int
brw_bo_map_gtt(struct brw_context *brw, drm_intel_bo *bo, const char *bo_name)
{
if (likely(!brw->perf_debug) || !drm_intel_bo_busy(bo))
return drm_intel_gem_bo_map_gtt(bo);
float start_time = get_time();
int ret = drm_intel_gem_bo_map_gtt(bo);
perf_debug("GTT mapping a busy %s BO stalled and took %.03f ms.\n",
bo_name, (get_time() - start_time) * 1000);
return ret;
}
/**
* Interface for getting memory for uploading streamed data to the GPU
*
* In most cases, streamed data (for GPU state structures, for example) is
* uploaded through brw_state_batch(), since that interface allows relocations
* from the streamed space returned to other BOs. However, that interface has
* the restriction that the amount of space allocated has to be "small" (see
* estimated_max_prim_size in brw_draw.c).
*
* This interface, on the other hand, is able to handle arbitrary sized
* allocation requests, though it will batch small allocations into the same
* BO for efficiency and reduced memory footprint.
*
* \note The returned pointer is valid only until intel_upload_finish(), which
* will happen at batch flush or the next
* intel_upload_space()/intel_upload_data().
*
* \param out_bo Pointer to a BO, which must point to a valid BO or NULL on
* entry, and will have a reference to the new BO containing the state on
* return.
*
* \param out_offset Offset within the buffer object that the data will land.
*/
void *
intel_upload_space(struct brw_context *brw,
uint32_t size,
uint32_t alignment,
drm_intel_bo **out_bo,
uint32_t *out_offset)
{
uint32_t offset;
offset = ALIGN_NPOT(brw->upload.next_offset, alignment);
if (brw->upload.bo && offset + size > brw->upload.bo->size) {
intel_upload_finish(brw);
offset = 0;
}
if (!brw->upload.bo) {
brw->upload.bo = drm_intel_bo_alloc(brw->bufmgr, "streamed data",
MAX2(INTEL_UPLOAD_SIZE, size), 4096);
if (brw->has_llc)
drm_intel_bo_map(brw->upload.bo, true);
else
drm_intel_gem_bo_map_gtt(brw->upload.bo);
}
brw->upload.next_offset = offset + size;
*out_offset = offset;
if (*out_bo != brw->upload.bo) {
drm_intel_bo_unreference(*out_bo);
*out_bo = brw->upload.bo;
drm_intel_bo_reference(brw->upload.bo);
}
return brw->upload.bo->virtual + offset;
}
/* Creates a new VS constant buffer reflecting the current VS program's
* constants, if needed by the VS program.
*
* Otherwise, constants go through the CURBEs using the brw_constant_buffer
* state atom.
*/
static void
brw_upload_vs_pull_constants(struct brw_context *brw)
{
struct gl_context *ctx = &brw->intel.ctx;
struct intel_context *intel = &brw->intel;
/* BRW_NEW_VERTEX_PROGRAM */
struct brw_vertex_program *vp =
(struct brw_vertex_program *) brw->vertex_program;
const struct gl_program_parameter_list *params = vp->program.Base.Parameters;
int i;
if (vp->program.IsNVProgram)
_mesa_load_tracked_matrices(ctx);
/* Updates the ParamaterValues[i] pointers for all parameters of the
* basic type of PROGRAM_STATE_VAR.
*/
_mesa_load_state_parameters(&brw->intel.ctx, vp->program.Base.Parameters);
/* CACHE_NEW_VS_PROG */
if (!brw->vs.prog_data->nr_pull_params) {
if (brw->vs.const_bo) {
drm_intel_bo_unreference(brw->vs.const_bo);
brw->vs.const_bo = NULL;
brw->bind.surf_offset[SURF_INDEX_VERT_CONST_BUFFER] = 0;
brw->state.dirty.brw |= BRW_NEW_VS_CONSTBUF;
}
return;
}
/* _NEW_PROGRAM_CONSTANTS */
drm_intel_bo_unreference(brw->vs.const_bo);
brw->vs.const_bo = drm_intel_bo_alloc(intel->bufmgr, "vp_const_buffer",
brw->vs.prog_data->nr_pull_params * 4,
64);
drm_intel_gem_bo_map_gtt(brw->vs.const_bo);
for (i = 0; i < brw->vs.prog_data->nr_pull_params; i++) {
memcpy(brw->vs.const_bo->virtual + i * 4,
brw->vs.prog_data->pull_param[i],
4);
}
if (0) {
for (i = 0; i < params->NumParameters; i++) {
float *row = (float *)brw->vs.const_bo->virtual + i * 4;
printf("vs const surface %3d: %4.3f %4.3f %4.3f %4.3f\n",
i, row[0], row[1], row[2], row[3]);
}
}
drm_intel_gem_bo_unmap_gtt(brw->vs.const_bo);
const int surf = SURF_INDEX_VERT_CONST_BUFFER;
intel->vtbl.create_constant_surface(brw, brw->vs.const_bo,
params->NumParameters,
&brw->bind.surf_offset[surf]);
brw->state.dirty.brw |= BRW_NEW_VS_CONSTBUF;
}
static void *
i915_drm_buffer_map(struct i915_winsys *iws,
struct i915_winsys_buffer *buffer,
boolean write)
{
struct i915_drm_buffer *buf = i915_drm_buffer(buffer);
drm_intel_bo *bo = intel_bo(buffer);
int ret = 0;
assert(bo);
if (buf->map_count)
goto out;
ret = drm_intel_gem_bo_map_gtt(bo);
buf->ptr = bo->virtual;
assert(ret == 0);
out:
if (ret)
return NULL;
buf->map_count++;
return buf->ptr;
}
/* XXX: Thread safety?
