void r600_upload_const_buffer(struct r600_context *rctx, struct si_resource **rbuffer,
const uint8_t *ptr, unsigned size,
uint32_t *const_offset)
{
*rbuffer = NULL;
if (R600_BIG_ENDIAN) {
uint32_t *tmpPtr;
unsigned i;
if (!(tmpPtr = malloc(size))) {
R600_ERR("Failed to allocate BE swap buffer.\n");
return;
}
for (i = 0; i < size / 4; ++i) {
tmpPtr[i] = bswap_32(((uint32_t *)ptr)[i]);
}
u_upload_data(rctx->uploader, 0, size, tmpPtr, const_offset,
(struct pipe_resource**)rbuffer);
free(tmpPtr);
} else {
u_upload_data(rctx->uploader, 0, size, ptr, const_offset,
(struct pipe_resource**)rbuffer);
}
}
static void
finalize_cbuf_state(struct ilo_context *ilo,
struct ilo_cbuf_state *cbuf,
const struct ilo_shader_state *sh)
{
uint32_t upload_mask = cbuf->enabled_mask;
/* skip CBUF0 if the kernel does not need it */
upload_mask &=
~ilo_shader_get_kernel_param(sh, ILO_KERNEL_SKIP_CBUF0_UPLOAD);
while (upload_mask) {
const enum pipe_format elem_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
unsigned offset, i;
i = u_bit_scan(&upload_mask);
/* no need to upload */
if (cbuf->cso[i].resource)
continue;
u_upload_data(ilo->uploader, 0, cbuf->cso[i].user_buffer_size,
cbuf->cso[i].user_buffer, &offset, &cbuf->cso[i].resource);
ilo_gpe_init_view_surface_for_buffer(ilo->dev,
ilo_buffer(cbuf->cso[i].resource),
offset, cbuf->cso[i].user_buffer_size,
util_format_get_blocksize(elem_format), elem_format,
false, false, &cbuf->cso[i].surface);
ilo->state_vector.dirty |= ILO_DIRTY_CBUF;
}
}
static void
setup_index_buffer(struct st_context *st,
const struct _mesa_index_buffer *ib,
struct pipe_index_buffer *ibuffer)
{
struct gl_buffer_object *bufobj = ib->obj;
ibuffer->index_size = vbo_sizeof_ib_type(ib->type);
/* get/create the index buffer object */
if (_mesa_is_bufferobj(bufobj)) {
/* indices are in a real VBO */
ibuffer->buffer = st_buffer_object(bufobj)->buffer;
ibuffer->offset = pointer_to_offset(ib->ptr);
}
else if (st->indexbuf_uploader) {
u_upload_data(st->indexbuf_uploader, 0, ib->count * ibuffer->index_size,
ib->ptr, &ibuffer->offset, &ibuffer->buffer);
u_upload_unmap(st->indexbuf_uploader);
}
else {
/* indices are in user space memory */
ibuffer->user_buffer = ib->ptr;
}
cso_set_index_buffer(st->cso_context, ibuffer);
}
/* As above, but upload the full contents of a buffer. Useful for
* uploading user buffers, avoids generating an explosion of GPU
* buffers if you have an app that does lots of small vertex buffer
* renders or DrawElements calls.
*/
enum pipe_error u_upload_buffer( struct u_upload_mgr *upload,
unsigned offset,
unsigned size,
struct pipe_buffer *inbuf,
unsigned *out_offset,
struct pipe_buffer **outbuf )
{
enum pipe_error ret = PIPE_OK;
const char *map = NULL;
map = (const char *)pipe_buffer_map(
upload->screen, inbuf, PIPE_BUFFER_USAGE_CPU_READ );
if (map == NULL) {
ret = PIPE_ERROR_OUT_OF_MEMORY;
goto done;
}
if (0)
debug_printf("upload ptr %p ofs %d sz %d\n", map, offset, size);
ret = u_upload_data( upload,
size,
map + offset,
out_offset,
outbuf );
if (ret)
goto done;
done:
if (map)
pipe_buffer_unmap( upload->screen, inbuf );
return ret;
}
/**
* Basically, translate Mesa's index buffer information into
* a pipe_index_buffer object.
* \return TRUE or FALSE for success/failure
*/
static boolean
setup_index_buffer(struct st_context *st,
const struct _mesa_index_buffer *ib,
struct pipe_index_buffer *ibuffer)
{
struct gl_buffer_object *bufobj = ib->obj;
ibuffer->index_size = vbo_sizeof_ib_type(ib->type);
/* get/create the index buffer object */
if (_mesa_is_bufferobj(bufobj)) {
/* indices are in a real VBO */
ibuffer->buffer = st_buffer_object(bufobj)->buffer;
ibuffer->offset = pointer_to_offset(ib->ptr);
}
else if (st->indexbuf_uploader) {
/* upload indexes from user memory into a real buffer */
if (u_upload_data(st->indexbuf_uploader, 0,
ib->count * ibuffer->index_size, ib->ptr,
&ibuffer->offset, &ibuffer->buffer) != PIPE_OK) {
/* out of memory */
return FALSE;
}
u_upload_unmap(st->indexbuf_uploader);
}
else {
/* indices are in user space memory */
ibuffer->user_buffer = ib->ptr;
}
cso_set_index_buffer(st->cso_context, ibuffer);
return TRUE;
}
/**
* Pass the given program parameters to the graphics pipe as a
* constant buffer.
