[Mesa-dev] [PATCH 03/10] i965/skl: Enable fast color clears on SKL
Ben Widawsky
ben at bwidawsk.net
Tue Nov 3 14:59:17 PST 2015
On Fri, Oct 16, 2015 at 04:10:02PM -0700, Chad Versace wrote:
> But this patch doesn't enable fast clears! The reverts in pathches 6 and
> 7 need to be folded into this patch, otherwise the patch does not do
> what it claims.
>
> Also, you can't enable fast clears before patches 4 and 5 without introducing
> regressions. Patches 4 and 5 must precede this patch.
>
I did mention this in a later patch. I should have at least added that comment
here as well and probably changed the short summary. To repeat what I had there
though, I really liked having the differing SKL functionality be split out in
the separate patches and couldn't do it in my earlier patches because I didn't
have the tiny one liners that you added for strictly disabling the fast clears.
I still prefer having these split out, and I consider that a pretty strong
distinction given that I usually opt for fewer patches.
I've modified this locally to summarize the above. I haven't read your review
comments for the patches after this yet. It's likely that you noticed this, and
we can continue the discussion there.
> On Tue 13 Oct 2015, Ben Widawsky wrote:
> > Based on a patch originally from Kristian. Skylake has extended capabilities
> > with regard to fast clears, but that is saved for another patch.
> >
> > The same effect could be acheived with the following, however I think the way
> > I've done it is more in line with how the docs explain it.
> > --- a/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
> > +++ b/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
> > @@ -150,9 +150,13 @@ intel_get_non_msrt_mcs_alignment(struct brw_context *brw,
> > /* In release builds, fall through */
> > case I915_TILING_Y:
> > *width_px = 32 / mt->cpp;
> > - *height = 4;
> > + if (brw->gen >= 9)
> > + *height = 2;
> > + else
> > + *height = 4;
> >
> > Signed-off-by: Ben Widawsky <ben at bwidawsk.net>
> > ---
> > src/mesa/drivers/dri/i965/brw_meta_fast_clear.c | 54 +++++++++++++++++--------
> > src/mesa/drivers/dri/i965/gen8_surface_state.c | 34 ++++++++++++----
> > src/mesa/drivers/dri/i965/intel_mipmap_tree.c | 9 +++++
> > src/mesa/drivers/dri/i965/intel_mipmap_tree.h | 7 +++-
> > 4 files changed, 78 insertions(+), 26 deletions(-)
> >
> > diff --git a/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c b/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c
> > index fbde3f0..7bf52f0 100644
> > --- a/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c
> > +++ b/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c
> > @@ -204,7 +204,7 @@ brw_draw_rectlist(struct gl_context *ctx, struct rect *rect, int num_instances)
> > }
> >
> > static void
> > -get_fast_clear_rect(struct gl_framebuffer *fb,
> > +get_fast_clear_rect(struct brw_context *brw, struct gl_framebuffer *fb,
> > struct intel_renderbuffer *irb, struct rect *rect)
> > {
> > unsigned int x_align, y_align;
> > @@ -228,7 +228,14 @@ get_fast_clear_rect(struct gl_framebuffer *fb,
> > */
> > intel_get_non_msrt_mcs_alignment(irb->mt, &x_align, &y_align);
> > x_align *= 16;
> > - y_align *= 32;
> > +
> > + /* SKL+ line alignment requirement for Y-tiled are half those of the prior
> > + * generations.
> > + */
> > + if (brw->gen >= 9)
> > + y_align *= 16;
> > + else
> > + y_align *= 32;
> >
> > /* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
> > * Target(s)", beneath the "Fast Color Clear" bullet (p327):
> > @@ -265,8 +272,10 @@ get_fast_clear_rect(struct gl_framebuffer *fb,
> > * terms of (width,height) of the RT.
> > *
> > * MSAA Width of Clear Rect Height of Clear Rect
> > + * 2X Ceil(1/8*width) Ceil(1/2*height)
> > * 4X Ceil(1/8*width) Ceil(1/2*height)
> > * 8X Ceil(1/2*width) Ceil(1/2*height)
> > + * 16X width Ceil(1/2*height)
> > *
> > * The text "with upper left co-ordinate to coincide with actual
> > * rectangle being cleared" is a little confusing--it seems to imply
> > @@ -289,6 +298,9 @@ get_fast_clear_rect(struct gl_framebuffer *fb,
> > case 8:
> > x_scaledown = 2;
> > break;
> > + case 16:
> > + x_scaledown = 1;
> > + break;
> > default:
> > unreachable("Unexpected sample count for fast clear");
> > }
> > @@ -358,18 +370,24 @@ is_color_fast_clear_compatible(struct brw_context *brw,
> >
> > /**
> > * Convert the given color to a bitfield suitable for ORing into DWORD 7 of
> > - * SURFACE_STATE.
