static void *
nvfx_sampler_state_create(struct pipe_context *pipe,
const struct pipe_sampler_state *cso)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
struct nvfx_sampler_state *ps;
ps = MALLOC(sizeof(struct nvfx_sampler_state));
/* on nv30, we use this as an internal flag */
ps->fmt = cso->normalized_coords ? 0 : NV40_3D_TEX_FORMAT_RECT;
ps->en = 0;
ps->filt = nvfx_tex_filter(cso) | 0x2000; /*voodoo*/
ps->wrap = (nvfx_tex_wrap_mode(cso->wrap_s) << NV30_3D_TEX_WRAP_S__SHIFT) |
(nvfx_tex_wrap_mode(cso->wrap_t) << NV30_3D_TEX_WRAP_T__SHIFT) |
(nvfx_tex_wrap_mode(cso->wrap_r) << NV30_3D_TEX_WRAP_R__SHIFT);
ps->compare = FALSE;
if(cso->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE)
{
ps->wrap |= nvfx_tex_wrap_compare_mode(cso->compare_func);
ps->compare = TRUE;
}
ps->bcol = nvfx_tex_border_color(cso->border_color.f);
if(nvfx->is_nv4x)
nv40_sampler_state_init(pipe, ps, cso);
else
nv30_sampler_state_init(pipe, ps, cso);
return (void *)ps;
}
void
nvfx_draw_vbo_swtnl(struct pipe_context *pipe, const struct pipe_draw_info* info)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
unsigned i;
void *map;
if (!nvfx_state_validate_swtnl(nvfx))
return;
nvfx_state_emit(nvfx);
/* these must be passed without adding the offsets */
for (i = 0; i < nvfx->vtxbuf_nr; i++) {
map = nvfx_buffer(nvfx->vtxbuf[i].buffer)->data;
draw_set_mapped_vertex_buffer(nvfx->draw, i, map);
}
map = NULL;
if (info->indexed && nvfx->idxbuf.buffer)
map = nvfx_buffer(nvfx->idxbuf.buffer)->data;
draw_set_mapped_index_buffer(nvfx->draw, map);
if (nvfx->constbuf[PIPE_SHADER_VERTEX]) {
const unsigned nr = nvfx->constbuf_nr[PIPE_SHADER_VERTEX];
map = nvfx_buffer(nvfx->constbuf[PIPE_SHADER_VERTEX])->data;
draw_set_mapped_constant_buffer(nvfx->draw, PIPE_SHADER_VERTEX, 0,
map, nr);
}
draw_vbo(nvfx->draw, info);
draw_flush(nvfx->draw);
}
static void
nvfx_vtxelts_state_bind(struct pipe_context *pipe, void *hwcso)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
nvfx->vtxelt = hwcso;
nvfx->use_vertex_buffers = -1;
nvfx->draw_dirty |= NVFX_NEW_ARRAYS;
}
void nvfx_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
unsigned upload_mode = 0;
if (!nvfx->vtxelt->needs_translate)
upload_mode = nvfx_decide_upload_mode(pipe, info);
nvfx->use_index_buffer = upload_mode > 1;
if ((upload_mode > 0) != nvfx->use_vertex_buffers)
{
nvfx->use_vertex_buffers = (upload_mode > 0);
nvfx->dirty |= NVFX_NEW_ARRAYS;
nvfx->draw_dirty |= NVFX_NEW_ARRAYS;
}
if (upload_mode > 0)
{
for (unsigned i = 0; i < nvfx->vtxelt->num_per_vertex_buffer_infos; i++)
{
struct nvfx_per_vertex_buffer_info* vbi = &nvfx->vtxelt->per_vertex_buffer_info[i];
struct pipe_vertex_buffer *vb = &nvfx->vtxbuf[vbi->vertex_buffer_index];
nvfx_buffer_upload(nvfx_buffer(vb->buffer));
}
if (upload_mode > 1)
{
nvfx_buffer_upload(nvfx_buffer(nvfx->idxbuf.buffer));