* \param shader_type either PIPE_SHADER_VERTEX or PIPE_SHADER_FRAGMENT
*/
void st_upload_constants( struct st_context *st,
struct gl_program_parameter_list *params,
unsigned shader_type)
{
assert(shader_type == PIPE_SHADER_VERTEX ||
shader_type == PIPE_SHADER_FRAGMENT ||
shader_type == PIPE_SHADER_GEOMETRY ||
shader_type == PIPE_SHADER_TESS_CTRL ||
shader_type == PIPE_SHADER_TESS_EVAL);
/* update constants */
if (params && params->NumParameters) {
struct pipe_constant_buffer cb;
const uint paramBytes = params->NumParameters * sizeof(GLfloat) * 4;
/* Update the constants which come from fixed-function state, such as
* transformation matrices, fog factors, etc. The rest of the values in
* the parameters list are explicitly set by the user with glUniform,
* glProgramParameter(), etc.
*/
if (params->StateFlags)
_mesa_load_state_parameters(st->ctx, params);
/* We always need to get a new buffer, to keep the drivers simple and
* avoid gratuitous rendering synchronization.
* Let's use a user buffer to avoid an unnecessary copy.
*/
if (st->constbuf_uploader) {
cb.buffer = NULL;
cb.user_buffer = NULL;
u_upload_data(st->constbuf_uploader, 0, paramBytes,
params->ParameterValues, &cb.buffer_offset, &cb.buffer);
u_upload_unmap(st->constbuf_uploader);
} else {
cb.buffer = NULL;
cb.user_buffer = params->ParameterValues;
cb.buffer_offset = 0;
}
cb.buffer_size = paramBytes;
if (ST_DEBUG & DEBUG_CONSTANTS) {
debug_printf("%s(shader=%d, numParams=%d, stateFlags=0x%x)\n",
__func__, shader_type, params->NumParameters,
params->StateFlags);
_mesa_print_parameter_list(params);
}
cso_set_constant_buffer(st->cso_context, shader_type, 0, &cb);
pipe_resource_reference(&cb.buffer, NULL);
st->state.constants[shader_type].ptr = params->ParameterValues;
st->state.constants[shader_type].size = paramBytes;
}
else if (st->state.constants[shader_type].ptr) {
/* Unbind. */
st->state.constants[shader_type].ptr = NULL;
st->state.constants[shader_type].size = 0;
cso_set_constant_buffer(st->cso_context, shader_type, 0, NULL);
}
}
static void
finalize_index_buffer(struct ilo_context *ilo)
{
struct ilo_state_vector *vec = &ilo->state_vector;
const bool need_upload = (vec->draw->indexed &&
(vec->ib.user_buffer || vec->ib.offset % vec->ib.index_size));
struct pipe_resource *current_hw_res = NULL;
if (!(vec->dirty & ILO_DIRTY_IB) && !need_upload)
return;
pipe_resource_reference(¤t_hw_res, vec->ib.hw_resource);
if (need_upload) {
const unsigned offset = vec->ib.index_size * vec->draw->start;
const unsigned size = vec->ib.index_size * vec->draw->count;
unsigned hw_offset;
if (vec->ib.user_buffer) {
u_upload_data(ilo->uploader, 0, size,
vec->ib.user_buffer + offset, &hw_offset, &vec->ib.hw_resource);
}
else {
u_upload_buffer(ilo->uploader, 0, vec->ib.offset + offset, size,
vec->ib.buffer, &hw_offset, &vec->ib.hw_resource);
}
/* the HW offset should be aligned */
assert(hw_offset % vec->ib.index_size == 0);
vec->ib.draw_start_offset = hw_offset / vec->ib.index_size;
/*
* INDEX[vec->draw->start] in the original buffer is INDEX[0] in the HW
* resource
*/
vec->ib.draw_start_offset -= vec->draw->start;
}
else {
pipe_resource_reference(&vec->ib.hw_resource, vec->ib.buffer);
/* note that index size may be zero when the draw is not indexed */
if (vec->draw->indexed)
vec->ib.draw_start_offset = vec->ib.offset / vec->ib.index_size;
else
vec->ib.draw_start_offset = 0;
}
/* treat the IB as clean if the HW states do not change */
if (vec->ib.hw_resource == current_hw_res &&
vec->ib.hw_index_size == vec->ib.index_size)
vec->dirty &= ~ILO_DIRTY_IB;
else
vec->ib.hw_index_size = vec->ib.index_size;
pipe_resource_reference(¤t_hw_res, NULL);
}
static void
iris_update_grid_size_resource(struct iris_context *ice,
const struct pipe_grid_info *grid)
{
const struct iris_screen *screen = (void *) ice->ctx.screen;
const struct isl_device *isl_dev = &screen->isl_dev;
struct iris_state_ref *grid_ref = &ice->state.grid_size;
struct iris_state_ref *state_ref = &ice->state.grid_surf_state;
// XXX: if the shader doesn't actually care about the grid info,
// don't bother uploading the surface?
if (grid->indirect) {
pipe_resource_reference(&grid_ref->res, grid->indirect);
grid_ref->offset = grid->indirect_offset;
/* Zero out the grid size so that the next non-indirect grid launch will
* re-upload it properly.
