[Mesa-dev] [PATCH 8/9] freedreno: implement a20x hw binning
Jonathan Marek
jonathan at marek.ca
Tue Nov 13 20:04:10 UTC 2018
Not in this patch: emitting the hw binning variant and filling the
"draw_patches". That is part of the ir2 patch.
Signed-off-by: Jonathan Marek <jonathan at marek.ca>
---
src/gallium/drivers/freedreno/a2xx/fd2_draw.c | 47 +++++--
src/gallium/drivers/freedreno/a2xx/fd2_emit.c | 8 +-
src/gallium/drivers/freedreno/a2xx/fd2_emit.h | 3 +-
src/gallium/drivers/freedreno/a2xx/fd2_gmem.c | 118 ++++++++++++++++++
.../drivers/freedreno/freedreno_gmem.c | 29 +++--
.../drivers/freedreno/freedreno_gmem.h | 1 +
6 files changed, 187 insertions(+), 19 deletions(-)
diff --git a/src/gallium/drivers/freedreno/a2xx/fd2_draw.c b/src/gallium/drivers/freedreno/a2xx/fd2_draw.c
index 49df1daa59..46c76df807 100644
--- a/src/gallium/drivers/freedreno/a2xx/fd2_draw.c
+++ b/src/gallium/drivers/freedreno/a2xx/fd2_draw.c
@@ -75,11 +75,13 @@ emit_vertexbufs(struct fd_context *ctx)
// CONST(20,0) (or CONST(26,0) in soliv_vp)
fd2_emit_vertex_bufs(ctx->batch->draw, 0x78, bufs, vtx->num_elements);
+ fd2_emit_vertex_bufs(ctx->batch->binning, 0x78, bufs, vtx->num_elements);
}
static void
draw_impl(struct fd_context *ctx, const struct pipe_draw_info *info,
- struct fd_ringbuffer *ring, unsigned index_offset)
+ struct fd_ringbuffer *ring, unsigned index_offset,
+ bool binning)
{
enum pc_di_vis_cull_mode vismode;
@@ -87,9 +89,15 @@ draw_impl(struct fd_context *ctx, const struct pipe_draw_info *info,
OUT_RING(ring, CP_REG(REG_A2XX_VGT_INDX_OFFSET));
OUT_RING(ring, info->index_size ? 0 : info->start);
- OUT_PKT3(ring, CP_SET_CONSTANT, 2);
- OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
- OUT_RING(ring, is_a20x(ctx->screen) ? 0x00000002 : 0x0000003b);
+ /* in the binning batch, this value is set once in fd2_emit_tile_init */
+ if (!binning) {
+ OUT_PKT3(ring, CP_SET_CONSTANT, 2);
+ OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
+ /* XXX do this for every REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL write ?
+ * if set to 0x3b on a20x, clipping is broken
+ */
+ OUT_RING(ring, is_a20x(ctx->screen) ? 0x00000002 : 0x0000003b);
+ }
OUT_PKT0(ring, REG_A2XX_TC_CNTL_STATUS, 1);
OUT_RING(ring, A2XX_TC_CNTL_STATUS_L2_INVALIDATE);
@@ -125,8 +133,26 @@ draw_impl(struct fd_context *ctx, const struct pipe_draw_info *info,
OUT_RING(ring, info->min_index); /* VGT_MIN_VTX_INDX */
}
+ /* binning shader will take offset from C64 */
+ if (binning && is_a20x(ctx->screen)) {
+ OUT_PKT3(ring, CP_SET_CONSTANT, 5);
+ OUT_RING(ring, 0x00000180);
+ OUT_RING(ring, fui(ctx->batch->num_vertices));
+ OUT_RING(ring, fui(0.0f));
+ OUT_RING(ring, fui(0.0f));
+ OUT_RING(ring, fui(0.0f));
+ }
+
+ vismode = binning ? IGNORE_VISIBILITY : USE_VISIBILITY;
+ /* a22x hw binning not implemented */
+ if (binning || !is_a20x(ctx->screen) || (fd_mesa_debug & FD_DBG_NOBIN))
