/* 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;
}
/**
* Write parameter array for the given vertex program into dst.
* Return the total number of components written.
*/
static int r300VertexProgUpdateParams(GLcontext * ctx, struct r300_vertex_program *vp, float *dst)
{
int i;
if (vp->Base->IsNVProgram) {
_mesa_load_tracked_matrices(ctx);
} else {
if (vp->Base->Base.Parameters) {
_mesa_load_state_parameters(ctx, vp->Base->Base.Parameters);
}
}
for(i = 0; i < vp->code.constants.Count; ++i) {
const float * src = 0;
const struct rc_constant * constant = &vp->code.constants.Constants[i];
switch(constant->Type) {
case RC_CONSTANT_EXTERNAL:
if (vp->Base->IsNVProgram) {
src = ctx->VertexProgram.Parameters[constant->u.External];
} else {
src = vp->Base->Base.Parameters->ParameterValues[constant->u.External];
}
break;
case RC_CONSTANT_IMMEDIATE:
src = constant->u.Immediate;
break;
}
dst[4*i] = src[0];
dst[4*i + 1] = src[1];
dst[4*i + 2] = src[2];
dst[4*i + 3] = src[3];
}
return 4 * vp->code.constants.Count;
}
/**
* This function executes vertex programs
*/
static GLboolean
run_vp( GLcontext *ctx, struct tnl_pipeline_stage *stage )
{
TNLcontext *tnl = TNL_CONTEXT(ctx);
struct vp_stage_data *store = VP_STAGE_DATA(stage);
struct vertex_buffer *VB = &tnl->vb;
struct gl_vertex_program *program = ctx->VertexProgram._Current;
struct gl_program_machine machine;
GLuint outputs[VERT_RESULT_MAX], numOutputs;
GLuint i, j;
if (!program)
return GL_TRUE;
if (program->IsNVProgram) {
_mesa_load_tracked_matrices(ctx);
}
else {
/* ARB program or vertex shader */
_mesa_load_state_parameters(ctx, program->Base.Parameters);
}
numOutputs = 0;
for (i = 0; i < VERT_RESULT_MAX; i++) {
if (program->Base.OutputsWritten & (1 << i)) {
outputs[numOutputs++] = i;
}
}
for (i = 0; i < VB->Count; i++) {
GLuint attr;
init_machine(ctx, &machine);
#if 0
printf("Input %d: %f, %f, %f, %f\n", i,
VB->AttribPtr[0]->data[i][0],
VB->AttribPtr[0]->data[i][1],
VB->AttribPtr[0]->data[i][2],
VB->AttribPtr[0]->data[i][3]);
printf(" color: %f, %f, %f, %f\n",
VB->AttribPtr[3]->data[i][0],
VB->AttribPtr[3]->data[i][1],
VB->AttribPtr[3]->data[i][2],
VB->AttribPtr[3]->data[i][3]);
printf(" normal: %f, %f, %f, %f\n",
VB->AttribPtr[2]->data[i][0],
VB->AttribPtr[2]->data[i][1],
VB->AttribPtr[2]->data[i][2],
VB->AttribPtr[2]->data[i][3]);
#endif
/* the vertex array case */
for (attr = 0; attr < VERT_ATTRIB_MAX; attr++) {
if (program->Base.InputsRead & (1 << attr)) {
const GLubyte *ptr = (const GLubyte*) VB->AttribPtr[attr]->data;
const GLuint size = VB->AttribPtr[attr]->size;
const GLuint stride = VB->AttribPtr[attr]->stride;
const GLfloat *data = (GLfloat *) (ptr + stride * i);
COPY_CLEAN_4V(machine.VertAttribs[attr], size, data);
}
}
/* execute the program */
_mesa_execute_program(ctx, &program->Base, &machine);
/* copy the output registers into the VB->attribs arrays */
for (j = 0; j < numOutputs; j++) {
const GLuint attr = outputs[j];
COPY_4V(store->results[attr].data[i], machine.Outputs[attr]);
}
#if 0
printf("HPOS: %f %f %f %f\n",
machine.Outputs[0][0],
machine.Outputs[0][1],
machine.Outputs[0][2],
machine.Outputs[0][3]);
#endif
}
/* Fixup fog and point size results if needed */
if (program->IsNVProgram) {
if (ctx->Fog.Enabled &&
(program->Base.OutputsWritten & (1 << VERT_RESULT_FOGC)) == 0) {
for (i = 0; i < VB->Count; i++) {
store->results[VERT_RESULT_FOGC].data[i][0] = 1.0;
}
}
if (ctx->VertexProgram.PointSizeEnabled &&
(program->Base.OutputsWritten & (1 << VERT_RESULT_PSIZ)) == 0) {
for (i = 0; i < VB->Count; i++) {
store->results[VERT_RESULT_PSIZ].data[i][0] = ctx->Point.Size;
}
}
}
/* Setup the VB pointers so that the next pipeline stages get
* their data from the right place (the program output arrays).