*/
void *
intel_region_map(struct intel_context *intel, struct intel_region *region,
GLbitfield mode)
{
/* We have the region->map_refcount controlling mapping of the BO because
* in software fallbacks we may end up mapping the same buffer multiple
* times on Mesa's behalf, so we refcount our mappings to make sure that
* the pointer stays valid until the end of the unmap chain. However, we
* must not emit any batchbuffers between the start of mapping and the end
* of unmapping, or further use of the map will be incoherent with the GPU
* rendering done by that batchbuffer. Hence we assert in
* intel_batchbuffer_flush() that that doesn't happen, which means that the
* flush is only needed on first map of the buffer.
*/
_DBG("%s %p\n", __FUNCTION__, region);
if (!region->map_refcount) {
intel_flush(&intel->ctx);
if (region->tiling != I915_TILING_NONE)
drm_intel_gem_bo_map_gtt(region->bo);
else
drm_intel_bo_map(region->bo, true);
region->map = region->bo->virtual;
}
static int
map_bo(struct buffer *my_buf)
{
if (drm_intel_gem_bo_map_gtt(my_buf->bo) != 0)
return 0;
my_buf->mmap = my_buf->bo->virtual;
return 1;
}
void
brw_upload_vec4_pull_constants(struct brw_context *brw,
GLbitfield brw_new_constbuf,
const struct gl_program *prog,
struct brw_stage_state *stage_state,
const struct brw_vec4_prog_data *prog_data)
{
int i;
uint32_t surf_index = prog_data->base.binding_table.pull_constants_start;
/* Updates the ParamaterValues[i] pointers for all parameters of the
* basic type of PROGRAM_STATE_VAR.
*/
_mesa_load_state_parameters(&brw->ctx, prog->Parameters);
if (!prog_data->nr_pull_params) {
if (stage_state->const_bo) {
drm_intel_bo_unreference(stage_state->const_bo);
stage_state->const_bo = NULL;
stage_state->surf_offset[surf_index] = 0;
brw->state.dirty.brw |= brw_new_constbuf;
}
return;
}
/* _NEW_PROGRAM_CONSTANTS */
drm_intel_bo_unreference(stage_state->const_bo);
uint32_t size = prog_data->nr_pull_params * 4;
stage_state->const_bo = drm_intel_bo_alloc(brw->bufmgr, "vec4_const_buffer",
size, 64);
drm_intel_gem_bo_map_gtt(stage_state->const_bo);
for (i = 0; i < prog_data->nr_pull_params; i++) {
memcpy(stage_state->const_bo->virtual + i * 4,
prog_data->pull_param[i],
4);
}
if (0) {
for (i = 0; i < ALIGN(prog_data->nr_pull_params, 4) / 4; i++) {
float *row = (float *)stage_state->const_bo->virtual + i * 4;
printf("const surface %3d: %4.3f %4.3f %4.3f %4.3f\n",
i, row[0], row[1], row[2], row[3]);
}
}
drm_intel_gem_bo_unmap_gtt(stage_state->const_bo);
brw_create_constant_surface(brw, stage_state->const_bo, 0, size,
&stage_state->surf_offset[surf_index],
false);
brw->state.dirty.brw |= brw_new_constbuf;
}
/**
* Called via glMapBufferARB().