> > + * SURFACE_STATE (DWORD 12-15 on SKL+).
> > */
> > -static uint32_t
> > -compute_fast_clear_color_bits(const union gl_color_union *color)
> > +static void
> > +set_fast_clear_color(struct brw_context *brw,
> > + struct intel_mipmap_tree *mt,
> > + const union gl_color_union *color)
> > {
> > - uint32_t bits = 0;
> > - for (int i = 0; i < 4; i++) {
> > - /* Testing for non-0 works for integer and float colors */
> > - if (color->f[i] != 0.0f)
> > - bits |= 1 << (GEN7_SURFACE_CLEAR_COLOR_SHIFT + (3 - i));
> > + if (brw->gen >= 9) {
> > + mt->gen9_fast_clear_color = *color;
> > + } else {
> > + mt->fast_clear_color_value = 0;
> > + for (int i = 0; i < 4; i++) {
> > + /* Testing for non-0 works for integer and float colors */
> > + if (color->f[i] != 0.0f)
> > + mt->fast_clear_color_value |=
> > + 1 << (GEN7_SURFACE_CLEAR_COLOR_SHIFT + (3 - i));
>
> Please put braces round the multi-line if-statement.
Got it.
>
> > + }
> > }
> > - return bits;
> > }
> >
> > static const uint32_t fast_clear_color[4] = { ~0, ~0, ~0, ~0 };
> > @@ -504,8 +522,7 @@ brw_meta_fast_clear(struct brw_context *brw, struct gl_framebuffer *fb,
> >
> > switch (clear_type) {
> > case FAST_CLEAR:
> > - irb->mt->fast_clear_color_value =
> > - compute_fast_clear_color_bits(&ctx->Color.ClearColor);
> > + set_fast_clear_color(brw, irb->mt, &ctx->Color.ClearColor);
> > irb->need_downsample = true;
> >
> > /* If the buffer is already in INTEL_FAST_CLEAR_STATE_CLEAR, the
> > @@ -521,7 +538,7 @@ brw_meta_fast_clear(struct brw_context *brw, struct gl_framebuffer *fb,
> > irb->mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_RESOLVED;
> > irb->need_downsample = true;
> > fast_clear_buffers |= 1 << index;
> > - get_fast_clear_rect(fb, irb, &fast_clear_rect);
> > + get_fast_clear_rect(brw, fb, irb, &fast_clear_rect);
> > break;
> >
> > case REP_CLEAR:
> > @@ -654,8 +671,9 @@ get_resolve_rect(struct brw_context *brw,
> > *
> > * The scaledown factors in the table that follows are related to the
> > * alignment size returned by intel_get_non_msrt_mcs_alignment() by a
> > - * multiplier. For IVB and HSW, we divide by two, for BDW we multiply
> > - * by 8 and 16 and 8 and 8 for SKL.
> > + * multiplier. For IVB and HSW, we divide by two, for BDW we multiply
> > + * by 8 and 16. Similar to the fast clear, SKL eases the BDW vertical scaling
> > + * by a factor of 2.
> > */
> >
> > intel_get_non_msrt_mcs_alignment(mt, &x_align, &y_align);
> > @@ -701,6 +719,10 @@ brw_meta_resolve_color(struct brw_context *brw,
> >
> > brw_bind_rep_write_shader(brw, (float *) fast_clear_color);
> >
> > + /* SKL+ also has a resolve mode for compressed render targets and thus more
> > + * bits to let us select the type of resolve. For fast clear resolves, it
> > + * turns out we can use the same value as pre-SKL though.
