Skip to content

L
linux
Commit c5e35891 authored by Jay Cornwall's avatar Jay Cornwall Committed by Alex Deucher
Browse files
Options

drm/amdkfd: Replace deprecated gfx12 trap handler instructions 

Newer assemblers reject S_WAITCNT. All instances of S_WAITCNT can be
replaced by S_WAITCNT 0 (< gfx12) or S_WAIT_IDLE (>= gfx12) since
there is no concurrency of different memory instruction classes.
Signed-off-by: default avatarJay Cornwall <jay.cornwall@amd.com>
Reviewed-by: default avatarLancelot Six <lancelot.six@amd.com>
Acked-by: default avatarAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: default avatarAlex Deucher <alexander.deucher@amd.com>
parent dec4f2d2
Show whitespace changes
Inline Side-by-side
Showing with 97 additions and 95 deletions
drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler.h
View file @ c5e35891
...@@ -711,12 +711,12 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -711,12 +711,12 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0xbf0d8f7b, 0xbf840002, 0xbf0d8f7b, 0xbf840002,
0x887bff7b, 0xffff0000, 0x887bff7b, 0xffff0000,
0xf4011bbd, 0xfa000010, 0xf4011bbd, 0xfa000010,
0xbf8cc07f, 0x8f6e976e, 0xbf8c0000, 0x8f6e976e,
0x8a77ff77, 0x00800000, 0x8a77ff77, 0x00800000,
0x88776e77, 0xf4051bbd, 0x88776e77, 0xf4051bbd,
0xfa000000, 0xbf8cc07f, 0xfa000000, 0xbf8c0000,
0xf4051ebd, 0xfa000008, 0xf4051ebd, 0xfa000008,
0xbf8cc07f, 0x87ee6e6e, 0xbf8c0000, 0x87ee6e6e,
0xbf840001, 0xbe80206e, 0xbf840001, 0xbe80206e,
0x876eff6d, 0x00ff0000, 0x876eff6d, 0x00ff0000,
0xbf850008, 0x876eff6d, 0xbf850008, 0x876eff6d,
...@@ -1185,7 +1185,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1185,7 +1185,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0x785d0000, 0xe0304080, 0x785d0000, 0xe0304080,
0x785d0100, 0xe0304100, 0x785d0100, 0xe0304100,
0x785d0200, 0xe0304180, 0x785d0200, 0xe0304180,
0x785d0300, 0xbf8c3f70, 0x785d0300, 0xbf8c0000,
0x7e008500, 0x7e028501, 0x7e008500, 0x7e028501,
0x7e048502, 0x7e068503, 0x7e048502, 0x7e068503,
0x807c847c, 0x8078ff78, 0x807c847c, 0x8078ff78,
...@@ -1194,7 +1194,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1194,7 +1194,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0x6e5d0000, 0xe0304080, 0x6e5d0000, 0xe0304080,
0x6e5d0100, 0xe0304100, 0x6e5d0100, 0xe0304100,
0x6e5d0200, 0xe0304180, 0x6e5d0200, 0xe0304180,
0x6e5d0300, 0xbf8c3f70, 0x6e5d0300, 0xbf8c0000,
0xbf820034, 0xbef603ff, 0xbf820034, 0xbef603ff,
0x01000000, 0xbeee0378, 0x01000000, 0xbeee0378,
0x8078ff78, 0x00000400, 0x8078ff78, 0x00000400,
...@@ -1203,7 +1203,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1203,7 +1203,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0x785d0000, 0xe0304100, 0x785d0000, 0xe0304100,
0x785d0100, 0xe0304200, 0x785d0100, 0xe0304200,
0x785d0200, 0xe0304300, 0x785d0200, 0xe0304300,
0x785d0300, 0xbf8c3f70, 0x785d0300, 0xbf8c0000,
0x7e008500, 0x7e028501, 0x7e008500, 0x7e028501,
0x7e048502, 0x7e068503, 0x7e048502, 0x7e068503,
0x807c847c, 0x8078ff78, 0x807c847c, 0x8078ff78,
...@@ -1213,7 +1213,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1213,7 +1213,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0x8f6f836f, 0x806f7c6f, 0x8f6f836f, 0x806f7c6f,
0xbefe03c1, 0xbeff0380, 0xbefe03c1, 0xbeff0380,
0xe0304000, 0x785d0000, 0xe0304000, 0x785d0000,
0xbf8c3f70, 0x7e008500, 0xbf8c0000, 0x7e008500,
0x807c817c, 0x8078ff78, 0x807c817c, 0x8078ff78,
0x00000080, 0xbf0a6f7c, 0x00000080, 0xbf0a6f7c,
0xbf85fff7, 0xbeff03c1, 0xbf85fff7, 0xbeff03c1,
...@@ -1221,7 +1221,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1221,7 +1221,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0xe0304100, 0x6e5d0100, 0xe0304100, 0x6e5d0100,
0xe0304200, 0x6e5d0200, 0xe0304200, 0x6e5d0200,
0xe0304300, 0x6e5d0300, 0xe0304300, 0x6e5d0300,
0xbf8c3f70, 0xb9783a05, 0xbf8c0000, 0xb9783a05,
0x80788178, 0xbf0d9972, 0x80788178, 0xbf0d9972,
0xbf850002, 0x8f788978, 0xbf850002, 0x8f788978,
0xbf820001, 0x8f788a78, 0xbf820001, 0x8f788a78,
...@@ -1232,16 +1232,16 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1232,16 +1232,16 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0x01000000, 0xbefc03ff, 0x01000000, 0xbefc03ff,
0x0000006c, 0x80f89078, 0x0000006c, 0x80f89078,
0xf429003a, 0xf0000000, 0xf429003a, 0xf0000000,
0xbf8cc07f, 0x80fc847c, 0xbf8c0000, 0x80fc847c,
0xbf800000, 0xbe803100, 0xbf800000, 0xbe803100,
0xbe823102, 0x80f8a078, 0xbe823102, 0x80f8a078,
0xf42d003a, 0xf0000000, 0xf42d003a, 0xf0000000,
0xbf8cc07f, 0x80fc887c, 0xbf8c0000, 0x80fc887c,
0xbf800000, 0xbe803100, 0xbf800000, 0xbe803100,
0xbe823102, 0xbe843104, 0xbe823102, 0xbe843104,
0xbe863106, 0x80f8c078, 0xbe863106, 0x80f8c078,
0xf431003a, 0xf0000000, 0xf431003a, 0xf0000000,
0xbf8cc07f, 0x80fc907c, 0xbf8c0000, 0x80fc907c,
0xbf800000, 0xbe803100, 0xbf800000, 0xbe803100,
0xbe823102, 0xbe843104, 0xbe823102, 0xbe843104,
0xbe863106, 0xbe883108, 0xbe863106, 0xbe883108,