if (unlikely(info->index_bias != nvfx->base_vertex))
{
nvfx->base_vertex = info->index_bias;
nvfx->dirty |= NVFX_NEW_ARRAYS;
}
}
else
{
if (unlikely(info->start < nvfx->base_vertex && nvfx->base_vertex))
{
nvfx->base_vertex = 0;
nvfx->dirty |= NVFX_NEW_ARRAYS;
}
}
}
if (nvfx->screen->force_swtnl || !nvfx_state_validate(nvfx))
nvfx_draw_vbo_swtnl(pipe, info);
else
nvfx_push_vbo(pipe, info);
}
static void
nvfx_sampler_state_bind(struct pipe_context *pipe, unsigned nr, void **sampler)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
unsigned unit;
for (unit = 0; unit < nr; unit++) {
nvfx->tex_sampler[unit] = sampler[unit];
nvfx->dirty_samplers |= (1 << unit);
}
for (unit = nr; unit < nvfx->nr_samplers; unit++) {
nvfx->tex_sampler[unit] = NULL;
nvfx->dirty_samplers |= (1 << unit);
}
nvfx->nr_samplers = nr;
nvfx->dirty |= NVFX_NEW_SAMPLER;
}
static void
nvfx_set_vertex_buffers(struct pipe_context *pipe, unsigned count,
const struct pipe_vertex_buffer *vb)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
for(unsigned i = 0; i < count; ++i)
{
pipe_resource_reference(&nvfx->vtxbuf[i].buffer, vb[i].buffer);
nvfx->vtxbuf[i].buffer_offset = vb[i].buffer_offset;
nvfx->vtxbuf[i].max_index = vb[i].max_index;
nvfx->vtxbuf[i].stride = vb[i].stride;
}
for(unsigned i = count; i < nvfx->vtxbuf_nr; ++i)
pipe_resource_reference(&nvfx->vtxbuf[i].buffer, 0);
nvfx->vtxbuf_nr = count;
nvfx->use_vertex_buffers = -1;
nvfx->draw_dirty |= NVFX_NEW_ARRAYS;
}
static void
nvfx_set_index_buffer(struct pipe_context *pipe,
const struct pipe_index_buffer *ib)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
if(ib)
{
pipe_resource_reference(&nvfx->idxbuf.buffer, ib->buffer);
nvfx->idxbuf.index_size = ib->index_size;
nvfx->idxbuf.offset = ib->offset;
}
else
{
pipe_resource_reference(&nvfx->idxbuf.buffer, 0);
nvfx->idxbuf.index_size = 0;
nvfx->idxbuf.offset = 0;
}
nvfx->dirty |= NVFX_NEW_INDEX;
nvfx->draw_dirty |= NVFX_NEW_INDEX;
}
static void
nvfx_set_fragment_sampler_views(struct pipe_context *pipe,
unsigned nr,
struct pipe_sampler_view **views)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
unsigned unit;
for (unit = 0; unit < nr; unit++) {
pipe_sampler_view_reference(&nvfx->fragment_sampler_views[unit],
views[unit]);
nvfx->dirty_samplers |= (1 << unit);
}
for (unit = nr; unit < nvfx->nr_textures; unit++) {
pipe_sampler_view_reference(&nvfx->fragment_sampler_views[unit],
NULL);
nvfx->dirty_samplers |= (1 << unit);
}
nvfx->nr_textures = nr;
nvfx->dirty |= NVFX_NEW_SAMPLER;
}
static struct pipe_sampler_view *
nvfx_create_sampler_view(struct pipe_context *pipe,
struct pipe_resource *pt,
const struct pipe_sampler_view *templ)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
struct nvfx_sampler_view *sv = CALLOC_STRUCT(nvfx_sampler_view);
