AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

KFDTest: Make shader compatiable for gfx9 and gfx10

Browse Source 
Remove the CHIP name from the shader ISA and add wave_size(32) to make the same
shader can be  used for both  GFX9 and GFX10

Change-Id: I16ea72f87980c3d9c11298e20c06a0a073fe9a28
Signed-off-by: shaoyunl <shaoyun.liu@amd.com>


[ROCm/ROCR-Runtime commit: 78e754ca5b]
This commit is contained in:
shaoyunl
2019-07-25 12:02:46 -04:00
committed by Shaoyun Liu
parent e6eb9e9821
commit a7086d5bee
4 changed files with 18 additions and 149 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocr-runtime/tests/kfdtest/src/KFDCWSRTest.cpp
+2 -29
View File
@@ -52,35 +52,10 @@ LOOP:\n\
end\n\
";
//This shader can be used by gfx9 and gfx10
static const char* iterate_isa_gfx9 = \
"\
shader iterate_isa\n\
asic(GFX9)\n\
type(CS)\n\
/*copy the parameters from scalar registers to vector registers*/\n\
v_mov_b32 v0, s0\n\
v_mov_b32 v1, s1\n\
v_mov_b32 v2, s2\n\
v_mov_b32 v3, s3\n\
flat_load_dword v4, v[0:1] slc /*load target iteration value*/\n\
s_waitcnt vmcnt(0)&lgkmcnt(0)\n\
v_mov_b32 v5, 0\n\
LOOP:\n\
v_add_co_u32 v5, vcc, 1, v5\n\
s_waitcnt vmcnt(0)&lgkmcnt(0)\n\
/*compare the result value (v5) to iteration value (v4), and jump if equal (i.e. if VCC is not zero after the comparison)*/\n\
v_cmp_lt_u32 vcc, v5, v4\n\
s_cbranch_vccnz LOOP\n\
flat_store_dword v[2,3], v5\n\
s_waitcnt vmcnt(0)&lgkmcnt(0)\n\
s_endpgm\n\
end\n\
";
static const char* iterate_isa_gfx10 = \
"\
shader iterate_isa\n\
asic(GFX10)\n\
wave_size(32)\n\
type(CS)\n\
/*copy the parameters from scalar registers to vector registers*/\n\
@@ -160,10 +135,8 @@ TEST_F(KFDCWSRTest, BasicTest) {
if (m_FamilyId < FAMILY_AI)
pIterateIsa = iterate_isa_gfx8;
else if (m_FamilyId < FAMILY_NV)
pIterateIsa = iterate_isa_gfx9;
else
pIterateIsa = iterate_isa_gfx10;
pIterateIsa = iterate_isa_gfx9;
m_pIsaGen->CompileShader(pIterateIsa, "iterate_isa", isaBuffer);
projects/rocr-runtime/tests/kfdtest/src/KFDEvictTest.cpp
+6 -58
View File
@@ -364,11 +364,15 @@ TEST_F(KFDEvictTest, BasicTest) {
* v[2:3] - address of corresponding local buf address offset: s[0:1] + v0 * 8
* v[4:5] - corresponding output buf address: s[2:3] + v0 * 4
* v[6:7] - local buf address used for read test
*
* This shader can be used by gfx9 and gfx10
*
*/
static const char* gfx9_ReadMemory =
"\
shader ReadMemory\n\
asic(GFX9)\n\
wave_size(32)\n\
type(CS)\n\
\n\
// compute address of corresponding output buffer\n\
@@ -469,67 +473,11 @@ L_QUIT:\n\
end\n\
";
static const char* gfx10_ReadMemory =
"\
shader ReadMemory\n\
asic(GFX10)\n\
wave_size(32)\n\
type(CS)\n\
\n\
// compute address of corresponding output buffer\n\
v_mov_b32 v0, s4 // use workgroup id as index\n\
v_lshlrev_b32 v0, 2, v0 // v0 *= 4\n\
v_add_co_u32 v4, vcc, s2, v0 // v[4:5] = s[2:3] + v0 * 4\n\
v_mov_b32 v5, s3\n\
v_add_co_u32 v5, vcc, v5, vcc_lo\n\
\n\
