Christoph Bumiller
2009-Oct-14 19:23 UTC
[PATCH 3/7] nv50: submit user vbo data through the fifo
Requesting a new real buffer from the kernel and
copying all the data is wasteful e.g. if only a
few (but widely spread) vertices are accessed.
---
src/gallium/drivers/nv50/nv50_context.h | 3 +
src/gallium/drivers/nv50/nv50_vbo.c | 409 +++++++++++++++++++++++++++++--
2 files changed, 394 insertions(+), 18 deletions(-)
diff --git a/src/gallium/drivers/nv50/nv50_context.h
b/src/gallium/drivers/nv50/nv50_context.h
index be53990..8e2d695 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -120,6 +120,7 @@ struct nv50_state {
struct nouveau_stateobj *vtxfmt;
struct nouveau_stateobj *vtxbuf;
struct nouveau_stateobj *vtxattr;
+ unsigned vtxelt_nr;
};
struct nv50_context {
@@ -152,6 +153,8 @@ struct nv50_context {
unsigned sampler_nr;
struct nv50_miptree *miptree[PIPE_MAX_SAMPLERS];
unsigned miptree_nr;
+
+ uint16_t vbo_fifo;
};
static INLINE struct nv50_context *
diff --git a/src/gallium/drivers/nv50/nv50_vbo.c
b/src/gallium/drivers/nv50/nv50_vbo.c
index 8b0fbf0..697a807 100644
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -26,6 +26,18 @@
#include "nv50_context.h"
+static boolean
+nv50_push_elements_u08(struct nv50_context *, uint8_t *, unsigned);
+
+static boolean
+nv50_push_elements_u16(struct nv50_context *, uint16_t *, unsigned);
+
+static boolean
+nv50_push_elements_u32(struct nv50_context *, uint32_t *, unsigned);
+
+static boolean
+nv50_push_arrays(struct nv50_context *, unsigned, unsigned);
+
static INLINE unsigned
nv50_prim(unsigned mode)
{
@@ -132,6 +144,7 @@ nv50_draw_arrays(struct pipe_context *pipe, unsigned mode,
unsigned start,
struct nv50_context *nv50 = nv50_context(pipe);
struct nouveau_channel *chan = nv50->screen->tesla->channel;
struct nouveau_grobj *tesla = nv50->screen->tesla;
+ boolean ret;
nv50_state_validate(nv50);
@@ -142,17 +155,22 @@ nv50_draw_arrays(struct pipe_context *pipe, unsigned mode,
unsigned start,
BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BEGIN, 1);
OUT_RING (chan, nv50_prim(mode));
- BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BUFFER_FIRST, 2);
- OUT_RING (chan, start);
- OUT_RING (chan, count);
+
+ if (nv50->vbo_fifo)
+ ret = nv50_push_arrays(nv50, start, count);
+ else {
+ BEGIN_RING(chan, tesla, NV50TCL_VERTEX_BUFFER_FIRST, 2);
+ OUT_RING (chan, start);
+ OUT_RING (chan, count);
+ ret = TRUE;
+ }
BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
OUT_RING (chan, 0);
- pipe->flush(pipe, 0, NULL);
- return TRUE;
+ return ret;
}
-static INLINE void
+static INLINE boolean
nv50_draw_elements_inline_u08(struct nv50_context *nv50, uint8_t *map,
unsigned start, unsigned count)
{
@@ -161,6 +179,9 @@ nv50_draw_elements_inline_u08(struct nv50_context *nv50,
uint8_t *map,
map += start;
+ if (nv50->vbo_fifo)
+ return nv50_push_elements_u08(nv50, map, count);
+
if (count & 1) {
BEGIN_RING(chan, tesla, 0x15e8, 1);
