NVIDIA drivers perform too aggressive optimization involving change of computation order (aka 'fastmath') by default. The setting helps to workaround bugs like https://bugs.winehq.org/show_bug.cgi?id=35207.
Signed-off-by: Paul Gofman gofmanp@gmail.com --- dlls/wined3d/adapter_gl.c | 17 +++++++++++++++++ dlls/wined3d/glsl_shader.c | 3 +++ dlls/wined3d/wined3d_main.c | 8 ++++++++ dlls/wined3d/wined3d_private.h | 2 ++ 4 files changed, 30 insertions(+)
diff --git a/dlls/wined3d/adapter_gl.c b/dlls/wined3d/adapter_gl.c index 751e1672e2..ae1680605a 100644 --- a/dlls/wined3d/adapter_gl.c +++ b/dlls/wined3d/adapter_gl.c @@ -832,6 +832,13 @@ static BOOL match_broken_viewport_subpixel_bits(const struct wined3d_gl_info *gl return !wined3d_caps_gl_ctx_test_viewport_subpixel_bits(ctx); }
+static BOOL match_disable_nvidia_fastmath(const struct wined3d_gl_info *gl_info, + struct wined3d_caps_gl_ctx *ctx, const char *gl_renderer, enum wined3d_gl_vendor gl_vendor, + enum wined3d_pci_vendor card_vendor, enum wined3d_pci_device device) +{ + return wined3d_settings.disable_nvidia_fastmath && gl_vendor == GL_VENDOR_NVIDIA; +} + static void quirk_apple_glsl_constants(struct wined3d_gl_info *gl_info) { /* MacOS needs uniforms for relative addressing offsets. This can @@ -986,6 +993,11 @@ static void quirk_broken_viewport_subpixel_bits(struct wined3d_gl_info *gl_info) } }
+static void quirk_disable_nvidia_fastmath(struct wined3d_gl_info *gl_info) +{ + gl_info->quirks |= WINED3D_QUIRK_DISABLE_NVIDIA_FASTMATH; +} + static const struct wined3d_gpu_description *query_gpu_description(const struct wined3d_gl_info *gl_info, UINT64 *vram_bytes) { @@ -1130,6 +1142,11 @@ static void fixup_extensions(struct wined3d_gl_info *gl_info, struct wined3d_cap quirk_broken_viewport_subpixel_bits, "NVIDIA viewport subpixel bits bug" }, + { + match_disable_nvidia_fastmath, + quirk_disable_nvidia_fastmath, + "Disabling fastmath optimization on NVIDIA." + }, };
for (i = 0; i < ARRAY_SIZE(quirk_table); ++i) diff --git a/dlls/wined3d/glsl_shader.c b/dlls/wined3d/glsl_shader.c index 2f65b75084..8a042a82d4 100644 --- a/dlls/wined3d/glsl_shader.c +++ b/dlls/wined3d/glsl_shader.c @@ -2916,6 +2916,9 @@ static void shader_generate_glsl_declarations(const struct wined3d_context *cont shader_addline(buffer, ";\n"); } } + + if (gl_info->quirks & WINED3D_QUIRK_DISABLE_NVIDIA_FASTMATH) + shader_addline(buffer, "#pragma optionNV(fastmath off)\n"); }
/* Prototypes */ diff --git a/dlls/wined3d/wined3d_main.c b/dlls/wined3d/wined3d_main.c index feab178e24..b9db439d59 100644 --- a/dlls/wined3d/wined3d_main.c +++ b/dlls/wined3d/wined3d_main.c @@ -83,6 +83,7 @@ struct wined3d_settings wined3d_settings = TRUE, /* Prefer multisample textures to multisample renderbuffers. */ ~0u, /* Don't force a specific sample count by default. */ FALSE, /* Don't range check relative addressing indices in float constants. */ + FALSE, /* Do not disable fastmath optimization on NVIDIA. */ ~0U, /* No VS shader model limit by default. */ ~0U, /* No HS shader model limit by default. */ ~0U, /* No DS shader model limit by default. */ @@ -297,6 +298,13 @@ static BOOL wined3d_dll_init(HINSTANCE hInstDLL) TRACE("Checking relative addressing indices in float constants.\n"); wined3d_settings.check_float_constants = TRUE; } + if (!get_config_key(hkey, appkey, "DisableNVIDIAFastmath", buffer, size) + && !strcmp(buffer, "enabled")) + { + TRACE("Disabling fastmath optimization on NVIDIA.\n"); + wined3d_settings.disable_nvidia_fastmath = TRUE; + } + if (!get_config_key_dword(hkey, appkey, "MaxShaderModelVS", &wined3d_settings.max_sm_vs)) TRACE("Limiting VS shader model to %u.\n", wined3d_settings.max_sm_vs); if (!get_config_key_dword(hkey, appkey, "MaxShaderModelHS", &wined3d_settings.max_sm_hs)) diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h index a6bfd132f4..76b31f61e4 100644 --- a/dlls/wined3d/wined3d_private.h +++ b/dlls/wined3d/wined3d_private.h @@ -71,6 +71,7 @@ #define WINED3D_QUIRK_INFO_LOG_SPAM 0x00000080 #define WINED3D_QUIRK_LIMITED_TEX_FILTERING 0x00000100 #define WINED3D_QUIRK_BROKEN_ARB_FOG 0x00000200 +#define WINED3D_QUIRK_DISABLE_NVIDIA_FASTMATH 0x00000400
struct fragment_pipeline; struct wined3d_context; @@ -396,6 +397,7 @@ struct wined3d_settings unsigned int multisample_textures; unsigned int sample_count; BOOL check_float_constants; + BOOL disable_nvidia_fastmath; unsigned int max_sm_vs; unsigned int max_sm_hs; unsigned int max_sm_ds;
On Mon, 25 Feb 2019 at 19:44, Paul Gofman gofmanp@gmail.com wrote:
NVIDIA drivers perform too aggressive optimization involving change of computation order (aka 'fastmath') by default. The setting helps to workaround bugs like https://bugs.winehq.org/show_bug.cgi?id=35207.
Signed-off-by: Paul Gofman gofmanp@gmail.com
dlls/wined3d/adapter_gl.c | 17 +++++++++++++++++ dlls/wined3d/glsl_shader.c | 3 +++ dlls/wined3d/wined3d_main.c | 8 ++++++++ dlls/wined3d/wined3d_private.h | 2 ++ 4 files changed, 30 insertions(+)
Does the attached patch also work for you?
On 2/26/19 20:44, Henri Verbeet wrote:
On Mon, 25 Feb 2019 at 19:44, Paul Gofman gofmanp@gmail.com wrote:
NVIDIA drivers perform too aggressive optimization involving change of computation order (aka 'fastmath') by default. The setting helps to workaround bugs like https://bugs.winehq.org/show_bug.cgi?id=35207.
Signed-off-by: Paul Gofman gofmanp@gmail.com
dlls/wined3d/adapter_gl.c | 17 +++++++++++++++++ dlls/wined3d/glsl_shader.c | 3 +++ dlls/wined3d/wined3d_main.c | 8 ++++++++ dlls/wined3d/wined3d_private.h | 2 ++ 4 files changed, 30 insertions(+)
Does the attached patch also work for you?
Yes, sure, it does.
-
Henri Verbeet -
Paul Gofman