kfdtest: Add coherency tests for Aqua Vanjaram
Aqua Vanjaram is intended to have fine-grained coherency from anywhere to anywhere else using read-acquire and write-release primitives. Add a test that writes to memory covered by five different cache lines, then write-releases, while another thread read-acquires, then reads those five locations in memory. There are nine variations of the test to cover CPU-GPU, same-GPU and across-GPU, vector instructions and scalar instructions, and data local to the acquirer or receiver. Signed-off-by: David Francis <David.Francis@amd.com> Change-Id: I20d2db5c53bd280e971479aad7e61df6ed5d3623
This commit is contained in:
David Francis
committed by
Graham Sider
Graham Sider
parent
0696f06c16
commit
30b1f23f7a
@@ -400,6 +400,215 @@ const char *WriteAndSignalIsa =
|
||||
s_endpgm
|
||||
)";
|
||||
|
||||
/* Input:
|
||||
* s[0:1], A buffer of at least 64 * 6 bytes
|
||||
*
|
||||
* Store the value 0x77 at the 5 addresses 0x40,
|
||||
* 0x80, ..., 0x140 in the buffer
|
||||
*
|
||||
* Aqua Vanjaram only
|
||||
*/
|
||||
const char *FlushBufferForAcquireReleaseIsa =
|
||||
SHADER_START
|
||||
R"(
|
||||
.if (.amdgcn.gfx_generation_number == 9 && .amdgcn.gfx_generation_minor == 4)
|
||||
s_mov_b32 s11, 0x77
|
||||
s_mov_b32 s12, 0x0
|
||||
// Store some data on 5 different cache lines
|
||||
s_store_dword s12, s[0:1], 0x0 glc
|
||||
s_store_dword s11, s[0:1], 0x40 glc
|
||||
s_store_dword s11, s[0:1], 0x80 glc
|
||||
s_store_dword s11, s[0:1], 0xc0 glc
|
||||
s_store_dword s11, s[0:1], 0x100 glc
|
||||
s_store_dword s11, s[0:1], 0x140 glc
|
||||
s_waitcnt lgkmcnt(0)
|
||||
.endif
|
||||
s_endpgm
|
||||
)";
|
||||
|
||||
/* Input:
|
||||
* s[0:1], A buffer of at least 64 * 6 bytes,
|
||||
* shared with the acquiring shader
|
||||
*
|
||||
* Store the values 1 - 5 at the 5 addresses 0x40,
|
||||
* 0x80, ..., 0x140 in the buffer, then signal
|
||||
* the flag at address 0x0 in the buffer.
|
||||
*
|
||||
* Uses vector stores
|
||||
*
|
||||
* Aqua Vanjaram only
|
||||
*/
|
||||
const char *WriteReleaseVectorIsa =
|
||||
SHADER_START
|
||||
R"(
|
||||
.if (.amdgcn.gfx_generation_number == 9 && .amdgcn.gfx_generation_minor == 4)
|
||||
v_mov_b32 v11, 0x1
|
||||
v_mov_b32 v12, 0x2
|
||||
v_mov_b32 v13, 0x3
|
||||
v_mov_b32 v14, 0x4
|
||||
v_mov_b32 v15, 0x5
|
||||
v_mov_b32 v21, 0x40
|
||||
v_mov_b32 v22, 0x80
|
||||
v_mov_b32 v23, 0xc0
|
||||
v_mov_b32 v24, 0x100
|
||||
v_mov_b32 v25, 0x140
|
||||
// Store some data on 5 different cache lines
|
||||
global_store_dword v21, v11, s[0:1]
|
||||
global_store_dword v22, v12, s[0:1]
|
||||
global_store_dword v23, v13, s[0:1]
|
||||
global_store_dword v24, v14, s[0:1]
|
||||
global_store_dword v25, v15, s[0:1] nt sc1 sc0
|
||||
s_waitcnt vmcnt(0)
|
||||
// Write-Release
|
||||
s_mov_b32 s16, 0x1
|
||||
buffer_wbl2 sc1 sc0
|
||||
s_waitcnt vmcnt(0) & lgkmcnt(0)
|
||||
s_store_dword s16, s[0:1], 0x0 glc
|
||||
.endif
|
||||
s_endpgm
|
||||
)";
|
||||
|
||||
/* Input:
|
||||
* s[0:1], A buffer of at least 64 * 6 bytes,
|
||||
* shared with the acquiring shader
|
||||
*
|
||||
* Store the values 6 - 10 at the 5 addresses 0x40,
|
||||
* 0x80, ..., 0x140 in the buffer, then signal
|
||||
* the flag at address 0x0 in the buffer.
|
||||
*
|
||||
* Uses scalar stores
|
||||
*
|
||||
* Aqua Vanjaram only
|
||||
*/
|
||||
const char *WriteReleaseScalarIsa =
|
||||
SHADER_START
|
||||
R"(
|
||||
.if (.amdgcn.gfx_generation_number == 9 && .amdgcn.gfx_generation_minor == 4)
|
||||
s_mov_b32 s11, 0x6
|
||||
s_mov_b32 s12, 0x7
|
||||
s_mov_b32 s13, 0x8
|
||||
s_mov_b32 s14, 0x9
|
||||
s_mov_b32 s15, 0xa
|
||||
// Store some data on 5 different cache lines
|
||||
s_store_dword s11, s[0:1], 0x40
|
||||
s_store_dword s12, s[0:1], 0x80
|
||||
s_store_dword s13, s[0:1], 0xc0
|
||||
s_store_dword s14, s[0:1], 0x100
|
||||
s_store_dword s15, s[0:1], 0x140 glc
|
||||
s_waitcnt lgkmcnt(0)
|
||||
// Write-Release
|
||||
s_dcache_wb // WB Scalar L1 cache
|
||||
s_mov_b32 s16, 0x1
|
||||
buffer_wbl2 sc1 sc0
|
||||
s_waitcnt vmcnt(0) & lgkmcnt(0)
|
||||
s_store_dword s16, s[0:1], 0x0 glc
|
||||
s_waitcnt lgkmcnt(0)
|
||||
.endif
|
||||
s_endpgm
|
||||
)";
|
||||
|
||||
/* Input:
|
||||
* s[0:1], A buffer of at least 64 * 6 bytes,
|
||||
* shared with the releasing shader
|
||||
* s[2:3], A buffer of at least 64 * 6 bytes,
|
||||
* accessible by the CPU, used for output
|
||||
*
|
||||
* Polls the flag at address 0x0 in the shared buffer.