// compute input buffer offset used to store corresponding local buffer address\n\
v_lshlrev_b32 v0, 1, v0 // v0 *= 8\n\
v_add_co_u32 v2, vcc, s0, v0 // v[2:3] = s[0:1] + v0 * 8\n\
v_mov_b32 v3, s1\n\
v_add_co_u32 v3, vcc, v3, vcc_lo\n\
\n\
// load 64bit local buffer address stored at v[2:3] to v[6:7]\n\
flat_load_dwordx2 v[6:7], v[2:3] slc\n\
s_waitcnt vmcnt(0) & lgkmcnt(0) // wait for memory reads to finish\n\
\n\
v_mov_b32 v8, 0x5678\n\
s_movk_i32 s8, 0x5678\n\
L_REPEAT:\n\
s_load_dword s16, s[0:1], 0x0 glc\n\
s_waitcnt vmcnt(0) & lgkmcnt(0) // wait for memory reads to finish\n\
s_cmp_eq_i32 s16, s8\n\
s_cbranch_scc1 L_QUIT // if notified to quit by host\n\
// loop read 64M local buffer starting at v[6:7]\n\
// every 4k page only read once\n\
v_mov_b32 v9, 0\n\
v_mov_b32 v10, 0x1000 // 4k page\n\
v_mov_b32 v11, 0x4000000 // 64M size\n\
v_mov_b32 v12, v6\n\
v_mov_b32 v13, v7\n\
L_LOOP_READ:\n\
flat_load_dwordx2 v[14:15], v[12:13] slc\n\
v_add_co_u32 v9, vcc, v9, v10 \n\
v_add_co_u32 v12, vcc, v12, v10\n\
v_add_co_u32 v13, vcc, v13, vcc_lo\n\
v_cmp_lt_u32 vcc, v9, v11\n\
s_cbranch_vccnz L_LOOP_READ\n\
s_branch L_REPEAT\n\
L_QUIT:\n\
flat_store_dword v[4:5], v8\n\
s_waitcnt vmcnt(0) & lgkmcnt(0) // wait for memory writes to finish\n\
s_endpgm\n\
end\n\
";
std::string KFDEvictTest::CreateShader() {
if (m_FamilyId < FAMILY_AI)
return gfx8_ReadMemory;
else if (m_FamilyId < FAMILY_NV)
return gfx9_ReadMemory;
else
return gfx10_ReadMemory;
return gfx9_ReadMemory;
}
/* Evict and restore queue test
projects/rocr-runtime/tests/kfdtest/src/KFDMemoryTest.cpp
+8 -62
View File
@@ -118,23 +118,7 @@ end\n\
const char* gfx9_PollMemory =
"\
shader ReadMemory\n\
asic(GFX9)\n\
type(CS)\n\
/* Assume src address in s0, s1 and dst address in s2, s3*/\n\
s_movk_i32 s18, 0x5678\n\
LOOP:\n\
s_load_dword s16, s[0:1], 0x0 glc\n\
s_cmp_eq_i32 s16, s18\n\
s_cbranch_scc0 LOOP\n\
s_store_dword s18, s[2:3], 0x0 glc\n\
s_endpgm\n\
end\n\
";
const char* gfx10_PollMemory =
"\
shader ReadMemory\n\
asic(GFX10)\n\
wave_size(32)\n\
type(CS)\n\
/* Assume src address in s0, s1 and dst address in s2, s3*/\n\
s_movk_i32 s18, 0x5678\n\
@@ -159,7 +143,7 @@ type(CS)\n\
const char* gfx9_CopyOnSignal =
"\
shader CopyOnSignal\n\
asic(GFX9)\n\
wave_size(32)\n\
type(CS)\n\
/* Assume input buffer in s0, s1 */\n\
s_mov_b32 s18, 0xcafe\n\
@@ -175,27 +159,6 @@ POLLSIGNAL:\n\
end\n\
";
const char* gfx10_CopyOnSignal =
"\
shader CopyOnSignal\n\
asic(GFX10)\n\
type(CS)\n\
/* Assume input buffer in s0, s1 */\n\
s_mov_b32 s18, 0xcafe\n\
POLLSIGNAL:\n\
s_load_dword s16, s[0:1], 0x0 glc\n\
s_cmp_eq_i32 s16, s18\n\
s_cbranch_scc0 POLLSIGNAL\n\
s_load_dword s17, s[0:1], 0x4 glc\n\
s_waitcnt vmcnt(0) & lgkmcnt(0)\n\
s_store_dword s17, s[0:1], 0x8 glc\n\
s_waitcnt vmcnt(0) & lgkmcnt(0)\n\
s_endpgm\n\
end\n\
";
/* Input0: A buffer of at least 2 dwords.
* DW0: used as a signal. Write 0xcafe to signal
* DW1: Write to this buffer for other device to read.