*/
memset(ice->state.last_grid, 0, sizeof(ice->state.last_grid));
} else {
/* If the size is the same, we don't need to upload anything. */
if (memcmp(ice->state.last_grid, grid->grid, sizeof(grid->grid)) == 0)
return;
memcpy(ice->state.last_grid, grid->grid, sizeof(grid->grid));
u_upload_data(ice->state.dynamic_uploader, 0, sizeof(grid->grid), 4,
grid->grid, &grid_ref->offset, &grid_ref->res);
}
void *surf_map = NULL;
u_upload_alloc(ice->state.surface_uploader, 0, isl_dev->ss.size,
isl_dev->ss.align, &state_ref->offset, &state_ref->res,
&surf_map);
state_ref->offset +=
iris_bo_offset_from_base_address(iris_resource_bo(state_ref->res));
isl_buffer_fill_state(&screen->isl_dev, surf_map,
.address = grid_ref->offset +
iris_resource_bo(grid_ref->res)->gtt_offset,
.size_B = sizeof(grid->grid),
.format = ISL_FORMAT_RAW,
.stride_B = 1,
.mocs = 4); // XXX: MOCS
ice->state.dirty |= IRIS_DIRTY_BINDINGS_CS;
}
static void r300_render_draw_elements(struct vbuf_render* render,
const ushort* indices,
uint count)
{
struct r300_render* r300render = r300_render(render);
struct r300_context* r300 = r300render->r300;
unsigned max_index = (r300->vbo->size - r300->draw_vbo_offset) /
(r300render->r300->vertex_info.size * 4) - 1;
struct pipe_resource *index_buffer = NULL;
unsigned index_buffer_offset;
CS_LOCALS(r300);
DBG(r300, DBG_DRAW, "r300: render_draw_elements (count: %d)\n", count);
u_upload_data(r300->uploader, 0, count * 2, indices,
&index_buffer_offset, &index_buffer);
if (!index_buffer) {
return;
}
if (!r300_prepare_for_rendering(r300,
PREP_EMIT_STATES |
PREP_EMIT_VARRAYS_SWTCL | PREP_INDEXED,
index_buffer, 12, 0, 0, -1)) {
pipe_resource_reference(&index_buffer, NULL);
return;
}
BEGIN_CS(12);
OUT_CS_REG(R300_GA_COLOR_CONTROL,
r300_provoking_vertex_fixes(r300, r300render->prim));
OUT_CS_REG(R300_VAP_VF_MAX_VTX_INDX, max_index);
OUT_CS_PKT3(R300_PACKET3_3D_DRAW_INDX_2, 0);
OUT_CS(R300_VAP_VF_CNTL__PRIM_WALK_INDICES | (count << 16) |
r300render->hwprim);
OUT_CS_PKT3(R300_PACKET3_INDX_BUFFER, 2);
OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2));
OUT_CS(index_buffer_offset);
OUT_CS((count + 1) / 2);
OUT_CS_RELOC(r300_resource(index_buffer));
END_CS;
pipe_resource_reference(&index_buffer, NULL);
}
void r300_upload_index_buffer(struct r300_context *r300,
struct pipe_resource **index_buffer,
unsigned index_size, unsigned *start,
unsigned count, uint8_t *ptr)
{
unsigned index_offset;
*index_buffer = NULL;
u_upload_data(r300->uploader,
0, count * index_size,
ptr + (*start * index_size),
&index_offset,
index_buffer);
*start = index_offset / index_size;
}
/* As above, but upload the full contents of a buffer. Useful for
* uploading user buffers, avoids generating an explosion of GPU
* buffers if you have an app that does lots of small vertex buffer
* renders or DrawElements calls.
*/
enum pipe_error u_upload_buffer( struct u_upload_mgr *upload,
unsigned min_out_offset,
unsigned offset,
unsigned size,
struct pipe_resource *inbuf,
unsigned *out_offset,
struct pipe_resource **outbuf,
boolean *flushed )
{
enum pipe_error ret = PIPE_OK;
struct pipe_transfer *transfer = NULL;
const char *map = NULL;
map = (const char *)pipe_buffer_map_range(upload->pipe,
inbuf,
offset, size,
PIPE_TRANSFER_READ,
&transfer);
if (map == NULL) {
ret = PIPE_ERROR_OUT_OF_MEMORY;
goto done;
}
if (0)
debug_printf("upload ptr %p ofs %d sz %d\n", map, offset, size);
ret = u_upload_data( upload,
min_out_offset,
size,
map + offset,
out_offset,
outbuf, flushed );
done:
if (map)
pipe_buffer_unmap( upload->pipe, transfer );
return ret;
}
static void
vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct vc4_context *vc4 = vc4_context(pctx);
if (info->mode >= PIPE_PRIM_QUADS) {
util_primconvert_save_index_buffer(vc4->primconvert, &vc4->indexbuf);
util_primconvert_save_rasterizer_state(vc4->primconvert, &vc4->rasterizer->base);
util_primconvert_draw_vbo(vc4->primconvert, info);
perf_debug("Fallback conversion for %d %s vertices\n",
info->count, u_prim_name(info->mode));
return;
}
/* Before setting up the draw, do any fixup blits necessary. */
vc4_update_shadow_textures(pctx, &vc4->verttex);
vc4_update_shadow_textures(pctx, &vc4->fragtex);
vc4_hw_2116_workaround(pctx);
vc4_get_draw_cl_space(vc4);
if (vc4->prim_mode != info->mode) {
vc4->prim_mode = info->mode;
vc4->dirty |= VC4_DIRTY_PRIM_MODE;
}
vc4_start_draw(vc4);
vc4_update_compiled_shaders(vc4, info->mode);
vc4_emit_state(pctx);
if ((vc4->dirty & (VC4_DIRTY_VTXBUF |
VC4_DIRTY_VTXSTATE |
VC4_DIRTY_PRIM_MODE |
VC4_DIRTY_RASTERIZER |
VC4_DIRTY_COMPILED_CS |
VC4_DIRTY_COMPILED_VS |
VC4_DIRTY_COMPILED_FS |
vc4->prog.cs->uniform_dirty_bits |
vc4->prog.vs->uniform_dirty_bits |
vc4->prog.fs->uniform_dirty_bits)) ||
vc4->last_index_bias != info->index_bias) {
vc4_emit_gl_shader_state(vc4, info);
}
vc4->dirty = 0;
/* Note that the primitive type fields match with OpenGL/gallium
* definitions, up to but not including QUADS.
