[Mesa-dev] [PATCH 15/15] radeonsi: use higher subpixel precision (QUANT_MODE) for smaller viewports
Marek Olšák
maraeo at gmail.com
Tue Oct 2 22:35:47 UTC 2018
From: Marek Olšák <marek.olsak at amd.com>
---
src/gallium/drivers/radeonsi/si_get.c | 4 +-
src/gallium/drivers/radeonsi/si_pipe.h | 8 +++
.../drivers/radeonsi/si_state_viewport.c | 50 ++++++++++++++++---
3 files changed, 53 insertions(+), 9 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_get.c b/src/gallium/drivers/radeonsi/si_get.c
index a87cb3cbc8a..ac302b8a946 100644
--- a/src/gallium/drivers/radeonsi/si_get.c
+++ b/src/gallium/drivers/radeonsi/si_get.c
@@ -328,21 +328,23 @@ static int si_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
}
}
static float si_get_paramf(struct pipe_screen* pscreen, enum pipe_capf param)
{
switch (param) {
case PIPE_CAPF_MAX_LINE_WIDTH:
case PIPE_CAPF_MAX_LINE_WIDTH_AA:
case PIPE_CAPF_MAX_POINT_WIDTH:
case PIPE_CAPF_MAX_POINT_WIDTH_AA:
- return 8192.0f;
+ /* This depends on the quant mode, though the precise interactions
+ * are unknown. */
+ return 2048;
case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
return 16.0f;
case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
return 16.0f;
case PIPE_CAPF_MIN_CONSERVATIVE_RASTER_DILATE:
case PIPE_CAPF_MAX_CONSERVATIVE_RASTER_DILATE:
case PIPE_CAPF_CONSERVATIVE_RASTER_DILATE_GRANULARITY:
return 0.0f;
}
return 0.0f;
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 93082e262d6..7e15412ef87 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -596,25 +596,33 @@ struct si_framebuffer {
ubyte uncompressed_cb_mask;
ubyte color_is_int8;
ubyte color_is_int10;
ubyte dirty_cbufs;
bool dirty_zsbuf;
bool any_dst_linear;
bool CB_has_shader_readable_metadata;
bool DB_has_shader_readable_metadata;
};
+enum si_quant_mode {
+ /* This is the list we want to support. */
+ SI_QUANT_MODE_16_8_FIXED_POINT_1_256TH,
+ SI_QUANT_MODE_14_10_FIXED_POINT_1_1024TH,
+ SI_QUANT_MODE_12_12_FIXED_POINT_1_4096TH,
+};
+
struct si_signed_scissor {
int minx;
int miny;
int maxx;
int maxy;
+ enum si_quant_mode quant_mode;
};
struct si_scissors {
unsigned dirty_mask;
struct pipe_scissor_state states[SI_MAX_VIEWPORTS];
};
struct si_viewports {
unsigned dirty_mask;
unsigned depth_range_dirty_mask;
diff --git a/src/gallium/drivers/radeonsi/si_state_viewport.c b/src/gallium/drivers/radeonsi/si_state_viewport.c
index c69a56dffae..819c773ba8e 100644
--- a/src/gallium/drivers/radeonsi/si_state_viewport.c
+++ b/src/gallium/drivers/radeonsi/si_state_viewport.c
@@ -100,20 +100,21 @@ static void si_clip_scissor(struct pipe_scissor_state *out,
out->maxy = MIN2(out->maxy, clip->maxy);
}
static void si_scissor_make_union(struct si_signed_scissor *out,
struct si_signed_scissor *in)
{
out->minx = MIN2(out->minx, in->minx);
out->miny = MIN2(out->miny, in->miny);
out->maxx = MAX2(out->maxx, in->maxx);
out->maxy = MAX2(out->maxy, in->maxy);
+ out->quant_mode = MIN2(out->quant_mode, in->quant_mode);
}
static void si_emit_one_scissor(struct si_context *ctx,
struct radeon_cmdbuf *cs,
struct si_signed_scissor *vp_scissor,
struct pipe_scissor_state *scissor)
{
struct pipe_scissor_state final;
if (ctx->vs_disables_clipping_viewport) {