+ vismode = IGNORE_VISIBILITY;
+
+ if (info->mode == PIPE_PRIM_POINTS)
+ vismode = IGNORE_VISIBILITY;
+
fd_draw_emit(ctx->batch, ring, ctx->primtypes[info->mode],
- IGNORE_VISIBILITY, info, index_offset);
+ vismode, info, index_offset);
if (is_a20x(ctx->screen)) {
/* not sure why this is required, but it fixes some hangs */
@@ -152,6 +178,7 @@ fd2_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *pinfo,
emit_vertexbufs(ctx);
fd2_emit_state(ctx, ctx->batch->draw, ctx->dirty);
+ fd2_emit_state(ctx, ctx->batch->binning, ctx->dirty);
/* a20x can draw only 65535 vertices at once
* however, using a limit of 32k fixes an unexplained hang
@@ -171,17 +198,23 @@ fd2_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *pinfo,
struct pipe_draw_info info = *pinfo;
unsigned count = info.count;
unsigned step = step_tbl[info.mode];
+ unsigned num_vertices = ctx->batch->num_vertices;
if (!step)
return false;
for (; count + step > 32766; count -= step) {
info.count = MIN2(count, 32766);
- draw_impl(ctx, &info, ctx->batch->draw, index_offset);
+ draw_impl(ctx, &info, ctx->batch->draw, index_offset, false);
+ draw_impl(ctx, &info, ctx->batch->binning, index_offset, true);
info.start += step;
+ ctx->batch->num_vertices += step;
}
+ /* changing this value is a hack, restore it */
+ ctx->batch->num_vertices = num_vertices;
} else {
- draw_impl(ctx, pinfo, ctx->batch->draw, index_offset);
+ draw_impl(ctx, pinfo, ctx->batch->draw, index_offset, false);
+ draw_impl(ctx, pinfo, ctx->batch->binning, index_offset, true);
}
fd_context_all_clean(ctx);
diff --git a/src/gallium/drivers/freedreno/a2xx/fd2_emit.c b/src/gallium/drivers/freedreno/a2xx/fd2_emit.c
index 60bc9fad4c..4e93cb39b0 100644
--- a/src/gallium/drivers/freedreno/a2xx/fd2_emit.c
+++ b/src/gallium/drivers/freedreno/a2xx/fd2_emit.c
@@ -186,11 +186,12 @@ fd2_emit_vertex_bufs(struct fd_ringbuffer *ring, uint32_t val,
}
void
-fd2_emit_state(struct fd_context *ctx, const enum fd_dirty_3d_state dirty)
+fd2_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
+ const enum fd_dirty_3d_state dirty)
{
struct fd2_blend_stateobj *blend = fd2_blend_stateobj(ctx->blend);
struct fd2_zsa_stateobj *zsa = fd2_zsa_stateobj(ctx->zsa);
- struct fd_ringbuffer *ring = ctx->batch->draw;
+ bool is_binning = (ring == ctx->batch->binning);
/* NOTE: we probably want to eventually refactor this so each state
* object handles emitting it's own state.. although the mapping of
@@ -228,7 +229,8 @@ fd2_emit_state(struct fd_context *ctx, const enum fd_dirty_3d_state dirty)
OUT_RING(ring, CP_REG(REG_A2XX_PA_CL_CLIP_CNTL));
OUT_RING(ring, rasterizer->pa_cl_clip_cntl);
OUT_RING(ring, rasterizer->pa_su_sc_mode_cntl |
- A2XX_PA_SU_SC_MODE_CNTL_VTX_WINDOW_OFFSET_ENABLE);
+ A2XX_PA_SU_SC_MODE_CNTL_VTX_WINDOW_OFFSET_ENABLE |
+ COND(is_binning, A2XX_PA_SU_SC_MODE_CNTL_FACE_KILL_ENABLE));