@@ -204,7 +167,7 @@ POLLSIGNAL:\n\
const char* gfx9_WriteAndSignal =
"\
shader WriteAndSignal\n\
asic(GFX9)\n\
wave_size(32)\n\
type(CS)\n\
/* Assume input buffer in s0, s1 */\n\
s_mov_b32 s18, 0xbeef\n\
@@ -217,21 +180,7 @@ type(CS)\n\
end\n\
";
const char* gfx10_WriteAndSignal =
"\
shader WriteAndSignal\n\
asic(GFX10)\n\
type(CS)\n\
/* Assume input buffer in s0, s1 */\n\
s_mov_b32 s18, 0xbeef\n\
s_store_dword s18, s[0:1], 0x4 glc\n\
s_mov_b32 s18, 0x1\n\
s_store_dword s18, s[2:3], 0 glc\n\
s_mov_b32 s18, 0xcafe\n\
s_store_dword s18, s[0:1], 0x0 glc\n\
s_endpgm\n\
end\n\
";
//These gfx9_PullMemory, gfx9_CopyOnSignal, gfx9_WriteAndSignal shaders can be used by both gfx9 and gfx10
void KFDMemoryTest::SetUp() {
ROUTINE_START
@@ -364,7 +313,7 @@ TEST_F(KFDMemoryTest, MapUnmapToNodes) {
HsaMemoryBuffer srcBuffer(PAGE_SIZE, defaultGPUNode);
HsaMemoryBuffer dstBuffer(PAGE_SIZE, defaultGPUNode);
m_pIsaGen->CompileShader((m_FamilyId < FAMILY_NV) ? gfx9_PollMemory : gfx10_PollMemory, "ReadMemory", isaBuffer);
m_pIsaGen->CompileShader(gfx9_PollMemory, "ReadMemory", isaBuffer);
PM4Queue pm4Queue;
ASSERT_SUCCESS(pm4Queue.Create(defaultGPUNode));
@@ -1961,8 +1910,7 @@ TEST_F(KFDMemoryTest, HostHdpFlush) {
PM4Queue queue;
ASSERT_SUCCESS(queue.Create(defaultGPUNode));
HsaMemoryBuffer isaBuffer(PAGE_SIZE, defaultGPUNode, true/*zero*/, false/*local*/, true/*exec*/);
m_pIsaGen->CompileShader((m_FamilyId < FAMILY_NV) ? gfx9_CopyOnSignal : gfx10_CopyOnSignal,
"CopyOnSignal", isaBuffer);
m_pIsaGen->CompileShader(gfx9_CopyOnSignal, "CopyOnSignal", isaBuffer);
Dispatch dispatch0(isaBuffer);
dispatch0.SetArgs(buffer, NULL);
dispatch0.Submit(queue);
@@ -2081,8 +2029,7 @@ TEST_F(KFDMemoryTest, DeviceHdpFlush) {
PM4Queue queue;
ASSERT_SUCCESS(queue.Create(nodes[0]));
HsaMemoryBuffer isaBuffer(PAGE_SIZE, nodes[0], true/*zero*/, false/*local*/, true/*exec*/);
m_pIsaGen->CompileShader((queue.GetFamilyId() < FAMILY_NV) ? gfx9_CopyOnSignal : gfx10_CopyOnSignal,
"CopyOnSignal", isaBuffer);
m_pIsaGen->CompileShader(gfx9_CopyOnSignal, "CopyOnSignal", isaBuffer);
Dispatch dispatch(isaBuffer);
dispatch.SetArgs(buffer, NULL);
dispatch.Submit(queue);
@@ -2090,8 +2037,7 @@ TEST_F(KFDMemoryTest, DeviceHdpFlush) {
PM4Queue queue0;
ASSERT_SUCCESS(queue0.Create(nodes[1]));
HsaMemoryBuffer isaBuffer0(PAGE_SIZE, nodes[1], true/*zero*/, false/*local*/, true/*exec*/);
m_pIsaGen->CompileShader((queue0.GetFamilyId() < FAMILY_NV) ? gfx9_WriteAndSignal : gfx10_WriteAndSignal,
"WriteAndSignal", isaBuffer0);
m_pIsaGen->CompileShader(gfx9_WriteAndSignal, "WriteAndSignal", isaBuffer0);
Dispatch dispatch0(isaBuffer0);
dispatch0.SetArgs(buffer, mmioBase);
dispatch0.Submit(queue0);
projects/rocr-runtime/tests/kfdtest/src/KFDQMTest.cpp
+2
View File
@@ -489,10 +489,12 @@ TEST_F(KFDQMTest, OverSubscribeCpQueues) {
/* A simple isa loop program with dense mathematic operations
* s1 controls the number iterations of the loop
* This shader can be used by GFX8, GFX9 and GFX10
*/
static const char *loop_isa = \
"\
shader loop_isa\n\
wave_size(32)\n\
type(CS)\n\
s_movk_i32 s0, 0x0008\n\
s_movk_i32 s1, 0x00ff\n\