struct nvfx_texture_format *tf = &nvfx_texture_formats[templ->format];
unsigned txf;
if (!sv)
return NULL;
sv->base = *templ;
sv->base.reference.count = 1;
sv->base.texture = NULL;
pipe_resource_reference(&sv->base.texture, pt);
sv->base.context = pipe;
txf = NV30_3D_TEX_FORMAT_NO_BORDER;
switch (pt->target) {
case PIPE_TEXTURE_CUBE:
txf |= NV30_3D_TEX_FORMAT_CUBIC;
/* fall-through */
case PIPE_TEXTURE_2D:
case PIPE_TEXTURE_RECT:
txf |= NV30_3D_TEX_FORMAT_DIMS_2D;
break;
case PIPE_TEXTURE_3D:
txf |= NV30_3D_TEX_FORMAT_DIMS_3D;
break;
case PIPE_TEXTURE_1D:
txf |= NV30_3D_TEX_FORMAT_DIMS_1D;
break;
default:
assert(0);
}
sv->u.init_fmt = txf;
sv->swizzle = 0
| (tf->src[sv->base.swizzle_r] << NV30_3D_TEX_SWIZZLE_S0_Z__SHIFT)
| (tf->src[sv->base.swizzle_g] << NV30_3D_TEX_SWIZZLE_S0_Y__SHIFT)
| (tf->src[sv->base.swizzle_b] << NV30_3D_TEX_SWIZZLE_S0_X__SHIFT)
| (tf->src[sv->base.swizzle_a] << NV30_3D_TEX_SWIZZLE_S0_W__SHIFT)
| (tf->comp[sv->base.swizzle_r] << NV30_3D_TEX_SWIZZLE_S1_Z__SHIFT)
| (tf->comp[sv->base.swizzle_g] << NV30_3D_TEX_SWIZZLE_S1_Y__SHIFT)
| (tf->comp[sv->base.swizzle_b] << NV30_3D_TEX_SWIZZLE_S1_X__SHIFT)
| (tf->comp[sv->base.swizzle_a] << NV30_3D_TEX_SWIZZLE_S1_W__SHIFT);
sv->filt = tf->sign;
sv->wrap = tf->wrap;
sv->wrap_mask = ~0;
if (pt->target == PIPE_TEXTURE_CUBE)
{
sv->offset = 0;
sv->npot_size = (pt->width0 << NV30_3D_TEX_NPOT_SIZE_W__SHIFT) | pt->height0;
}
else
{
sv->offset = nvfx_subresource_offset(pt, 0, sv->base.u.tex.first_level, 0);
sv->npot_size = (u_minify(pt->width0, sv->base.u.tex.first_level) << NV30_3D_TEX_NPOT_SIZE_W__SHIFT) | u_minify(pt->height0, sv->base.u.tex.first_level);
/* apparently, we need to ignore the t coordinate for 1D textures to fix piglit tex1d-2dborder */
if(pt->target == PIPE_TEXTURE_1D)
{
sv->wrap_mask &=~ NV30_3D_TEX_WRAP_T__MASK;
sv->wrap |= NV30_3D_TEX_WRAP_T_REPEAT;
}
}
if(nvfx->is_nv4x)
nv40_sampler_view_init(pipe, sv);
else
nv30_sampler_view_init(pipe, sv);
return &sv->base;
}
void
nvfx_push_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
{
struct nvfx_context *nvfx = nvfx_context(pipe);
struct nouveau_channel *chan = nvfx->screen->base.channel;
struct push_context ctx;
struct util_split_prim s;
unsigned instances_left = info->instance_count;
int vtx_value;
unsigned hw_mode = nvgl_primitive(info->mode);
int i;
struct
{
uint8_t* map;
unsigned step;
} per_instance[16];
unsigned p_overhead = 64 /* magic fix */
+ 4 /* begin/end */
+ 4; /* potential edgeflag enable/disable */
ctx.chan = nvfx->screen->base.channel;
ctx.translate = nvfx->vtxelt->translate;
ctx.idxbuf = NULL;
ctx.vertex_length = nvfx->vtxelt->vertex_length;
ctx.max_vertices_per_packet = nvfx->vtxelt->max_vertices_per_packet;
ctx.edgeflag = 0.5f;
// TODO: figure out if we really want to handle this, and do so in that case
