wine/wine-staging
0
0
Fork 0
mirror of https://gitlab.winehq.org/wine/wine-staging.git synced 2025-01-28 22:04:43 -08:00
Code Issues Packages Projects Releases Wiki Activity

nvcuda-CUDA_Support: Update patchset with various improvements.

Browse Source 
* Add static to a couple of functions
* TLS callbacks should be declared as CDECL
* Do not unlock the critical section while processing TLS callbacks
This commit is contained in:
Sebastian Lackner 2015-01-21 03:09:54 +01:00
parent ddbbcb6c0f
commit f64437a901
4 changed files with 54 additions and 56 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

76
patches/nvcuda-CUDA_Support/0003-nvcuda-First-implementation.patch
View File

@ -1,28 +1,32 @@
From 81fe2ed75ea9e0f282c496f0c8183a4e316fb78d Mon Sep 17 00:00:00 2001
From 3a96e7ad09c6ad0ad31313634688c5c374d5d89f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Michael=20M=C3=BCller?= <michael@fds-team.de>
Date: Sat, 3 Jan 2015 03:39:11 +0100
Subject: nvcuda: First implementation.
Subject: nvcuda: First implementation. (rev 2)
Changes by Sebastian Lackner <sebastian@fds-team.de>:
* Add static to a couple of functions
* TLS callbacks should be declared as CDECL
* Do not unlock the critical section while processing TLS callbacks
---
configure.ac | 1 +
dlls/nvcuda/Makefile.in | 3 +-
dlls/nvcuda/internal.c | 521 ++++++++++
dlls/nvcuda/internal.c | 515 ++++++++++
dlls/nvcuda/nvcuda.c | 2097 ++++++++++++++++++++++++++++++++++++++++-
dlls/nvcuda/nvcuda.h | 29 +
dlls/nvcuda/nvcuda.spec | 492 +++++-----
dlls/nvcuda/tests/Makefile.in | 4 +
dlls/nvcuda/tests/nvcuda.c | 168 ++++
8 files changed, 3063 insertions(+), 252 deletions(-)
8 files changed, 3057 insertions(+), 252 deletions(-)
create mode 100755 dlls/nvcuda/internal.c
create mode 100644 dlls/nvcuda/nvcuda.h
create mode 100644 dlls/nvcuda/tests/Makefile.in
create mode 100644 dlls/nvcuda/tests/nvcuda.c
diff --git a/configure.ac b/configure.ac
index a4248d0..73e67bb 100644
index 48bd58c..dc03595 100644
--- a/configure.ac
+++ b/configure.ac
@@ -3126,6 +3126,7 @@ WINE_CONFIG_DLL(ntoskrnl.exe,,[implib])
@@ -3127,6 +3127,7 @@ WINE_CONFIG_DLL(ntoskrnl.exe,,[implib])
WINE_CONFIG_DLL(ntprint)
WINE_CONFIG_TEST(dlls/ntprint/tests)
WINE_CONFIG_DLL(nvcuda)
@ -45,10 +49,10 @@ index 4b33278..6322fb2 100644
RC_SRCS = nvcuda.rc
diff --git a/dlls/nvcuda/internal.c b/dlls/nvcuda/internal.c
new file mode 100755
index 0000000..4e733eb
index 0000000..611e310
--- /dev/null
+++ b/dlls/nvcuda/internal.c
@@ -0,0 +1,521 @@
@@ -0,0 +1,515 @@
+/*
+ * Copyright (C) 2014-2015 Michael Müller
+ * Copyright (C) 2014-2015 Sebastian Lackner
@ -87,7 +91,7 @@ index 0000000..4e733eb
+struct tls_callback_entry
+{
+ struct list entry;
+ void (*callback)(DWORD, void *);
+ void (CDECL *callback)(DWORD, void *);
+ void *userdata;
+ ULONG count;
+};
@ -105,7 +109,6 @@ index 0000000..4e733eb
+
+void cuda_process_tls_callbacks(DWORD reason)
+{
+ struct tls_callback_entry *to_free = NULL;
+ struct list *ptr;
+
+ TRACE("(%d)\n", reason);
@ -119,24 +122,19 @@ index 0000000..4e733eb
+ {
+ struct tls_callback_entry *callback = LIST_ENTRY( ptr, struct tls_callback_entry, entry );