*/
VB->ClipPtr = &store->results[VERT_RESULT_HPOS];
VB->ClipPtr->size = 4;
VB->ClipPtr->count = VB->Count;
VB->ColorPtr[0] = &store->results[VERT_RESULT_COL0];
VB->ColorPtr[1] = &store->results[VERT_RESULT_BFC0];
VB->SecondaryColorPtr[0] = &store->results[VERT_RESULT_COL1];
VB->SecondaryColorPtr[1] = &store->results[VERT_RESULT_BFC1];
VB->FogCoordPtr = &store->results[VERT_RESULT_FOGC];
VB->AttribPtr[VERT_ATTRIB_COLOR0] = &store->results[VERT_RESULT_COL0];
VB->AttribPtr[VERT_ATTRIB_COLOR1] = &store->results[VERT_RESULT_COL1];
VB->AttribPtr[VERT_ATTRIB_FOG] = &store->results[VERT_RESULT_FOGC];
VB->AttribPtr[_TNL_ATTRIB_POINTSIZE] = &store->results[VERT_RESULT_PSIZ];
for (i = 0; i < ctx->Const.MaxTextureCoordUnits; i++) {
VB->TexCoordPtr[i] =
VB->AttribPtr[_TNL_ATTRIB_TEX0 + i]
= &store->results[VERT_RESULT_TEX0 + i];
}
for (i = 0; i < ctx->Const.MaxVarying; i++) {
if (program->Base.OutputsWritten & (1 << (VERT_RESULT_VAR0 + i))) {
/* Note: varying results get put into the generic attributes */
VB->AttribPtr[VERT_ATTRIB_GENERIC0+i]
= &store->results[VERT_RESULT_VAR0 + i];
}
}
/* Cliptest and perspective divide. Clip functions must clear
* the clipmask.
*/
store->ormask = 0;
store->andmask = CLIP_FRUSTUM_BITS;
if (tnl->NeedNdcCoords) {
VB->NdcPtr =
_mesa_clip_tab[VB->ClipPtr->size]( VB->ClipPtr,
&store->ndcCoords,
store->clipmask,
&store->ormask,
&store->andmask );
}
else {
VB->NdcPtr = NULL;
_mesa_clip_np_tab[VB->ClipPtr->size]( VB->ClipPtr,
NULL,
store->clipmask,
&store->ormask,
&store->andmask );
}
if (store->andmask) /* All vertices are outside the frustum */
return GL_FALSE;
/* This is where we'd do clip testing against the user-defined
* clipping planes, but they're not supported by vertex programs.
*/
VB->ClipOrMask = store->ormask;
VB->ClipMask = store->clipmask;
return GL_TRUE;
}
/**
* This function executes vertex programs
*/
static GLboolean
run_vp( struct gl_context *ctx, struct tnl_pipeline_stage *stage )
{
TNLcontext *tnl = TNL_CONTEXT(ctx);
struct vp_stage_data *store = VP_STAGE_DATA(stage);
struct vertex_buffer *VB = &tnl->vb;
struct gl_vertex_program *program = ctx->VertexProgram._Current;
struct gl_program_machine *machine = &store->machine;
GLuint outputs[VERT_RESULT_MAX], numOutputs;
GLuint i, j;
if (!program)
return GL_TRUE;
if (program->IsNVProgram) {
_mesa_load_tracked_matrices(ctx);
}
else {
/* ARB program or vertex shader */
_mesa_load_state_parameters(ctx, program->Base.Parameters);
}
/* make list of outputs to save some time below */
numOutputs = 0;
for (i = 0; i < VERT_RESULT_MAX; i++) {
if (program->Base.OutputsWritten & BITFIELD64_BIT(i)) {
outputs[numOutputs++] = i;
}
}
/* Allocate result vectors. We delay this until now to avoid allocating
* memory that would never be used if we don't run the software tnl pipeline.