ctx.edgeflag_attr = 0xff; // nvfx->vertprog->cfg.edgeflag_in;
if(!nvfx->use_vertex_buffers)
{
for(i = 0; i < nvfx->vtxelt->num_per_vertex_buffer_infos; ++i)
{
struct nvfx_per_vertex_buffer_info* vbi = &nvfx->vtxelt->per_vertex_buffer_info[i];
struct pipe_vertex_buffer *vb = &nvfx->vtxbuf[vbi->vertex_buffer_index];
uint8_t* data = nvfx_buffer(vb->buffer)->data + vb->buffer_offset;
if(info->indexed)
data += info->index_bias * vb->stride;
ctx.translate->set_buffer(ctx.translate, i, data, vb->stride, ~0);
}
if(ctx.edgeflag_attr < 16)
vtx_value = -(ctx.vertex_length + 3); /* vertex data and edgeflag header and value */
else
{
p_overhead += 1; /* initial vertex_data header */
vtx_value = -ctx.vertex_length; /* vertex data and edgeflag header and value */
}
if (info->indexed) {
// XXX: this case and is broken and probably need a new VTX_ATTR push path
if (nvfx->idxbuf.index_size == 1)
s.emit = emit_vertices_lookup8;
else if (nvfx->idxbuf.index_size == 2)
s.emit = emit_vertices_lookup16;
else
s.emit = emit_vertices_lookup32;
} else
s.emit = emit_vertices;
}
else
{
if(!info->indexed || nvfx->use_index_buffer)
{
s.emit = info->indexed ? emit_ib_ranges : emit_vb_ranges;
p_overhead += 3;
vtx_value = 0;
}
else if (nvfx->idxbuf.index_size == 4)
{
s.emit = emit_elt32;
p_overhead += 1;
vtx_value = 8;
}
else
{
s.emit = (nvfx->idxbuf.index_size == 2) ? emit_elt16 : emit_elt8;
p_overhead += 3;
vtx_value = 7;
}
}
ctx.idxbias = info->index_bias;
if(nvfx->use_vertex_buffers)
ctx.idxbias -= nvfx->base_vertex;
/* map index buffer, if present */
if (info->indexed && !nvfx->use_index_buffer)
ctx.idxbuf = nvfx_buffer(nvfx->idxbuf.buffer)->data + nvfx->idxbuf.offset;
s.priv = &ctx;
s.edge = emit_edgeflag;
for (i = 0; i < nvfx->vtxelt->num_per_instance; ++i)
{
struct nvfx_per_instance_element *ve = &nvfx->vtxelt->per_instance[i];
struct pipe_vertex_buffer *vb = &nvfx->vtxbuf[ve->base.vertex_buffer_index];
float v[4];
per_instance[i].step = info->start_instance % ve->instance_divisor;
per_instance[i].map = nvfx_buffer(vb->buffer)->data + vb->buffer_offset + ve->base.src_offset;
nvfx->vtxelt->per_instance[i].base.fetch_rgba_float(v, per_instance[i].map, 0, 0);
WAIT_RING(chan, 5);
nvfx_emit_vtx_attr(chan, nvfx->vtxelt->per_instance[i].base.idx, v, nvfx->vtxelt->per_instance[i].base.ncomp);
}
/* per-instance loop */
while (instances_left--) {
int max_verts;
boolean done;
util_split_prim_init(&s, info->mode, info->start, info->count);
nvfx_state_emit(nvfx);
for(;;) {
max_verts = AVAIL_RING(chan);
max_verts -= p_overhead;
/* if vtx_value < 0, each vertex is -vtx_value words long
* otherwise, each vertex is 2^(vtx_value) / 255 words long (this is an approximation)
*/
if(vtx_value < 0)
{
max_verts /= -vtx_value;
max_verts -= (max_verts >> 10); /* vertex data headers */
}
else
{
if(max_verts >= (1 << 23)) /* avoid overflow here */