OUT_RING (chan, map[0]);
@@ -179,9 +200,10 @@ nv50_draw_elements_inline_u08(struct nv50_context *nv50,
uint8_t *map,
count -= nr;
map += nr;
}
+ return TRUE;
}
-static INLINE void
+static INLINE boolean
nv50_draw_elements_inline_u16(struct nv50_context *nv50, uint16_t *map,
unsigned start, unsigned count)
{
@@ -190,6 +212,9 @@ nv50_draw_elements_inline_u16(struct nv50_context *nv50,
uint16_t *map,
map += start;
+ if (nv50->vbo_fifo)
+ return nv50_push_elements_u16(nv50, map, count);
+
if (count & 1) {
BEGIN_RING(chan, tesla, 0x15e8, 1);
OUT_RING (chan, map[0]);
@@ -208,9 +233,10 @@ nv50_draw_elements_inline_u16(struct nv50_context *nv50,
uint16_t *map,
count -= nr;
map += nr;
}
+ return TRUE;
}
-static INLINE void
+static INLINE boolean
nv50_draw_elements_inline_u32(struct nv50_context *nv50, uint32_t *map,
unsigned start, unsigned count)
{
@@ -219,6 +245,9 @@ nv50_draw_elements_inline_u32(struct nv50_context *nv50,
uint32_t *map,
map += start;
+ if (nv50->vbo_fifo)
+ return nv50_push_elements_u32(nv50, map, count);
+
while (count) {
unsigned nr = count > 2047 ? 2047 : count;
@@ -228,6 +257,7 @@ nv50_draw_elements_inline_u32(struct nv50_context *nv50,
uint32_t *map,
count -= nr;
map += nr;
}
+ return TRUE;
}
boolean
@@ -240,6 +270,7 @@ nv50_draw_elements(struct pipe_context *pipe,
struct nouveau_grobj *tesla = nv50->screen->tesla;
struct pipe_screen *pscreen = pipe->screen;
void *map;
+ boolean ret;
map = pipe_buffer_map(pscreen, indexBuffer, PIPE_BUFFER_USAGE_CPU_READ);
@@ -254,23 +285,25 @@ nv50_draw_elements(struct pipe_context *pipe,
OUT_RING (chan, nv50_prim(mode));
switch (indexSize) {
case 1:
- nv50_draw_elements_inline_u08(nv50, map, start, count);
+ ret = nv50_draw_elements_inline_u08(nv50, map, start, count);
break;
case 2:
- nv50_draw_elements_inline_u16(nv50, map, start, count);
+ ret = nv50_draw_elements_inline_u16(nv50, map, start, count);
break;
case 4:
- nv50_draw_elements_inline_u32(nv50, map, start, count);
+ ret = nv50_draw_elements_inline_u32(nv50, map, start, count);
break;
default:
assert(0);
+ ret = FALSE;
+ break;
}
BEGIN_RING(chan, tesla, NV50TCL_VERTEX_END, 1);
OUT_RING (chan, 0);
pipe_buffer_unmap(pscreen, indexBuffer);
- pipe->flush(pipe, 0, NULL);
- return TRUE;
+
+ return ret;
}
static INLINE boolean
@@ -337,17 +370,24 @@ nv50_vbo_validate(struct nv50_context *nv50)
{
struct nouveau_grobj *tesla = nv50->screen->tesla;
struct nouveau_stateobj *vtxbuf, *vtxfmt, *vtxattr;
- unsigned i;
+ unsigned i, n_ve;
/* don't validate if Gallium took away our buffers */
if (nv50->vtxbuf_nr == 0)
return;
+ nv50->vbo_fifo = 0;
+
+ for (i = 0; i < nv50->vtxbuf_nr; ++i)
+ if (nv50->vtxbuf[i].stride &&
+ !(nv50->vtxbuf[i].buffer->usage & PIPE_BUFFER_USAGE_VERTEX))
+ nv50->vbo_fifo = 0xffff;
+
+ n_ve = MAX2(nv50->vtxelt_nr, nv50->state.vtxelt_nr);
vtxattr = NULL;
- vtxbuf = so_new(nv50->vtxelt_nr * 7, nv50->vtxelt_nr * 4);
- vtxfmt = so_new(nv50->vtxelt_nr + 1, 0);
