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

kfdtest: Update KFDCWSRTest to LLVM Asm

Browse Source 
- Reformat shaders for legibility
- Move assembly processes to from IsaGen (CompileShader) to Assembler
(RunAssembleBuf)

Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I174f1ea5332c499440b30d9bcf06836274428a0f
This commit is contained in:
Graham Sider
2021-09-24 18:47:23 -04:00
committed by Harish Kasiviswanathan
parent 08d38fb140
commit c845b976d0
2 changed files with 54 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/kfdtest/src/KFDCWSRTest.cpp
+53 -69
View File
@@ -38,76 +38,63 @@
* v6 - counter
*/
static const char* iterate_isa_gfx8 = \
"\
shader iterate_isa\n\
wave_size(32)\n\
type(CS)\n\
// copy the parameters from scalar registers to vector registers\n\
v_mov_b32 v2, s0 // v[2:3] = s[0:1] \n\
v_mov_b32 v3, s1 // v[2:3] = s[0:1] \n\
v_mov_b32 v0, s4 // use workgroup id as index \n\
v_lshlrev_b32 v0, 2, v0 // v0 *= 4 \n\
v_add_u32 v4, vcc, s2, v0 // v[4:5] = s[2:3] + v0 * 4 \n\
v_mov_b32 v5, s3 // v[4:5] = s[2:3] + v0 * 4 \n\
v_add_u32 v5, vcc, v5, vcc_lo // v[4:5] = s[2:3] + v0 * 4 \n\
v_mov_b32 v6, 0 \n\
LOOP: \n\
v_add_u32 v6, vcc, 1, v6 \n\
// compare the result value (v6) to iteration value (v2), and \n\
// jump if equal (i.e. if VCC is not zero after the comparison) \n\
v_cmp_lt_u32 vcc, v6, v2 \n\
s_cbranch_vccnz LOOP \n\
flat_store_dword v[4:5], v6 \n\
s_waitcnt vmcnt(0)&lgkmcnt(0) \n\
s_endpgm \n\
end \n\
";
static const char* IterateIsa_gfx8 = R"(
.text
// Copy the parameters from scalar registers to vector registers
v_mov_b32 v2, s0 // v[2:3] = s[0:1]
v_mov_b32 v3, s1 // v[2:3] = s[0:1]
v_mov_b32 v0, s4 // use workgroup id as index
v_lshlrev_b32 v0, 2, v0 // v0 *= 4
v_add_u32 v4, vcc, s2, v0 // v[4:5] = s[2:3] + v0 * 4
v_mov_b32 v5, s3 // v[4:5] = s[2:3] + v0 * 4
v_add_u32 v5, vcc, v5, vcc_lo // v[4:5] = s[2:3] + v0 * 4
v_mov_b32 v6, 0
LOOP:
v_add_u32 v6, vcc, 1, v6
// Compare the result value (v6) to iteration value (v2), and
// jump if equal (i.e. if VCC is not zero after the comparison)
v_cmp_lt_u32 vcc, v6, v2
s_cbranch_vccnz LOOP
flat_store_dword v[4:5], v6
s_waitcnt vmcnt(0) & lgkmcnt(0)
s_endpgm
)";
//This shader can be used by gfx9 and gfx10
static const char* iterate_isa_gfx9 = \
"\
shader iterate_isa\n\
wave_size(32)\n\
type(CS)\n\
// copy the parameters from scalar registers to vector registers\n\
v_mov_b32 v2, s0 // v[2:3] = s[0:1] \n\
v_mov_b32 v3, s1 // v[2:3] = s[0:1] \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 // v[4:5] = s[2:3] + v0 * 4 \n\
v_add_co_u32 v5, vcc, v5, vcc_lo // v[4:5] = s[2:3] + v0 * 4 \n\
v_mov_b32 v6, 0 \n\
LOOP: \n\
v_add_co_u32 v6, vcc, 1, v6 \n\
// compare the result value (v6) to iteration value (v2), and \n\
// jump if equal (i.e. if VCC is not zero after the comparison) \n\
v_cmp_lt_u32 vcc, v6, v2 \n\
s_cbranch_vccnz LOOP \n\