*/
struct vc4_cl_out *bcl = cl_start(&vc4->bcl);
if (info->indexed) {
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
struct pipe_resource *prsc;
if (vc4->indexbuf.index_size == 4) {
prsc = vc4_get_shadow_index_buffer(pctx, &vc4->indexbuf,
info->count, &offset);
index_size = 2;
} else {
if (vc4->indexbuf.user_buffer) {
prsc = NULL;
u_upload_data(vc4->uploader, 0,
info->count * index_size, 4,
vc4->indexbuf.user_buffer,
&offset, &prsc);
} else {
prsc = vc4->indexbuf.buffer;
}
}
struct vc4_resource *rsc = vc4_resource(prsc);
cl_start_reloc(&vc4->bcl, &bcl, 1);
cl_u8(&bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
cl_u8(&bcl,
info->mode |
(index_size == 2 ?
VC4_INDEX_BUFFER_U16:
VC4_INDEX_BUFFER_U8));
cl_u32(&bcl, info->count);
cl_reloc(vc4, &vc4->bcl, &bcl, rsc->bo, offset);
cl_u32(&bcl, vc4->max_index);
if (vc4->indexbuf.index_size == 4 || vc4->indexbuf.user_buffer)
pipe_resource_reference(&prsc, NULL);
} else {
cl_u8(&bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
cl_u8(&bcl, info->mode);
cl_u32(&bcl, info->count);
cl_u32(&bcl, info->start);
}
cl_end(&vc4->bcl, bcl);
if (vc4->zsa && vc4->zsa->base.depth.enabled) {
vc4->resolve |= PIPE_CLEAR_DEPTH;
}
if (vc4->zsa && vc4->zsa->base.stencil[0].enabled)
vc4->resolve |= PIPE_CLEAR_STENCIL;
vc4->resolve |= PIPE_CLEAR_COLOR0;
vc4->shader_rec_count++;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(pctx);
}
void r600_upload_index_buffer(struct r600_context *rctx,
struct pipe_index_buffer *ib, unsigned count)
{
u_upload_data(rctx->uploader, 0, count * ib->index_size,
ib->user_buffer, &ib->offset, &ib->buffer);
}
static void
u_vbuf_upload_buffers(struct u_vbuf_priv *mgr,
int min_index, int max_index,
unsigned instance_count)
{
unsigned i;
unsigned count = max_index + 1 - min_index;
unsigned nr_velems = mgr->ve->count;
unsigned nr_vbufs = mgr->b.nr_vertex_buffers;
struct pipe_vertex_element *velems =
mgr->fallback_ve ? mgr->fallback_velems : mgr->ve->ve;
unsigned start_offset[PIPE_MAX_ATTRIBS];
unsigned end_offset[PIPE_MAX_ATTRIBS] = {0};
/* Determine how much data needs to be uploaded. */
for (i = 0; i < nr_velems; i++) {
struct pipe_vertex_element *velem = &velems[i];
unsigned index = velem->vertex_buffer_index;
struct pipe_vertex_buffer *vb = &mgr->b.vertex_buffer[index];
unsigned instance_div, first, size;
/* Skip the buffer generated by translate. */
if (index == mgr->fallback_vb_slot) {
continue;
}
assert(vb->buffer);
if (!u_vbuf_resource(vb->buffer)->user_ptr) {
continue;
}
instance_div = velem->instance_divisor;
first = vb->buffer_offset + velem->src_offset;
if (!vb->stride) {
/* Constant attrib. */
size = mgr->ve->src_format_size[i];
} else if (instance_div) {
/* Per-instance attrib. */
unsigned count = (instance_count + instance_div - 1) / instance_div;
size = vb->stride * (count - 1) + mgr->ve->src_format_size[i];
} else {
/* Per-vertex attrib. */
first += vb->stride * min_index;
size = vb->stride * (count - 1) + mgr->ve->src_format_size[i];
}
/* Update offsets. */
if (!end_offset[index]) {
start_offset[index] = first;
end_offset[index] = first + size;
} else {
if (first < start_offset[index])
start_offset[index] = first;
if (first + size > end_offset[index])
end_offset[index] = first + size;
}
}
/* Upload buffers. */
for (i = 0; i < nr_vbufs; i++) {
unsigned start, end = end_offset[i];
boolean flushed;
struct pipe_vertex_buffer *real_vb;
uint8_t *ptr;
if (!end) {
continue;
}
start = start_offset[i];
assert(start < end);
real_vb = &mgr->b.real_vertex_buffer[i];
ptr = u_vbuf_resource(mgr->b.vertex_buffer[i].buffer)->user_ptr;
u_upload_data(mgr->b.uploader, start, end - start, ptr + start,
&real_vb->buffer_offset, &real_vb->buffer, &flushed);
real_vb->buffer_offset -= start;
}
}
static void
vc4_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct vc4_context *vc4 = vc4_context(pctx);
if (info->mode >= PIPE_PRIM_QUADS) {
util_primconvert_save_index_buffer(vc4->primconvert, &vc4->indexbuf);
util_primconvert_save_rasterizer_state(vc4->primconvert, &vc4->rasterizer->base);