- so_method(vtxfmt, tesla, NV50TCL_VERTEX_ARRAY_ATTRIB(0),
- nv50->vtxelt_nr);
+ vtxbuf = so_new(n_ve * 7, nv50->vtxelt_nr * 4);
+ vtxfmt = so_new(n_ve + 1, 0);
+ so_method(vtxfmt, tesla, NV50TCL_VERTEX_ARRAY_ATTRIB(0), n_ve);
for (i = 0; i < nv50->vtxelt_nr; i++) {
struct pipe_vertex_element *ve = &nv50->vtxelt[i];
@@ -363,10 +403,19 @@ nv50_vbo_validate(struct nv50_context *nv50)
so_method(vtxbuf, tesla,
NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
so_data (vtxbuf, 0);
+
+ nv50->vbo_fifo &= ~(1 << i);
continue;
}
so_data(vtxfmt, hw | i);
+ if (nv50->vbo_fifo) {
+ so_method(vtxbuf, tesla,
+ NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
+ so_data (vtxbuf, 0);
+ continue;
+ }
+
so_method(vtxbuf, tesla, NV50TCL_VERTEX_ARRAY_FORMAT(i), 3);
so_data (vtxbuf, 0x20000000 | vb->stride);
so_reloc (vtxbuf, bo, vb->buffer_offset +
@@ -385,6 +434,13 @@ nv50_vbo_validate(struct nv50_context *nv50)
NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD |
NOUVEAU_BO_LOW, 0, 0);
}
+ for (; i < n_ve; ++i) {
+ so_data (vtxfmt, 0x7e080010);
+
+ so_method(vtxbuf, tesla, NV50TCL_VERTEX_ARRAY_FORMAT(i), 1);
+ so_data (vtxbuf, 0);
+ }
+ nv50->state.vtxelt_nr = nv50->vtxelt_nr;
so_ref (vtxfmt, &nv50->state.vtxfmt);
so_ref (vtxbuf, &nv50->state.vtxbuf);
@@ -394,3 +450,320 @@ nv50_vbo_validate(struct nv50_context *nv50)
so_ref (NULL, &vtxattr);
}
+typedef void (*pfn_push)(struct nouveau_channel *, void *);
+
+struct nv50_vbo_emitctx
+{
+ pfn_push push[16];
+ void *map[16];
+ unsigned stride[16];
+ unsigned nr_ve;
+ unsigned vtx_dwords;
+ unsigned vtx_max;
+};
+
+static INLINE void
+emit_vtx_next(struct nouveau_channel *chan, struct nv50_vbo_emitctx *emit)
+{
+ unsigned i;
+
+ for (i = 0; i < emit->nr_ve; ++i) {
+ emit->push[i](chan, emit->map[i]);
+ emit->map[i] += emit->stride[i];
+ }
+}
+
+static INLINE void
+emit_vtx(struct nouveau_channel *chan, struct nv50_vbo_emitctx *emit,
+ uint32_t vi)
+{
+ unsigned i;
+
+ for (i = 0; i < emit->nr_ve; ++i)
+ emit->push[i](chan, emit->map[i] + emit->stride[i] * vi);
+}
+
+static INLINE boolean
+nv50_map_vbufs(struct nv50_context *nv50)
+{
+ int i;
+
+ for (i = 0; i < nv50->vtxbuf_nr; ++i) {
+ struct pipe_vertex_buffer *vb = &nv50->vtxbuf[i];
+ unsigned size, delta;
+
+ if (nouveau_bo(vb->buffer)->map)
+ continue;
+
+ size = vb->stride * (vb->max_index + 1);
+ delta = vb->buffer_offset;
+
+ if (!size)
+ size = vb->buffer->size - vb->buffer_offset;
+
+ if (nouveau_bo_map_range(nouveau_bo(vb->buffer),
+ delta, size, NOUVEAU_BO_RD))
+ break;
+ }
+
+ if (i == nv50->vtxbuf_nr)
+ return TRUE;
+ for (; i >= 0; --i)
+ nouveau_bo_unmap(nouveau_bo(nv50->vtxbuf[i].buffer));
+ return FALSE;
+}
+
+static INLINE void
+nv50_unmap_vbufs(struct nv50_context *nv50)
+{
+ unsigned i;
+
+ for (i = 0; i < nv50->vtxbuf_nr; ++i)
+ if (nouveau_bo(nv50->vtxbuf[i].buffer)->map)
+
nouveau_bo_unmap(nouveau_bo(nv50->vtxbuf[i].buffer));
+}
+
+static void
