/**
* Handy interface to upload some data to temporary GPU memory quickly.
*
* References to this memory should not be retained across batch flushes.
*/
void
intel_upload_data(struct brw_context *brw,
const void *data,
uint32_t size,
uint32_t alignment,
drm_intel_bo **out_bo,
uint32_t *out_offset)
{
void *dst = intel_upload_space(brw, size, alignment, out_bo, out_offset);
memcpy(dst, data, size);
}
/**
* Creates a temporary BO containing the pull constant data for the shader
* stage, and the SURFACE_STATE struct that points at it.
*
* Pull constants are GLSL uniforms (and other constant data) beyond what we
* could fit as push constants, or that have variable-index array access
* (which is easiest to support using pull constants, and avoids filling
* register space with mostly-unused data).
*
* Compare this path to brw_curbe.c for gen4/5 push constants, and
* gen6_vs_state.c for gen6+ push constants.
*/
void
brw_upload_pull_constants(struct brw_context *brw,
GLbitfield brw_new_constbuf,
const struct gl_program *prog,
struct brw_stage_state *stage_state,
const struct brw_stage_prog_data *prog_data,
bool dword_pitch)
{
int i;
uint32_t surf_index = prog_data->binding_table.pull_constants_start;
if (!prog_data->nr_pull_params) {
if (stage_state->surf_offset[surf_index]) {
stage_state->surf_offset[surf_index] = 0;
brw->ctx.NewDriverState |= brw_new_constbuf;
}
return;
}
/* Updates the ParamaterValues[i] pointers for all parameters of the
* basic type of PROGRAM_STATE_VAR.
*/
_mesa_load_state_parameters(&brw->ctx, prog->Parameters);
/* BRW_NEW_*_PROG_DATA | _NEW_PROGRAM_CONSTANTS */
uint32_t size = prog_data->nr_pull_params * 4;
drm_intel_bo *const_bo = NULL;
uint32_t const_offset;
gl_constant_value *constants = intel_upload_space(brw, size, 64,
&const_bo, &const_offset);
STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
for (i = 0; i < prog_data->nr_pull_params; i++) {
constants[i] = *prog_data->pull_param[i];
}
if (0) {
for (i = 0; i < ALIGN(prog_data->nr_pull_params, 4) / 4; i++) {
const gl_constant_value *row = &constants[i * 4];
fprintf(stderr, "const surface %3d: %4.3f %4.3f %4.3f %4.3f\n",
i, row[0].f, row[1].f, row[2].f, row[3].f);
}
}
brw_create_constant_surface(brw, const_bo, const_offset, size,
&stage_state->surf_offset[surf_index],
dword_pitch);
drm_intel_bo_unreference(const_bo);
brw->ctx.NewDriverState |= brw_new_constbuf;
}
static void
copy_array_to_vbo_array(struct brw_context *brw,
struct brw_vertex_element *element,
int min, int max,
struct brw_vertex_buffer *buffer,
GLuint dst_stride)
{
const int src_stride = element->glarray->StrideB;
/* If the source stride is zero, we just want to upload the current
* attribute once and set the buffer's stride to 0. There's no need
* to replicate it out.
*/
if (src_stride == 0) {
intel_upload_data(brw, element->glarray->Ptr,
element->glarray->_ElementSize,
element->glarray->_ElementSize,
&buffer->bo, &buffer->offset);
buffer->stride = 0;
buffer->size = element->glarray->_ElementSize;
return;
}
const unsigned char *src = element->glarray->Ptr + min * src_stride;
int count = max - min + 1;
GLuint size = count * dst_stride;
uint8_t *dst = intel_upload_space(brw, size, dst_stride,
&buffer->bo, &buffer->offset);
if (dst_stride == src_stride) {
memcpy(dst, src, size);
} else {
while (count--) {
memcpy(dst, src, dst_stride);
src += src_stride;
dst += dst_stride;
}
}
buffer->stride = dst_stride;
buffer->size = size;
}
/**
* Gathers together all the uniform values into a block of memory to be
* uploaded into the CURBE, then emits the state packet telling the hardware
* the new location.