> > + */
> > set_fast_clear_op(brw, GEN7_PS_RENDER_TARGET_RESOLVE_ENABLE);
> >
> > mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_RESOLVED;
> > diff --git a/src/mesa/drivers/dri/i965/gen8_surface_state.c b/src/mesa/drivers/dri/i965/gen8_surface_state.c
> > index e70c15b..995b4dd 100644
> > --- a/src/mesa/drivers/dri/i965/gen8_surface_state.c
> > +++ b/src/mesa/drivers/dri/i965/gen8_surface_state.c
> > @@ -183,6 +183,28 @@ gen8_emit_buffer_surface_state(struct brw_context *brw,
> > }
> >
> > static void
> > +gen8_emit_fast_clear_color(struct brw_context *brw,
> > + struct intel_mipmap_tree *mt,
> > + uint32_t *surf)
> > +{
> > + if (brw->gen >= 9) {
> > +#define check_fast_clear_val(x) \
> > + assert(mt->gen9_fast_clear_color.f[x] == 0.0 || \
> > + mt->gen9_fast_clear_color.f[x] == 1.0)
> > + check_fast_clear_val(0);
> > + check_fast_clear_val(1);
> > + check_fast_clear_val(2);
> > + check_fast_clear_val(3);
> > +#undef check_fast_clear_val
>
> Replacing the above with a simple loop accomplishes the same. The code
> is shorder, cleaner, and the compiler will unroll it anyway.
>
> for (int i = 0; ++i; i < 4) {
> >assert(mt->gen9_fast_clear_color.f[x] == 0.0 ||
> mt->gen9_fast_clear_color.f[x] == 1.0)
> }
>
> Anyways, We'll need to remove this check very soon when enabling
> full-color clears. So I prefer that we just omit it now. But it's not
> a big deal either way.
>
I'm also not of a strong opinion here - The asserts are more for clarity than to
catch bugs. As for loop vs. open coded, I matched the way the colors are set
just below. Since it should be a fairly transient hunk like you mentioned, I
would prefer to just keep it.
> > + surf[12] = mt->gen9_fast_clear_color.ui[0];
> > + surf[13] = mt->gen9_fast_clear_color.ui[1];
> > + surf[14] = mt->gen9_fast_clear_color.ui[2];
> > + surf[15] = mt->gen9_fast_clear_color.ui[3];
> > + } else
> > + surf[7] |= mt->fast_clear_color_value;
> > +}
> > +
> > +static void
> > gen8_emit_texture_surface_state(struct brw_context *brw,
> > struct intel_mipmap_tree *mt,
> > GLenum target,
> > @@ -286,7 +308,9 @@ gen8_emit_texture_surface_state(struct brw_context *brw,
> > aux_mode;
> > }
> >
> > - surf[7] = mt->fast_clear_color_value |
> > + gen8_emit_fast_clear_color(brw, mt, surf);
> > +
> > + surf[7] |=
> > SET_FIELD(swizzle_to_scs(GET_SWZ(swizzle, 0)), GEN7_SURFACE_SCS_R) |
> > SET_FIELD(swizzle_to_scs(GET_SWZ(swizzle, 1)), GEN7_SURFACE_SCS_G) |
> > SET_FIELD(swizzle_to_scs(GET_SWZ(swizzle, 2)), GEN7_SURFACE_SCS_B) |
> > @@ -384,12 +408,6 @@ gen8_emit_null_surface_state(struct brw_context *brw,
> > SET_FIELD(height - 1, GEN7_SURFACE_HEIGHT);
> > }
> >
> > -static void
> > -gen8_emit_fast_clear_color(struct intel_mipmap_tree *mt, uint32_t *surf)
> > -{
> > - surf[7] |= mt->fast_clear_color_value;
> > -}
> > -
> > /**
> > * Sets up a surface state structure to point at the given region.
> > * While it is only used for the front/back buffer currently, it should be
> > @@ -516,7 +534,7 @@ gen8_update_renderbuffer_surface(struct brw_context *brw,
> > aux_mode;
> > }
> >
> > - gen8_emit_fast_clear_color(mt, surf);
> > + gen8_emit_fast_clear_color(brw, mt, surf);
> > surf[7] |= SET_FIELD(HSW_SCS_RED, GEN7_SURFACE_SCS_R) |
> > SET_FIELD(HSW_SCS_GREEN, GEN7_SURFACE_SCS_G) |
> > SET_FIELD(HSW_SCS_BLUE, GEN7_SURFACE_SCS_B) |
> > diff --git a/src/mesa/drivers/dri/i965/intel_mipmap_tree.c b/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
> > index b6e3520..32f0ff7 100644
> > --- a/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
> > +++ b/src/mesa/drivers/dri/i965/intel_mipmap_tree.c
> > @@ -192,6 +192,12 @@ intel_tiling_supports_non_msrt_mcs(struct brw_context *brw, unsigned tiling)
> > *
> > * - MCS buffer for non-MSRT is supported only for RT formats 32bpp,
> > * 64bpp, and 128bpp.