*/
static void *
intel_bufferobj_map(struct gl_context * ctx,
GLenum target,
GLenum access, struct gl_buffer_object *obj)
{
struct intel_context *intel = intel_context(ctx);
struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
GLboolean read_only = (access == GL_READ_ONLY_ARB);
GLboolean write_only = (access == GL_WRITE_ONLY_ARB);
assert(intel_obj);
if (intel_obj->sys_buffer) {
if (!read_only && intel_obj->source) {
release_buffer(intel_obj);
}
if (!intel_obj->buffer || intel_obj->source) {
obj->Pointer = intel_obj->sys_buffer;
obj->Length = obj->Size;
obj->Offset = 0;
return obj->Pointer;
}
free(intel_obj->sys_buffer);
intel_obj->sys_buffer = NULL;
}
/* Flush any existing batchbuffer that might reference this data. */
if (drm_intel_bo_references(intel->batch.bo, intel_obj->buffer))
intel_flush(ctx);
if (intel_obj->region)
intel_bufferobj_cow(intel, intel_obj);
if (intel_obj->buffer == NULL) {
obj->Pointer = NULL;
return NULL;
}
if (write_only) {
drm_intel_gem_bo_map_gtt(intel_obj->buffer);
intel_obj->mapped_gtt = GL_TRUE;
} else {
drm_intel_bo_map(intel_obj->buffer, !read_only);
intel_obj->mapped_gtt = GL_FALSE;
}
obj->Pointer = intel_obj->buffer->virtual;
obj->Length = obj->Size;
obj->Offset = 0;
return obj->Pointer;
}
void *
intel_bo_map_gtt(struct intel_bo *bo)
{
int err;
err = drm_intel_gem_bo_map_gtt(gem_bo(bo));
if (err) {
debug_error("failed to map bo");
return NULL;
}
return gem_bo(bo)->virtual;
}
void intelFlushBatch(void)
{
dri_bo *bo;
i965_end_batch();
drm_intel_bo_exec(xvmc_driver->batch.buf,
xvmc_driver->batch.ptr - xvmc_driver->batch.init_ptr,
0, 0, 0);
bo = drm_intel_bo_alloc(xvmc_driver->bufmgr,
"batch buffer", BATCH_SIZE, 0x1000);
if (bo != NULL && drm_intel_gem_bo_map_gtt(bo) == 0) {
drm_intel_bo_unreference(xvmc_driver->batch.buf);
xvmc_driver->batch.buf = bo;
} else {
if (bo != NULL)
drm_intel_bo_unreference(bo);
drm_intel_gem_bo_map_gtt(xvmc_driver->batch.buf);
}
reset_batch();
}
Bool intelInitBatchBuffer(void)
{
if ((xvmc_driver->batch.buf =
drm_intel_bo_alloc(xvmc_driver->bufmgr,
"batch buffer", BATCH_SIZE, 0x1000)) == NULL) {
fprintf(stderr, "unable to alloc batch buffer\n");
return False;
}
if (drm_intel_gem_bo_map_gtt(xvmc_driver->batch.buf)) {
drm_intel_bo_unreference(xvmc_driver->batch.buf);
return False;
}
reset_batch();
return True;
}
/* XXX: Thread safety?
*/
GLubyte *
intel_region_map(struct intel_context *intel, struct intel_region *region)
{
intelFlush(&intel->ctx);
_DBG("%s %p\n", __FUNCTION__, region);
if (!region->map_refcount++) {
if (region->pbo)
intel_region_cow(intel, region);
if (region->tiling != I915_TILING_NONE &&
intel->intelScreen->kernel_exec_fencing)
drm_intel_gem_bo_map_gtt(region->buffer);
else
dri_bo_map(region->buffer, GL_TRUE);
region->map = region->buffer->virtual;
}
return region->map;
}
void *
intel_miptree_map_raw(struct intel_context *intel, struct intel_mipmap_tree *mt)
{
drm_intel_bo *bo = mt->region->bo;
if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
if (drm_intel_bo_busy(bo)) {
perf_debug("Mapping a busy BO, causing a stall on the GPU.\n");
}
}
intel_flush(&intel->ctx);
if (mt->region->tiling != I915_TILING_NONE)
drm_intel_gem_bo_map_gtt(bo);
else
drm_intel_bo_map(bo, true);
return bo->virtual;
}
void
intel_renderbuffer_map(struct intel_context *intel, struct gl_renderbuffer *rb)
{
struct intel_renderbuffer *irb = intel_renderbuffer(rb);
if (irb == NULL || irb->region == NULL)
return;
drm_intel_gem_bo_map_gtt(irb->region->buffer);
rb->Data = irb->region->buffer->virtual;
rb->RowStride = irb->region->pitch;
/* Flip orientation if it's the window system buffer */
if (!rb->Name) {
rb->Data += rb->RowStride * (irb->region->height - 1) * irb->region->cpp;
rb->RowStride = -rb->RowStride;
}
intel_set_span_functions(intel, rb);
}
static void init_buffer(struct scratch_buf *buf, unsigned size)
{
buf->bo = drm_intel_bo_alloc(bufmgr, "tiled bo", size, 4096);
assert(buf->bo);
buf->tiling = I915_TILING_NONE;
buf->stride = 4096;
sanitize_stride(buf);
if (options.no_hw)
buf->data = malloc(size);
else {
if (options.use_cpu_maps)
drm_intel_bo_map(buf->bo, 1);
else
drm_intel_gem_bo_map_gtt(buf->bo);
buf->data = buf->bo->virtual;
}
buf->num_tiles = options.tiles_per_buf;
}
static void
copy_array_to_vbo_array( struct brw_context *brw,
struct brw_vertex_element *element,
GLuint dst_stride)
{
struct intel_context *intel = &brw->intel;
GLuint size = element->count * dst_stride;
get_space(brw, size, &element->bo, &element->offset);
if (element->glarray->StrideB == 0) {
assert(element->count == 1);
element->stride = 0;
} else {
element->stride = dst_stride;
}
if (dst_stride == element->glarray->StrideB) {
if (intel->intelScreen->kernel_exec_fencing) {
drm_intel_gem_bo_map_gtt(element->bo);
memcpy((char *)element->bo->virtual + element->offset,
element->glarray->Ptr, size);
drm_intel_gem_bo_unmap_gtt(element->bo);
} else {
/**
* Called via glMapBufferRange().
*
* The goal of this extension is to allow apps to accumulate their rendering
* at the same time as they accumulate their buffer object. Without it,
* you'd end up blocking on execution of rendering every time you mapped
* the buffer to put new data in.
*
* We support it in 3 ways: If unsynchronized, then don't bother
* flushing the batchbuffer before mapping the buffer, which can save blocking
* in many cases. If we would still block, and they allow the whole buffer
* to be invalidated, then just allocate a new buffer to replace the old one.