...@@ -1271,9 +1271,9 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1271,9 +1271,9 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0xf4211cfa, 0xf0000000, 0xf4211cfa, 0xf0000000,
0x80788478, 0xf4211bba, 0x80788478, 0xf4211bba,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xbf8cc07f, 0xb9eef814, 0xbf8c0000, 0xb9eef814,
0xf4211bba, 0xf0000000, 0xf4211bba, 0xf0000000,
0x80788478, 0xbf8cc07f, 0x80788478, 0xbf8c0000,
0xb9eef815, 0xbefc036f, 0xb9eef815, 0xbefc036f,
0xbefe0370, 0xbeff0371, 0xbefe0370, 0xbeff0371,
0xb9f9f816, 0xb9fbf803, 0xb9f9f816, 0xb9fbf803,
...@@ -1288,7 +1288,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = { ...@@ -1288,7 +1288,7 @@ static const uint32_t cwsr_trap_nv1x_hex[] = {
0x0000ffff, 0xf4091c37, 0x0000ffff, 0xf4091c37,
0xfa000050, 0xf4091d37, 0xfa000050, 0xf4091d37,
0xfa000060, 0xf4011e77, 0xfa000060, 0xf4011e77,
0xfa000074, 0xbf8cc07f, 0xfa000074, 0xbf8c0000,
0x906e8977, 0x876fff6e, 0x906e8977, 0x876fff6e,
0x003f8000, 0x906e8677, 0x003f8000, 0x906e8677,
0x876eff6e, 0x02000000, 0x876eff6e, 0x02000000,
...@@ -2299,12 +2299,12 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2299,12 +2299,12 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0xbf0d8f7b, 0xbf840002, 0xbf0d8f7b, 0xbf840002,
0x887bff7b, 0xffff0000, 0x887bff7b, 0xffff0000,
0xf4011bbd, 0xfa000010, 0xf4011bbd, 0xfa000010,
0xbf8cc07f, 0x8f6e976e, 0xbf8c0000, 0x8f6e976e,
0x8a77ff77, 0x00800000, 0x8a77ff77, 0x00800000,
0x88776e77, 0xf4051bbd, 0x88776e77, 0xf4051bbd,
0xfa000000, 0xbf8cc07f, 0xfa000000, 0xbf8c0000,
0xf4051ebd, 0xfa000008, 0xf4051ebd, 0xfa000008,
0xbf8cc07f, 0x87ee6e6e, 0xbf8c0000, 0x87ee6e6e,
0xbf840001, 0xbe80206e, 0xbf840001, 0xbe80206e,
0x876eff6d, 0x00ff0000, 0x876eff6d, 0x00ff0000,
0xbf850008, 0x876eff6d, 0xbf850008, 0x876eff6d,
...@@ -2319,7 +2319,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2319,7 +2319,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0x0000ffff, 0xbefa0380, 0x0000ffff, 0xbefa0380,
0xb9fa0283, 0xbeee037e, 0xb9fa0283, 0xbeee037e,
0xbeef037f, 0xbefe0480, 0xbeef037f, 0xbefe0480,
0xbf900004, 0xbf8cc07f, 0xbf900004, 0xbf8c0000,
0x877aff7f, 0x04000000, 0x877aff7f, 0x04000000,
0x8f7a857a, 0x886d7a6d, 0x8f7a857a, 0x886d7a6d,
0x7e008200, 0xbefa037e, 0x7e008200, 0xbefa037e,
...@@ -2595,7 +2595,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2595,7 +2595,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0xe0304080, 0x785d0100, 0xe0304080, 0x785d0100,
0xe0304100, 0x785d0200, 0xe0304100, 0x785d0200,
0xe0304180, 0x785d0300, 0xe0304180, 0x785d0300,
0xbf8c3f70, 0x7e008500, 0xbf8c0000, 0x7e008500,
0x7e028501, 0x7e048502, 0x7e028501, 0x7e048502,
0x7e068503, 0x807c847c, 0x7e068503, 0x807c847c,
0x8078ff78, 0x00000200, 0x8078ff78, 0x00000200,
...@@ -2604,7 +2604,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2604,7 +2604,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0xe0304080, 0x6e5d0100, 0xe0304080, 0x6e5d0100,
0xe0304100, 0x6e5d0200, 0xe0304100, 0x6e5d0200,
0xe0304180, 0x6e5d0300, 0xe0304180, 0x6e5d0300,
0xbf8c3f70, 0xbf820034, 0xbf8c0000, 0xbf820034,
0xbef603ff, 0x01000000, 0xbef603ff, 0x01000000,
0xbeee0378, 0x8078ff78, 0xbeee0378, 0x8078ff78,
0x00000400, 0xbefc0384, 0x00000400, 0xbefc0384,
...@@ -2613,7 +2613,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2613,7 +2613,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0xe0304100, 0x785d0100, 0xe0304100, 0x785d0100,
0xe0304200, 0x785d0200, 0xe0304200, 0x785d0200,
0xe0304300, 0x785d0300, 0xe0304300, 0x785d0300,
0xbf8c3f70, 0x7e008500, 0xbf8c0000, 0x7e008500,
0x7e028501, 0x7e048502, 0x7e028501, 0x7e048502,
0x7e068503, 0x807c847c, 0x7e068503, 0x807c847c,
0x8078ff78, 0x00000400, 0x8078ff78, 0x00000400,
...@@ -2622,7 +2622,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2622,7 +2622,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0xbf84000e, 0x8f6f836f, 0xbf84000e, 0x8f6f836f,
0x806f7c6f, 0xbefe03c1, 0x806f7c6f, 0xbefe03c1,
0xbeff0380, 0xe0304000, 0xbeff0380, 0xe0304000,
0x785d0000, 0xbf8c3f70, 0x785d0000, 0xbf8c0000,
0x7e008500, 0x807c817c, 0x7e008500, 0x807c817c,
0x8078ff78, 0x00000080, 0x8078ff78, 0x00000080,
0xbf0a6f7c, 0xbf85fff7, 0xbf0a6f7c, 0xbf85fff7,
...@@ -2630,7 +2630,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2630,7 +2630,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0x6e5d0000, 0xe0304100, 0x6e5d0000, 0xe0304100,
0x6e5d0100, 0xe0304200, 0x6e5d0100, 0xe0304200,
0x6e5d0200, 0xe0304300, 0x6e5d0200, 0xe0304300,
0x6e5d0300, 0xbf8c3f70, 0x6e5d0300, 0xbf8c0000,
0xb9783a05, 0x80788178, 0xb9783a05, 0x80788178,
0xbf0d9972, 0xbf850002, 0xbf0d9972, 0xbf850002,
0x8f788978, 0xbf820001, 0x8f788978, 0xbf820001,
...@@ -2641,16 +2641,16 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2641,16 +2641,16 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0xbef603ff, 0x01000000, 0xbef603ff, 0x01000000,
0xbefc03ff, 0x0000006c, 0xbefc03ff, 0x0000006c,