+ callback->count++;
+ LeaveCriticalSection( &tls_callback_section );
+ HeapFree( GetProcessHeap(), 0, to_free );
+ to_free = NULL;
+
+ TRACE("calling handler %p(0, %p)\n", callback->callback, callback->userdata);
+ callback->callback(0, callback->userdata);
+ TRACE("handler %p returned\n", callback->callback);
+
+ EnterCriticalSection( &tls_callback_section );
+ ptr = list_next( &tls_callbacks, ptr );
+ if (!--callback->count) /* removed during execution */
+ {
+ list_remove( &callback->entry );
+ to_free = callback;
+ HeapFree( GetProcessHeap(), 0, callback );
+ }
+ }
+ LeaveCriticalSection( &tls_callback_section );
+ HeapFree( GetProcessHeap(), 0, to_free );
+}
+
+static const CUuuid UUID_Unknown1 = {{0x6B, 0xD5, 0xFB, 0x6C, 0x5B, 0xF4, 0xE7, 0x4A,
@ -261,31 +259,31 @@ index 0000000..4e733eb
+} *Unknown5_orig = NULL;
+
+
+void* WINAPI Unknown1_func0_relay(void *param0, void *param1)
+static void* WINAPI Unknown1_func0_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown1_orig->func0(param0, param1);
+}
+
+void* WINAPI Unknown1_func1_relay(void *param0, void *param1)
+static void* WINAPI Unknown1_func1_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown1_orig->func1(param0, param1);
+}
+
+void* WINAPI Unknown1_func2_relay(void *param0, void *param1)
+static void* WINAPI Unknown1_func2_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown1_orig->func2(param0, param1);
+}
+
+void* WINAPI Unknown1_func3_relay(void *param0, void *param1)
+static void* WINAPI Unknown1_func3_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown1_orig->func3(param0, param1);
+}
+
+void* WINAPI Unknown1_func4_relay(void *param0)
+static void* WINAPI Unknown1_func4_relay(void *param0)
+{
+ TRACE("(%p)\n", param0);
+ return Unknown1_orig->func4(param0);
@ -301,37 +299,37 @@ index 0000000..4e733eb
+ Unknown1_func4_relay,
+};
+
+void* WINAPI Unknown2_func0_relay(void *param0, void *param1)
+static void* WINAPI Unknown2_func0_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown2_orig->func0(param0, param1);
+}
+
+void* WINAPI Unknown2_func1_relay(void *param0, void *param1)
+static void* WINAPI Unknown2_func1_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown2_orig->func1(param0, param1);
+}
+
+void* WINAPI Unknown2_func2_relay(void *param0, void *param1, void *param2)
+static void* WINAPI Unknown2_func2_relay(void *param0, void *param1, void *param2)
+{
+ TRACE("(%p, %p, %p)\n", param0, param1, param2);
+ return Unknown2_orig->func2(param0, param1, param2);
+}
+
+void* WINAPI Unknown2_func3_relay(void *param0, void *param1)
+static void* WINAPI Unknown2_func3_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown2_orig->func3(param0, param1);
+}
+
+void* WINAPI Unknown2_func4_relay(void *param0, void *param1)
+static void* WINAPI Unknown2_func4_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown2_orig->func4(param0, param1);
+}
+
+void* WINAPI Unknown2_func5_relay(void *param0, void *param1)
+static void* WINAPI Unknown2_func5_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown2_orig->func5(param0, param1);
@ -348,38 +346,38 @@ index 0000000..4e733eb