* If not, and we'd block, and they allow the subrange of the buffer to be
* invalidated, then we can make a new little BO, let them write into that,
* and blit it into the real BO at unmap time.
*/
static void *
intel_bufferobj_map_range(GLcontext * ctx,
GLenum target, GLintptr offset, GLsizeiptr length,
GLbitfield access, struct gl_buffer_object *obj)
{
struct intel_context *intel = intel_context(ctx);
struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
assert(intel_obj);
/* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
* internally uses our functions directly.
*/
obj->Offset = offset;
obj->Length = length;
obj->AccessFlags = access;
if (intel_obj->sys_buffer) {
obj->Pointer = intel_obj->sys_buffer + offset;
return obj->Pointer;
}
if (intel_obj->region)
intel_bufferobj_cow(intel, intel_obj);
/* If the mapping is synchronized with other GL operations, flush
* the batchbuffer so that GEM knows about the buffer access for later
* syncing.
*/
if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
drm_intel_bo_references(intel->batch->buf, intel_obj->buffer))
intelFlush(ctx);
if (intel_obj->buffer == NULL) {
obj->Pointer = NULL;
return NULL;
}
/* If the user doesn't care about existing buffer contents and mapping
* would cause us to block, then throw out the old buffer.
*/
if (!(access & GL_MAP_UNSYNCHRONIZED_BIT) &&
(access & GL_MAP_INVALIDATE_BUFFER_BIT) &&
drm_intel_bo_busy(intel_obj->buffer)) {
drm_intel_bo_unreference(intel_obj->buffer);
intel_obj->buffer = dri_bo_alloc(intel->bufmgr, "bufferobj",
intel_obj->Base.Size, 64);
}
/* If the user is mapping a range of an active buffer object but
* doesn't require the current contents of that range, make a new
* BO, and we'll copy what they put in there out at unmap or
* FlushRange time.
*/
if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
drm_intel_bo_busy(intel_obj->buffer)) {
if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
intel_obj->range_map_buffer = _mesa_malloc(length);
obj->Pointer = intel_obj->range_map_buffer;
} else {
intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
"range map",
length, 64);
if (!(access & GL_MAP_READ_BIT) &&
intel->intelScreen->kernel_exec_fencing) {
drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
intel_obj->mapped_gtt = GL_TRUE;
} else {
drm_intel_bo_map(intel_obj->range_map_bo,
(access & GL_MAP_WRITE_BIT) != 0);
intel_obj->mapped_gtt = GL_FALSE;
}
obj->Pointer = intel_obj->range_map_bo->virtual;
}
return obj->Pointer;
}
if (!(access & GL_MAP_READ_BIT) &&
intel->intelScreen->kernel_exec_fencing) {
drm_intel_gem_bo_map_gtt(intel_obj->buffer);
intel_obj->mapped_gtt = GL_TRUE;
} else {
drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
intel_obj->mapped_gtt = GL_FALSE;
}
obj->Pointer = intel_obj->buffer->virtual + offset;
return obj->Pointer;
}
/* Upload a new set of constants. Too much variability to go into the
* cache mechanism, but maybe would benefit from a comparison against
* the current uploaded set of constants.
*/
static void prepare_constant_buffer(struct brw_context *brw)
{
struct gl_context *ctx = &brw->intel.ctx;
const struct brw_vertex_program *vp =
brw_vertex_program_const(brw->vertex_program);
const GLuint sz = brw->curbe.total_size;
const GLuint bufsz = sz * 16 * sizeof(GLfloat);
GLfloat *buf;
GLuint i;
if (sz == 0) {
brw->curbe.last_bufsz = 0;
return;
}
buf = brw->curbe.next_buf;
/* fragment shader constants */
if (brw->curbe.wm_size) {
GLuint offset = brw->curbe.wm_start * 16;
/* copy float constants */
for (i = 0; i < brw->wm.prog_data->nr_params; i++) {
buf[offset + i] = convert_param(brw->wm.prog_data->param_convert[i],
*brw->wm.prog_data->param[i]);
}
}
/* The clipplanes are actually delivered to both CLIP and VS units.
* VS uses them to calculate the outcode bitmasks.
*/
if (brw->curbe.clip_size) {
GLuint offset = brw->curbe.clip_start * 16;
GLuint j;
/* If any planes are going this way, send them all this way:
*/
for (i = 0; i < 6; i++) {
buf[offset + i * 4 + 0] = fixed_plane[i][0];
buf[offset + i * 4 + 1] = fixed_plane[i][1];
buf[offset + i * 4 + 2] = fixed_plane[i][2];
buf[offset + i * 4 + 3] = fixed_plane[i][3];
}
/* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
* clip-space:
*/
assert(MAX_CLIP_PLANES == 6);
for (j = 0; j < MAX_CLIP_PLANES; j++) {
if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
buf[offset + i * 4 + 0] = ctx->Transform._ClipUserPlane[j][0];
buf[offset + i * 4 + 1] = ctx->Transform._ClipUserPlane[j][1];
buf[offset + i * 4 + 2] = ctx->Transform._ClipUserPlane[j][2];
buf[offset + i * 4 + 3] = ctx->Transform._ClipUserPlane[j][3];
i++;
}
}
}
/* vertex shader constants */
if (brw->curbe.vs_size) {
GLuint offset = brw->curbe.vs_start * 16;
GLuint nr = brw->vs.prog_data->nr_params / 4;
/* Load the subset of push constants that will get used when
* we also have a pull constant buffer.