@@ -138,43 +139,47 @@ static void si_emit_one_scissor(struct si_context *ctx,
return;
}
radeon_emit(cs, S_028250_TL_X(final.minx) |
S_028250_TL_Y(final.miny) |
S_028250_WINDOW_OFFSET_DISABLE(1));
radeon_emit(cs, S_028254_BR_X(final.maxx) |
S_028254_BR_Y(final.maxy));
}
-/* the range is [-MAX, MAX] */
-#define SI_MAX_VIEWPORT_RANGE 32768
-
static void si_emit_guardband(struct si_context *ctx)
{
const struct si_state_rasterizer *rs = ctx->queued.named.rasterizer;
struct si_signed_scissor vp_as_scissor;
struct pipe_viewport_state vp;
float left, top, right, bottom, max_range, guardband_x, guardband_y;
float discard_x, discard_y;
if (ctx->vs_writes_viewport_index) {
/* Shaders can draw to any viewport. Make a union of all
* viewports. */
vp_as_scissor = ctx->viewports.as_scissor[0];
for (unsigned i = 1; i < SI_MAX_VIEWPORTS; i++) {
si_scissor_make_union(&vp_as_scissor,
&ctx->viewports.as_scissor[i]);
}
} else {
vp_as_scissor = ctx->viewports.as_scissor[0];
}
+ /* Blits don't set the viewport state. The vertex shader determines
+ * the viewport size by scaling the coordinates, so we don't know
+ * how large the viewport is. Assume the worst case.
+ */
+ if (ctx->vs_disables_clipping_viewport)
+ vp_as_scissor.quant_mode = SI_QUANT_MODE_16_8_FIXED_POINT_1_256TH;
+
/* Determine the optimal hardware screen offset to center the viewport
* within the viewport range in order to maximize the guardband size.
*/
int hw_screen_offset_x = (vp_as_scissor.maxx - vp_as_scissor.minx) / 2;
int hw_screen_offset_y = (vp_as_scissor.maxy - vp_as_scissor.miny) / 2;
const unsigned hw_screen_offset_max = 8176;
/* SI-CI need to align the offset to an ubertile consisting of all SEs. */
const unsigned hw_screen_offset_alignment =
ctx->chip_class >= VI ? 16 : MAX2(ctx->screen->se_tile_repeat, 16);
@@ -204,23 +209,25 @@ static void si_emit_guardband(struct si_context *ctx)
if (vp_as_scissor.miny == vp_as_scissor.maxy)
vp.scale[1] = 0.5;
/* Find the biggest guard band that is inside the supported viewport
* range. The guard band is specified as a horizontal and vertical
* distance from (0,0) in clip space.
*
* This is done by applying the inverse viewport transformation
* on the viewport limits to get those limits in clip space.
*
- * Use a limit one pixel smaller to allow for some precision error.
+ * The viewport range is [-max_viewport_size/2, max_viewport_size/2].
*/
- max_range = SI_MAX_VIEWPORT_RANGE - 1;
+ static unsigned max_viewport_size[] = {65535, 16383, 4095};
+ assert(vp_as_scissor.quant_mode < ARRAY_SIZE(max_viewport_size));
+ max_range = max_viewport_size[vp_as_scissor.quant_mode] / 2;
left = (-max_range - vp.translate[0]) / vp.scale[0];
right = ( max_range - vp.translate[0]) / vp.scale[0];
top = (-max_range - vp.translate[1]) / vp.scale[1];
bottom = ( max_range - vp.translate[1]) / vp.scale[1];
assert(left <= -1 && top <= -1 && right >= 1 && bottom >= 1);
guardband_x = MIN2(-left, right);
guardband_y = MIN2(-top, bottom);
@@ -255,21 +262,22 @@ static void si_emit_guardband(struct si_context *ctx)
SI_TRACKED_PA_CL_GB_VERT_CLIP_ADJ,
fui(guardband_y), fui(discard_y),
fui(guardband_x), fui(discard_x));
radeon_opt_set_context_reg(ctx, R_028234_PA_SU_HARDWARE_SCREEN_OFFSET,
SI_TRACKED_PA_SU_HARDWARE_SCREEN_OFFSET,