+ Unknown2_func5_relay,
+};
+
+void* WINAPI Unknown3_func0_relay(void *param0)
+static void* WINAPI Unknown3_func0_relay(void *param0)
+{
+ TRACE("(%p)\n", param0);
+ return Unknown3_orig->func0(param0);
+}
+
+void* WINAPI Unknown3_func1_relay(void *param0)
+static void* WINAPI Unknown3_func1_relay(void *param0)
+{
+ TRACE("(%p)\n", param0);
+ return Unknown3_orig->func1(param0);
+}
+
+struct Unknown3_table Unknown3_Impl =
+static struct Unknown3_table Unknown3_Impl =
+{
+ sizeof(struct Unknown3_table),
+ Unknown3_func0_relay,
+ Unknown3_func1_relay,
+};
+
+void* WINAPI Unknown4_func0_relay(void *param0, void *param1, void *param2, void *param3)
+static void* WINAPI Unknown4_func0_relay(void *param0, void *param1, void *param2, void *param3)
+{
+ TRACE("(%p, %p, %p, %p)\n", param0, param1, param2, param3);
+ return Unknown4_orig->func0(param0, param1, param2, param3);
+}
+
+void* WINAPI Unknown4_func1_relay(void *param0, void *param1)
+static void* WINAPI Unknown4_func1_relay(void *param0, void *param1)
+{
+ TRACE("(%p, %p)\n", param0, param1);
+ return Unknown4_orig->func1(param0, param1);
+}
+
+void* WINAPI Unknown4_func2_relay(void *param0, void *param1, void *param2)
+static void* WINAPI Unknown4_func2_relay(void *param0, void *param1, void *param2)
+{
+ TRACE("(%p, %p, %p)\n", param0, param1, param2);
+ return Unknown4_orig->func2(param0, param1, param2);
@ -392,7 +390,7 @@ index 0000000..4e733eb
+ Unknown4_func2_relay,
+};
+
+void* WINAPI Unknown5_func0_relay(void *param0, void *param1, void *param2)
+static void* WINAPI Unknown5_func0_relay(void *param0, void *param1, void *param2)
+{
+ TRACE("(%p, %p, %p)\n", param0, param1, param2);
+ return Unknown5_orig->func0(param0, param1, param2);
@ -404,7 +402,7 @@ index 0000000..4e733eb
+ Unknown5_func0_relay,
+};
+
+CUresult WINAPI TlsNotifyInterface_Set(void **handle, void *callback, void *userdata)
+static CUresult WINAPI TlsNotifyInterface_Set(void **handle, void *callback, void *userdata)
+{
+ struct tls_callback_entry *new_entry;
+
@ -426,7 +424,7 @@ index 0000000..4e733eb
+ return CUDA_SUCCESS;
+}
+
+CUresult WINAPI TlsNotifyInterface_Remove(void *handle, void *param1)
+static CUresult WINAPI TlsNotifyInterface_Remove(void *handle, void *param1)
+{
+ CUresult ret = CUDA_ERROR_INVALID_VALUE;
+ struct tls_callback_entry *to_free = NULL;
@ -571,7 +569,7 @@ index 0000000..4e733eb
+ return CUDA_ERROR_UNKNOWN;
+}
diff --git a/dlls/nvcuda/nvcuda.c b/dlls/nvcuda/nvcuda.c
index 8b7eb9c..5b9eed6 100644
index 8b7eb9c..3ef7c35 100644
--- a/dlls/nvcuda/nvcuda.c
+++ b/dlls/nvcuda/nvcuda.c
@@ -1,4 +1,5 @@
@ -3251,7 +3249,7 @@ index 0000000..81f9290
+ nvcuda.c
diff --git a/dlls/nvcuda/tests/nvcuda.c b/dlls/nvcuda/tests/nvcuda.c
new file mode 100644
index 0000000..3b00c5d
index 0000000..fc8f300
--- /dev/null
+++ b/dlls/nvcuda/tests/nvcuda.c
@@ -0,0 +1,168 @@
@ -3329,7 +3327,7 @@ index 0000000..3b00c5d
+ DWORD threadid;
+};
+
+static void tls_callback_test(DWORD reason, void *data)
+static void CDECL tls_callback_test(DWORD reason, void *data)
+{
+ struct tls_test_data *test_data = data;
+ trace("reason: %d, data: %p\n", reason, data);

