summaryrefslogtreecommitdiffstats
path: root/drivers/gpu/drm/amd/amdkfd
diff options
context:
space:
mode:
authorJay Cornwall <Jay.Cornwall@amd.com>2018-05-01 17:56:02 -0400
committerOded Gabbay <oded.gabbay@gmail.com>2018-05-01 17:56:02 -0400
commit87e6d4e0777daf774ed9aa59ed25b6ebaaad7052 (patch)
tree1062fa04a237eef7021afd3401bea92ddb5015f3 /drivers/gpu/drm/amd/amdkfd
parent24f48a42038f5baaae49b181b64782ecfb703a9c (diff)
downloadop-kernel-dev-87e6d4e0777daf774ed9aa59ed25b6ebaaad7052.zip
op-kernel-dev-87e6d4e0777daf774ed9aa59ed25b6ebaaad7052.tar.gz
drm/amdkfd: Reduce priority of context-saving waves before spin-wait
Synchronization between context-saving wavefronts is achieved by sending a SAVEWAVE message to the SPI and then spin-waiting for a response. These spin-waiting wavefronts may inhibit the progress of other wavefronts in the context save handler, leading to the synchronization condition never being achieved. Before spin-waiting reduce the priority of each wavefront to guarantee foward progress in the others. Signed-off-by: Jay Cornwall <Jay.Cornwall@amd.com> Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com> Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com> Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd')
-rw-r--r--drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm10
-rw-r--r--drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm8
2 files changed, 15 insertions, 3 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm
index 997a383d..34eabcd 100644
--- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm
+++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx8.asm
@@ -98,6 +98,7 @@ var SWIZZLE_EN = 0 //whether we use swi
/**************************************************************************/
var SQ_WAVE_STATUS_INST_ATC_SHIFT = 23
var SQ_WAVE_STATUS_INST_ATC_MASK = 0x00800000
+var SQ_WAVE_STATUS_SPI_PRIO_SHIFT = 1
var SQ_WAVE_STATUS_SPI_PRIO_MASK = 0x00000006
var SQ_WAVE_LDS_ALLOC_LDS_SIZE_SHIFT = 12
@@ -319,6 +320,10 @@ end
s_sendmsg sendmsg(MSG_SAVEWAVE) //send SPI a message and wait for SPI's write to EXEC
end
+ // Set SPI_PRIO=2 to avoid starving instruction fetch in the waves we're waiting for.
+ s_or_b32 s_save_tmp, s_save_status, (2 << SQ_WAVE_STATUS_SPI_PRIO_SHIFT)
+ s_setreg_b32 hwreg(HW_REG_STATUS), s_save_tmp
+
L_SLEEP:
s_sleep 0x2 // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause SQ hang, since the 7,8th wave could not get arbit to exec inst, while other waves are stuck into the sleep-loop and waiting for wrexec!=0
@@ -1132,7 +1137,7 @@ end
#endif
static const uint32_t cwsr_trap_gfx8_hex[] = {
- 0xbf820001, 0xbf820123,
+ 0xbf820001, 0xbf820125,
0xb8f4f802, 0x89748674,
0xb8f5f803, 0x8675ff75,
0x00000400, 0xbf850011,
@@ -1158,7 +1163,8 @@ static const uint32_t cwsr_trap_gfx8_hex[] = {
0x867aff7a, 0x00007fff,
0xb97af807, 0xbef2007e,
0xbef3007f, 0xbefe0180,
- 0xbf900004, 0xbf8e0002,
+ 0xbf900004, 0x877a8474,
+ 0xb97af802, 0xbf8e0002,
0xbf88fffe, 0xbef8007e,
0x8679ff7f, 0x0000ffff,
0x8779ff79, 0x00040000,
diff --git a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm
index 033580c..cac8d49 100644
--- a/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm
+++ b/drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx9.asm
@@ -97,6 +97,7 @@ var ACK_SQC_STORE = 1 //workaround for suspected SQC store bug causing
/**************************************************************************/
var SQ_WAVE_STATUS_INST_ATC_SHIFT = 23
var SQ_WAVE_STATUS_INST_ATC_MASK = 0x00800000
+var SQ_WAVE_STATUS_SPI_PRIO_SHIFT = 1
var SQ_WAVE_STATUS_SPI_PRIO_MASK = 0x00000006
var SQ_WAVE_STATUS_HALT_MASK = 0x2000
@@ -362,6 +363,10 @@ end
s_sendmsg sendmsg(MSG_SAVEWAVE) //send SPI a message and wait for SPI's write to EXEC
end
+ // Set SPI_PRIO=2 to avoid starving instruction fetch in the waves we're waiting for.
+ s_or_b32 s_save_tmp, s_save_status, (2 << SQ_WAVE_STATUS_SPI_PRIO_SHIFT)
+ s_setreg_b32 hwreg(HW_REG_STATUS), s_save_tmp
+
L_SLEEP:
s_sleep 0x2 // sleep 1 (64clk) is not enough for 8 waves per SIMD, which will cause SQ hang, since the 7,8th wave could not get arbit to exec inst, while other waves are stuck into the sleep-loop and waiting for wrexec!=0
@@ -1210,7 +1215,7 @@ end
#endif
static const uint32_t cwsr_trap_gfx9_hex[] = {
- 0xbf820001, 0xbf820158,
+ 0xbf820001, 0xbf82015a,
0xb8f8f802, 0x89788678,
0xb8f1f803, 0x866eff71,
0x00000400, 0xbf850034,
@@ -1249,6 +1254,7 @@ static const uint32_t cwsr_trap_gfx9_hex[] = {
0x00007fff, 0xb970f807,
0xbeee007e, 0xbeef007f,
0xbefe0180, 0xbf900004,
+ 0x87708478, 0xb970f802,
0xbf8e0002, 0xbf88fffe,
0xb8f02a05, 0x80708170,
0x8e708a70, 0xb8f11605,
OpenPOWER on IntegriCloud