Mesa (master): freedreno/a5xx: transform-feedback support
Rob Clark
robclark at kemper.freedesktop.org
Tue Dec 27 22:03:30 UTC 2016
Module: Mesa
Branch: master
Commit: d10c5a24818b39585acfa60ceb6dbbae22b5be09
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=d10c5a24818b39585acfa60ceb6dbbae22b5be09
Author: Rob Clark <robdclark at gmail.com>
Date: Sat Dec 24 14:44:26 2016 -0500
freedreno/a5xx: transform-feedback support
We'll need to revisit when adding hw binning pass support, whether we
can still do this in main draw step, as we do w/ a3xx/a4xx, or if we
needed to move it to the binning stage.
Still some failing piglits but most tests pass and the common cases seem
to work.
Signed-off-by: Rob Clark <robdclark at gmail.com>
---
src/gallium/drivers/freedreno/a5xx/fd5_draw.c | 11 ++
src/gallium/drivers/freedreno/a5xx/fd5_emit.c | 56 ++++++-
src/gallium/drivers/freedreno/a5xx/fd5_emit.h | 2 +
src/gallium/drivers/freedreno/a5xx/fd5_program.c | 167 +++++++++++++++++----
src/gallium/drivers/freedreno/a5xx/fd5_program.h | 3 +-
.../drivers/freedreno/ir3/ir3_compiler_nir.c | 6 +-
src/gallium/drivers/freedreno/ir3/ir3_shader.c | 2 +
7 files changed, 209 insertions(+), 38 deletions(-)
diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_draw.c b/src/gallium/drivers/freedreno/a5xx/fd5_draw.c
index e6b42bf..ffd3e3b 100644
--- a/src/gallium/drivers/freedreno/a5xx/fd5_draw.c
+++ b/src/gallium/drivers/freedreno/a5xx/fd5_draw.c
@@ -173,6 +173,17 @@ fd5_draw_vbo(struct fd_context *ctx, const struct pipe_draw_info *info)
// emit.fp = NULL;
// draw_impl(ctx, ctx->batch->binning, &emit);
+ if (emit.streamout_mask) {
+ struct fd_ringbuffer *ring = ctx->batch->draw;
+
+ for (unsigned i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
+ if (emit.streamout_mask & (1 << i)) {
+ OUT_PKT7(ring, CP_EVENT_WRITE, 1);
+ OUT_RING(ring, FLUSH_SO_0 + i);
+ }
+ }
+ }
+
return true;
}
diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_emit.c b/src/gallium/drivers/freedreno/a5xx/fd5_emit.c
index 756ffb7..2404389 100644
--- a/src/gallium/drivers/freedreno/a5xx/fd5_emit.c
+++ b/src/gallium/drivers/freedreno/a5xx/fd5_emit.c
@@ -526,19 +526,63 @@ fd5_emit_state(struct fd_context *ctx, struct fd_ringbuffer *ring,
OUT_RING(ring, A5XX_GRAS_CL_VPORT_ZSCALE_0(ctx->viewport.scale[2]));
}
- if (dirty & (FD_DIRTY_PROG | FD_DIRTY_FRAMEBUFFER)) {
+ if (dirty & FD_DIRTY_PROG)
+ fd5_program_emit(ring, emit);
+
+ if (dirty & (FD_DIRTY_FRAMEBUFFER | FD_DIRTY_RASTERIZER)) {
struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
- unsigned n = pfb->nr_cbufs;
- /* if we have depth/stencil, we need at least on MRT: */
- if (pfb->zsbuf)
- n = MAX2(1, n);
- fd5_program_emit(ring, emit, n, pfb->cbufs);
+ uint32_t posz_regid = ir3_find_output_regid(fp, FRAG_RESULT_DEPTH);
+ unsigned nr = pfb->nr_cbufs;
+
+ if (emit->key.binning_pass)
+ nr = 0;
+ else if (ctx->rasterizer->rasterizer_discard)
+ nr = 0;
+
+ OUT_PKT4(ring, REG_A5XX_RB_FS_OUTPUT_CNTL, 1);
+ OUT_RING(ring, A5XX_RB_FS_OUTPUT_CNTL_MRT(nr) |
+ COND(fp->writes_pos, A5XX_RB_FS_OUTPUT_CNTL_FRAG_WRITES_Z));
+
+ OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_CNTL, 1);
+ OUT_RING(ring, A5XX_SP_FS_OUTPUT_CNTL_MRT(nr) |
+ A5XX_SP_FS_OUTPUT_CNTL_DEPTH_REGID(posz_regid) |
+ A5XX_SP_FS_OUTPUT_CNTL_SAMPLEMASK_REGID(regid(63, 0)));
}
if (emit->prog == &ctx->prog) { /* evil hack to deal sanely with clear path */
ir3_emit_consts(vp, ring, ctx, emit->info, dirty);
if (!emit->key.binning_pass)
ir3_emit_consts(fp, ring, ctx, emit->info, dirty);