OUT_PKT3(ring, CP_SET_CONSTANT, 5);
OUT_RING(ring, CP_REG(REG_A2XX_PA_SU_POINT_SIZE));
diff --git a/src/gallium/drivers/freedreno/a2xx/fd2_emit.h b/src/gallium/drivers/freedreno/a2xx/fd2_emit.h
index 5e4bddd1fa..38660e733e 100644
--- a/src/gallium/drivers/freedreno/a2xx/fd2_emit.h
+++ b/src/gallium/drivers/freedreno/a2xx/fd2_emit.h
@@ -40,7 +40,8 @@ struct fd2_vertex_buf {
void fd2_emit_vertex_bufs(struct fd_ringbuffer *ring, uint32_t val,
struct fd2_vertex_buf *vbufs, uint32_t n);
-void fd2_emit_state(struct fd_context *ctx, enum fd_dirty_3d_state dirty);
+void fd2_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
+ const enum fd_dirty_3d_state dirty);
void fd2_emit_restore(struct fd_context *ctx, struct fd_ringbuffer *ring);
void fd2_emit_init(struct pipe_context *pctx);
diff --git a/src/gallium/drivers/freedreno/a2xx/fd2_gmem.c b/src/gallium/drivers/freedreno/a2xx/fd2_gmem.c
index e98ae7334a..b6b7050026 100644
--- a/src/gallium/drivers/freedreno/a2xx/fd2_gmem.c
+++ b/src/gallium/drivers/freedreno/a2xx/fd2_gmem.c
@@ -375,6 +375,107 @@ fd2_emit_tile_init(struct fd_batch *batch)
if (pfb->zsbuf)
reg |= A2XX_RB_DEPTH_INFO_DEPTH_FORMAT(fd_pipe2depth(pfb->zsbuf->format));
OUT_RING(ring, reg); /* RB_DEPTH_INFO */
+
+ /* set to zero, for some reason hardware doesn't certain values */
+ OUT_PKT3(ring, CP_SET_CONSTANT, 2);
+ OUT_RING(ring, CP_REG(REG_A2XX_VGT_CURRENT_BIN_ID_MIN));
+ OUT_RING(ring, 0);
+
+ OUT_PKT3(ring, CP_SET_CONSTANT, 2);
+ OUT_RING(ring, CP_REG(REG_A2XX_VGT_CURRENT_BIN_ID_MAX));
+ OUT_RING(ring, 0);
+
+ if (is_a20x(ctx->screen) && !(fd_mesa_debug & FD_DBG_NOBIN) &&
+ gmem->num_vsc_pipes) {
+ /* patch out unneeded memory exports by changing EXEC CF to EXEC_END
+ *
+ * in the shader compiler, we guarantee that the shader ends with
+ * a specific pattern of ALLOC/EXEC CF pairs for the hw binning exports
+ *
+ * the since patches point only to dwords and CFs are 1.5 dwords
+ * the patch is aligned and might point to a ALLOC CF
+ */
+ for (int i = 0; i < fd_patch_num_elements(&batch->draw_patches); i++) {
+ struct fd_cs_patch *patch =
+ fd_patch_element(&batch->draw_patches, i);
+ *patch->cs = patch->val;
+
+ instr_cf_t *cf = (instr_cf_t*) patch->cs;
+ if (cf->opc == ALLOC)
+ cf++;
+ assert(cf->opc == EXEC);
+ assert(cf[ctx->screen->num_vsc_pipes*2-2].opc == EXEC_END);
+ cf[2*(gmem->num_vsc_pipes-1)].opc = EXEC_END;
+ }
+
+ /* initialize shader constants for the binning memexport */
+ OUT_PKT3(ring, CP_SET_CONSTANT, 1 + gmem->num_vsc_pipes * 4);
+ OUT_RING(ring, 0x0000000C);
+
+ for (int i = 0; i < gmem->num_vsc_pipes; i++) {
+ struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i];
+
+ /* XXX we know how large this needs to be..
+ * should do some sort of realloc
+ * it should be ctx->batch->num_vertices bytes large
+ * with this size it will break with more than 256k vertices..
