static GLboolean brwProgramStringNotify( GLcontext *ctx,
GLenum target,
struct gl_program *prog )
{
struct brw_context *brw = brw_context(ctx);
int i;
if (target == GL_FRAGMENT_PROGRAM_ARB) {
struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
struct brw_fragment_program *newFP = brw_fragment_program(fprog);
const struct brw_fragment_program *curFP =
brw_fragment_program_const(brw->fragment_program);
if (fprog->FogOption) {
_mesa_append_fog_code(ctx, fprog);
fprog->FogOption = GL_NONE;
}
if (newFP == curFP)
brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
newFP->id = brw->program_id++;
newFP->isGLSL = brw_wm_is_glsl(fprog);
}
else if (target == GL_VERTEX_PROGRAM_ARB) {
struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
struct brw_vertex_program *newVP = brw_vertex_program(vprog);
const struct brw_vertex_program *curVP =
brw_vertex_program_const(brw->vertex_program);
if (newVP == curVP)
brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
if (newVP->program.IsPositionInvariant) {
_mesa_insert_mvp_code(ctx, &newVP->program);
}
newVP->id = brw->program_id++;
/* Also tell tnl about it:
*/
_tnl_program_string(ctx, target, prog);
}
/* Reject programs with subroutines, which are totally broken at the moment
* (all program flows return when any program flow returns, and
* the VS also hangs if a function call calls a function.
*
* See piglit glsl-{vs,fs}-functions-[23] tests.
*/
for (i = 0; i < prog->NumInstructions; i++) {
if (prog->Instructions[i].Opcode == OPCODE_CAL) {
shader_error(ctx, prog,
"i965 driver doesn't yet support uninlined function "
"calls. Move to using a single return statement at "
"the end of the function to work around it.");
return GL_FALSE;
}
}
return GL_TRUE;
}
static void brwProgramStringNotify( GLcontext *ctx,
GLenum target,
struct gl_program *prog )
{
struct brw_context *brw = brw_context(ctx);
if (target == GL_FRAGMENT_PROGRAM_ARB) {
struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
struct brw_fragment_program *newFP = brw_fragment_program(fprog);
const struct brw_fragment_program *curFP =
brw_fragment_program_const(brw->fragment_program);
if (fprog->FogOption) {
_mesa_append_fog_code(ctx, fprog);
fprog->FogOption = GL_NONE;
}
if (newFP == curFP)
brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
newFP->id = brw->program_id++;
newFP->isGLSL = brw_wm_is_glsl(fprog);
}
else if (target == GL_VERTEX_PROGRAM_ARB) {
struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
struct brw_vertex_program *newVP = brw_vertex_program(vprog);
const struct brw_vertex_program *curVP =
brw_vertex_program_const(brw->vertex_program);
if (newVP == curVP)
brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
if (newVP->program.IsPositionInvariant) {
_mesa_insert_mvp_code(ctx, &newVP->program);
}
newVP->id = brw->program_id++;
/* Also tell tnl about it:
*/
_tnl_program_string(ctx, target, prog);
}
}
static void
gen6_upload_vs_push_constants(struct brw_context *brw)
{
struct brw_stage_state *stage_state = &brw->vs.base;
/* _BRW_NEW_VERTEX_PROGRAM */
const struct brw_vertex_program *vp =
brw_vertex_program_const(brw->vertex_program);
/* BRW_NEW_VS_PROG_DATA */
const struct brw_stage_prog_data *prog_data = &brw->vs.prog_data->base.base;
gen6_upload_push_constants(brw, &vp->program.Base, prog_data,
stage_state, AUB_TRACE_VS_CONSTANTS);
if (brw->gen >= 7) {
if (brw->gen == 7 && !brw->is_haswell && !brw->is_baytrail)
gen7_emit_vs_workaround_flush(brw);
gen7_upload_constant_state(brw, stage_state, true /* active */,
_3DSTATE_CONSTANT_VS);
}
}
static GLboolean
brwProgramStringNotify(struct gl_context *ctx,
GLenum target,
struct gl_program *prog)
{
struct brw_context *brw = brw_context(ctx);
const struct brw_compiler *compiler = brw->intelScreen->compiler;
switch (target) {
case GL_FRAGMENT_PROGRAM_ARB: {
struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
struct brw_fragment_program *newFP = brw_fragment_program(fprog);
const struct brw_fragment_program *curFP =
brw_fragment_program_const(brw->fragment_program);
if (newFP == curFP)
brw->ctx.NewDriverState |= BRW_NEW_FRAGMENT_PROGRAM;
newFP->id = get_new_program_id(brw->intelScreen);
brw_add_texrect_params(prog);
prog->nir = brw_create_nir(brw, NULL, prog, MESA_SHADER_FRAGMENT, true);
brw_fs_precompile(ctx, NULL, prog);
break;
}
case GL_VERTEX_PROGRAM_ARB: {
struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