0x80f89078, 0xf429003a, 0x80f89078, 0xf429003a,
0xf0000000, 0xbf8cc07f, 0xf0000000, 0xbf8c0000,
0x80fc847c, 0xbf800000, 0x80fc847c, 0xbf800000,
0xbe803100, 0xbe823102, 0xbe803100, 0xbe823102,
0x80f8a078, 0xf42d003a, 0x80f8a078, 0xf42d003a,
0xf0000000, 0xbf8cc07f, 0xf0000000, 0xbf8c0000,
0x80fc887c, 0xbf800000, 0x80fc887c, 0xbf800000,
0xbe803100, 0xbe823102, 0xbe803100, 0xbe823102,
0xbe843104, 0xbe863106, 0xbe843104, 0xbe863106,
0x80f8c078, 0xf431003a, 0x80f8c078, 0xf431003a,
0xf0000000, 0xbf8cc07f, 0xf0000000, 0xbf8c0000,
0x80fc907c, 0xbf800000, 0x80fc907c, 0xbf800000,
0xbe803100, 0xbe823102, 0xbe803100, 0xbe823102,
0xbe843104, 0xbe863106, 0xbe843104, 0xbe863106,
...@@ -2680,10 +2680,10 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2680,10 +2680,10 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0x80788478, 0xf4211cfa, 0x80788478, 0xf4211cfa,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xf4211bba, 0xf0000000, 0xf4211bba, 0xf0000000,
0x80788478, 0xbf8cc07f, 0x80788478, 0xbf8c0000,
0xb9eef814, 0xf4211bba, 0xb9eef814, 0xf4211bba,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xbf8cc07f, 0xb9eef815, 0xbf8c0000, 0xb9eef815,
0xbefc036f, 0xbefe0370, 0xbefc036f, 0xbefe0370,
0xbeff0371, 0xb9fbf803, 0xbeff0371, 0xb9fbf803,
0xb9f3f801, 0xb96e3a05, 0xb9f3f801, 0xb96e3a05,
...@@ -2697,7 +2697,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = { ...@@ -2697,7 +2697,7 @@ static const uint32_t cwsr_trap_gfx10_hex[] = {
0x0000ffff, 0xf4091c37, 0x0000ffff, 0xf4091c37,
0xfa000050, 0xf4091d37, 0xfa000050, 0xf4091d37,
0xfa000060, 0xf4011e77, 0xfa000060, 0xf4011e77,
0xfa000074, 0xbf8cc07f, 0xfa000074, 0xbf8c0000,
0x876dff6d, 0x0000ffff, 0x876dff6d, 0x0000ffff,
0x87fe7e7e, 0x87ea6a6a, 0x87fe7e7e, 0x87ea6a6a,
0xb9faf802, 0xbe80226c, 0xb9faf802, 0xbe80226c,
...@@ -2731,16 +2731,16 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -2731,16 +2731,16 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0x8b6eff6e, 0x00000800, 0x8b6eff6e, 0x00000800,
0xbfa20003, 0x8b6eff7b, 0xbfa20003, 0x8b6eff7b,
0x00000400, 0xbfa2002a, 0x00000400, 0xbfa2002a,
0xbefa4d82, 0xbf89fc07, 0xbefa4d82, 0xbf890000,
0x84fa887a, 0xbf0d8f7b, 0x84fa887a, 0xbf0d8f7b,
0xbfa10002, 0x8c7bff7b, 0xbfa10002, 0x8c7bff7b,
0xffff0000, 0xf4005bbd, 0xffff0000, 0xf4005bbd,
0xf8000010, 0xbf89fc07, 0xf8000010, 0xbf890000,
0x846e976e, 0x9177ff77, 0x846e976e, 0x9177ff77,
0x00800000, 0x8c776e77, 0x00800000, 0x8c776e77,
0xf4045bbd, 0xf8000000, 0xf4045bbd, 0xf8000000,
0xbf89fc07, 0xf4045ebd, 0xbf890000, 0xf4045ebd,
0xf8000008, 0xbf89fc07, 0xf8000008, 0xbf890000,
0x8bee6e6e, 0xbfa10001, 0x8bee6e6e, 0xbfa10001,
0xbe80486e, 0x8b6eff6d, 0xbe80486e, 0x8b6eff6d,
0x00ff0000, 0xbfa20008, 0x00ff0000, 0xbfa20008,
...@@ -2756,7 +2756,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -2756,7 +2756,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0xbefa0080, 0xb97a0283, 0xbefa0080, 0xb97a0283,
0xbeee007e, 0xbeef007f, 0xbeee007e, 0xbeef007f,
0xbefe0180, 0xbefe4d84, 0xbefe0180, 0xbefe4d84,
0xbf89fc07, 0x8b7aff7f, 0xbf890000, 0x8b7aff7f,
0x04000000, 0x847a857a, 0x04000000, 0x847a857a,
0x8c6d7a6d, 0xbefa007e, 0x8c6d7a6d, 0xbefa007e,
0x8b7bff7f, 0x0000ffff, 0x8b7bff7f, 0x0000ffff,
...@@ -3007,13 +3007,13 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3007,13 +3007,13 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0x857d9972, 0x8b7d817d, 0x857d9972, 0x8b7d817d,
0xbf06817d, 0xbefd0080, 0xbf06817d, 0xbefd0080,
0xbfa2000c, 0xe0500000, 0xbfa2000c, 0xe0500000,
0x781d0000, 0xbf8903f7, 0x781d0000, 0xbf890000,
0xdac00000, 0x00000000, 0xdac00000, 0x00000000,
0x807dff7d, 0x00000080, 0x807dff7d, 0x00000080,
0x8078ff78, 0x00000080, 0x8078ff78, 0x00000080,
0xbf0a6f7d, 0xbfa2fff5, 0xbf0a6f7d, 0xbfa2fff5,
0xbfa0000b, 0xe0500000, 0xbfa0000b, 0xe0500000,
0x781d0000, 0xbf8903f7, 0x781d0000, 0xbf890000,
0xdac00000, 0x00000000, 0xdac00000, 0x00000000,
0x807dff7d, 0x00000100, 0x807dff7d, 0x00000100,
0x8078ff78, 0x00000100, 0x8078ff78, 0x00000100,
...@@ -3034,7 +3034,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3034,7 +3034,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0xe0505080, 0x781d0100, 0xe0505080, 0x781d0100,
0xe0505100, 0x781d0200, 0xe0505100, 0x781d0200,
0xe0505180, 0x781d0300, 0xe0505180, 0x781d0300,
0xbf8903f7, 0x7e008500, 0xbf890000, 0x7e008500,
0x7e028501, 0x7e048502, 0x7e028501, 0x7e048502,
0x7e068503, 0x807d847d, 0x7e068503, 0x807d847d,
0x8078ff78, 0x00000200, 0x8078ff78, 0x00000200,
...@@ -3043,7 +3043,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3043,7 +3043,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0xe0505080, 0x6e1d0100, 0xe0505080, 0x6e1d0100,
0xe0505100, 0x6e1d0200, 0xe0505100, 0x6e1d0200,
0xe0505180, 0x6e1d0300, 0xe0505180, 0x6e1d0300,
0xbf8903f7, 0xbfa00034, 0xbf890000, 0xbfa00034,
0xbef600ff, 0x01000000, 0xbef600ff, 0x01000000,
0xbeee0078, 0x8078ff78, 0xbeee0078, 0x8078ff78,