+ */
+ if (!pipe->bo) {
+ pipe->bo = fd_bo_new(ctx->dev, 0x40000,
+ DRM_FREEDRENO_GEM_TYPE_KMEM);
+ }
+
+ /* memory export address (export32):
+ * .x: (base_address >> 2) | 0x40000000 (?)
+ * .y: index (float) - set by shader
+ * .z: 0x4B00D000 (?)
+ * .w: 0x4B000000 (?) | max_index (?)
+ */
+ OUT_RELOCW(ring, pipe->bo, 0, 0x40000000, -2);
+ OUT_RING(ring, 0x00000000);
+ OUT_RING(ring, 0x4B00D000);
+ OUT_RING(ring, 0x4B000000 | 0x40000);
+ }
+
+ OUT_PKT3(ring, CP_SET_CONSTANT, 1 + gmem->num_vsc_pipes * 8);
+ OUT_RING(ring, 0x0000018C);
+
+ for (int i = 0; i < gmem->num_vsc_pipes; i++) {
+ struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i];
+ float off_x, off_y, mul_x, mul_y;
+
+ /* const to tranform from [-1,1] to bin coordinates for this pipe
+ * for x/y, [0,256/2040] = 0, [256/2040,512/2040] = 1, etc
+ * 8 possible values on x/y axis,
+ * to clip at binning stage: only use center 6x6
+ * TODO: set the z parameters too so that hw binning
+ * can clip primitives in Z too
+ */
+
+ mul_x = 1.0f / (float) (gmem->bin_w * 8);
+ mul_y = 1.0f / (float) (gmem->bin_h * 8);
+ off_x = -pipe->x * (1.0/8.0f) + 0.125f;
+ off_y = -pipe->y * (1.0/8.0f) + 0.125f;
+
+ OUT_RING(ring, fui(off_x * (256.0f/255.0f)));
+ OUT_RING(ring, fui(off_y * (256.0f/255.0f)));
+ OUT_RING(ring, 0x3f000000);
+ OUT_RING(ring, fui(0.0f));
+
+ OUT_RING(ring, fui(mul_x * (256.0f/255.0f)));
+ OUT_RING(ring, fui(mul_y * (256.0f/255.0f)));
+ OUT_RING(ring, fui(0.0f));
+ OUT_RING(ring, fui(0.0f));
+ }
+
+ OUT_PKT3(ring, CP_SET_CONSTANT, 2);
+ OUT_RING(ring, CP_REG(REG_A2XX_VGT_VERTEX_REUSE_BLOCK_CNTL));
+ OUT_RING(ring, 0);
+
+ ctx->emit_ib(ring, batch->binning);
+ }
+
+ util_dynarray_resize(&batch->draw_patches, 0);
}
/* before mem2gmem */
@@ -403,6 +504,7 @@ fd2_emit_tile_prep(struct fd_batch *batch, struct fd_tile *tile)
static void
fd2_emit_tile_renderprep(struct fd_batch *batch, struct fd_tile *tile)
{
+ struct fd_context *ctx = batch->ctx;
struct fd_ringbuffer *ring = batch->gmem;
struct pipe_framebuffer_state *pfb = &batch->framebuffer;
enum pipe_format format = pipe_surface_format(pfb->cbufs[0]);
@@ -419,6 +521,22 @@ fd2_emit_tile_renderprep(struct fd_batch *batch, struct fd_tile *tile)
OUT_RING(ring, CP_REG(REG_A2XX_PA_SC_WINDOW_OFFSET));
OUT_RING(ring, A2XX_PA_SC_WINDOW_OFFSET_X(-tile->xoff) |
A2XX_PA_SC_WINDOW_OFFSET_Y(-tile->yoff));
+
+ if (is_a20x(ctx->screen) && !(fd_mesa_debug & FD_DBG_NOBIN)) {
+ struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[tile->p];
+
+ OUT_PKT3(ring, CP_SET_CONSTANT, 2);
+ OUT_RING(ring, CP_REG(REG_A2XX_VGT_CURRENT_BIN_ID_MIN));
+ OUT_RING(ring, tile->n);
+