*/
for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) {
if (brw->vs.constant_map[i] != -1) {
assert(brw->vs.constant_map[i] <= nr);
memcpy(buf + offset + brw->vs.constant_map[i] * 4,
vp->program.Base.Parameters->ParameterValues[i],
4 * sizeof(float));
}
}
}
if (0) {
for (i = 0; i < sz*16; i+=4)
printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
printf("last_buf %p buf %p sz %d/%d cmp %d\n",
brw->curbe.last_buf, buf,
bufsz, brw->curbe.last_bufsz,
brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
}
if (brw->curbe.curbe_bo != NULL &&
bufsz == brw->curbe.last_bufsz &&
memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
/* constants have not changed */
} else {
/* Update the record of what our last set of constants was. We
* don't just flip the pointers because we don't fill in the
* data in the padding between the entries.
*/
memcpy(brw->curbe.last_buf, buf, bufsz);
brw->curbe.last_bufsz = bufsz;
if (brw->curbe.curbe_bo != NULL &&
brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size)
{
drm_intel_gem_bo_unmap_gtt(brw->curbe.curbe_bo);
drm_intel_bo_unreference(brw->curbe.curbe_bo);
brw->curbe.curbe_bo = NULL;
}
if (brw->curbe.curbe_bo == NULL) {
/* Allocate a single page for CURBE entries for this batchbuffer.
* They're generally around 64b.
*/
brw->curbe.curbe_bo = drm_intel_bo_alloc(brw->intel.bufmgr, "CURBE",
4096, 1 << 6);
brw->curbe.curbe_next_offset = 0;
drm_intel_gem_bo_map_gtt(brw->curbe.curbe_bo);
assert(bufsz < 4096);
}
brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
brw->curbe.curbe_next_offset += bufsz;
brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
/* Copy data to the buffer:
*/
memcpy(brw->curbe.curbe_bo->virtual + brw->curbe.curbe_offset,
buf,
bufsz);
}
brw_add_validated_bo(brw, brw->curbe.curbe_bo);
/* Because this provokes an action (ie copy the constants into the
* URB), it shouldn't be shortcircuited if identical to the
* previous time - because eg. the urb destination may have
* changed, or the urb contents different to last time.
*
* Note that the data referred to is actually copied internally,
* not just used in place according to passed pointer.
*
* It appears that the CS unit takes care of using each available
* URB entry (Const URB Entry == CURBE) in turn, and issuing
* flushes as necessary when doublebuffering of CURBEs isn't
* possible.
*/
}
/* Ensure the supplied VA surface has valid storage for decoding the
current picture */
VAStatus
avc_ensure_surface_bo(
VADriverContextP ctx,
struct decode_state *decode_state,
struct object_surface *obj_surface,
const VAPictureParameterBufferH264 *pic_param
)
{
VAStatus va_status;
uint32_t hw_fourcc, fourcc, subsample, chroma_format;
/* Validate chroma format */
switch (pic_param->seq_fields.bits.chroma_format_idc) {
case 0: // Grayscale
fourcc = VA_FOURCC_Y800;
subsample = SUBSAMPLE_YUV400;
chroma_format = VA_RT_FORMAT_YUV400;
break;
case 1: // YUV 4:2:0
fourcc = VA_FOURCC_NV12;
subsample = SUBSAMPLE_YUV420;
chroma_format = VA_RT_FORMAT_YUV420;
break;
default:
return VA_STATUS_ERROR_UNSUPPORTED_RT_FORMAT;
}
/* Determine the HW surface format, bound to VA config needs */
if ((decode_state->base.chroma_formats & chroma_format) == chroma_format)
hw_fourcc = fourcc;
else {
hw_fourcc = 0;
switch (fourcc) {
case VA_FOURCC_Y800: // Implement with an NV12 surface
if (decode_state->base.chroma_formats & VA_RT_FORMAT_YUV420) {
hw_fourcc = VA_FOURCC_NV12;
subsample = SUBSAMPLE_YUV420;
}
break;
}
}
if (!hw_fourcc)
return VA_STATUS_ERROR_UNSUPPORTED_RT_FORMAT;
/* (Re-)allocate the underlying surface buffer store, if necessary */
if (!obj_surface->bo || obj_surface->fourcc != hw_fourcc) {
struct i965_driver_data * const i965 = i965_driver_data(ctx);
i965_destroy_surface_storage(obj_surface);
va_status = i965_check_alloc_surface_bo(ctx, obj_surface,
i965->codec_info->has_tiled_surface, hw_fourcc, subsample);
if (va_status != VA_STATUS_SUCCESS)
return va_status;
}
/* Fake chroma components if grayscale is implemented on top of NV12 */
if (fourcc == VA_FOURCC_Y800 && hw_fourcc == VA_FOURCC_NV12) {