+
+ struct pipe_stream_output_info *info = &vp->shader->stream_output;
+ if (info->num_outputs) {
+ struct fd_streamout_stateobj *so = &ctx->streamout;
+
+ for (unsigned i = 0; i < so->num_targets; i++) {
+ struct pipe_stream_output_target *target = so->targets[i];
+
+ if (!target)
+ continue;
+
+ unsigned offset = (so->offsets[i] * info->stride[i] * 4) +
+ target->buffer_offset;
+
+ OUT_PKT4(ring, REG_A5XX_VPC_SO_BUFFER_BASE_LO(i), 3);
+ /* VPC_SO[i].BUFFER_BASE_LO: */
+ OUT_RELOCW(ring, fd_resource(target->buffer)->bo, 0, 0, 0);
+ OUT_RING(ring, target->buffer_size + offset);
+
+ OUT_PKT4(ring, REG_A5XX_VPC_SO_BUFFER_OFFSET(i), 3);
+ OUT_RING(ring, offset);
+ /* VPC_SO[i].FLUSH_BASE_LO/HI: */
+ // TODO just give hw a dummy addr for now.. we should
+ // be using this an then CP_MEM_TO_REG to set the
+ // VPC_SO[i].BUFFER_OFFSET for the next draw..
+ OUT_RELOCW(ring, fd5_context(ctx)->blit_mem, 0x100, 0, 0);
+
+ emit->streamout_mask |= (1 << i);
+ }
+ }
}
if ((dirty & FD_DIRTY_BLEND)) {
diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_emit.h b/src/gallium/drivers/freedreno/a5xx/fd5_emit.h
index 2c6b717..0525b3e 100644
--- a/src/gallium/drivers/freedreno/a5xx/fd5_emit.h
+++ b/src/gallium/drivers/freedreno/a5xx/fd5_emit.h
@@ -54,6 +54,8 @@ struct fd5_emit {
/* cached to avoid repeated lookups of same variants: */
const struct ir3_shader_variant *vp, *fp;
/* TODO: other shader stages.. */
+
+ unsigned streamout_mask;
};
static inline enum a5xx_color_fmt fd5_emit_format(struct pipe_surface *surf)
diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_program.c b/src/gallium/drivers/freedreno/a5xx/fd5_program.c
index 27e4e43..983d1bc 100644
--- a/src/gallium/drivers/freedreno/a5xx/fd5_program.c
+++ b/src/gallium/drivers/freedreno/a5xx/fd5_program.c
@@ -132,6 +132,122 @@ emit_shader(struct fd_ringbuffer *ring, const struct ir3_shader_variant *so)
}
}
+/* Add any missing varyings needed for stream-out. Otherwise varyings not
+ * used by fragment shader will be stripped out.
+ */
+static void
+link_stream_out(struct ir3_shader_linkage *l, const struct ir3_shader_variant *v)
+{
+ const struct pipe_stream_output_info *strmout = &v->shader->stream_output;
+
+ /*
+ * First, any stream-out varyings not already in linkage map (ie. also
+ * consumed by frag shader) need to be added:
+ */
+ for (unsigned i = 0; i < strmout->num_outputs; i++) {
+ const struct pipe_stream_output *out = &strmout->output[i];
+ unsigned k = out->register_index;
+ unsigned compmask =
+ (1 << (out->num_components + out->start_component)) - 1;
+ unsigned idx, nextloc = 0;
+
+ /* psize/pos need to be the last entries in linkage map, and will
+ * get added link_stream_out, so skip over them:
+ */
+ if ((v->outputs[k].slot == VARYING_SLOT_PSIZ) ||
+ (v->outputs[k].slot == VARYING_SLOT_POS))
+ continue;
+
+ for (idx = 0; idx < l->cnt; idx++) {
+ if (l->var[idx].regid == v->outputs[k].regid)
+ break;
+ nextloc = MAX2(nextloc, l->var[idx].loc + 4);
+ }
+
+ /* add if not already in linkage map: */
+ if (idx == l->cnt)
+ ir3_link_add(l, v->outputs[k].regid, compmask, nextloc);
+
+ /* expand component-mask if needed, ie streaming out all components
+ * but frag shader doesn't consume all components:
+ */
+ if (compmask & ~l->var[idx].compmask) {
+ l->var[idx].compmask |= compmask;
+ l->max_loc = MAX2(l->max_loc,
+ l->var[idx].loc + util_last_bit(l->var[idx].compmask));
+ }
+ }
+}
+
+/* TODO maybe some of this we could pre-compute once rather than having
+ * so much draw-time logic?