S_028234_HW_SCREEN_OFFSET_X(hw_screen_offset_x >> 4) |
S_028234_HW_SCREEN_OFFSET_Y(hw_screen_offset_y >> 4));
radeon_opt_set_context_reg(ctx, R_028BE4_PA_SU_VTX_CNTL,
SI_TRACKED_PA_SU_VTX_CNTL,
S_028BE4_PIX_CENTER(rs->half_pixel_center) |
- S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH));
+ S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH +
+ vp_as_scissor.quant_mode));
}
static void si_emit_scissors(struct si_context *ctx)
{
struct radeon_cmdbuf *cs = ctx->gfx_cs;
struct pipe_scissor_state *states = ctx->scissors.states;
unsigned mask = ctx->scissors.dirty_mask;
bool scissor_enabled = ctx->queued.named.rasterizer->scissor_enable;
/* The simple case: Only 1 viewport is active. */
@@ -304,24 +312,47 @@ static void si_set_viewport_states(struct pipe_context *pctx,
unsigned start_slot,
unsigned num_viewports,
const struct pipe_viewport_state *state)
{
struct si_context *ctx = (struct si_context *)pctx;
unsigned mask;
int i;
for (i = 0; i < num_viewports; i++) {
unsigned index = start_slot + i;
+ struct si_signed_scissor *scissor = &ctx->viewports.as_scissor[index];
ctx->viewports.states[index] = state[i];
- si_get_scissor_from_viewport(ctx, &state[i],
- &ctx->viewports.as_scissor[index]);
+
+ si_get_scissor_from_viewport(ctx, &state[i], scissor);
+
+ unsigned w = scissor->maxx - scissor->minx;
+ unsigned h = scissor->maxy - scissor->miny;
+ unsigned max_extent = MAX2(w, h);
+
+ /* Determine the best quantization mode (subpixel precision),
+ * but also leave enough space for the guardband.
+ *
+ * Note that primitive binning requires QUANT_MODE == 16_8 on Vega10
+ * and Raven1. What we do depends on the chip:
+ * - Vega10: Never use primitive binning.
+ * - Raven1: Always use QUANT_MODE == 16_8.
+ */
+ if (ctx->family == CHIP_RAVEN)
+ max_extent = 16384; /* Use QUANT_MODE == 16_8. */
+
+ if (max_extent <= 1024) /* 4K scanline area for guardband */
+ scissor->quant_mode = SI_QUANT_MODE_12_12_FIXED_POINT_1_4096TH;
+ else if (max_extent <= 4096) /* 16K scanline area for guardband */
+ scissor->quant_mode = SI_QUANT_MODE_14_10_FIXED_POINT_1_1024TH;
+ else /* 64K scanline area for guardband */
+ scissor->quant_mode = SI_QUANT_MODE_16_8_FIXED_POINT_1_256TH;
}
mask = ((1 << num_viewports) - 1) << start_slot;
ctx->viewports.dirty_mask |= mask;
ctx->viewports.depth_range_dirty_mask |= mask;
ctx->scissors.dirty_mask |= mask;
si_mark_atom_dirty(ctx, &ctx->atoms.s.viewports);
si_mark_atom_dirty(ctx, &ctx->atoms.s.guardband);
si_mark_atom_dirty(ctx, &ctx->atoms.s.scissors);
}
@@ -558,11 +589,14 @@ static void si_set_window_rectangles(struct pipe_context *ctx,
void si_init_viewport_functions(struct si_context *ctx)
{
ctx->atoms.s.guardband.emit = si_emit_guardband;
ctx->atoms.s.scissors.emit = si_emit_scissors;
ctx->atoms.s.viewports.emit = si_emit_viewport_states;
ctx->atoms.s.window_rectangles.emit = si_emit_window_rectangles;
ctx->b.set_scissor_states = si_set_scissor_states;
ctx->b.set_viewport_states = si_set_viewport_states;
ctx->b.set_window_rectangles = si_set_window_rectangles;
+
+ for (unsigned i = 0; i < 16; i++)
+ ctx->viewports.as_scissor[i].quant_mode = SI_QUANT_MODE_16_8_FIXED_POINT_1_256TH;
}
--
2.17.1
More information about the mesa-dev
mailing list