0x00000400, 0xbefd0084, 0x00000400, 0xbefd0084,
...@@ -3052,7 +3052,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3052,7 +3052,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0xe0505100, 0x781d0100, 0xe0505100, 0x781d0100,
0xe0505200, 0x781d0200, 0xe0505200, 0x781d0200,
0xe0505300, 0x781d0300, 0xe0505300, 0x781d0300,
0xbf8903f7, 0x7e008500, 0xbf890000, 0x7e008500,
0x7e028501, 0x7e048502, 0x7e028501, 0x7e048502,
0x7e068503, 0x807d847d, 0x7e068503, 0x807d847d,
0x8078ff78, 0x00000400, 0x8078ff78, 0x00000400,
...@@ -3061,7 +3061,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3061,7 +3061,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0xbfa1000e, 0x846f836f, 0xbfa1000e, 0x846f836f,
0x806f7d6f, 0xbefe00c1, 0x806f7d6f, 0xbefe00c1,
0xbeff0080, 0xe0505000, 0xbeff0080, 0xe0505000,
0x781d0000, 0xbf8903f7, 0x781d0000, 0xbf890000,
0x7e008500, 0x807d817d, 0x7e008500, 0x807d817d,
0x8078ff78, 0x00000080, 0x8078ff78, 0x00000080,
0xbf0a6f7d, 0xbfa2fff7, 0xbf0a6f7d, 0xbfa2fff7,
...@@ -3069,7 +3069,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3069,7 +3069,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0x6e1d0000, 0xe0505100, 0x6e1d0000, 0xe0505100,
0x6e1d0100, 0xe0505200, 0x6e1d0100, 0xe0505200,
0x6e1d0200, 0xe0505300, 0x6e1d0200, 0xe0505300,
0x6e1d0300, 0xbf8903f7, 0x6e1d0300, 0xbf890000,
0xb8f83b05, 0x80788178, 0xb8f83b05, 0x80788178,
0xbf0d9972, 0xbfa20002, 0xbf0d9972, 0xbfa20002,
0x84788978, 0xbfa00001, 0x84788978, 0xbfa00001,
...@@ -3080,16 +3080,16 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3080,16 +3080,16 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0xbef600ff, 0x01000000, 0xbef600ff, 0x01000000,
0xbefd00ff, 0x0000006c, 0xbefd00ff, 0x0000006c,
0x80f89078, 0xf428403a, 0x80f89078, 0xf428403a,
0xf0000000, 0xbf89fc07, 0xf0000000, 0xbf890000,
0x80fd847d, 0xbf800000, 0x80fd847d, 0xbf800000,
0xbe804300, 0xbe824302, 0xbe804300, 0xbe824302,
0x80f8a078, 0xf42c403a, 0x80f8a078, 0xf42c403a,
0xf0000000, 0xbf89fc07, 0xf0000000, 0xbf890000,
0x80fd887d, 0xbf800000, 0x80fd887d, 0xbf800000,
0xbe804300, 0xbe824302, 0xbe804300, 0xbe824302,
0xbe844304, 0xbe864306, 0xbe844304, 0xbe864306,
0x80f8c078, 0xf430403a, 0x80f8c078, 0xf430403a,
0xf0000000, 0xbf89fc07, 0xf0000000, 0xbf890000,
0x80fd907d, 0xbf800000, 0x80fd907d, 0xbf800000,
0xbe804300, 0xbe824302, 0xbe804300, 0xbe824302,
0xbe844304, 0xbe864306, 0xbe844304, 0xbe864306,
...@@ -3119,10 +3119,10 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3119,10 +3119,10 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0x80788478, 0xf4205cfa, 0x80788478, 0xf4205cfa,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xf4205bba, 0xf0000000, 0xf4205bba, 0xf0000000,
0x80788478, 0xbf89fc07, 0x80788478, 0xbf890000,
0xb96ef814, 0xf4205bba, 0xb96ef814, 0xf4205bba,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xbf89fc07, 0xb96ef815, 0xbf890000, 0xb96ef815,
0xbefd006f, 0xbefe0070, 0xbefd006f, 0xbefe0070,
0xbeff0071, 0xb97bf803, 0xbeff0071, 0xb97bf803,
0xb973f801, 0xb8ee3b05, 0xb973f801, 0xb8ee3b05,
...@@ -3136,7 +3136,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = { ...@@ -3136,7 +3136,7 @@ static const uint32_t cwsr_trap_gfx11_hex[] = {
0x0000ffff, 0xf4085c37, 0x0000ffff, 0xf4085c37,
0xf8000050, 0xf4085d37, 0xf8000050, 0xf4085d37,
0xf8000060, 0xf4005e77, 0xf8000060, 0xf4005e77,
0xf8000074, 0xbf89fc07, 0xf8000074, 0xbf890000,
0x8b6dff6d, 0x0000ffff, 0x8b6dff6d, 0x0000ffff,
0x8bfe7e7e, 0x8bea6a6a, 0x8bfe7e7e, 0x8bea6a6a,
0xb8eef802, 0xbf0d866e, 0xb8eef802, 0xbf0d866e,
...@@ -3657,16 +3657,16 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -3657,16 +3657,16 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x8b6fff6f, 0x00000200, 0x8b6fff6f, 0x00000200,
0xbfa20002, 0x8b6ea07b, 0xbfa20002, 0x8b6ea07b,
0xbfa2002b, 0xbefa4d82, 0xbfa2002b, 0xbefa4d82,
0xbf89fc07, 0x84fa887a, 0xbf8a0000, 0x84fa887a,
0xbf0d8f7b, 0xbfa10002, 0xbf0d8f7b, 0xbfa10002,
0x8c7bff7b, 0xffff0000, 0x8c7bff7b, 0xffff0000,
0xf4601bbd, 0xf8000010, 0xf4601bbd, 0xf8000010,
0xbf89fc07, 0x846e976e, 0xbf8a0000, 0x846e976e,
0x9177ff77, 0x00800000, 0x9177ff77, 0x00800000,
0x8c776e77, 0xf4603bbd, 0x8c776e77, 0xf4603bbd,
0xf8000000, 0xbf89fc07, 0xf8000000, 0xbf8a0000,
0xf4603ebd, 0xf8000008, 0xf4603ebd, 0xf8000008,
0xbf89fc07, 0x8bee6e6e, 0xbf8a0000, 0x8bee6e6e,
0xbfa10001, 0xbe80486e, 0xbfa10001, 0xbe80486e,
0x8b6eff6d, 0xf0000000, 0x8b6eff6d, 0xf0000000,
0xbfa20009, 0xb8eef811, 0xbfa20009, 0xb8eef811,
...@@ -3682,7 +3682,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -3682,7 +3682,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0xbefa0080, 0xb97a0151, 0xbefa0080, 0xb97a0151,
0xbeee007e, 0xbeef007f, 0xbeee007e, 0xbeef007f,
0xbefe0180, 0xbefe4d84, 0xbefe0180, 0xbefe4d84,
0xbf89fc07, 0x8b7aff7f, 0xbf8a0000, 0x8b7aff7f,
0x04000000, 0x847a857a, 0x04000000, 0x847a857a,
0x8c6d7a6d, 0xbefa007e, 0x8c6d7a6d, 0xbefa007e,