flat_store_dword v[4:5], v6 \n\
s_waitcnt vmcnt(0)&lgkmcnt(0) \n\
s_endpgm \n\
end \n\
";
// This shader can be used by gfx9 and gfx10
static const char* IterateIsa_gfx9 = R"(
.text
// Copy the parameters from scalar registers to vector registers
v_mov_b32 v2, s0 // v[2:3] = s[0:1]
v_mov_b32 v3, s1 // v[2:3] = s[0:1]
v_mov_b32 v0, s4 // use workgroup id as index
v_lshlrev_b32 v0, 2, v0 // v0 *= 4
v_add_co_u32 v4, vcc, s2, v0 // v[4:5] = s[2:3] + v0 * 4
v_mov_b32 v5, s3 // v[4:5] = s[2:3] + v0 * 4
v_add_co_u32 v5, vcc, v5, vcc_lo // v[4:5] = s[2:3] + v0 * 4
v_mov_b32 v6, 0
LOOP:
v_add_co_u32 v6, vcc, 1, v6
// Compare the result value (v6) to iteration value (v2), and
// jump if equal (i.e. if VCC is not zero after the comparison)
v_cmp_lt_u32 vcc, v6, v2
s_cbranch_vccnz LOOP
flat_store_dword v[4:5], v6
s_waitcnt vmcnt(0) & lgkmcnt(0)
s_endpgm
)";
static const char* infinite_isa = \
"\
shader infinite_isa \n\
wave_size(32) \n\
type(CS) \n\
LOOP: \n\
s_branch LOOP \n\
end \n\
";
static const char* InfiniteIsa = R"(
.text
LOOP:
s_branch LOOP
s_endpgm
)";
void KFDCWSRTest::SetUp() {
ROUTINE_START
KFDBaseComponentTest::SetUp();
m_pIsaGen = IsaGenerator::Create(m_FamilyId);
wave_number = 1;
ROUTINE_END
@@ -115,9 +102,6 @@ void KFDCWSRTest::SetUp() {
void KFDCWSRTest::TearDown() {
ROUTINE_START
if (m_pIsaGen)
delete m_pIsaGen;
m_pIsaGen = NULL;
KFDBaseComponentTest::TearDown();
@@ -159,9 +143,9 @@ TEST_F(KFDCWSRTest, BasicTest) {
uint64_t count1 = 400000000;
if (m_FamilyId < FAMILY_AI)
pIterateIsa = iterate_isa_gfx8;
pIterateIsa = IterateIsa_gfx8;
else
pIterateIsa = iterate_isa_gfx9;
pIterateIsa = IterateIsa_gfx9;
if (isOnEmulator()) {
// Divide the iterator times by 10000 so that the test can
@@ -172,7 +156,7 @@ TEST_F(KFDCWSRTest, BasicTest) {
unsigned int* result1 = resultBuf1.As<unsigned int*>();
m_pIsaGen->CompileShader(pIterateIsa, "iterate_isa", isaBuffer);
ASSERT_SUCCESS(m_pAsm->RunAssembleBuf(pIterateIsa, isaBuffer.As<char*>()));
PM4Queue queue1;
@@ -236,7 +220,7 @@ TEST_F(KFDCWSRTest, InterruptRestore) {
if ((m_FamilyId >= FAMILY_VI) && (checkCWSREnabled())) {
HsaMemoryBuffer isaBuffer(PAGE_SIZE, defaultGPUNode, true/*zero*/, false/*local*/, true/*exec*/);
m_pIsaGen->CompileShader(infinite_isa, "infinite_isa", isaBuffer);
ASSERT_SUCCESS(m_pAsm->RunAssembleBuf(InfiniteIsa, isaBuffer.As<char*>()));
PM4Queue queue1, queue2, queue3;
tests/kfdtest/src/KFDCWSRTest.hpp
+1 -3
View File
@@ -27,12 +27,11 @@
#include <gtest/gtest.h>
#include "PM4Queue.hpp"
#include "IsaGenerator.hpp"
#include "KFDBaseComponentTest.hpp"
class KFDCWSRTest : public KFDBaseComponentTest {
public:
KFDCWSRTest() :m_pIsaGen(NULL) {}
KFDCWSRTest() {}
~KFDCWSRTest() {}
protected:
@@ -41,7 +40,6 @@ class KFDCWSRTest : public KFDBaseComponentTest {
protected: // Members
unsigned wave_number;
IsaGenerator* m_pIsaGen;
};
#endif // __KFD_CWSR_TEST__H__