max_verts = (1 << 23);
max_verts = (max_verts * 255) >> vtx_value;
}
//printf("avail %u max_verts %u\n", AVAIL_RING(chan), max_verts);
if(max_verts >= 16)
{
/* XXX: any command a lot of times seems to (mostly) fix corruption that would otherwise happen */
/* this seems to cause issues on nv3x, and also be unneeded there */
if(nvfx->is_nv4x)
{
int i;
for(i = 0; i < 32; ++i)
{
OUT_RING(chan, RING_3D(0x1dac, 1));
OUT_RING(chan, 0);
}
}
OUT_RING(chan, RING_3D(NV30_3D_VERTEX_BEGIN_END, 1));
OUT_RING(chan, hw_mode);
done = util_split_prim_next(&s, max_verts);
OUT_RING(chan, RING_3D(NV30_3D_VERTEX_BEGIN_END, 1));
OUT_RING(chan, 0);
if(done)
break;
}
FIRE_RING(chan);
nvfx_state_emit(nvfx);
}
static unsigned nvfx_decide_upload_mode(struct pipe_context *pipe, const struct pipe_draw_info *info)
{
struct nvfx_context* nvfx = nvfx_context(pipe);
unsigned hardware_cost = 0;
unsigned inline_cost = 0;
unsigned unique_vertices;
unsigned upload_mode;
float best_index_cost_for_hardware_vertices_as_inline_cost;
boolean prefer_hardware_indices;
unsigned index_inline_cost;
unsigned index_hardware_cost;
if (info->indexed)
unique_vertices = util_guess_unique_indices_count(info->mode, info->count);
else
unique_vertices = info->count;
/* Here we try to figure out if we are better off writing vertex data directly on the FIFO,
* or create hardware buffer objects and pointing the hardware to them.
*
* This is done by computing the total memcpy cost of each option, ignoring uploads
* if we think that the buffer is static and thus the upload cost will be amortized over
* future draw calls.
*
* For instance, if everything looks static, we will always create buffer objects, while if
* everything is a user buffer and we are not doing indexed drawing, we never do.
*
* Other interesting cases are where a small user vertex buffer, but a huge user index buffer,
* where we will upload the vertex buffer, so that we can use hardware index lookup, and
* the opposite case, where we instead do index lookup in software to avoid uploading
* a huge amount of vertex data that is not going to be used.
*
* Otherwise, we generally move to the GPU the after it has been pushed
* NVFX_STATIC_BUFFER_MIN_REUSE_TIMES times to the GPU without having
* been updated with a transfer (or just the buffer having been destroyed).
*
* There is no special handling for user buffers, since applications can use
* OpenGL VBOs in a one-shot fashion. OpenGL 3/4 core profile forces this
* by the way.
*
* Note that currently we don't support only putting some data on the FIFO, and
* some on vertex buffers (constant and instanced data is independent from this).
*
* nVidia doesn't seem to do this either, even though it should be at least
* doable with VTX_ATTR and possibly with VERTEX_DATA too if not indexed.