+ */
+static void
+emit_stream_out(struct fd_ringbuffer *ring, const struct ir3_shader_variant *v,
+ struct ir3_shader_linkage *l)
+{
+ const struct pipe_stream_output_info *strmout = &v->shader->stream_output;
+ unsigned ncomp[PIPE_MAX_SO_BUFFERS] = {0};
+ unsigned prog[align(l->max_loc, 2) / 2];
+
+ memset(prog, 0, sizeof(prog));
+
+ for (unsigned i = 0; i < strmout->num_outputs; i++) {
+ const struct pipe_stream_output *out = &strmout->output[i];
+ unsigned k = out->register_index;
+ unsigned idx;
+
+ ncomp[out->output_buffer] += out->num_components;
+
+ /* linkage map sorted by order frag shader wants things, so
+ * a bit less ideal here..
+ */
+ for (idx = 0; idx < l->cnt; idx++)
+ if (l->var[idx].regid == v->outputs[k].regid)
+ break;
+
+ debug_assert(idx < l->cnt);
+
+ for (unsigned j = 0; j < out->num_components; j++) {
+ unsigned c = j + out->start_component;
+ unsigned loc = l->var[idx].loc + c;
+ unsigned off = j + out->dst_offset; /* in dwords */
+
+ if (loc & 1) {
+ prog[loc/2] |= A5XX_VPC_SO_PROG_B_EN |
+ A5XX_VPC_SO_PROG_B_BUF(out->output_buffer) |
+ A5XX_VPC_SO_PROG_B_OFF(off * 4);
+ } else {
+ prog[loc/2] |= A5XX_VPC_SO_PROG_A_EN |
+ A5XX_VPC_SO_PROG_A_BUF(out->output_buffer) |
+ A5XX_VPC_SO_PROG_A_OFF(off * 4);
+ }
+ }
+ }
+
+ OUT_PKT7(ring, CP_CONTEXT_REG_BUNCH, 12 + (2 * ARRAY_SIZE(prog)));
+ OUT_RING(ring, REG_A5XX_VPC_SO_BUF_CNTL);
+ OUT_RING(ring, A5XX_VPC_SO_BUF_CNTL_ENABLE |
+ COND(ncomp[0] > 0, A5XX_VPC_SO_BUF_CNTL_BUF0) |
+ COND(ncomp[1] > 0, A5XX_VPC_SO_BUF_CNTL_BUF1) |
+ COND(ncomp[2] > 0, A5XX_VPC_SO_BUF_CNTL_BUF2) |
+ COND(ncomp[3] > 0, A5XX_VPC_SO_BUF_CNTL_BUF3));
+ OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(0));
+ OUT_RING(ring, ncomp[0]);
+ OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(1));
+ OUT_RING(ring, ncomp[1]);
+ OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(2));
+ OUT_RING(ring, ncomp[2]);
+ OUT_RING(ring, REG_A5XX_VPC_SO_NCOMP(3));
+ OUT_RING(ring, ncomp[3]);
+ OUT_RING(ring, REG_A5XX_VPC_SO_CNTL);
+ OUT_RING(ring, A5XX_VPC_SO_CNTL_ENABLE);
+ for (unsigned i = 0; i < ARRAY_SIZE(prog); i++) {
+ OUT_RING(ring, REG_A5XX_VPC_SO_PROG);
+ OUT_RING(ring, prog[i]);
+ }
+}
+
struct stage {
const struct ir3_shader_variant *v;
const struct ir3_info *i;
@@ -214,24 +330,17 @@ setup_stages(struct fd5_emit *emit, struct stage *s)
}
void
-fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
- int nr, struct pipe_surface **bufs)
+fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit)
{
struct stage s[MAX_STAGES];
- uint32_t pos_regid, posz_regid, psize_regid, color_regid[8];
+ uint32_t pos_regid, psize_regid, color_regid[8];
uint32_t face_regid, coord_regid, zwcoord_regid;
uint32_t vcoord_regid, vertex_regid, instance_regid;
int i, j;
- debug_assert(nr <= ARRAY_SIZE(color_regid));
-
- if (emit->key.binning_pass)
- nr = 0;
-
setup_stages(emit, s);