util_primconvert_draw_vbo(vc4->primconvert, info);
perf_debug("Fallback conversion for %d %s vertices\n",
info->count, u_prim_name(info->mode));
return;
}
/* Before setting up the draw, do any fixup blits necessary. */
vc4_predraw_check_textures(pctx, &vc4->verttex);
vc4_predraw_check_textures(pctx, &vc4->fragtex);
vc4_hw_2116_workaround(pctx, info->count);
struct vc4_job *job = vc4_get_job_for_fbo(vc4);
vc4_get_draw_cl_space(job, info->count);
if (vc4->prim_mode != info->mode) {
vc4->prim_mode = info->mode;
vc4->dirty |= VC4_DIRTY_PRIM_MODE;
}
vc4_start_draw(vc4);
if (!vc4_update_compiled_shaders(vc4, info->mode)) {
debug_warn_once("shader compile failed, skipping draw call.\n");
return;
}
vc4_emit_state(pctx);
if ((vc4->dirty & (VC4_DIRTY_VTXBUF |
VC4_DIRTY_VTXSTATE |
VC4_DIRTY_PRIM_MODE |
VC4_DIRTY_RASTERIZER |
VC4_DIRTY_COMPILED_CS |
VC4_DIRTY_COMPILED_VS |
VC4_DIRTY_COMPILED_FS |
vc4->prog.cs->uniform_dirty_bits |
vc4->prog.vs->uniform_dirty_bits |
vc4->prog.fs->uniform_dirty_bits)) ||
vc4->last_index_bias != info->index_bias) {
vc4_emit_gl_shader_state(vc4, info, 0);
}
vc4->dirty = 0;
/* Note that the primitive type fields match with OpenGL/gallium
* definitions, up to but not including QUADS.
*/
struct vc4_cl_out *bcl = cl_start(&job->bcl);
if (info->indexed) {
uint32_t offset = vc4->indexbuf.offset;
uint32_t index_size = vc4->indexbuf.index_size;
struct pipe_resource *prsc;
if (vc4->indexbuf.index_size == 4) {
prsc = vc4_get_shadow_index_buffer(pctx, &vc4->indexbuf,
info->count, &offset);
index_size = 2;
} else {
if (vc4->indexbuf.user_buffer) {
prsc = NULL;
u_upload_data(vc4->uploader, 0,
info->count * index_size, 4,
vc4->indexbuf.user_buffer,
&offset, &prsc);
} else {
prsc = vc4->indexbuf.buffer;
}
}
struct vc4_resource *rsc = vc4_resource(prsc);
cl_start_reloc(&job->bcl, &bcl, 1);
cl_u8(&bcl, VC4_PACKET_GL_INDEXED_PRIMITIVE);
cl_u8(&bcl,
info->mode |
(index_size == 2 ?
VC4_INDEX_BUFFER_U16:
VC4_INDEX_BUFFER_U8));
cl_u32(&bcl, info->count);
cl_reloc(job, &job->bcl, &bcl, rsc->bo, offset);
cl_u32(&bcl, vc4->max_index);
job->draw_calls_queued++;
if (vc4->indexbuf.index_size == 4 || vc4->indexbuf.user_buffer)
pipe_resource_reference(&prsc, NULL);
} else {
uint32_t count = info->count;
uint32_t start = info->start;
uint32_t extra_index_bias = 0;
while (count) {
uint32_t this_count = count;
uint32_t step = count;
static const uint32_t max_verts = 65535;
/* GFXH-515 / SW-5891: The binner emits 16 bit indices
* for drawarrays, which means that if start + count >
* 64k it would truncate the top bits. Work around
* this by emitting a limited number of primitives at
* a time and reemitting the shader state pointing
* farther down the vertex attribute arrays.
*
* To do this properly for line loops or trifans, we'd
* need to make a new VB containing the first vertex
* plus whatever remainder.
*/
if (extra_index_bias) {
cl_end(&job->bcl, bcl);
vc4_emit_gl_shader_state(vc4, info,
extra_index_bias);
bcl = cl_start(&job->bcl);
}
if (start + count > max_verts) {
switch (info->mode) {
case PIPE_PRIM_POINTS:
this_count = step = max_verts;
break;
case PIPE_PRIM_LINES:
this_count = step = max_verts - (max_verts % 2);
break;
case PIPE_PRIM_LINE_STRIP:
this_count = max_verts;
step = max_verts - 1;
break;
case PIPE_PRIM_LINE_LOOP:
this_count = max_verts;
step = max_verts - 1;
debug_warn_once("unhandled line loop "
"looping behavior with "
">65535 verts\n");
break;
case PIPE_PRIM_TRIANGLES:
this_count = step = max_verts - (max_verts % 3);
break;
case PIPE_PRIM_TRIANGLE_STRIP:
this_count = max_verts;
step = max_verts - 2;
break;
default:
debug_warn_once("unhandled primitive "
"max vert count, truncating\n");
this_count = step = max_verts;
}
}
cl_u8(&bcl, VC4_PACKET_GL_ARRAY_PRIMITIVE);
cl_u8(&bcl, info->mode);
cl_u32(&bcl, this_count);
cl_u32(&bcl, start);
job->draw_calls_queued++;
count -= step;
extra_index_bias += start + step;
start = 0;
}
}
cl_end(&job->bcl, bcl);
/* We shouldn't have tripped the HW_2116 bug with the GFXH-515
* workaround.