const uint32_t uv_offset = obj_surface->width * obj_surface->height;
const uint32_t uv_size = obj_surface->width * obj_surface->height / 2;
drm_intel_gem_bo_map_gtt(obj_surface->bo);
memset(obj_surface->bo->virtual + uv_offset, 0x80, uv_size);
drm_intel_gem_bo_unmap_gtt(obj_surface->bo);
}
return VA_STATUS_SUCCESS;
}
/* Copy BitBlt
*/
GLboolean
intelEmitCopyBlit(struct intel_context *intel,
GLuint cpp,
GLshort src_pitch,
dri_bo *src_buffer,
GLuint src_offset,
uint32_t src_tiling,
GLshort dst_pitch,
dri_bo *dst_buffer,
GLuint dst_offset,
uint32_t dst_tiling,
GLshort src_x, GLshort src_y,
GLshort dst_x, GLshort dst_y,
GLshort w, GLshort h,
GLenum logic_op)
{
GLuint CMD, BR13, pass = 0;
int dst_y2 = dst_y + h;
int dst_x2 = dst_x + w;
dri_bo *aper_array[3];
BATCH_LOCALS;
/* Blits are in a different ringbuffer so we don't use them. */
if (intel->gen >= 6)
return GL_FALSE;
if (dst_tiling != I915_TILING_NONE) {
if (dst_offset & 4095)
return GL_FALSE;
if (dst_tiling == I915_TILING_Y)
return GL_FALSE;
}
if (src_tiling != I915_TILING_NONE) {
if (src_offset & 4095)
return GL_FALSE;
if (src_tiling == I915_TILING_Y)
return GL_FALSE;
}
/* do space check before going any further */
do {
aper_array[0] = intel->batch->buf;
aper_array[1] = dst_buffer;
aper_array[2] = src_buffer;
if (dri_bufmgr_check_aperture_space(aper_array, 3) != 0) {
intel_batchbuffer_flush(intel->batch);
pass++;
} else
break;
} while (pass < 2);
if (pass >= 2) {
drm_intel_gem_bo_map_gtt(dst_buffer);
drm_intel_gem_bo_map_gtt(src_buffer);
_mesa_copy_rect((GLubyte *)dst_buffer->virtual + dst_offset,
cpp,
dst_pitch,
dst_x, dst_y,
w, h,
(GLubyte *)src_buffer->virtual + src_offset,
src_pitch,
src_x, src_y);
drm_intel_gem_bo_unmap_gtt(src_buffer);
drm_intel_gem_bo_unmap_gtt(dst_buffer);
return GL_TRUE;
}
static void
upload_vs_state(struct brw_context *brw)
{
struct intel_context *intel = &brw->intel;
GLcontext *ctx = &intel->ctx;
const struct brw_vertex_program *vp =
brw_vertex_program_const(brw->vertex_program);
unsigned int nr_params = vp->program.Base.Parameters->NumParameters;
drm_intel_bo *constant_bo;
int i;
if (vp->use_const_buffer || nr_params == 0) {
/* Disable the push constant buffers. */
BEGIN_BATCH(5);
OUT_BATCH(CMD_3D_CONSTANT_VS_STATE << 16 | (5 - 2));
OUT_BATCH(0);
OUT_BATCH(0);
OUT_BATCH(0);
OUT_BATCH(0);
ADVANCE_BATCH();
} else {
if (brw->vertex_program->IsNVProgram)
_mesa_load_tracked_matrices(ctx);
/* Updates the ParamaterValues[i] pointers for all parameters of the
* basic type of PROGRAM_STATE_VAR.
*/
_mesa_load_state_parameters(ctx, vp->program.Base.Parameters);
constant_bo = drm_intel_bo_alloc(intel->bufmgr, "VS constant_bo",
nr_params * 4 * sizeof(float),
4096);
drm_intel_gem_bo_map_gtt(constant_bo);
for (i = 0; i < nr_params; i++) {
memcpy((char *)constant_bo->virtual + i * 4 * sizeof(float),
vp->program.Base.Parameters->ParameterValues[i],
4 * sizeof(float));
}
drm_intel_gem_bo_unmap_gtt(constant_bo);
BEGIN_BATCH(5);
OUT_BATCH(CMD_3D_CONSTANT_VS_STATE << 16 |
GEN6_CONSTANT_BUFFER_0_ENABLE |
(5 - 2));
OUT_RELOC(constant_bo,
I915_GEM_DOMAIN_RENDER, 0, /* XXX: bad domain */
ALIGN(nr_params, 2) / 2 - 1);
OUT_BATCH(0);
OUT_BATCH(0);
OUT_BATCH(0);
ADVANCE_BATCH();
drm_intel_bo_unreference(constant_bo);
}
intel_batchbuffer_emit_mi_flush(intel->batch);
BEGIN_BATCH(6);
OUT_BATCH(CMD_3D_VS_STATE << 16 | (6 - 2));
OUT_RELOC(brw->vs.prog_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
OUT_BATCH((0 << GEN6_VS_SAMPLER_COUNT_SHIFT) |
(brw->vs.nr_surfaces << GEN6_VS_BINDING_TABLE_ENTRY_COUNT_SHIFT));
OUT_BATCH(0); /* scratch space base offset */
OUT_BATCH((1 << GEN6_VS_DISPATCH_START_GRF_SHIFT) |
(brw->vs.prog_data->urb_read_length << GEN6_VS_URB_READ_LENGTH_SHIFT) |
(0 << GEN6_VS_URB_ENTRY_READ_OFFSET_SHIFT));
OUT_BATCH((0 << GEN6_VS_MAX_THREADS_SHIFT) |
GEN6_VS_STATISTICS_ENABLE);
ADVANCE_BATCH();
intel_batchbuffer_emit_mi_flush(intel->batch);
}
/* Upload a new set of constants. Too much variability to go into the
* cache mechanism, but maybe would benefit from a comparison against
* the current uploaded set of constants.