*/
if (!store->results[0].storage) {
for (i = 0; i < VERT_RESULT_MAX; i++) {
assert(!store->results[i].storage);
_mesa_vector4f_alloc( &store->results[i], 0, VB->Size, 32 );
store->results[i].size = 4;
}
}
map_textures(ctx, program);
for (i = 0; i < VB->Count; i++) {
GLuint attr;
init_machine(ctx, machine, tnl->CurInstance);
#if 0
printf("Input %d: %f, %f, %f, %f\n", i,
VB->AttribPtr[0]->data[i][0],
VB->AttribPtr[0]->data[i][1],
VB->AttribPtr[0]->data[i][2],
VB->AttribPtr[0]->data[i][3]);
printf(" color: %f, %f, %f, %f\n",
VB->AttribPtr[3]->data[i][0],
VB->AttribPtr[3]->data[i][1],
VB->AttribPtr[3]->data[i][2],
VB->AttribPtr[3]->data[i][3]);
printf(" normal: %f, %f, %f, %f\n",
VB->AttribPtr[2]->data[i][0],
VB->AttribPtr[2]->data[i][1],
VB->AttribPtr[2]->data[i][2],
VB->AttribPtr[2]->data[i][3]);
#endif
/* the vertex array case */
for (attr = 0; attr < VERT_ATTRIB_MAX; attr++) {
if (program->Base.InputsRead & BITFIELD64_BIT(attr)) {
const GLubyte *ptr = (const GLubyte*) VB->AttribPtr[attr]->data;
const GLuint size = VB->AttribPtr[attr]->size;
const GLuint stride = VB->AttribPtr[attr]->stride;
const GLfloat *data = (GLfloat *) (ptr + stride * i);
#ifdef NAN_CHECK
check_float(data[0]);
check_float(data[1]);
check_float(data[2]);
check_float(data[3]);
#endif
COPY_CLEAN_4V(machine->VertAttribs[attr], size, data);
}
}
/* execute the program */
_mesa_execute_program(ctx, &program->Base, machine);
/* copy the output registers into the VB->attribs arrays */
for (j = 0; j < numOutputs; j++) {
const GLuint attr = outputs[j];
#ifdef NAN_CHECK
check_float(machine->Outputs[attr][0]);
check_float(machine->Outputs[attr][1]);
check_float(machine->Outputs[attr][2]);
check_float(machine->Outputs[attr][3]);
#endif
COPY_4V(store->results[attr].data[i], machine->Outputs[attr]);
}
/* FOGC is a special case. Fragment shader expects (f,0,0,1) */
if (program->Base.OutputsWritten & BITFIELD64_BIT(VERT_RESULT_FOGC)) {
store->results[VERT_RESULT_FOGC].data[i][1] = 0.0;
store->results[VERT_RESULT_FOGC].data[i][2] = 0.0;
store->results[VERT_RESULT_FOGC].data[i][3] = 1.0;
}
#ifdef NAN_CHECK
ASSERT(machine->Outputs[0][3] != 0.0F);
#endif
#if 0
printf("HPOS: %f %f %f %f\n",
machine->Outputs[0][0],
machine->Outputs[0][1],
machine->Outputs[0][2],
machine->Outputs[0][3]);
#endif
}
unmap_textures(ctx, program);
/* Fixup fog and point size results if needed */
if (program->IsNVProgram) {
if (ctx->Fog.Enabled &&
(program->Base.OutputsWritten & BITFIELD64_BIT(VERT_RESULT_FOGC)) == 0) {
for (i = 0; i < VB->Count; i++) {
store->results[VERT_RESULT_FOGC].data[i][0] = 1.0;
}
}
if (ctx->VertexProgram.PointSizeEnabled &&
(program->Base.OutputsWritten & BITFIELD64_BIT(VERT_RESULT_PSIZ)) == 0) {
for (i = 0; i < VB->Count; i++) {
store->results[VERT_RESULT_PSIZ].data[i][0] = ctx->Point.Size;
}
}
}
if (program->IsPositionInvariant) {
/* We need the exact same transform as in the fixed function path here
* to guarantee invariance, depending on compiler optimization flags
* results could be different otherwise.