> > + *
> > + * From the Skylake documentation, it is made clear that X-tiling is no longer
> > + * supported:
> > + *
> > + * - MCS and Lossless compression is supported for TiledY/TileYs/TileYf
> > + * non-MSRTs only.
> > */
> > static bool
> > intel_miptree_supports_non_msrt_fast_clear(struct brw_context *brw,
> > @@ -1485,6 +1491,9 @@ intel_miptree_alloc_non_msrt_mcs(struct brw_context *brw,
> > intel_get_non_msrt_mcs_alignment(mt, &block_width_px, &block_height);
> > unsigned width_divisor = block_width_px * 4;
> > unsigned height_divisor = block_height * 8;
>
> > + if (brw->gen >= 9)
> > + height_divisor /= 2;
> > +
>
> This hunk really needs a comment. Please explain why Skylake's MCS is
> twice as tall as previous generations'.
>
To my knowledge we've yet to actually articulate in a comment what the MCS
buffer's layout is (perhaps I just don't see it). Therefore I am uncertain what
comment I could provide here which isn't obvious from the code. I had a
different path to finding this than you did which I believe gives me some bias
(I just needed a reminder that divisor needs to be halved instead of doubled to
obtain twice the height).
If you suggest something, I will happily add it.
> > unsigned mcs_width =
> > ALIGN(mt->logical_width0, width_divisor) / width_divisor;
> > unsigned mcs_height =
> > diff --git a/src/mesa/drivers/dri/i965/intel_mipmap_tree.h b/src/mesa/drivers/dri/i965/intel_mipmap_tree.h
> > index 805cd71..d9cdba0 100644
> > --- a/src/mesa/drivers/dri/i965/intel_mipmap_tree.h
> > +++ b/src/mesa/drivers/dri/i965/intel_mipmap_tree.h
> > @@ -634,14 +634,17 @@ struct intel_mipmap_tree
> > * color mipmap tree, if any.
> > *
> > * This value will only ever contain ones in bits 28-31, so it is safe to
> > - * OR into dword 7 of SURFACE_STATE.
> > + * OR into dword 7 of SURFACE_STATE. (dword 12-15 on SKL)
>
> The comment is wrong. "This value will only ever contain ones in bits
> 28-31" applies only to fast_clear_color_value but not to
> gen9_fast_clear_color. And it doesn't make sense to bit-or
> gen9_fast_color_clear to dwords 12-15, as gen9_fast_color_clear is
> exactly 4 dwords big, and all the bits are significant.
>
How does this sound?
}
diff --git a/src/mesa/drivers/dri/i965/intel_mipmap_tree.h b/src/mesa/drivers/dri/i965/intel_mipmap_tree.h
index d9cdba0..64f73ea 100644
--- a/src/mesa/drivers/dri/i965/intel_mipmap_tree.h
+++ b/src/mesa/drivers/dri/i965/intel_mipmap_tree.h
@@ -633,8 +633,12 @@ struct intel_mipmap_tree
* The SURFACE_STATE bits associated with the last fast color clear to this
* color mipmap tree, if any.
*
- * This value will only ever contain ones in bits 28-31, so it is safe to
- * OR into dword 7 of SURFACE_STATE. (dword 12-15 on SKL)
+ * Prior to GEN9 there is a single bit for RGBA clear values which gives you
+ * the option of 2^4 clear colors. Each bit determines if the color channel
+ * is fully saturated or unsaturated (Cherryview does add a 32b value per
+ * channel, but it is globally applied instead of being part of the render
+ * surface state). Starting with GEN9, the surface state accepts a 32b value
+ * for each color channel.
*
* @see RENDER_SURFACE_STATE.RedClearColor
* @see RENDER_SURFACE_STATE.GreenClearColor
> > *
> > * @see RENDER_SURFACE_STATE.RedClearColor
> > * @see RENDER_SURFACE_STATE.GreenClearColor
> > * @see RENDER_SURFACE_STATE.BlueClearColor
> > * @see RENDER_SURFACE_STATE.AlphaClearColor
> > */
> > - uint32_t fast_clear_color_value;
> > + union {
> > + uint32_t fast_clear_color_value;
> > + union gl_color_union gen9_fast_clear_color;
> > + };
> >
> > /**
> > * Disable allocation of auxiliary buffers, such as the HiZ buffer and MCS
> > --
> > 2.6.1
> >
More information about the mesa-dev
mailing list