/* 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;
}
static void *r600_buffer_transfer_map(struct pipe_context *ctx,
struct pipe_resource *resource,
unsigned level,
unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **ptransfer)
{
struct r600_common_context *rctx = (struct r600_common_context*)ctx;
struct r600_common_screen *rscreen = (struct r600_common_screen*)ctx->screen;
struct r600_resource *rbuffer = r600_resource(resource);
uint8_t *data;
assert(box->x + box->width <= resource->width0);
/* See if the buffer range being mapped has never been initialized,
* in which case it can be mapped unsynchronized. */
if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
usage & PIPE_TRANSFER_WRITE &&
!util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) {
usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
}
/* If discarding the entire range, discard the whole resource instead. */
if (usage & PIPE_TRANSFER_DISCARD_RANGE &&
box->x == 0 && box->width == resource->width0) {
usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
}
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE &&
!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
assert(usage & PIPE_TRANSFER_WRITE);
/* Check if mapping this buffer would cause waiting for the GPU. */
if (r600_rings_is_buffer_referenced(rctx, rbuffer->cs_buf, RADEON_USAGE_READWRITE) ||
rctx->ws->buffer_is_busy(rbuffer->buf, RADEON_USAGE_READWRITE)) {
rctx->invalidate_buffer(&rctx->b, &rbuffer->b.b);
}
/* At this point, the buffer is always idle. */
usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
}
else if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
!(rscreen->debug_flags & DBG_NO_DISCARD_RANGE) &&
r600_can_dma_copy_buffer(rctx, box->x, 0, box->width)) {
assert(usage & PIPE_TRANSFER_WRITE);
/* Check if mapping this buffer would cause waiting for the GPU. */
if (r600_rings_is_buffer_referenced(rctx, rbuffer->cs_buf, RADEON_USAGE_READWRITE) ||
rctx->ws->buffer_is_busy(rbuffer->buf, RADEON_USAGE_READWRITE)) {
/* Do a wait-free write-only transfer using a temporary buffer. */
unsigned offset;
struct r600_resource *staging = NULL;
u_upload_alloc(rctx->uploader, 0, box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT),
&offset, (struct pipe_resource**)&staging, (void**)&data);
if (staging) {
data += box->x % R600_MAP_BUFFER_ALIGNMENT;
return r600_buffer_get_transfer(ctx, resource, level, usage, box,
ptransfer, data, staging, offset);
} else {
return NULL; /* error, shouldn't occur though */
}
}
/* At this point, the buffer is always idle (we checked it above). */
usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
}
/* Using a staging buffer in GTT for larger reads is much faster. */
else if ((usage & PIPE_TRANSFER_READ) &&
!(usage & PIPE_TRANSFER_WRITE) &&
rbuffer->domains == RADEON_DOMAIN_VRAM &&
r600_can_dma_copy_buffer(rctx, 0, box->x, box->width)) {
struct r600_resource *staging;
staging = (struct r600_resource*) pipe_buffer_create(
ctx->screen, PIPE_BIND_TRANSFER_READ, PIPE_USAGE_STAGING,
box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT));
if (staging) {
/* Copy the VRAM buffer to the staging buffer. */
rctx->dma_copy(ctx, &staging->b.b, 0,
box->x % R600_MAP_BUFFER_ALIGNMENT,
0, 0, resource, level, box);
data = r600_buffer_map_sync_with_rings(rctx, staging, PIPE_TRANSFER_READ);
data += box->x % R600_MAP_BUFFER_ALIGNMENT;
return r600_buffer_get_transfer(ctx, resource, level, usage, box,
ptransfer, data, staging, 0);
}
}
data = r600_buffer_map_sync_with_rings(rctx, rbuffer, usage);
if (!data) {
return NULL;
}
data += box->x;
return r600_buffer_get_transfer(ctx, resource, level, usage, box,
ptransfer, data, NULL, 0);
}
static enum pipe_error
emit_constbuf_vgpu10(struct svga_context *svga, unsigned shader)
{
const struct pipe_constant_buffer *cbuf;
struct pipe_resource *dst_buffer = NULL;
enum pipe_error ret = PIPE_OK;
struct pipe_transfer *src_transfer;
struct svga_winsys_surface *dst_handle;
float extras[MAX_EXTRA_CONSTS][4];
unsigned extra_count, extra_size, extra_offset;
unsigned new_buf_size;
void *src_map = NULL, *dst_map;
unsigned offset;
const struct svga_shader_variant *variant;
assert(shader == PIPE_SHADER_VERTEX ||
shader == PIPE_SHADER_GEOMETRY ||
shader == PIPE_SHADER_FRAGMENT);
cbuf = &svga->curr.constbufs[shader][0];
switch (shader) {
case PIPE_SHADER_VERTEX:
variant = svga->state.hw_draw.vs;
extra_count = svga_get_extra_vs_constants(svga, (float *) extras);
break;
case PIPE_SHADER_FRAGMENT:
variant = svga->state.hw_draw.fs;
extra_count = svga_get_extra_fs_constants(svga, (float *) extras);
break;
case PIPE_SHADER_GEOMETRY:
variant = svga->state.hw_draw.gs;
extra_count = svga_get_extra_gs_constants(svga, (float *) extras);
break;
default:
assert(!"Unexpected shader type");
/* Don't return an error code since we don't want to keep re-trying
* this function and getting stuck in an infinite loop.