0x8b7bff7f, 0x0000ffff, 0x8b7bff7f, 0x0000ffff,
...@@ -3869,7 +3869,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -3869,7 +3869,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x00000080, 0xbf800000, 0x00000080, 0xbf800000,
0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000,
0xd8d80000, 0x01000000, 0xd8d80000, 0x01000000,
0xbf890000, 0xc4068070, 0xbf8a0000, 0xc4068070,
0x008ce801, 0x00000000, 0x008ce801, 0x00000000,
0x807d037d, 0x80700370, 0x807d037d, 0x80700370,
0xd5250000, 0x0001ff00, 0xd5250000, 0x0001ff00,
...@@ -3878,7 +3878,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -3878,7 +3878,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0xbe8300ff, 0x00000100, 0xbe8300ff, 0x00000100,
0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000,
0xbf800000, 0xd8d80000, 0xbf800000, 0xd8d80000,
0x01000000, 0xbf890000, 0x01000000, 0xbf8a0000,
0xc4068070, 0x008ce801, 0xc4068070, 0x008ce801,
0x00000000, 0x807d037d, 0x00000000, 0x807d037d,
0x80700370, 0xd5250000, 0x80700370, 0xd5250000,
...@@ -3954,14 +3954,14 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -3954,14 +3954,14 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x8b7d817d, 0xbf06817d, 0x8b7d817d, 0xbf06817d,
0xbefd0080, 0xbfa2000d, 0xbefd0080, 0xbfa2000d,
0xc4050078, 0x0080e800, 0xc4050078, 0x0080e800,
0x00000000, 0xbf8903f7, 0x00000000, 0xbf8a0000,
0xdac00000, 0x00000000, 0xdac00000, 0x00000000,
0x807dff7d, 0x00000080, 0x807dff7d, 0x00000080,
0x8078ff78, 0x00000080, 0x8078ff78, 0x00000080,
0xbf0a6f7d, 0xbfa2fff4, 0xbf0a6f7d, 0xbfa2fff4,
0xbfa0000c, 0xc4050078, 0xbfa0000c, 0xc4050078,
0x0080e800, 0x00000000, 0x0080e800, 0x00000000,
0xbf8903f7, 0xdac00000, 0xbf8a0000, 0xdac00000,
0x00000000, 0x807dff7d, 0x00000000, 0x807dff7d,
0x00000100, 0x8078ff78, 0x00000100, 0x8078ff78,
0x00000100, 0xbf0a6f7d, 0x00000100, 0xbf0a6f7d,
...@@ -3983,7 +3983,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -3983,7 +3983,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x00008000, 0xc4050078, 0x00008000, 0xc4050078,
0x008ce802, 0x00010000, 0x008ce802, 0x00010000,
0xc4050078, 0x008ce803, 0xc4050078, 0x008ce803,
0x00018000, 0xbf8903f7, 0x00018000, 0xbf8a0000,
0x7e008500, 0x7e028501, 0x7e008500, 0x7e028501,
0x7e048502, 0x7e068503, 0x7e048502, 0x7e068503,
0x807d847d, 0x8078ff78, 0x807d847d, 0x8078ff78,
...@@ -3994,7 +3994,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -3994,7 +3994,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x00008000, 0xc405006e, 0x00008000, 0xc405006e,
0x008ce802, 0x00010000, 0x008ce802, 0x00010000,
0xc405006e, 0x008ce803, 0xc405006e, 0x008ce803,
0x00018000, 0xbf8903f7, 0x00018000, 0xbf8a0000,
0xbfa0003d, 0xbef600ff, 0xbfa0003d, 0xbef600ff,
0x01000000, 0xbeee0078, 0x01000000, 0xbeee0078,
0x8078ff78, 0x00000400, 0x8078ff78, 0x00000400,
...@@ -4005,7 +4005,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -4005,7 +4005,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x00010000, 0xc4050078, 0x00010000, 0xc4050078,
0x008ce802, 0x00020000, 0x008ce802, 0x00020000,
0xc4050078, 0x008ce803, 0xc4050078, 0x008ce803,
0x00030000, 0xbf8903f7, 0x00030000, 0xbf8a0000,
0x7e008500, 0x7e028501, 0x7e008500, 0x7e028501,
0x7e048502, 0x7e068503, 0x7e048502, 0x7e068503,
0x807d847d, 0x8078ff78, 0x807d847d, 0x8078ff78,
...@@ -4015,7 +4015,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -4015,7 +4015,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x846f836f, 0x806f7d6f, 0x846f836f, 0x806f7d6f,
0xbefe00c1, 0xbeff0080, 0xbefe00c1, 0xbeff0080,
0xc4050078, 0x008ce800, 0xc4050078, 0x008ce800,
0x00000000, 0xbf8903f7, 0x00000000, 0xbf8a0000,
0x7e008500, 0x807d817d, 0x7e008500, 0x807d817d,
0x8078ff78, 0x00000080, 0x8078ff78, 0x00000080,
0xbf0a6f7d, 0xbfa2fff6, 0xbf0a6f7d, 0xbfa2fff6,
...@@ -4025,7 +4025,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -4025,7 +4025,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x00010000, 0xc405006e, 0x00010000, 0xc405006e,
0x008ce802, 0x00020000, 0x008ce802, 0x00020000,
0xc405006e, 0x008ce803, 0xc405006e, 0x008ce803,
0x00030000, 0xbf8903f7, 0x00030000, 0xbf8a0000,
0xb8f83b05, 0x80788178, 0xb8f83b05, 0x80788178,
0xbf0d9972, 0xbfa20002, 0xbf0d9972, 0xbfa20002,
0x84788978, 0xbfa00001, 0x84788978, 0xbfa00001,
...@@ -4036,16 +4036,16 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -4036,16 +4036,16 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0xbef600ff, 0x01000000, 0xbef600ff, 0x01000000,
0xbefd00ff, 0x0000006c, 0xbefd00ff, 0x0000006c,
0x80f89078, 0xf462403a, 0x80f89078, 0xf462403a,
0xf0000000, 0xbf89fc07, 0xf0000000, 0xbf8a0000,
0x80fd847d, 0xbf800000, 0x80fd847d, 0xbf800000,
0xbe804300, 0xbe824302, 0xbe804300, 0xbe824302,
0x80f8a078, 0xf462603a, 0x80f8a078, 0xf462603a,
0xf0000000, 0xbf89fc07, 0xf0000000, 0xbf8a0000,
0x80fd887d, 0xbf800000, 0x80fd887d, 0xbf800000,
0xbe804300, 0xbe824302, 0xbe804300, 0xbe824302,