*/
assert(job->draw_calls_queued <= VC4_HW_2116_COUNT);
if (vc4->zsa && vc4->framebuffer.zsbuf) {
struct vc4_resource *rsc =
vc4_resource(vc4->framebuffer.zsbuf->texture);
if (vc4->zsa->base.depth.enabled) {
job->resolve |= PIPE_CLEAR_DEPTH;
rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
}
if (vc4->zsa->base.stencil[0].enabled) {
job->resolve |= PIPE_CLEAR_STENCIL;
rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
}
}
job->resolve |= PIPE_CLEAR_COLOR0;
if (vc4_debug & VC4_DEBUG_ALWAYS_FLUSH)
vc4_flush(pctx);
}
/**
* Pass the given program parameters to the graphics pipe as a
* constant buffer.
* \param shader_type either PIPE_SHADER_VERTEX or PIPE_SHADER_FRAGMENT
*/
void st_upload_constants( struct st_context *st,
struct gl_program_parameter_list *params,
gl_shader_stage stage)
{
enum pipe_shader_type shader_type = st_shader_stage_to_ptarget(stage);
assert(shader_type == PIPE_SHADER_VERTEX ||
shader_type == PIPE_SHADER_FRAGMENT ||
shader_type == PIPE_SHADER_GEOMETRY ||
shader_type == PIPE_SHADER_TESS_CTRL ||
shader_type == PIPE_SHADER_TESS_EVAL ||
shader_type == PIPE_SHADER_COMPUTE);
/* update the ATI constants before rendering */
if (shader_type == PIPE_SHADER_FRAGMENT && st->fp->ati_fs) {
struct ati_fragment_shader *ati_fs = st->fp->ati_fs;
unsigned c;
for (c = 0; c < MAX_NUM_FRAGMENT_CONSTANTS_ATI; c++) {
if (ati_fs->LocalConstDef & (1 << c))
memcpy(params->ParameterValues[c],
ati_fs->Constants[c], sizeof(GLfloat) * 4);
else
memcpy(params->ParameterValues[c],
st->ctx->ATIFragmentShader.GlobalConstants[c], sizeof(GLfloat) * 4);
}
}
/* update constants */
if (params && params->NumParameters) {
struct pipe_constant_buffer cb;
const uint paramBytes = params->NumParameters * sizeof(GLfloat) * 4;
/* Update the constants which come from fixed-function state, such as
* transformation matrices, fog factors, etc. The rest of the values in
* the parameters list are explicitly set by the user with glUniform,
* glProgramParameter(), etc.
*/
if (params->StateFlags)
_mesa_load_state_parameters(st->ctx, params);
_mesa_shader_write_subroutine_indices(st->ctx, stage);
/* We always need to get a new buffer, to keep the drivers simple and
* avoid gratuitous rendering synchronization.
* Let's use a user buffer to avoid an unnecessary copy.
*/
if (st->constbuf_uploader) {
cb.buffer = NULL;
cb.user_buffer = NULL;
u_upload_data(st->constbuf_uploader, 0, paramBytes,
st->ctx->Const.UniformBufferOffsetAlignment,
params->ParameterValues, &cb.buffer_offset, &cb.buffer);
u_upload_unmap(st->constbuf_uploader);
} else {
cb.buffer = NULL;
cb.user_buffer = params->ParameterValues;
cb.buffer_offset = 0;
}
cb.buffer_size = paramBytes;
if (ST_DEBUG & DEBUG_CONSTANTS) {
debug_printf("%s(shader=%d, numParams=%d, stateFlags=0x%x)\n",
__func__, shader_type, params->NumParameters,
params->StateFlags);
_mesa_print_parameter_list(params);
}
cso_set_constant_buffer(st->cso_context, shader_type, 0, &cb);
pipe_resource_reference(&cb.buffer, NULL);
st->state.constants[shader_type].ptr = params->ParameterValues;
st->state.constants[shader_type].size = paramBytes;
}
else if (st->state.constants[shader_type].ptr) {
/* Unbind. */
st->state.constants[shader_type].ptr = NULL;
st->state.constants[shader_type].size = 0;
cso_set_constant_buffer(st->cso_context, shader_type, 0, NULL);
}
}
static void
u_vbuf_upload_buffers(struct u_vbuf *mgr,
int start_vertex, unsigned num_vertices,
int start_instance, unsigned num_instances)
{
unsigned i;
unsigned nr_velems = mgr->ve->count;
unsigned nr_vbufs = mgr->nr_vertex_buffers;
struct pipe_vertex_element *velems =
mgr->using_translate ? mgr->fallback_velems : mgr->ve->ve;
unsigned start_offset[PIPE_MAX_ATTRIBS];
unsigned end_offset[PIPE_MAX_ATTRIBS] = {0};
/* Determine how much data needs to be uploaded. */
for (i = 0; i < nr_velems; i++) {
struct pipe_vertex_element *velem = &velems[i];
unsigned index = velem->vertex_buffer_index;
struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[index];
unsigned instance_div, first, size;
/* Skip the buffers generated by translate. */
if (index == mgr->fallback_vbs[VB_VERTEX] ||
index == mgr->fallback_vbs[VB_INSTANCE] ||
index == mgr->fallback_vbs[VB_CONST]) {
continue;
}
if (!vb->user_buffer) {
continue;
}
instance_div = velem->instance_divisor;
first = vb->buffer_offset + velem->src_offset;
if (!vb->stride) {
/* Constant attrib. */
size = mgr->ve->src_format_size[i];
} else if (instance_div) {
/* Per-instance attrib. */
unsigned count = (num_instances + instance_div - 1) / instance_div;
first += vb->stride * start_instance;
size = vb->stride * (count - 1) + mgr->ve->src_format_size[i];
} else {
/* Per-vertex attrib. */
first += vb->stride * start_vertex;
size = vb->stride * (num_vertices - 1) + mgr->ve->src_format_size[i];
}
/* Update offsets. */
if (!end_offset[index]) {
start_offset[index] = first;
end_offset[index] = first + size;
} else {
if (first < start_offset[index])
start_offset[index] = first;
if (first + size > end_offset[index])
end_offset[index] = first + size;
}
}