*/
static void
brw_upload_constant_buffer(struct brw_context *brw)
{
struct intel_context *intel = &brw->intel;
struct gl_context *ctx = &intel->ctx;
const struct brw_vertex_program *vp =
brw_vertex_program_const(brw->vertex_program);
const GLuint sz = brw->curbe.total_size;
const GLuint bufsz = sz * 16 * sizeof(GLfloat);
GLfloat *buf;
GLuint i;
gl_clip_plane *clip_planes;
if (sz == 0) {
brw->curbe.last_bufsz = 0;
goto emit;
}
buf = brw->curbe.next_buf;
/* fragment shader constants */
if (brw->curbe.wm_size) {
GLuint offset = brw->curbe.wm_start * 16;
/* copy float constants */
for (i = 0; i < brw->wm.prog_data->nr_params; i++) {
buf[offset + i] = convert_param(brw->wm.prog_data->param_convert[i],
brw->wm.prog_data->param[i]);
}
}
/* When using the old VS backend, the clipplanes are actually delivered to
* both CLIP and VS units. VS uses them to calculate the outcode bitmasks.
*
* When using the new VS backend, it is responsible for setting up its own
* clipplane constants if it needs them. This results in a slight waste of
* of curbe space, but the advantage is that the new VS backend can use its
* general-purpose uniform layout code to store the clipplanes.
*/
if (brw->curbe.clip_size) {
GLuint offset = brw->curbe.clip_start * 16;
GLuint j;
/* If any planes are going this way, send them all this way:
*/
for (i = 0; i < 6; i++) {
buf[offset + i * 4 + 0] = fixed_plane[i][0];
buf[offset + i * 4 + 1] = fixed_plane[i][1];
buf[offset + i * 4 + 2] = fixed_plane[i][2];
buf[offset + i * 4 + 3] = fixed_plane[i][3];
}
/* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
* clip-space:
*/
clip_planes = brw_select_clip_planes(ctx);
for (j = 0; j < MAX_CLIP_PLANES; j++) {
if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
buf[offset + i * 4 + 0] = clip_planes[j][0];
buf[offset + i * 4 + 1] = clip_planes[j][1];
buf[offset + i * 4 + 2] = clip_planes[j][2];
buf[offset + i * 4 + 3] = clip_planes[j][3];
i++;
}
}
}
/* vertex shader constants */
if (brw->curbe.vs_size) {
GLuint offset = brw->curbe.vs_start * 16;
GLuint nr = brw->vs.prog_data->nr_params / 4;
if (brw->vs.prog_data->uses_new_param_layout) {
for (i = 0; i < brw->vs.prog_data->nr_params; i++) {
buf[offset + i] = *brw->vs.prog_data->param[i];
}
} else {
/* Load the subset of push constants that will get used when
* we also have a pull constant buffer.
*/
for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) {
if (brw->vs.constant_map[i] != -1) {
assert(brw->vs.constant_map[i] <= nr);
memcpy(buf + offset + brw->vs.constant_map[i] * 4,
vp->program.Base.Parameters->ParameterValues[i],
4 * sizeof(float));
}
}
}
}
if (0) {
for (i = 0; i < sz*16; i+=4)
printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
printf("last_buf %p buf %p sz %d/%d cmp %d\n",
brw->curbe.last_buf, buf,
bufsz, brw->curbe.last_bufsz,
brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
}
if (brw->curbe.curbe_bo != NULL &&
bufsz == brw->curbe.last_bufsz &&
memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
/* constants have not changed */
} else {
/* Update the record of what our last set of constants was. We
* don't just flip the pointers because we don't fill in the
* data in the padding between the entries.
*/
memcpy(brw->curbe.last_buf, buf, bufsz);
brw->curbe.last_bufsz = bufsz;
if (brw->curbe.curbe_bo != NULL &&
brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size)
{
drm_intel_gem_bo_unmap_gtt(brw->curbe.curbe_bo);
drm_intel_bo_unreference(brw->curbe.curbe_bo);
brw->curbe.curbe_bo = NULL;
}
if (brw->curbe.curbe_bo == NULL) {
/* Allocate a single page for CURBE entries for this batchbuffer.
* They're generally around 64b.
*/
brw->curbe.curbe_bo = drm_intel_bo_alloc(brw->intel.bufmgr, "CURBE",
4096, 1 << 6);
brw->curbe.curbe_next_offset = 0;
drm_intel_gem_bo_map_gtt(brw->curbe.curbe_bo);
assert(bufsz < 4096);
}
brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
brw->curbe.curbe_next_offset += bufsz;
brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
/* Copy data to the buffer:
*/
memcpy(brw->curbe.curbe_bo->virtual + brw->curbe.curbe_offset,
buf,
bufsz);
}
/* Because this provokes an action (ie copy the constants into the
* URB), it shouldn't be shortcircuited if identical to the
* previous time - because eg. the urb destination may have
* changed, or the urb contents different to last time.