pos_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_POS);
- posz_regid = ir3_find_output_regid(s[FS].v, FRAG_RESULT_DEPTH);
psize_regid = ir3_find_output_regid(s[VS].v, VARYING_SLOT_PSIZ);
vertex_regid = ir3_find_output_regid(s[VS].v, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
instance_regid = ir3_find_output_regid(s[VS].v, SYSTEM_VALUE_INSTANCE_ID);
@@ -342,6 +451,10 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
struct ir3_shader_linkage l = {0};
ir3_link_shaders(&l, s[VS].v, s[FS].v);
+ if ((s[VS].v->shader->stream_output.num_outputs > 0) &&
+ !emit->key.binning_pass)
+ link_stream_out(&l, s[VS].v);
+
BITSET_DECLARE(varbs, 128) = {0};
uint32_t *varmask = (uint32_t *)varbs;
@@ -362,6 +475,17 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
if (psize_regid != regid(63,0))
ir3_link_add(&l, psize_regid, 0x1, l.max_loc);
+ if ((s[VS].v->shader->stream_output.num_outputs > 0) &&
+ !emit->key.binning_pass) {
+ emit_stream_out(ring, s[VS].v, &l);
+
+ OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1);
+ OUT_RING(ring, 0x00000000);
+ } else {
+ OUT_PKT4(ring, REG_A5XX_VPC_SO_OVERRIDE, 1);
+ OUT_RING(ring, A5XX_VPC_SO_OVERRIDE_SO_DISABLE);
+ }
+
for (i = 0, j = 0; (i < 16) && (j < l.cnt); i++) {
uint32_t reg = 0;
@@ -406,11 +530,6 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
OUT_RING(ring, 0x00000000); /* SP_FS_OBJ_START_LO */
OUT_RING(ring, 0x00000000); /* SP_FS_OBJ_START_HI */
} else {
- uint32_t stride_in_vpc = align(s[FS].v->total_in, 4) + 4;
-
- if (s[VS].v->writes_psize)
- stride_in_vpc++;
-
// TODO if some of these other bits depend on something other than
// program state we should probably move these next three regs:
@@ -418,12 +537,12 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
OUT_RING(ring, A5XX_SP_PRIMITIVE_CNTL_VSOUT(l.cnt));
OUT_PKT4(ring, REG_A5XX_VPC_CNTL_0, 1);
- OUT_RING(ring, A5XX_VPC_CNTL_0_STRIDE_IN_VPC(stride_in_vpc) |
+ OUT_RING(ring, A5XX_VPC_CNTL_0_STRIDE_IN_VPC(l.max_loc) |
COND(s[FS].v->total_in > 0, A5XX_VPC_CNTL_0_VARYING) |
0x10000); // XXX
OUT_PKT4(ring, REG_A5XX_PC_PRIMITIVE_CNTL, 1);
- OUT_RING(ring, A5XX_PC_PRIMITIVE_CNTL_STRIDE_IN_VPC(stride_in_vpc) |
+ OUT_RING(ring, A5XX_PC_PRIMITIVE_CNTL_STRIDE_IN_VPC(l.max_loc) |
0x400); // XXX
OUT_PKT4(ring, REG_A5XX_SP_FS_OBJ_START_LO, 2);
@@ -467,25 +586,17 @@ fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
A5XX_GRAS_CNTL_UNK3) |
COND(s[FS].v->frag_face, A5XX_GRAS_CNTL_UNK3));
- OUT_PKT4(ring, REG_A5XX_RB_RENDER_CONTROL0, 3);
- OUT_RING(ring,
- COND(s[FS].v->total_in > 0, A5XX_RB_RENDER_CONTROL0_VARYING) |
+ OUT_PKT4(ring, REG_A5XX_RB_RENDER_CONTROL0, 2);
+ OUT_RING(ring, COND(s[FS].v->total_in > 0, A5XX_RB_RENDER_CONTROL0_VARYING) |
COND(s[FS].v->frag_coord, A5XX_RB_RENDER_CONTROL0_XCOORD |
A5XX_RB_RENDER_CONTROL0_YCOORD |
A5XX_RB_RENDER_CONTROL0_ZCOORD |
A5XX_RB_RENDER_CONTROL0_WCOORD |
A5XX_RB_RENDER_CONTROL0_UNK3) |
COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL0_UNK3));
+ OUT_RING(ring, COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL1_FACENESS));
- OUT_RING(ring,
- COND(s[FS].v->frag_face, A5XX_RB_RENDER_CONTROL1_FACENESS));
- OUT_RING(ring, A5XX_RB_FS_OUTPUT_CNTL_MRT(nr) |
- COND(s[FS].v->writes_pos, A5XX_RB_FS_OUTPUT_CNTL_FRAG_WRITES_Z));
-
- OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_CNTL, 9);
- OUT_RING(ring, A5XX_SP_FS_OUTPUT_CNTL_MRT(nr) |
- A5XX_SP_FS_OUTPUT_CNTL_DEPTH_REGID(posz_regid) |
- A5XX_SP_FS_OUTPUT_CNTL_SAMPLEMASK_REGID(regid(63, 0)));
+ OUT_PKT4(ring, REG_A5XX_SP_FS_OUTPUT_REG(0), 8);
for (i = 0; i < 8; i++) {
OUT_RING(ring, A5XX_SP_FS_OUTPUT_REG_REGID(color_regid[i]) |
COND(emit->key.half_precision,
diff --git a/src/gallium/drivers/freedreno/a5xx/fd5_program.h b/src/gallium/drivers/freedreno/a5xx/fd5_program.h
index cd03bc5..adcd180 100644
--- a/src/gallium/drivers/freedreno/a5xx/fd5_program.h
+++ b/src/gallium/drivers/freedreno/a5xx/fd5_program.h
@@ -37,8 +37,7 @@ struct fd5_shader_stateobj {
struct fd5_emit;
-void fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit,
- int nr, struct pipe_surface **bufs);
+void fd5_program_emit(struct fd_ringbuffer *ring, struct fd5_emit *emit);
void fd5_prog_init(struct pipe_context *pctx);
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c b/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
index 095e7f2..bfc04a7 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_compiler_nir.c
@@ -218,7 +218,8 @@ compile_init(struct ir3_compiler *compiler,
so->constbase.driver_param = constoff;
constoff += align(IR3_DP_COUNT, 4) / 4;
- if (so->shader->stream_output.num_outputs > 0) {
+ if ((compiler->gpu_id < 500) &&
+ so->shader->stream_output.num_outputs > 0) {
so->constbase.tfbo = constoff;
constoff += align(PIPE_MAX_SO_BUFFERS * ptrsz, 4) / 4;
}
@@ -2001,7 +2002,8 @@ emit_function(struct ir3_compile *ctx, nir_function_impl *impl)
* out, we guarantee that all exit paths flow into the stream-
* out instructions.
*/
- if ((ctx->so->shader->stream_output.num_outputs > 0) &&
+ if ((ctx->compiler->gpu_id < 500) &&
+ (ctx->so->shader->stream_output.num_outputs > 0) &&
!ctx->so->key.binning_pass) {
debug_assert(ctx->so->type == SHADER_VERTEX);
emit_stream_out(ctx);
diff --git a/src/gallium/drivers/freedreno/ir3/ir3_shader.c b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
index 4da7246..ffe1b04 100644
--- a/src/gallium/drivers/freedreno/ir3/ir3_shader.c
+++ b/src/gallium/drivers/freedreno/ir3/ir3_shader.c
@@ -614,6 +614,8 @@ max_tf_vtx(struct fd_context *ctx, const struct ir3_shader_variant *v)
struct pipe_stream_output_info *info = &v->shader->stream_output;
uint32_t maxvtxcnt = 0x7fffffff;
+ if (ctx->screen->gpu_id >= 500)
+ return 0;
if (v->key.binning_pass)
return 0;
if (v->shader->stream_output.num_outputs == 0)
More information about the mesa-commit
mailing list