4
patches/nvcuda-CUDA_Support/0004-nvcuda-Implement-new-functions-added-in-CUDA-6.5.patch
View File

@ -1,4 +1,4 @@
From 4d10b5ffaadd216525cfef2d9c832a68268d8614 Mon Sep 17 00:00:00 2001
From e08627884ed59ebbac09a1182428718ed0ec1cbe Mon Sep 17 00:00:00 2001
From: Sebastian Lackner <sebastian@fds-team.de>
Date: Fri, 9 Jan 2015 04:39:49 +0100
Subject: nvcuda: Implement new functions added in CUDA 6.5.
@ -10,7 +10,7 @@ Subject: nvcuda: Implement new functions added in CUDA 6.5.
3 files changed, 125 insertions(+)
diff --git a/dlls/nvcuda/nvcuda.c b/dlls/nvcuda/nvcuda.c
index 5b9eed6..bd39547 100644
index 3ef7c35..3118c49 100644
--- a/dlls/nvcuda/nvcuda.c
+++ b/dlls/nvcuda/nvcuda.c
@@ -293,6 +293,20 @@ static CUresult (*pcuTexRefSetMipmapLevelBias)(CUtexref hTexRef, float bias);

28
patches/nvcuda-CUDA_Support/0005-nvcuda-Properly-wrap-undocumented-ContextStorage-int.patch
View File