*/
for (unsigned i = 0; i < nvfx->vtxelt->num_per_vertex_buffer_infos; i++)
{
struct nvfx_per_vertex_buffer_info* vbi = &nvfx->vtxelt->per_vertex_buffer_info[i];
struct pipe_vertex_buffer *vb = &nvfx->vtxbuf[vbi->vertex_buffer_index];
struct nvfx_buffer* buffer = nvfx_buffer(vb->buffer);
buffer->bytes_to_draw_until_static -= vbi->per_vertex_size * unique_vertices;
if (!nvfx_buffer_seems_static(buffer))
{
hardware_cost += buffer->dirty_end - buffer->dirty_begin;
if (!buffer->base.bo)
hardware_cost += nvfx->screen->buffer_allocation_cost;
}
inline_cost += vbi->per_vertex_size * info->count;
}
best_index_cost_for_hardware_vertices_as_inline_cost = 0.0f;
prefer_hardware_indices = FALSE;
index_inline_cost = 0;
index_hardware_cost = 0;
if (info->indexed)
{
index_inline_cost = nvfx->idxbuf.index_size * info->count;
if (nvfx->screen->index_buffer_reloc_flags
&& (nvfx->idxbuf.index_size == 2 || nvfx->idxbuf.index_size == 4)
&& !(nvfx->idxbuf.offset & (nvfx->idxbuf.index_size - 1)))
{
struct nvfx_buffer* buffer = nvfx_buffer(nvfx->idxbuf.buffer);
buffer->bytes_to_draw_until_static -= index_inline_cost;
prefer_hardware_indices = TRUE;
if (!nvfx_buffer_seems_static(buffer))
{
index_hardware_cost = buffer->dirty_end - buffer->dirty_begin;
if (!buffer->base.bo)
index_hardware_cost += nvfx->screen->buffer_allocation_cost;
}
if ((float) index_inline_cost < (float) index_hardware_cost * nvfx->screen->inline_cost_per_hardware_cost)
{
best_index_cost_for_hardware_vertices_as_inline_cost = (float) index_inline_cost;
}
else
{
best_index_cost_for_hardware_vertices_as_inline_cost = (float) index_hardware_cost * nvfx->screen->inline_cost_per_hardware_cost;
prefer_hardware_indices = TRUE;
}
}
}
/* let's finally figure out which of the 3 paths we want to take */
if ((float) (inline_cost + index_inline_cost) > ((float) hardware_cost * nvfx->screen->inline_cost_per_hardware_cost + best_index_cost_for_hardware_vertices_as_inline_cost))
upload_mode = 1 + prefer_hardware_indices;
else
upload_mode = 0;
#ifdef DEBUG
if (unlikely(nvfx->screen->trace_draw))
{
fprintf(stderr, "DRAW");
if (info->indexed)
{
fprintf(stderr, "_IDX%u", nvfx->idxbuf.index_size);
if (info->index_bias)
fprintf(stderr, " biased %u", info->index_bias);
fprintf(stderr, " idxrange %u -> %u", info->min_index, info->max_index);
}
if (info->instance_count > 1)
fprintf(stderr, " %u instances from %u", info->instance_count, info->indexed);
fprintf(stderr, " start %u count %u prim %u", info->start, info->count, info->mode);
if (!upload_mode)
fprintf(stderr, " -> inline vertex data");
else if (upload_mode == 2 || !info->indexed)
fprintf(stderr, " -> buffer range");
else
fprintf(stderr, " -> inline indices");
fprintf(stderr, " [ivtx %u hvtx %u iidx %u hidx %u bidx %f] <", inline_cost, hardware_cost, index_inline_cost, index_hardware_cost, best_index_cost_for_hardware_vertices_as_inline_cost);
for (unsigned i = 0; i < nvfx->vtxelt->num_per_vertex_buffer_infos; ++i)
{
struct nvfx_per_vertex_buffer_info* vbi = &nvfx->vtxelt->per_vertex_buffer_info[i];
struct pipe_vertex_buffer *vb = &nvfx->vtxbuf[vbi->vertex_buffer_index];
struct nvfx_buffer* buffer = nvfx_buffer(vb->buffer);
if (i)
fprintf(stderr, ", ");
fprintf(stderr, "%p%s left %Li", buffer, buffer->last_update_static ? " static" : "", buffer->bytes_to_draw_until_static);
}
fprintf(stderr, ">\n");
}
#endif
return upload_mode;
}