|
||||
* When the signal is received, read the values
|
||||
* at the 5 addresses 0x40, 0x80, ... 0x140,
|
||||
* and store them at the same locations in
|
||||
* the output buffer
|
||||
*
|
||||
* Uses vector loads
|
||||
*
|
||||
* Aqua Vanjaram only
|
||||
*/
|
||||
const char *ReadAcquireVectorIsa =
|
||||
SHADER_START
|
||||
R"(
|
||||
.if (.amdgcn.gfx_generation_number == 9 && .amdgcn.gfx_generation_minor == 4)
|
||||
// Read-Acquire
|
||||
s_mov_b32 s18, 0x1
|
||||
LOOP:
|
||||
s_load_dword s17, s[0:1], 0x0 glc
|
||||
s_waitcnt lgkmcnt(0)
|
||||
s_cmp_eq_i32 s17, s18
|
||||
s_cbranch_scc0 LOOP
|
||||
buffer_inv sc1 sc0
|
||||
// Load data
|
||||
v_mov_b32 v21, 0x40
|
||||
v_mov_b32 v22, 0x80
|
||||
v_mov_b32 v23, 0xc0
|
||||
v_mov_b32 v24, 0x100
|
||||
v_mov_b32 v25, 0x140
|
||||
global_load_dword v11, v21, s[0:1]
|
||||
global_load_dword v12, v22, s[0:1]
|
||||
global_load_dword v13, v23, s[0:1]
|
||||
global_load_dword v14, v24, s[0:1]
|
||||
global_load_dword v15, v25, s[0:1]
|
||||
s_waitcnt vmcnt(0)
|
||||
// Store data for output
|
||||
v_mov_b32 v21, 0x40
|
||||
v_mov_b32 v22, 0x80
|
||||
v_mov_b32 v23, 0xc0
|
||||
v_mov_b32 v24, 0x100
|
||||
v_mov_b32 v25, 0x140
|
||||
global_store_dword v21, v11, s[2:3] nt sc1 sc0
|
||||
global_store_dword v22, v12, s[2:3] nt sc1 sc0
|
||||
global_store_dword v23, v13, s[2:3] nt sc1 sc0
|
||||
global_store_dword v24, v14, s[2:3] nt sc1 sc0
|
||||
global_store_dword v25, v15, s[2:3] nt sc1 sc0
|
||||
s_waitcnt vmcnt(0)
|
||||
.endif
|
||||
s_endpgm
|
||||
)";
|
||||
|
||||
/* Input:
|
||||
* s[0:1], A buffer of at least 64 * 6 bytes,
|
||||
* shared with the releasing shader
|
||||
* s[2:3], A buffer of at least 64 * 6 bytes,
|
||||
* accessible by the CPU, used for output
|
||||
*
|
||||
* Polls the flag at address 0x0 in the shared buffer.
|
||||
* When the signal is received, read the values
|
||||
* at the 5 addresses 0x40, 0x80, ... 0x140,
|
||||
* and store them at the same locations in
|
||||
* the output buffer
|
||||
*
|
||||
* Uses scalar loads
|
||||
*
|
||||
* Aqua Vanjaram only
|
||||
*/
|
||||
const char *ReadAcquireScalarIsa =
|
||||
SHADER_START
|
||||
R"(
|
||||
.if (.amdgcn.gfx_generation_number == 9 && .amdgcn.gfx_generation_minor == 4)
|
||||
// Read-Acquire
|
||||
s_mov_b32 s18, 0x1
|
||||
LOOP:
|
||||
s_load_dword s17, s[0:1], 0x0 glc
|
||||
s_waitcnt lgkmcnt(0)
|
||||
s_cmp_eq_i32 s17, s18
|
||||
s_cbranch_scc0 LOOP
|
||||
buffer_inv sc1 sc0
|
||||
// Load data
|
||||
s_load_dword s21, s[0:1], 0x40
|
||||
s_load_dword s22, s[0:1], 0x80
|
||||
s_load_dword s23, s[0:1], 0xc0
|
||||
s_load_dword s24, s[0:1], 0x100
|
||||
s_load_dword s25, s[0:1], 0x140
|
||||
s_waitcnt lgkmcnt(0)
|
||||
// Store data for output
|
||||
s_store_dword s21, s[2:3], 0x40 glc
|
||||
s_store_dword s22, s[2:3], 0x80 glc
|
||||
s_store_dword s23, s[2:3], 0xc0 glc
|
||||
s_store_dword s24, s[2:3], 0x100 glc
|
||||
s_store_dword s25, s[2:3], 0x140 glc
|
||||
s_waitcnt lgkmcnt(0)
|
||||
.endif
|
||||
s_endpgm
|
||||
)";
|
||||
|
||||
/**
|
||||
* KFDQMTest
|
||||
*/
|
||||
|
||||
Reference in New Issue
Block a user