@ -1,4 +1,4 @@
From 9890ed7c2dcde12d86180d9c65d398d43a2cffca Mon Sep 17 00:00:00 2001
From 39568b59352627db2c0ec1235c71cfcba05141a8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Michael=20M=C3=BCller?= <michael@fds-team.de>
Date: Fri, 9 Jan 2015 07:42:32 +0100
Subject: nvcuda: Properly wrap undocumented 'ContextStorage' interface and add
@ -11,10 +11,10 @@ Subject: nvcuda: Properly wrap undocumented 'ContextStorage' interface and add
3 files changed, 319 insertions(+), 33 deletions(-)
diff --git a/dlls/nvcuda/internal.c b/dlls/nvcuda/internal.c
index 4e733eb..4ba42b3 100755
index 611e310..8a2dffd 100755
--- a/dlls/nvcuda/internal.c
+++ b/dlls/nvcuda/internal.c
@@ -94,7 +94,7 @@ static const CUuuid UUID_Unknown2 = {{0xA0, 0x94, 0x79, 0x8C,
@@ -88,7 +88,7 @@ static const CUuuid UUID_Unknown2 = {{0xA0, 0x94, 0x79, 0x8C,
0x93, 0xF2, 0x08, 0x00, 0x20, 0x0C, 0x0A, 0x66}};
static const CUuuid UUID_Unknown3 = {{0x42, 0xD8, 0x5A, 0x81, 0x23, 0xF6, 0xCB, 0x47,
0x82, 0x98, 0xF6, 0xE7, 0x8A, 0x3A, 0xEC, 0xDC}};
@ -23,7 +23,7 @@ index 4e733eb..4ba42b3 100755
0xA8, 0xC3, 0x68, 0xF3, 0x55, 0xD8, 0x95, 0x93}};
static const CUuuid UUID_Unknown5 = {{0x0C, 0xA5, 0x0B, 0x8C, 0x10, 0x04, 0x92, 0x9A,
0x89, 0xA7, 0xD0, 0xDF, 0x10, 0xE7, 0x72, 0x86}};
@@ -170,20 +170,20 @@ static const struct
@@ -164,20 +164,20 @@ static const struct
} *Unknown3_orig = NULL;
/*
@ -53,11 +53,11 @@ index 4e733eb..4ba42b3 100755
/*
* TlsNotifyInterface
@@ -316,29 +316,80 @@ struct Unknown3_table Unknown3_Impl =
@@ -310,29 +310,80 @@ static struct Unknown3_table Unknown3_Impl =
Unknown3_func1_relay,
};
-void* WINAPI Unknown4_func0_relay(void *param0, void *param1, void *param2, void *param3)
-static void* WINAPI Unknown4_func0_relay(void *param0, void *param1, void *param2, void *param3)
+struct context_storage
{
- TRACE("(%p, %p, %p, %p)\n", param0, param1, param2, param3);
@ -83,8 +83,8 @@ index 4e733eb..4ba42b3 100755
+ HeapFree( GetProcessHeap(), 0, storage );
}
-void* WINAPI Unknown4_func1_relay(void *param0, void *param1)
+CUresult WINAPI ContextStorage_Set(CUcontext ctx, void *key, void *value, void *callback)
-static void* WINAPI Unknown4_func1_relay(void *param0, void *param1)
+static CUresult WINAPI ContextStorage_Set(CUcontext ctx, void *key, void *value, void *callback)
{
- TRACE("(%p, %p)\n", param0, param1);
- return Unknown4_orig->func1(param0, param1);
@ -105,8 +105,8 @@ index 4e733eb..4ba42b3 100755
+ return ret;
}
-void* WINAPI Unknown4_func2_relay(void *param0, void *param1, void *param2)
+CUresult WINAPI ContextStorage_Remove(CUcontext ctx, void *key)
-static void* WINAPI Unknown4_func2_relay(void *param0, void *param1, void *param2)
+static CUresult WINAPI ContextStorage_Remove(CUcontext ctx, void *key)
{
- TRACE("(%p, %p, %p)\n", param0, param1, param2);
- return Unknown4_orig->func2(param0, param1, param2);
@ -123,7 +123,7 @@ index 4e733eb..4ba42b3 100755
+ return ContextStorage_orig->Remove(ctx, key);
+}
+
+CUresult WINAPI ContextStorage_Get(void **value, CUcontext ctx, void *key)
+static CUresult WINAPI ContextStorage_Get(void **value, CUcontext ctx, void *key)
+{
+ struct context_storage *storage;
+ CUresult ret;
@ -146,8 +146,8 @@ index 4e733eb..4ba42b3 100755
+ ContextStorage_Get,
};
void* WINAPI Unknown5_func0_relay(void *param0, void *param1, void *param2)
@@ -486,15 +537,15 @@ CUresult cuda_get_table(const void **table, const CUuuid *uuid, const void *orig
static void* WINAPI Unknown5_func0_relay(void *param0, void *param1, void *param2)
@@ -480,15 +531,15 @@ CUresult cuda_get_table(const void **table, const CUuuid *uuid, const void *orig
*table = (void *)&Unknown3_Impl;
return CUDA_SUCCESS;
}
@ -167,7 +167,7 @@ index 4e733eb..4ba42b3 100755
}
else if (cuda_equal_uuid(uuid, &UUID_Unknown5))
diff --git a/dlls/nvcuda/tests/nvcuda.c b/dlls/nvcuda/tests/nvcuda.c
index 3b00c5d..93aaa49 100644
index fc8f300..9c30c37 100644
--- a/dlls/nvcuda/tests/nvcuda.c
+++ b/dlls/nvcuda/tests/nvcuda.c
@@ -27,10 +27,24 @@

2
patches/patchinstall.sh
View File

@ -2207,7 +2207,7 @@ if test "$enable_nvcuda_CUDA_Support" -eq 1; then
(
echo '+ { "Sebastian Lackner", "include: Add cuda.h.h.", 1 },';
echo '+ { "Sebastian Lackner", "nvcuda: Add stub dll.", 1 },';
echo '+ { "Michael Müller", "nvcuda: First implementation.", 1 },';
echo '+ { "Michael Müller", "nvcuda: First implementation.", 2 },';
echo '+ { "Sebastian Lackner", "nvcuda: Implement new functions added in CUDA 6.5.", 1 },';
echo '+ { "Michael Müller", "nvcuda: Properly wrap undocumented '\''ContextStorage'\'' interface and add tests.", 1 },';
echo '+ { "Michael Müller", "nvcuda: Emulate two d3d9 initialization functions.", 1 },';