0xbe844304, 0xbe864306, 0xbe844304, 0xbe864306,
0x80f8c078, 0xf462803a, 0x80f8c078, 0xf462803a,
0xf0000000, 0xbf89fc07, 0xf0000000, 0xbf8a0000,
0x80fd907d, 0xbf800000, 0x80fd907d, 0xbf800000,
0xbe804300, 0xbe824302, 0xbe804300, 0xbe824302,
0xbe844304, 0xbe864306, 0xbe844304, 0xbe864306,
...@@ -4075,19 +4075,19 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -4075,19 +4075,19 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x80788478, 0xf4621cfa, 0x80788478, 0xf4621cfa,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xf4621bba, 0xf0000000, 0xf4621bba, 0xf0000000,
0x80788478, 0xbf89fc07, 0x80788478, 0xbf8a0000,
0xb96ef814, 0xf4621bba, 0xb96ef814, 0xf4621bba,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xbf89fc07, 0xb96ef815, 0xbf8a0000, 0xb96ef815,
0xf4621bba, 0xf0000000, 0xf4621bba, 0xf0000000,
0x80788478, 0xbf89fc07, 0x80788478, 0xbf8a0000,
0xb96ef812, 0xf4621bba, 0xb96ef812, 0xf4621bba,
0xf0000000, 0x80788478, 0xf0000000, 0x80788478,
0xbf89fc07, 0xb96ef813, 0xbf8a0000, 0xb96ef813,
0x8b6eff7f, 0x04000000, 0x8b6eff7f, 0x04000000,
0xbfa1000d, 0x80788478, 0xbfa1000d, 0x80788478,
0xf4621bba, 0xf0000000, 0xf4621bba, 0xf0000000,
0x80788478, 0xbf89fc07, 0x80788478, 0xbf8a0000,
0xbf0d806e, 0xbfa10006, 0xbf0d806e, 0xbfa10006,
0x856e906e, 0x8b6e6e6e, 0x856e906e, 0x8b6e6e6e,
0xbfa10003, 0xbe804ec1, 0xbfa10003, 0xbe804ec1,
...@@ -4106,7 +4106,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = { ...@@ -4106,7 +4106,7 @@ static const uint32_t cwsr_trap_gfx12_hex[] = {
0x0000ffff, 0xf4605c37, 0x0000ffff, 0xf4605c37,
0xf8000050, 0xf4605d37, 0xf8000050, 0xf4605d37,
0xf8000060, 0xf4601e77, 0xf8000060, 0xf4601e77,
0xf8000074, 0xbf89fc07, 0xf8000074, 0xbf8a0000,
0x8b6dff6d, 0x0000ffff, 0x8b6dff6d, 0x0000ffff,
0x8bfe7e7e, 0x8bea6a6a, 0x8bfe7e7e, 0x8bea6a6a,
0xb97af804, 0xbe804a6c, 0xb97af804, 0xbe804a6c,
......
drivers/gpu/drm/amd/amdkfd/cwsr_trap_handler_gfx10.asm
View file @ c5e35891
...@@ -55,9 +55,11 @@ ...@@ -55,9 +55,11 @@
#if ASIC_FAMILY < CHIP_GFX12 #if ASIC_FAMILY < CHIP_GFX12
#define S_COHERENCE glc:1 #define S_COHERENCE glc:1
#define V_COHERENCE slc:1 glc:1 #define V_COHERENCE slc:1 glc:1
#define S_WAITCNT_0 s_waitcnt 0
#else #else
#define S_COHERENCE scope:SCOPE_SYS #define S_COHERENCE scope:SCOPE_SYS
#define V_COHERENCE scope:SCOPE_SYS #define V_COHERENCE scope:SCOPE_SYS
#define S_WAITCNT_0 s_wait_idle
#define HW_REG_SHADER_FLAT_SCRATCH_LO HW_REG_WAVE_SCRATCH_BASE_LO #define HW_REG_SHADER_FLAT_SCRATCH_LO HW_REG_WAVE_SCRATCH_BASE_LO
#define HW_REG_SHADER_FLAT_SCRATCH_HI HW_REG_WAVE_SCRATCH_BASE_HI #define HW_REG_SHADER_FLAT_SCRATCH_HI HW_REG_WAVE_SCRATCH_BASE_HI
...@@ -364,7 +366,7 @@ L_FETCH_2ND_TRAP: ...@@ -364,7 +366,7 @@ L_FETCH_2ND_TRAP:
// ttmp12 holds SQ_WAVE_STATUS // ttmp12 holds SQ_WAVE_STATUS
#if HAVE_SENDMSG_RTN #if HAVE_SENDMSG_RTN
s_sendmsg_rtn_b64 [ttmp14, ttmp15], sendmsg(MSG_RTN_GET_TMA) s_sendmsg_rtn_b64 [ttmp14, ttmp15], sendmsg(MSG_RTN_GET_TMA)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
#else #else
s_getreg_b32 ttmp14, hwreg(HW_REG_SHADER_TMA_LO) s_getreg_b32 ttmp14, hwreg(HW_REG_SHADER_TMA_LO)
s_getreg_b32 ttmp15, hwreg(HW_REG_SHADER_TMA_HI) s_getreg_b32 ttmp15, hwreg(HW_REG_SHADER_TMA_HI)
...@@ -377,15 +379,15 @@ L_FETCH_2ND_TRAP: ...@@ -377,15 +379,15 @@ L_FETCH_2ND_TRAP:
L_NO_SIGN_EXTEND_TMA: L_NO_SIGN_EXTEND_TMA:
s_load_dword ttmp2, [ttmp14, ttmp15], 0x10 S_COHERENCE // debug trap enabled flag s_load_dword ttmp2, [ttmp14, ttmp15], 0x10 S_COHERENCE // debug trap enabled flag
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_lshl_b32 ttmp2, ttmp2, TTMP11_DEBUG_TRAP_ENABLED_SHIFT s_lshl_b32 ttmp2, ttmp2, TTMP11_DEBUG_TRAP_ENABLED_SHIFT
s_andn2_b32 ttmp11, ttmp11, TTMP11_DEBUG_TRAP_ENABLED_MASK s_andn2_b32 ttmp11, ttmp11, TTMP11_DEBUG_TRAP_ENABLED_MASK
s_or_b32 ttmp11, ttmp11, ttmp2 s_or_b32 ttmp11, ttmp11, ttmp2
s_load_dwordx2 [ttmp2, ttmp3], [ttmp14, ttmp15], 0x0 S_COHERENCE // second-level TBA s_load_dwordx2 [ttmp2, ttmp3], [ttmp14, ttmp15], 0x0 S_COHERENCE // second-level TBA
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_load_dwordx2 [ttmp14, ttmp15], [ttmp14, ttmp15], 0x8 S_COHERENCE // second-level TMA s_load_dwordx2 [ttmp14, ttmp15], [ttmp14, ttmp15], 0x8 S_COHERENCE // second-level TMA
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_and_b64 [ttmp2, ttmp3], [ttmp2, ttmp3], [ttmp2, ttmp3] s_and_b64 [ttmp2, ttmp3], [ttmp2, ttmp3], [ttmp2, ttmp3]
s_cbranch_scc0 L_NO_NEXT_TRAP // second-level trap handler not been set s_cbranch_scc0 L_NO_NEXT_TRAP // second-level trap handler not been set
...@@ -460,7 +462,7 @@ L_SLEEP: ...@@ -460,7 +462,7 @@ L_SLEEP:
s_sleep 0x2 s_sleep 0x2
s_cbranch_execz L_SLEEP s_cbranch_execz L_SLEEP
#else #else
s_waitcnt lgkmcnt(0) S_WAITCNT_0
#endif #endif
// Save first_wave flag so we can clear high bits of save address. // Save first_wave flag so we can clear high bits of save address.