+emit_b32_1(struct nouveau_channel *chan, void *data)
+{
+ uint32_t *v = data;
+
+ OUT_RING(chan, v[0]);
+}
+
+static void
+emit_b32_2(struct nouveau_channel *chan, void *data)
+{
+ uint32_t *v = data;
+
+ OUT_RING(chan, v[0]);
+ OUT_RING(chan, v[1]);
+}
+
+static void
+emit_b32_3(struct nouveau_channel *chan, void *data)
+{
+ uint32_t *v = data;
+
+ OUT_RING(chan, v[0]);
+ OUT_RING(chan, v[1]);
+ OUT_RING(chan, v[2]);
+}
+
+static void
+emit_b32_4(struct nouveau_channel *chan, void *data)
+{
+ uint32_t *v = data;
+
+ OUT_RING(chan, v[0]);
+ OUT_RING(chan, v[1]);
+ OUT_RING(chan, v[2]);
+ OUT_RING(chan, v[3]);
+}
+
+static void
+emit_b16_1(struct nouveau_channel *chan, void *data)
+{
+ uint16_t *v = data;
+
+ OUT_RING(chan, v[0]);
+}
+
+static void
+emit_b16_3(struct nouveau_channel *chan, void *data)
+{
+ uint16_t *v = data;
+
+ OUT_RING(chan, (v[1] << 16) | v[0]);
+ OUT_RING(chan, v[2]);
+}
+
+static void
+emit_b08_1(struct nouveau_channel *chan, void *data)
+{
+ uint8_t *v = data;
+
+ OUT_RING(chan, v[0]);
+}
+
+static void
+emit_b08_3(struct nouveau_channel *chan, void *data)
+{
+ uint8_t *v = data;
+
+ OUT_RING(chan, (v[2] << 16) | (v[1] << 8) | v[0]);
+}
+
+static boolean
+emit_prepare(struct nv50_context *nv50, struct nv50_vbo_emitctx *emit,
+ unsigned start)
+{
+ unsigned i;
+
+ if (nv50_map_vbufs(nv50) == FALSE)
+ return FALSE;
+
+ emit->nr_ve = 0;
+ emit->vtx_dwords = 0;
+
+ for (i = 0; i < nv50->vtxelt_nr; ++i) {
+ struct pipe_vertex_element *ve;
+ struct pipe_vertex_buffer *vb;
+ unsigned n, type, size;
+
+ ve = &nv50->vtxelt[i];
+ vb = &nv50->vtxbuf[ve->vertex_buffer_index];
+ if (!(nv50->vbo_fifo & (1 << i)))
+ continue;
+ n = emit->nr_ve++;
+
+ emit->stride[n] = vb->stride;
+ emit->map[n] = nouveau_bo(vb->buffer)->map +
+ (start * vb->stride + ve->src_offset);
+
+ type = pf_type(ve->src_format);
+ size = pf_size_x(ve->src_format) << pf_exp2(ve->src_format);
+
+ assert(ve->nr_components > 0 && ve->nr_components <= 4);
+
+ /* It shouldn't be necessary to push the implicit 1s
+ * for case 3 and size 8 cases 1, 2, 3.
+ */
+ switch (size) {
+ default:
+ NOUVEAU_ERR("unsupported vtxelt size: %u\n", size);
+ return FALSE;
+ case 32:
+ switch (ve->nr_components) {
+ case 1: emit->push[n] = emit_b32_1; break;
+ case 2: emit->push[n] = emit_b32_2; break;
+ case 3: emit->push[n] = emit_b32_3; break;
+ case 4: emit->push[n] = emit_b32_4; break;
+ }
+ emit->vtx_dwords += ve->nr_components;
+ break;
+ case 16:
+ switch (ve->nr_components) {
+ case 1: emit->push[n] = emit_b16_1; break;
+ case 2: emit->push[n] = emit_b32_1; break;
+ case 3: emit->push[n] = emit_b16_3; break;
+ case 4: emit->push[n] = emit_b32_2; break;
+ }
+ emit->vtx_dwords += (ve->nr_components + 1) >> 1;
+ break;
+ case 8:
+ switch (type) {
+ case 1: emit->push[n] = emit_b08_1; break;
+ case 2: emit->push[n] = emit_b16_1; break;
+ case 3: emit->push[n] = emit_b08_3; break;
+ case 4: emit->push[n] = emit_b32_1; break;
+ }