*
* Note that the data referred to is actually copied internally,
* not just used in place according to passed pointer.
*
* It appears that the CS unit takes care of using each available
* URB entry (Const URB Entry == CURBE) in turn, and issuing
* flushes as necessary when doublebuffering of CURBEs isn't
* possible.
*/
emit:
BEGIN_BATCH(2);
if (brw->curbe.total_size == 0) {
OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
OUT_BATCH(0);
} else {
OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
OUT_RELOC(brw->curbe.curbe_bo,
I915_GEM_DOMAIN_INSTRUCTION, 0,
(brw->curbe.total_size - 1) + brw->curbe.curbe_offset);
}
ADVANCE_BATCH();
}
/**
* Called via glMapBufferRange and glMapBuffer
*
* The goal of this extension is to allow apps to accumulate their rendering
* at the same time as they accumulate their buffer object. Without it,
* you'd end up blocking on execution of rendering every time you mapped
* the buffer to put new data in.
*
* We support it in 3 ways: If unsynchronized, then don't bother
* flushing the batchbuffer before mapping the buffer, which can save blocking
* in many cases. If we would still block, and they allow the whole buffer
* to be invalidated, then just allocate a new buffer to replace the old one.
* If not, and we'd block, and they allow the subrange of the buffer to be
* invalidated, then we can make a new little BO, let them write into that,
* and blit it into the real BO at unmap time.
*/
static void *
intel_bufferobj_map_range(struct gl_context * ctx,
GLintptr offset, GLsizeiptr length,
GLbitfield access, struct gl_buffer_object *obj)
{
struct intel_context *intel = intel_context(ctx);
struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
assert(intel_obj);
/* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
* internally uses our functions directly.
*/
obj->Offset = offset;
obj->Length = length;
obj->AccessFlags = access;
if (intel_obj->sys_buffer) {
const bool read_only =
(access & (GL_MAP_READ_BIT | GL_MAP_WRITE_BIT)) == GL_MAP_READ_BIT;
if (!read_only && intel_obj->source)
release_buffer(intel_obj);
if (!intel_obj->buffer || intel_obj->source) {
obj->Pointer = intel_obj->sys_buffer + offset;
return obj->Pointer;
}
free(intel_obj->sys_buffer);
intel_obj->sys_buffer = NULL;
}
if (intel_obj->buffer == NULL) {
obj->Pointer = NULL;
return NULL;
}
/* If the access is synchronized (like a normal buffer mapping), then get
* things flushed out so the later mapping syncs appropriately through GEM.
* If the user doesn't care about existing buffer contents and mapping would
* cause us to block, then throw out the old buffer.
*
* If they set INVALIDATE_BUFFER, we can pitch the current contents to
* achieve the required synchronization.
*/
if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
if (drm_intel_bo_references(intel->batch.bo, intel_obj->buffer)) {
if (access & GL_MAP_INVALIDATE_BUFFER_BIT) {
drm_intel_bo_unreference(intel_obj->buffer);
intel_bufferobj_alloc_buffer(intel, intel_obj);
} else {
perf_debug("Stalling on the GPU for mapping a busy buffer "
"object\n");
intel_flush(ctx);
}
} else if (drm_intel_bo_busy(intel_obj->buffer) &&
(access & GL_MAP_INVALIDATE_BUFFER_BIT)) {
drm_intel_bo_unreference(intel_obj->buffer);
intel_bufferobj_alloc_buffer(intel, intel_obj);
}
}
/* If the user is mapping a range of an active buffer object but
* doesn't require the current contents of that range, make a new
* BO, and we'll copy what they put in there out at unmap or
* FlushRange time.
*/
if ((access & GL_MAP_INVALIDATE_RANGE_BIT) &&
drm_intel_bo_busy(intel_obj->buffer)) {
if (access & GL_MAP_FLUSH_EXPLICIT_BIT) {
intel_obj->range_map_buffer = malloc(length);
obj->Pointer = intel_obj->range_map_buffer;
} else {
intel_obj->range_map_bo = drm_intel_bo_alloc(intel->bufmgr,
"range map",
length, 64);
if (!(access & GL_MAP_READ_BIT)) {
drm_intel_gem_bo_map_gtt(intel_obj->range_map_bo);
} else {
drm_intel_bo_map(intel_obj->range_map_bo,
(access & GL_MAP_WRITE_BIT) != 0);
}
obj->Pointer = intel_obj->range_map_bo->virtual;
}
return obj->Pointer;
}
if (access & GL_MAP_UNSYNCHRONIZED_BIT)
drm_intel_gem_bo_map_unsynchronized(intel_obj->buffer);
else if (!(access & GL_MAP_READ_BIT)) {
drm_intel_gem_bo_map_gtt(intel_obj->buffer);
} else {
drm_intel_bo_map(intel_obj->buffer, (access & GL_MAP_WRITE_BIT) != 0);
}
obj->Pointer = intel_obj->buffer->virtual + offset;
return obj->Pointer;
}