...@@ -794,7 +796,7 @@ L_SAVE_LDS_W32: ...@@ -794,7 +796,7 @@ L_SAVE_LDS_W32:
L_SAVE_LDS_LOOP_SQC_W32: L_SAVE_LDS_LOOP_SQC_W32:
ds_read_b32 v1, v0 ds_read_b32 v1, v0
s_waitcnt 0 S_WAITCNT_0
write_vgprs_to_mem_with_sqc_w32(v1, 1, s_save_buf_rsrc0, s_save_mem_offset) write_vgprs_to_mem_with_sqc_w32(v1, 1, s_save_buf_rsrc0, s_save_mem_offset)
...@@ -814,7 +816,7 @@ L_SAVE_LDS_WITH_TCP_W32: ...@@ -814,7 +816,7 @@ L_SAVE_LDS_WITH_TCP_W32:
s_nop 0 s_nop 0
L_SAVE_LDS_LOOP_W32: L_SAVE_LDS_LOOP_W32:
ds_read_b32 v1, v0 ds_read_b32 v1, v0
s_waitcnt 0 S_WAITCNT_0
buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset V_COHERENCE buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset V_COHERENCE
s_add_u32 m0, m0, s3 //every buffer_store_lds does 128 bytes s_add_u32 m0, m0, s3 //every buffer_store_lds does 128 bytes
...@@ -832,7 +834,7 @@ L_SAVE_LDS_W64: ...@@ -832,7 +834,7 @@ L_SAVE_LDS_W64:
L_SAVE_LDS_LOOP_SQC_W64: L_SAVE_LDS_LOOP_SQC_W64:
ds_read_b32 v1, v0 ds_read_b32 v1, v0
s_waitcnt 0 S_WAITCNT_0
write_vgprs_to_mem_with_sqc_w64(v1, 1, s_save_buf_rsrc0, s_save_mem_offset) write_vgprs_to_mem_with_sqc_w64(v1, 1, s_save_buf_rsrc0, s_save_mem_offset)
...@@ -852,7 +854,7 @@ L_SAVE_LDS_WITH_TCP_W64: ...@@ -852,7 +854,7 @@ L_SAVE_LDS_WITH_TCP_W64:
s_nop 0 s_nop 0
L_SAVE_LDS_LOOP_W64: L_SAVE_LDS_LOOP_W64:
ds_read_b32 v1, v0 ds_read_b32 v1, v0
s_waitcnt 0 S_WAITCNT_0
buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset V_COHERENCE buffer_store_dword v1, v0, s_save_buf_rsrc0, s_save_mem_offset V_COHERENCE
s_add_u32 m0, m0, s3 //every buffer_store_lds does 256 bytes s_add_u32 m0, m0, s3 //every buffer_store_lds does 256 bytes
...@@ -1073,7 +1075,7 @@ L_RESTORE_LDS_LOOP_W32: ...@@ -1073,7 +1075,7 @@ L_RESTORE_LDS_LOOP_W32:
buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW
#else #else
buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset
s_waitcnt vmcnt(0) S_WAITCNT_0
ds_store_addtid_b32 v0 ds_store_addtid_b32 v0
#endif #endif
s_add_u32 m0, m0, 128 // 128 DW s_add_u32 m0, m0, 128 // 128 DW
...@@ -1087,7 +1089,7 @@ L_RESTORE_LDS_LOOP_W64: ...@@ -1087,7 +1089,7 @@ L_RESTORE_LDS_LOOP_W64:
buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset lds:1 // first 64DW
#else #else
buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset
s_waitcnt vmcnt(0) S_WAITCNT_0
ds_store_addtid_b32 v0 ds_store_addtid_b32 v0
#endif #endif
s_add_u32 m0, m0, 256 // 256 DW s_add_u32 m0, m0, 256 // 256 DW
...@@ -1132,7 +1134,7 @@ L_RESTORE_VGPR_WAVE32_LOOP: ...@@ -1132,7 +1134,7 @@ L_RESTORE_VGPR_WAVE32_LOOP:
buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:128 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:128
buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:128*2 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:128*2
buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:128*3 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:128*3
s_waitcnt vmcnt(0) S_WAITCNT_0
v_movreld_b32 v0, v0 //v[0+m0] = v0 v_movreld_b32 v0, v0 //v[0+m0] = v0
v_movreld_b32 v1, v1 v_movreld_b32 v1, v1
v_movreld_b32 v2, v2 v_movreld_b32 v2, v2
...@@ -1147,7 +1149,7 @@ L_RESTORE_VGPR_WAVE32_LOOP: ...@@ -1147,7 +1149,7 @@ L_RESTORE_VGPR_WAVE32_LOOP:
buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:128 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:128
buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:128*2 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:128*2
buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:128*3 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:128*3
s_waitcnt vmcnt(0) S_WAITCNT_0
s_branch L_RESTORE_SGPR s_branch L_RESTORE_SGPR
...@@ -1166,7 +1168,7 @@ L_RESTORE_VGPR_WAVE64_LOOP: ...@@ -1166,7 +1168,7 @@ L_RESTORE_VGPR_WAVE64_LOOP:
buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:256 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:256
buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:256*2 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:256*2
buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:256*3 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE offset:256*3
s_waitcnt vmcnt(0) S_WAITCNT_0
v_movreld_b32 v0, v0 //v[0+m0] = v0 v_movreld_b32 v0, v0 //v[0+m0] = v0
v_movreld_b32 v1, v1 v_movreld_b32 v1, v1
v_movreld_b32 v2, v2 v_movreld_b32 v2, v2
...@@ -1189,7 +1191,7 @@ L_RESTORE_SHARED_VGPR: ...@@ -1189,7 +1191,7 @@ L_RESTORE_SHARED_VGPR:
s_mov_b32 exec_hi, 0x00000000 s_mov_b32 exec_hi, 0x00000000
L_RESTORE_SHARED_VGPR_WAVE64_LOOP: L_RESTORE_SHARED_VGPR_WAVE64_LOOP:
buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE buffer_load_dword v0, v0, s_restore_buf_rsrc0, s_restore_mem_offset V_COHERENCE
s_waitcnt vmcnt(0) S_WAITCNT_0
v_movreld_b32 v0, v0 //v[0+m0] = v0 v_movreld_b32 v0, v0 //v[0+m0] = v0
s_add_u32 m0, m0, 1 //next vgpr index s_add_u32 m0, m0, 1 //next vgpr index
s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128 s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 128
...@@ -1204,7 +1206,7 @@ L_RESTORE_V0: ...@@ -1204,7 +1206,7 @@ L_RESTORE_V0:
buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:256 buffer_load_dword v1, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:256
buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:256*2 buffer_load_dword v2, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:256*2
buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:256*3 buffer_load_dword v3, v0, s_restore_buf_rsrc0, s_restore_mem_offset_save V_COHERENCE offset:256*3
s_waitcnt vmcnt(0) S_WAITCNT_0
/* restore SGPRs */ /* restore SGPRs */
//will be 2+8+16*6 //will be 2+8+16*6
...@@ -1221,7 +1223,7 @@ L_RESTORE_SGPR: ...@@ -1221,7 +1223,7 @@ L_RESTORE_SGPR:
s_mov_b32 m0, s_sgpr_save_num s_mov_b32 m0, s_sgpr_save_num
read_4sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset) read_4sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_sub_u32 m0, m0, 4 // Restore from S[0] to S[104] s_sub_u32 m0, m0, 4 // Restore from S[0] to S[104]
s_nop 0 // hazard SALU M0=> S_MOVREL s_nop 0 // hazard SALU M0=> S_MOVREL
...@@ -1230,7 +1232,7 @@ L_RESTORE_SGPR: ...@@ -1230,7 +1232,7 @@ L_RESTORE_SGPR:
s_movreld_b64 s2, s2 s_movreld_b64 s2, s2
read_8sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset) read_8sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_sub_u32 m0, m0, 8 // Restore from S[0] to S[96] s_sub_u32 m0, m0, 8 // Restore from S[0] to S[96]
s_nop 0 // hazard SALU M0=> S_MOVREL s_nop 0 // hazard SALU M0=> S_MOVREL
...@@ -1242,7 +1244,7 @@ L_RESTORE_SGPR: ...@@ -1242,7 +1244,7 @@ L_RESTORE_SGPR:
L_RESTORE_SGPR_LOOP: L_RESTORE_SGPR_LOOP:
read_16sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset) read_16sgpr_from_mem(s0, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_sub_u32 m0, m0, 16 // Restore from S[n] to S[0] s_sub_u32 m0, m0, 16 // Restore from S[n] to S[0]
s_nop 0 // hazard SALU M0=> S_MOVREL s_nop 0 // hazard SALU M0=> S_MOVREL
...@@ -1291,22 +1293,22 @@ L_RESTORE_HWREG: ...@@ -1291,22 +1293,22 @@ L_RESTORE_HWREG:
read_hwreg_from_mem(s_restore_xnack_mask, s_restore_buf_rsrc0, s_restore_mem_offset) read_hwreg_from_mem(s_restore_xnack_mask, s_restore_buf_rsrc0, s_restore_mem_offset)
read_hwreg_from_mem(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset) read_hwreg_from_mem(s_restore_mode, s_restore_buf_rsrc0, s_restore_mem_offset)
read_hwreg_from_mem(s_restore_flat_scratch, s_restore_buf_rsrc0, s_restore_mem_offset) read_hwreg_from_mem(s_restore_flat_scratch, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_setreg_b32 hwreg(HW_REG_SHADER_FLAT_SCRATCH_LO), s_restore_flat_scratch s_setreg_b32 hwreg(HW_REG_SHADER_FLAT_SCRATCH_LO), s_restore_flat_scratch
read_hwreg_from_mem(s_restore_flat_scratch, s_restore_buf_rsrc0, s_restore_mem_offset) read_hwreg_from_mem(s_restore_flat_scratch, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) //from now on, it is safe to restore STATUS and IB_STS S_WAITCNT_0
s_setreg_b32 hwreg(HW_REG_SHADER_FLAT_SCRATCH_HI), s_restore_flat_scratch s_setreg_b32 hwreg(HW_REG_SHADER_FLAT_SCRATCH_HI), s_restore_flat_scratch
#if ASIC_FAMILY >= CHIP_GFX12 #if ASIC_FAMILY >= CHIP_GFX12
read_hwreg_from_mem(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset) read_hwreg_from_mem(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_setreg_b32 hwreg(HW_REG_WAVE_EXCP_FLAG_USER), s_restore_tmp s_setreg_b32 hwreg(HW_REG_WAVE_EXCP_FLAG_USER), s_restore_tmp
read_hwreg_from_mem(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset) read_hwreg_from_mem(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_setreg_b32 hwreg(HW_REG_WAVE_TRAP_CTRL), s_restore_tmp s_setreg_b32 hwreg(HW_REG_WAVE_TRAP_CTRL), s_restore_tmp
// Only the first wave needs to restore the workgroup barrier. // Only the first wave needs to restore the workgroup barrier.
...@@ -1317,7 +1319,7 @@ L_RESTORE_HWREG: ...@@ -1317,7 +1319,7 @@ L_RESTORE_HWREG:
s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 4 s_add_u32 s_restore_mem_offset, s_restore_mem_offset, 4
read_hwreg_from_mem(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset) read_hwreg_from_mem(s_restore_tmp, s_restore_buf_rsrc0, s_restore_mem_offset)
s_waitcnt lgkmcnt(0) S_WAITCNT_0
s_bitcmp1_b32 s_restore_tmp, BARRIER_STATE_VALID_OFFSET s_bitcmp1_b32 s_restore_tmp, BARRIER_STATE_VALID_OFFSET
s_cbranch_scc0 L_SKIP_BARRIER_RESTORE s_cbranch_scc0 L_SKIP_BARRIER_RESTORE
...@@ -1364,7 +1366,7 @@ L_SKIP_BARRIER_RESTORE: ...@@ -1364,7 +1366,7 @@ L_SKIP_BARRIER_RESTORE:
s_load_dwordx4 [ttmp4, ttmp5, ttmp6, ttmp7], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x50 S_COHERENCE s_load_dwordx4 [ttmp4, ttmp5, ttmp6, ttmp7], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x50 S_COHERENCE
s_load_dwordx4 [ttmp8, ttmp9, ttmp10, ttmp11], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x60 S_COHERENCE s_load_dwordx4 [ttmp8, ttmp9, ttmp10, ttmp11], [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x60 S_COHERENCE
s_load_dword ttmp13, [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x74 S_COHERENCE s_load_dword ttmp13, [s_restore_ttmps_lo, s_restore_ttmps_hi], 0x74 S_COHERENCE
s_waitcnt lgkmcnt(0) S_WAITCNT_0
#if HAVE_XNACK #if HAVE_XNACK
restore_ib_sts(s_restore_tmp, s_restore_m0) restore_ib_sts(s_restore_tmp, s_restore_m0)
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7