*/
VB->ClipPtr = TransformRaw( &store->results[0],
&ctx->_ModelProjectMatrix,
VB->AttribPtr[0] );
/* Drivers expect this to be clean to element 4...
*/
switch (VB->ClipPtr->size) {
case 1:
/* impossible */
case 2:
_mesa_vector4f_clean_elem( VB->ClipPtr, VB->Count, 2 );
/* fall-through */
case 3:
_mesa_vector4f_clean_elem( VB->ClipPtr, VB->Count, 3 );
/* fall-through */
case 4:
break;
}
}
else {
/* Setup the VB pointers so that the next pipeline stages get
* their data from the right place (the program output arrays).
*/
VB->ClipPtr = &store->results[VERT_RESULT_HPOS];
VB->ClipPtr->size = 4;
VB->ClipPtr->count = VB->Count;
}
VB->AttribPtr[VERT_ATTRIB_COLOR0] = &store->results[VERT_RESULT_COL0];
VB->AttribPtr[VERT_ATTRIB_COLOR1] = &store->results[VERT_RESULT_COL1];
VB->AttribPtr[VERT_ATTRIB_FOG] = &store->results[VERT_RESULT_FOGC];
VB->AttribPtr[_TNL_ATTRIB_POINTSIZE] = &store->results[VERT_RESULT_PSIZ];
VB->BackfaceColorPtr = &store->results[VERT_RESULT_BFC0];
VB->BackfaceSecondaryColorPtr = &store->results[VERT_RESULT_BFC1];
for (i = 0; i < ctx->Const.MaxTextureCoordUnits; i++) {
VB->AttribPtr[_TNL_ATTRIB_TEX0 + i]
= &store->results[VERT_RESULT_TEX0 + i];
}
for (i = 0; i < ctx->Const.MaxVarying; i++) {
if (program->Base.OutputsWritten & BITFIELD64_BIT(VERT_RESULT_VAR0 + i)) {
/* Note: varying results get put into the generic attributes */
VB->AttribPtr[VERT_ATTRIB_GENERIC0+i]
= &store->results[VERT_RESULT_VAR0 + i];
}
}
/* Perform NDC and cliptest operations:
*/
return do_ndc_cliptest(ctx, store);
}
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 prepare_constant_buffer(struct brw_context *brw)
{
GLcontext *ctx = &brw->intel.ctx;
const struct brw_vertex_program *vp =
brw_vertex_program_const(brw->vertex_program);
const struct brw_fragment_program *fp =
brw_fragment_program_const(brw->fragment_program);
const GLuint sz = brw->curbe.total_size;
const GLuint bufsz = sz * 16 * sizeof(GLfloat);
GLfloat *buf;
GLuint i;
if (sz == 0) {
if (brw->curbe.last_buf) {
free(brw->curbe.last_buf);
brw->curbe.last_buf = NULL;
brw->curbe.last_bufsz = 0;
}
return;
}
buf = (GLfloat *) calloc(1, bufsz);
/* fragment shader constants */
if (brw->curbe.wm_size) {
GLuint offset = brw->curbe.wm_start * 16;
_mesa_load_state_parameters(ctx, fp->program.Base.Parameters);
/* copy float constants */
for (i = 0; i < brw->wm.prog_data->nr_params; i++)
buf[offset + 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;
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);
if (vp->use_const_buffer) {
/* 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));
}
}
} else {
for (i = 0; i < nr; i++) {
memcpy(buf + offset + 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 &&
brw->curbe.last_buf &&
bufsz == brw->curbe.last_bufsz &&
memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
/* constants have not changed */
free(buf);
}
else {
/* constants have changed */
if (brw->curbe.last_buf)
free(brw->curbe.last_buf);
brw->curbe.last_buf = buf;
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);
dri_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 = dri_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);
}
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.
*/
}