struct brw_vertex_program *newVP = brw_vertex_program(vprog);
const struct brw_vertex_program *curVP =
brw_vertex_program_const(brw->vertex_program);
if (newVP == curVP)
brw->ctx.NewDriverState |= BRW_NEW_VERTEX_PROGRAM;
if (newVP->program.IsPositionInvariant) {
_mesa_insert_mvp_code(ctx, &newVP->program);
}
newVP->id = get_new_program_id(brw->intelScreen);
/* Also tell tnl about it:
*/
_tnl_program_string(ctx, target, prog);
brw_add_texrect_params(prog);
prog->nir = brw_create_nir(brw, NULL, prog, MESA_SHADER_VERTEX,
compiler->scalar_stage[MESA_SHADER_VERTEX]);
brw_vs_precompile(ctx, NULL, prog);
break;
}
default:
/*
* driver->ProgramStringNotify is only called for ARB programs, fixed
* function vertex programs, and ir_to_mesa (which isn't used by the
* i965 back-end). Therefore, even after geometry shaders are added,
* this function should only ever be called with a target of
* GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
*/
unreachable("Unexpected target in brwProgramStringNotify");
}
return true;
}
/* 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();
}
static GLboolean brwProgramStringNotify( struct gl_context *ctx,
GLenum target,
struct gl_program *prog )
{
struct brw_context *brw = brw_context(ctx);
int i;
if (target == GL_FRAGMENT_PROGRAM_ARB) {
struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
struct brw_fragment_program *newFP = brw_fragment_program(fprog);
const struct brw_fragment_program *curFP =
brw_fragment_program_const(brw->fragment_program);
struct gl_shader_program *shader_program;
if (newFP == curFP)
brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
newFP->id = brw->program_id++;
/* Don't reject fragment shaders for their Mesa IR state when we're
* using the new FS backend.
*/
shader_program = _mesa_lookup_shader_program(ctx, prog->Id);
if (shader_program
&& shader_program->_LinkedShaders[MESA_SHADER_FRAGMENT]) {
return GL_TRUE;
}
}
else if (target == GL_VERTEX_PROGRAM_ARB) {
struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
struct brw_vertex_program *newVP = brw_vertex_program(vprog);
const struct brw_vertex_program *curVP =
brw_vertex_program_const(brw->vertex_program);
if (newVP == curVP)
brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
if (newVP->program.IsPositionInvariant) {
_mesa_insert_mvp_code(ctx, &newVP->program);
}
newVP->id = brw->program_id++;
/* Also tell tnl about it:
*/
_tnl_program_string(ctx, target, prog);
}
/* Reject programs with subroutines, which are totally broken at the moment
* (all program flows return when any program flow returns, and
* the VS also hangs if a function call calls a function.
*
* See piglit glsl-{vs,fs}-functions-[23] tests.
*/
for (i = 0; i < prog->NumInstructions; i++) {
struct prog_instruction *inst = prog->Instructions + i;
int r;
if (prog->Instructions[i].Opcode == OPCODE_CAL) {
shader_error(ctx, prog,
"i965 driver doesn't yet support uninlined function "
"calls. Move to using a single return statement at "
"the end of the function to work around it.\n");
return GL_FALSE;
}
if (prog->Instructions[i].Opcode == OPCODE_RET) {
shader_error(ctx, prog,
"i965 driver doesn't yet support \"return\" "
"from main().\n");
return GL_FALSE;
}
for (r = 0; r < _mesa_num_inst_src_regs(inst->Opcode); r++) {
if (prog->Instructions[i].SrcReg[r].RelAddr &&
prog->Instructions[i].SrcReg[r].File == PROGRAM_INPUT) {
shader_error(ctx, prog,
"Variable indexing of shader inputs unsupported\n");
return GL_FALSE;
}
}
if (target == GL_FRAGMENT_PROGRAM_ARB &&
prog->Instructions[i].DstReg.RelAddr &&
prog->Instructions[i].DstReg.File == PROGRAM_OUTPUT) {
shader_error(ctx, prog,
"Variable indexing of FS outputs unsupported\n");
return GL_FALSE;
}
if (target == GL_FRAGMENT_PROGRAM_ARB) {
if ((prog->Instructions[i].DstReg.RelAddr &&
prog->Instructions[i].DstReg.File == PROGRAM_TEMPORARY) ||
(prog->Instructions[i].SrcReg[0].RelAddr &&
prog->Instructions[i].SrcReg[0].File == PROGRAM_TEMPORARY) ||
(prog->Instructions[i].SrcReg[1].RelAddr &&
prog->Instructions[i].SrcReg[1].File == PROGRAM_TEMPORARY) ||
(prog->Instructions[i].SrcReg[2].RelAddr &&
prog->Instructions[i].SrcReg[2].File == PROGRAM_TEMPORARY)) {
shader_error(ctx, prog,
"Variable indexing of variable arrays in the FS "
"unsupported\n");
return GL_FALSE;
}
}
}
return GL_TRUE;
}
/* 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.
*/
}
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);
}