/* Upload buffers. */
for (i = 0; i < nr_vbufs; i++) {
unsigned start, end = end_offset[i];
struct pipe_vertex_buffer *real_vb;
const uint8_t *ptr;
if (!end) {
continue;
}
start = start_offset[i];
assert(start < end);
real_vb = &mgr->real_vertex_buffer[i];
ptr = mgr->vertex_buffer[i].user_buffer;
u_upload_data(mgr->uploader, start, end - start, ptr + start,
&real_vb->buffer_offset, &real_vb->buffer);
real_vb->buffer_offset -= start;
}
}
void
st_setup_current(struct st_context *st,
const struct st_vertex_program *vp,
const struct st_vp_variant *vp_variant,
struct pipe_vertex_element *velements,
struct pipe_vertex_buffer *vbuffer, unsigned *num_vbuffers)
{
struct gl_context *ctx = st->ctx;
const GLbitfield inputs_read = vp_variant->vert_attrib_mask;
/* Process values that should have better been uniforms in the application */
GLbitfield curmask = inputs_read & _mesa_draw_current_bits(ctx);
if (curmask) {
/* vertex program validation must be done before this */
const struct st_vertex_program *vp = st->vp;
const ubyte *input_to_index = vp->input_to_index;
/* For each attribute, upload the maximum possible size. */
GLubyte data[VERT_ATTRIB_MAX * sizeof(GLdouble) * 4];
GLubyte *cursor = data;
const unsigned bufidx = (*num_vbuffers)++;
unsigned max_alignment = 1;
while (curmask) {
const gl_vert_attrib attr = u_bit_scan(&curmask);
const struct gl_array_attributes *const attrib
= _mesa_draw_current_attrib(ctx, attr);
const unsigned size = attrib->Format._ElementSize;
const unsigned alignment = util_next_power_of_two(size);
max_alignment = MAX2(max_alignment, alignment);
memcpy(cursor, attrib->Ptr, size);
if (alignment != size)
memset(cursor + size, 0, alignment - size);
init_velement_lowered(vp, velements, &attrib->Format, cursor - data, 0,
bufidx, input_to_index[attr]);
cursor += alignment;
}
vbuffer[bufidx].is_user_buffer = false;
vbuffer[bufidx].buffer.resource = NULL;
/* vbuffer[bufidx].buffer_offset is set below */
vbuffer[bufidx].stride = 0;
/* Use const_uploader for zero-stride vertex attributes, because
* it may use a better memory placement than stream_uploader.
* The reason is that zero-stride attributes can be fetched many
* times (thousands of times), so a better placement is going to
* perform better.
*/
struct u_upload_mgr *uploader = st->can_bind_const_buffer_as_vertex ?
st->pipe->const_uploader :
st->pipe->stream_uploader;
u_upload_data(uploader,
0, cursor - data, max_alignment, data,
&vbuffer[bufidx].buffer_offset,
&vbuffer[bufidx].buffer.resource);
/* Always unmap. The uploader might use explicit flushes. */
u_upload_unmap(uploader);
}
}
static enum pipe_error
u_vbuf_upload_buffers(struct u_vbuf *mgr,
int start_vertex, unsigned num_vertices,
int start_instance, unsigned num_instances)
{
unsigned i;
unsigned nr_velems = mgr->ve->count;
struct pipe_vertex_element *velems =
mgr->using_translate ? mgr->fallback_velems : mgr->ve->ve;
unsigned start_offset[PIPE_MAX_ATTRIBS];
unsigned end_offset[PIPE_MAX_ATTRIBS];
uint32_t buffer_mask = 0;
/* Determine how much data needs to be uploaded. */
for (i = 0; i < nr_velems; i++) {
struct pipe_vertex_element *velem = &velems[i];
unsigned index = velem->vertex_buffer_index;
struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[index];
unsigned instance_div, first, size, index_bit;
/* Skip the buffers generated by translate. */
if (index == mgr->fallback_vbs[VB_VERTEX] ||
index == mgr->fallback_vbs[VB_INSTANCE] ||
index == mgr->fallback_vbs[VB_CONST]) {
continue;
}
if (!vb->user_buffer) {
continue;
}
instance_div = velem->instance_divisor;
first = vb->buffer_offset + velem->src_offset;
if (!vb->stride) {
/* Constant attrib. */
size = mgr->ve->src_format_size[i];
} else if (instance_div) {
/* Per-instance attrib. */
unsigned count = (num_instances + instance_div - 1) / instance_div;
first += vb->stride * start_instance;
size = vb->stride * (count - 1) + mgr->ve->src_format_size[i];
} else {
/* Per-vertex attrib. */
first += vb->stride * start_vertex;
size = vb->stride * (num_vertices - 1) + mgr->ve->src_format_size[i];
}
index_bit = 1 << index;
/* Update offsets. */
if (!(buffer_mask & index_bit)) {
start_offset[index] = first;
end_offset[index] = first + size;
} else {
if (first < start_offset[index])
start_offset[index] = first;
if (first + size > end_offset[index])
end_offset[index] = first + size;
}
buffer_mask |= index_bit;
}
/* Upload buffers. */
while (buffer_mask) {
unsigned start, end;
struct pipe_vertex_buffer *real_vb;
const uint8_t *ptr;
i = u_bit_scan(&buffer_mask);
start = start_offset[i];
end = end_offset[i];
assert(start < end);
real_vb = &mgr->real_vertex_buffer[i];
ptr = mgr->vertex_buffer[i].user_buffer;
u_upload_data(mgr->pipe->stream_uploader, start, end - start, 4, ptr + start,
&real_vb->buffer_offset, &real_vb->buffer);
if (!real_vb->buffer)
return PIPE_ERROR_OUT_OF_MEMORY;
real_vb->buffer_offset -= start;
}
return PIPE_OK;
}
/* Setup all vertex pipeline state, rasterizer state, and fragment shader
* constants, and issue the draw call for PBO upload/download.