*/
return PIPE_OK;
}
assert(variant);
/* Compute extra constants size and offset in bytes */
extra_size = extra_count * 4 * sizeof(float);
extra_offset = 4 * sizeof(float) * variant->extra_const_start;
if (cbuf->buffer_size + extra_size == 0)
return PIPE_OK; /* nothing to do */
/* Typically, the cbuf->buffer here is a user-space buffer so mapping
* it is really cheap. If we ever get real HW buffers for constants
* we should void mapping and instead use a ResourceCopy command.
*/
if (cbuf->buffer_size > 0) {
src_map = pipe_buffer_map_range(&svga->pipe, cbuf->buffer,
cbuf->buffer_offset, cbuf->buffer_size,
PIPE_TRANSFER_READ, &src_transfer);
assert(src_map);
if (!src_map) {
return PIPE_ERROR_OUT_OF_MEMORY;
}
}
/* The new/dest buffer's size must be large enough to hold the original,
* user-specified constants, plus the extra constants.
* The size of the original constant buffer _should_ agree with what the
* shader is expecting, but it might not (it's not enforced anywhere by
* gallium).
*/
new_buf_size = MAX2(cbuf->buffer_size, extra_offset) + extra_size;
/* According to the DX10 spec, the constant buffer size must be
* in multiples of 16.
*/
new_buf_size = align(new_buf_size, 16);
u_upload_alloc(svga->const0_upload, 0, new_buf_size, &offset,
&dst_buffer, &dst_map);
if (!dst_map) {
if (src_map)
pipe_buffer_unmap(&svga->pipe, src_transfer);
return PIPE_ERROR_OUT_OF_MEMORY;
}
if (src_map) {
memcpy(dst_map, src_map, cbuf->buffer_size);
pipe_buffer_unmap(&svga->pipe, src_transfer);
}
if (extra_size) {
assert(extra_offset + extra_size <= new_buf_size);
memcpy((char *) dst_map + extra_offset, extras, extra_size);
}
u_upload_unmap(svga->const0_upload);
/* Issue the SetSingleConstantBuffer command */
dst_handle = svga_buffer_handle(svga, dst_buffer);
if (!dst_handle) {
pipe_resource_reference(&dst_buffer, NULL);
return PIPE_ERROR_OUT_OF_MEMORY;
}
assert(new_buf_size % 16 == 0);
ret = SVGA3D_vgpu10_SetSingleConstantBuffer(svga->swc,
0, /* index */
svga_shader_type(shader),
dst_handle,
offset,
new_buf_size);
if (ret != PIPE_OK) {
pipe_resource_reference(&dst_buffer, NULL);
return ret;
}
/* Save this const buffer until it's replaced in the future.
* Otherwise, all references to the buffer will go away after the
* command buffer is submitted, it'll get recycled and we will have
* incorrect constant buffer bindings.
*/
pipe_resource_reference(&svga->state.hw_draw.constbuf[shader], dst_buffer);
svga->state.hw_draw.default_constbuf_size[shader] = new_buf_size;
pipe_resource_reference(&dst_buffer, NULL);
return ret;
}
static void
setup_bitmap_vertex_data(struct st_context *st, bool normalized,
int x, int y, int width, int height,
float z, const float color[4],
struct pipe_resource **vbuf,
unsigned *vbuf_offset)
{
const GLfloat fb_width = (GLfloat)st->state.framebuffer.width;
const GLfloat fb_height = (GLfloat)st->state.framebuffer.height;
const GLfloat x0 = (GLfloat)x;
const GLfloat x1 = (GLfloat)(x + width);
const GLfloat y0 = (GLfloat)y;
const GLfloat y1 = (GLfloat)(y + height);
GLfloat sLeft = (GLfloat)0.0, sRight = (GLfloat)1.0;
GLfloat tTop = (GLfloat)0.0, tBot = (GLfloat)1.0 - tTop;
const GLfloat clip_x0 = (GLfloat)(x0 / fb_width * 2.0 - 1.0);
const GLfloat clip_y0 = (GLfloat)(y0 / fb_height * 2.0 - 1.0);
const GLfloat clip_x1 = (GLfloat)(x1 / fb_width * 2.0 - 1.0);
const GLfloat clip_y1 = (GLfloat)(y1 / fb_height * 2.0 - 1.0);
GLuint i;
float (*vertices)[3][4]; /**< vertex pos + color + texcoord */
if(!normalized)
{
sRight = (GLfloat) width;
tBot = (GLfloat) height;
}
if (u_upload_alloc(st->uploader, 0, 4 * sizeof(vertices[0]),
vbuf_offset, vbuf, (void **) &vertices) != PIPE_OK) {
return;
}
/* Positions are in clip coords since we need to do clipping in case
* the bitmap quad goes beyond the window bounds.
