002000,
0x906e8977, 0x876fff6e, 0x003f8000, 0x906e8677, 0x876eff6e, 0x02000000,
0x886e6f6e, 0xb9eef807, 0x87fe7e7e, 0x87ea6a6a, 0xb9f8f802, 0xbe80226c,
};
} // namespace amd
} // namespace rocr
projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp
+10 -2
Näytä tiedosto
@@ -202,6 +202,7 @@ void GpuAgent::AssembleShader(const char* func_name, AssembleTarget assemble_tar
ASICShader compute_7;
ASICShader compute_8;
ASICShader compute_9;
ASICShader compute_1010;
ASICShader compute_10;
};
@@ -211,6 +212,7 @@ void GpuAgent::AssembleShader(const char* func_name, AssembleTarget assemble_tar
{NULL, 0, 0, 0},
{kCodeTrapHandler8, sizeof(kCodeTrapHandler8), 2, 4},
{kCodeTrapHandler9, sizeof(kCodeTrapHandler9), 2, 4},
{kCodeTrapHandler1010, sizeof(kCodeTrapHandler1010), 2, 4},
{kCodeTrapHandler10, sizeof(kCodeTrapHandler10), 2, 4},
}},
{"CopyAligned",
@@ -219,6 +221,7 @@ void GpuAgent::AssembleShader(const char* func_name, AssembleTarget assemble_tar
{kCodeCopyAligned8, sizeof(kCodeCopyAligned8), 32, 12},
{kCodeCopyAligned8, sizeof(kCodeCopyAligned8), 32, 12},
{kCodeCopyAligned10, sizeof(kCodeCopyAligned10), 32, 12},
{kCodeCopyAligned10, sizeof(kCodeCopyAligned10), 32, 12},
}},
{"CopyMisaligned",
{
@@ -226,6 +229,7 @@ void GpuAgent::AssembleShader(const char* func_name, AssembleTarget assemble_tar
{kCodeCopyMisaligned8, sizeof(kCodeCopyMisaligned8), 23, 10},
{kCodeCopyMisaligned8, sizeof(kCodeCopyMisaligned8), 23, 10},
{kCodeCopyMisaligned10, sizeof(kCodeCopyMisaligned10), 23, 10},
{kCodeCopyMisaligned10, sizeof(kCodeCopyMisaligned10), 23, 10},
}},
{"Fill",
{
@@ -233,6 +237,7 @@ void GpuAgent::AssembleShader(const char* func_name, AssembleTarget assemble_tar
{kCodeFill8, sizeof(kCodeFill8), 19, 8},
{kCodeFill8, sizeof(kCodeFill8), 19, 8},
{kCodeFill10, sizeof(kCodeFill10), 19, 8},
{kCodeFill10, sizeof(kCodeFill10), 19, 8},
}}};
auto compiled_shader_it = compiled_shaders.find(func_name);
@@ -249,10 +254,13 @@ void GpuAgent::AssembleShader(const char* func_name, AssembleTarget assemble_tar
asic_shader = &compiled_shader_it->second.compute_8;
break;
case 9:
asic_shader = &compiled_shader_it->second.compute_9;
asic_shader = &compiled_shader_it->second.compute_9;
break;
case 10:
asic_shader = &compiled_shader_it->second.compute_10;
if(isa_->GetMinorVersion() == 1)
asic_shader = &compiled_shader_it->second.compute_1010;
else
asic_shader = &compiled_shader_it->second.compute_10;
break;
default:
assert(false && "Precompiled shader unavailable for target");