*
* The caller is responsible for saving and restoring state, as well as for
* setting other fragment shader state (fragment shader, samplers), and
* framebuffer/viewport/DSA/blend state.
*/
bool
st_pbo_draw(struct st_context *st, const struct st_pbo_addresses *addr,
unsigned surface_width, unsigned surface_height)
{
struct cso_context *cso = st->cso_context;
/* Setup vertex and geometry shaders */
if (!st->pbo.vs) {
st->pbo.vs = st_pbo_create_vs(st);
if (!st->pbo.vs)
return false;
}
if (addr->depth != 1 && st->pbo.use_gs && !st->pbo.gs) {
st->pbo.gs = st_pbo_create_gs(st);
if (!st->pbo.gs)
return false;
}
cso_set_vertex_shader_handle(cso, st->pbo.vs);
cso_set_geometry_shader_handle(cso, addr->depth != 1 ? st->pbo.gs : NULL);
cso_set_tessctrl_shader_handle(cso, NULL);
cso_set_tesseval_shader_handle(cso, NULL);
/* Upload vertices */
{
struct pipe_vertex_buffer vbo;
struct pipe_vertex_element velem;
float x0 = (float) addr->xoffset / surface_width * 2.0f - 1.0f;
float y0 = (float) addr->yoffset / surface_height * 2.0f - 1.0f;
float x1 = (float) (addr->xoffset + addr->width) / surface_width * 2.0f - 1.0f;
float y1 = (float) (addr->yoffset + addr->height) / surface_height * 2.0f - 1.0f;
float *verts = NULL;
vbo.user_buffer = NULL;
vbo.buffer = NULL;
vbo.stride = 2 * sizeof(float);
u_upload_alloc(st->uploader, 0, 8 * sizeof(float), 4,
&vbo.buffer_offset, &vbo.buffer, (void **) &verts);
if (!verts)
return false;
verts[0] = x0;
verts[1] = y0;
verts[2] = x0;
verts[3] = y1;
verts[4] = x1;
verts[5] = y0;
verts[6] = x1;
verts[7] = y1;
u_upload_unmap(st->uploader);
velem.src_offset = 0;
velem.instance_divisor = 0;
velem.vertex_buffer_index = cso_get_aux_vertex_buffer_slot(cso);
velem.src_format = PIPE_FORMAT_R32G32_FLOAT;
cso_set_vertex_elements(cso, 1, &velem);
cso_set_vertex_buffers(cso, velem.vertex_buffer_index, 1, &vbo);
pipe_resource_reference(&vbo.buffer, NULL);
}
/* Upload constants */
{
struct pipe_constant_buffer cb;
if (st->constbuf_uploader) {
cb.buffer = NULL;
cb.user_buffer = NULL;
u_upload_data(st->constbuf_uploader, 0, sizeof(addr->constants),
st->ctx->Const.UniformBufferOffsetAlignment,
&addr->constants, &cb.buffer_offset, &cb.buffer);
if (!cb.buffer)
return false;
u_upload_unmap(st->constbuf_uploader);
} else {
cb.buffer = NULL;
cb.user_buffer = &addr->constants;
cb.buffer_offset = 0;
}
cb.buffer_size = sizeof(addr->constants);
cso_set_constant_buffer(cso, PIPE_SHADER_FRAGMENT, 0, &cb);
pipe_resource_reference(&cb.buffer, NULL);
}
/* Rasterizer state */
cso_set_rasterizer(cso, &st->pbo.raster);
/* Disable stream output */
cso_set_stream_outputs(cso, 0, NULL, 0);
if (addr->depth == 1) {
cso_draw_arrays(cso, PIPE_PRIM_TRIANGLE_STRIP, 0, 4);
} else {
cso_draw_arrays_instanced(cso, PIPE_PRIM_TRIANGLE_STRIP,
0, 4, 0, addr->depth);
}
return true;
}