+ emit->vtx_dwords += 1;
+ break;
+ }
+ }
+
+ emit->vtx_max = 512 / emit->vtx_dwords;
+
+ return TRUE;
+}
+
+static boolean
+nv50_push_arrays(struct nv50_context *nv50, unsigned start, unsigned count)
+{
+ struct nouveau_channel *chan = nv50->screen->base.channel;
+ struct nouveau_grobj *tesla = nv50->screen->tesla;
+ struct nv50_vbo_emitctx emit;
+
+ if (emit_prepare(nv50, &emit, start) == FALSE)
+ return FALSE;
+
+ while (count) {
+ unsigned i, dw, nr = MIN2(count, emit.vtx_max);
+ dw = nr * emit.vtx_dwords;
+
+ BEGIN_RING(chan, tesla, NV50TCL_VERTEX_DATA | 0x40000000, dw);
+ for (i = 0; i < nr; ++i)
+ emit_vtx_next(chan, &emit);
+
+ count -= nr;
+ }
+ nv50_unmap_vbufs(nv50);
+
+ return TRUE;
+}
+
+static boolean
+nv50_push_elements_u32(struct nv50_context *nv50, uint32_t *map, unsigned
count)
+{
+ struct nouveau_channel *chan = nv50->screen->base.channel;
+ struct nouveau_grobj *tesla = nv50->screen->tesla;
+ struct nv50_vbo_emitctx emit;
+
+ if (emit_prepare(nv50, &emit, 0) == FALSE)
+ return FALSE;
+
+ while (count) {
+ unsigned i, dw, nr = MIN2(count, emit.vtx_max);
+ dw = nr * emit.vtx_dwords;
+
+ BEGIN_RING(chan, tesla, NV50TCL_VERTEX_DATA | 0x40000000, dw);
+ for (i = 0; i < nr; ++i)
+ emit_vtx(chan, &emit, *map++);
+
+ count -= nr;
+ }
+ nv50_unmap_vbufs(nv50);
+
+ return TRUE;
+}
+
+static boolean
+nv50_push_elements_u16(struct nv50_context *nv50, uint16_t *map, unsigned
count)
+{
+ struct nouveau_channel *chan = nv50->screen->base.channel;
+ struct nouveau_grobj *tesla = nv50->screen->tesla;
+ struct nv50_vbo_emitctx emit;
+
+ if (emit_prepare(nv50, &emit, 0) == FALSE)
+ return FALSE;
+
+ while (count) {
+ unsigned i, dw, nr = MIN2(count, emit.vtx_max);
+ dw = nr * emit.vtx_dwords;
+
+ BEGIN_RING(chan, tesla, NV50TCL_VERTEX_DATA | 0x40000000, dw);
+ for (i = 0; i < nr; ++i)
+ emit_vtx(chan, &emit, *map++);
+
+ count -= nr;
+ }
+ nv50_unmap_vbufs(nv50);
+
+ return TRUE;
+}
+
+static boolean
+nv50_push_elements_u08(struct nv50_context *nv50, uint8_t *map, unsigned count)
+{
+ struct nouveau_channel *chan = nv50->screen->base.channel;
+ struct nouveau_grobj *tesla = nv50->screen->tesla;
+ struct nv50_vbo_emitctx emit;
+
+ if (emit_prepare(nv50, &emit, 0) == FALSE)
+ return FALSE;
+
+ while (count) {
+ unsigned i, dw, nr = MIN2(count, emit.vtx_max);
+ dw = nr * emit.vtx_dwords;
+
+ BEGIN_RING(chan, tesla, NV50TCL_VERTEX_DATA | 0x40000000, dw);
+ for (i = 0; i < nr; ++i)
+ emit_vtx(chan, &emit, *map++);
+
+ count -= nr;
+ }
+ nv50_unmap_vbufs(nv50);
+
+ return TRUE;
+}
--
1.6.3.3
--=_3hwvn9kudmw4--
Possibly Parallel Threads
- [PATCH 6/8] nv50: support more vtxelt formats
- [Bug 56474] New: 3D app segfaults on NV46
- [PATCH] nv50: avoid using inline vertex data submit when gl_VertexID is used
- [PATCH 1/2] nv50: do an explicit flush on draw when there are persistent buffers
- [PATCH 1/3] nv50: remove vtxbuf stateobject after a referenced vtxbuf is mapped
Wisdom of the Ancients