*/
vertices[0][0][0] = clip_x0;
vertices[0][0][1] = clip_y0;
vertices[0][2][0] = sLeft;
vertices[0][2][1] = tTop;
vertices[1][0][0] = clip_x1;
vertices[1][0][1] = clip_y0;
vertices[1][2][0] = sRight;
vertices[1][2][1] = tTop;
vertices[2][0][0] = clip_x1;
vertices[2][0][1] = clip_y1;
vertices[2][2][0] = sRight;
vertices[2][2][1] = tBot;
vertices[3][0][0] = clip_x0;
vertices[3][0][1] = clip_y1;
vertices[3][2][0] = sLeft;
vertices[3][2][1] = tBot;
/* same for all verts: */
for (i = 0; i < 4; i++) {
vertices[i][0][2] = z;
vertices[i][0][3] = 1.0f;
vertices[i][1][0] = color[0];
vertices[i][1][1] = color[1];
vertices[i][1][2] = color[2];
vertices[i][1][3] = color[3];
vertices[i][2][2] = 0.0; /*R*/
vertices[i][2][3] = 1.0; /*Q*/
}
u_upload_unmap(st->uploader);
}
static bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
{
struct si_descriptors *desc = &sctx->vertex_buffers;
bool bound[SI_NUM_VERTEX_BUFFERS] = {};
unsigned i, count = sctx->vertex_elements->count;
uint64_t va;
uint32_t *ptr;
if (!sctx->vertex_buffers_dirty)
return true;
if (!count || !sctx->vertex_elements)
return true;
/* Vertex buffer descriptors are the only ones which are uploaded
* directly through a staging buffer and don't go through
* the fine-grained upload path.
*/
u_upload_alloc(sctx->b.uploader, 0, count * 16, &desc->buffer_offset,
(struct pipe_resource**)&desc->buffer, (void**)&ptr);
if (!desc->buffer)
return false;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
desc->buffer, RADEON_USAGE_READ,
RADEON_PRIO_DESCRIPTORS);
assert(count <= SI_NUM_VERTEX_BUFFERS);
for (i = 0; i < count; i++) {
struct pipe_vertex_element *ve = &sctx->vertex_elements->elements[i];
struct pipe_vertex_buffer *vb;
struct r600_resource *rbuffer;
unsigned offset;
uint32_t *desc = &ptr[i*4];
if (ve->vertex_buffer_index >= Elements(sctx->vertex_buffer)) {
memset(desc, 0, 16);
continue;
}
vb = &sctx->vertex_buffer[ve->vertex_buffer_index];
rbuffer = (struct r600_resource*)vb->buffer;
if (!rbuffer) {
memset(desc, 0, 16);
continue;
}
offset = vb->buffer_offset + ve->src_offset;
va = rbuffer->gpu_address + offset;
/* Fill in T# buffer resource description */
desc[0] = va;
desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) |
S_008F04_STRIDE(vb->stride);
if (sctx->b.chip_class <= CIK && vb->stride)
/* Round up by rounding down and adding 1 */
desc[2] = (vb->buffer->width0 - offset -
sctx->vertex_elements->format_size[i]) /
vb->stride + 1;
else
desc[2] = vb->buffer->width0 - offset;
desc[3] = sctx->vertex_elements->rsrc_word3[i];
if (!bound[ve->vertex_buffer_index]) {
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
(struct r600_resource*)vb->buffer,
RADEON_USAGE_READ, RADEON_PRIO_VERTEX_BUFFER);
bound[ve->vertex_buffer_index] = true;
}
}
/* Don't flush the const cache. It would have a very negative effect
* on performance (confirmed by testing). New descriptors are always
* uploaded to a fresh new buffer, so I don't think flushing the const
* cache is needed. */
desc->pointer_dirty = true;
si_mark_atom_dirty(sctx, &sctx->shader_userdata.atom);
sctx->vertex_buffers_dirty = false;
return true;
}