+ OUT_PKT3(ring, CP_SET_CONSTANT, 2);
+ OUT_RING(ring, CP_REG(REG_A2XX_VGT_CURRENT_BIN_ID_MAX));
+ OUT_RING(ring, tile->n);
+
+ /* TODO only emit this when tile->p changes */
+ OUT_PKT3(ring, CP_SET_DRAW_INIT_FLAGS, 1);
+ OUT_RELOC(ring, pipe->bo, 0, 0, 0);
+ }
}
void
diff --git a/src/gallium/drivers/freedreno/freedreno_gmem.c b/src/gallium/drivers/freedreno/freedreno_gmem.c
index bb15f0a3e1..d2483de1b0 100644
--- a/src/gallium/drivers/freedreno/freedreno_gmem.c
+++ b/src/gallium/drivers/freedreno/freedreno_gmem.c
@@ -214,12 +214,21 @@ calculate_tiles(struct fd_batch *batch)
#define div_round_up(v, a) (((v) + (a) - 1) / (a))
/* figure out number of tiles per pipe: */
- tpp_x = tpp_y = 1;
- while (div_round_up(nbins_y, tpp_y) > screen->num_vsc_pipes)
- tpp_y += 2;
- while ((div_round_up(nbins_y, tpp_y) *
- div_round_up(nbins_x, tpp_x)) > screen->num_vsc_pipes)
- tpp_x += 1;
+ if (is_a20x(ctx->screen)) {
+ /* for a20x we want to minimize the number of "pipes"
+ * binning data has 3 bits for x/y (8x8) but the edges are used to
+ * cull off-screen vertices with hw binning, so we have 6x6 pipes
+ */
+ tpp_x = 6;
+ tpp_y = 6;
+ } else {
+ tpp_x = tpp_y = 1;
+ while (div_round_up(nbins_y, tpp_y) > screen->num_vsc_pipes)
+ tpp_y += 2;
+ while ((div_round_up(nbins_y, tpp_y) *
+ div_round_up(nbins_x, tpp_x)) > screen->num_vsc_pipes)
+ tpp_x += 1;
+ }
gmem->maxpw = tpp_x;
gmem->maxph = tpp_y;
@@ -246,6 +255,9 @@ calculate_tiles(struct fd_batch *batch)
xoff += tpp_x;
}
+ /* number of pipes to use for a20x */
+ gmem->num_vsc_pipes = MAX2(1, i);
+
for (; i < npipes; i++) {
struct fd_vsc_pipe *pipe = &ctx->vsc_pipe[i];
pipe->x = pipe->y = pipe->w = pipe->h = 0;
@@ -280,11 +292,12 @@ calculate_tiles(struct fd_batch *batch)
/* pipe number: */
p = ((i / tpp_y) * div_round_up(nbins_x, tpp_x)) + (j / tpp_x);
+ assert(p < gmem->num_vsc_pipes);
/* clip bin width: */
bw = MIN2(bin_w, minx + width - xoff);
-
- tile->n = tile_n[p]++;
+ tile->n = !is_a20x(ctx->screen) ? tile_n[p]++ :
+ ((i % tpp_y + 1) << 3 | (j % tpp_x + 1));
tile->p = p;
tile->bin_w = bw;
tile->bin_h = bh;
diff --git a/src/gallium/drivers/freedreno/freedreno_gmem.h b/src/gallium/drivers/freedreno/freedreno_gmem.h
index b953999ff9..70641d62f3 100644
--- a/src/gallium/drivers/freedreno/freedreno_gmem.h
+++ b/src/gallium/drivers/freedreno/freedreno_gmem.h
@@ -57,6 +57,7 @@ struct fd_gmem_stateobj {
uint16_t minx, miny;
uint16_t width, height;
uint16_t maxpw, maxph; /* maximum pipe width/height */
+ uint8_t num_vsc_pipes; /* number of pipes for a20x */
};
struct fd_batch;
--
2.17.1
More information about the mesa-dev
mailing list