*/
static void
brw_upload_constant_buffer(struct brw_context *brw)
{
struct gl_context *ctx = &brw->ctx;
/* BRW_NEW_CURBE_OFFSETS */
const GLuint sz = brw->curbe.total_size;
const GLuint bufsz = sz * 16 * sizeof(GLfloat);
gl_constant_value *buf;
GLuint i;
gl_clip_plane *clip_planes;
if (sz == 0) {
goto emit;
}
buf = intel_upload_space(brw, bufsz, 64,
&brw->curbe.curbe_bo, &brw->curbe.curbe_offset);
STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
/* fragment shader constants */
if (brw->curbe.wm_size) {
_mesa_load_state_parameters(ctx, brw->fragment_program->Base.Parameters);
/* BRW_NEW_CURBE_OFFSETS */
GLuint offset = brw->curbe.wm_start * 16;
/* BRW_NEW_FS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */
for (i = 0; i < brw->wm.prog_data->base.nr_params; i++) {
buf[offset + i] = *brw->wm.prog_data->base.param[i];
}
}
/* clipper constants */
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].f = fixed_plane[i][0];
buf[offset + i * 4 + 1].f = fixed_plane[i][1];
buf[offset + i * 4 + 2].f = fixed_plane[i][2];
buf[offset + i * 4 + 3].f = 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].f = clip_planes[j][0];
buf[offset + i * 4 + 1].f = clip_planes[j][1];
buf[offset + i * 4 + 2].f = clip_planes[j][2];
buf[offset + i * 4 + 3].f = clip_planes[j][3];
i++;
}
}
}
/* vertex shader constants */
if (brw->curbe.vs_size) {
_mesa_load_state_parameters(ctx, brw->vertex_program->Base.Parameters);
GLuint offset = brw->curbe.vs_start * 16;
/* BRW_NEW_VS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */
for (i = 0; i < brw->vs.prog_data->base.base.nr_params; i++) {
buf[offset + i] = *brw->vs.prog_data->base.base.param[i];
}
}
if (0) {
for (i = 0; i < sz*16; i+=4)
fprintf(stderr, "curbe %d.%d: %f %f %f %f\n", i/8, i&4,
buf[i+0].f, buf[i+1].f, buf[i+2].f, buf[i+3].f);
}
/* 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:
/* Work around mysterious 965 hangs that appear to happen if you do
* two 3DPRIMITIVEs with only a CONSTANT_BUFFER inbetween. If we
* haven't already flushed for some other reason, explicitly do so.
*
* We've found no documented reason why this should be necessary.
*/
if (brw->gen == 4 && !brw->is_g4x &&
(brw->ctx.NewDriverState & (BRW_NEW_BATCH | BRW_NEW_PSP)) == 0) {
BEGIN_BATCH(1);
OUT_BATCH(MI_FLUSH);
ADVANCE_BATCH();
}
/* BRW_NEW_URB_FENCE: From the gen4 PRM, volume 1, section 3.9.8
* (CONSTANT_BUFFER (CURBE Load)):
*
* "Modifying the CS URB allocation via URB_FENCE invalidates any
* previous CURBE entries. Therefore software must subsequently
* [re]issue a CONSTANT_BUFFER command before CURBE data can be used
* in the pipeline."
*/
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();
}
/**
* Gathers together all the uniform values into a block of memory to be
* uploaded into the CURBE, then emits the state packet telling the hardware
* the new location.
*/
static void
brw_upload_constant_buffer(struct brw_context *brw)
{
struct gl_context *ctx = &brw->ctx;
/* BRW_NEW_CURBE_OFFSETS */
const GLuint sz = brw->curbe.total_size;
const GLuint bufsz = sz * 16 * sizeof(GLfloat);
gl_constant_value *buf;
GLuint i;
gl_clip_plane *clip_planes;
if (sz == 0) {
goto emit;
}
buf = intel_upload_space(brw, bufsz, 64,
&brw->curbe.curbe_bo, &brw->curbe.curbe_offset);
STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
/* fragment shader constants */
if (brw->curbe.wm_size) {
_mesa_load_state_parameters(ctx, brw->fragment_program->Parameters);
/* BRW_NEW_CURBE_OFFSETS */
GLuint offset = brw->curbe.wm_start * 16;
/* BRW_NEW_FS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */
for (i = 0; i < brw->wm.base.prog_data->nr_params; i++) {
buf[offset + i] = *brw->wm.base.prog_data->param[i];
}
}
/* clipper constants */
if (brw->curbe.clip_size) {
GLuint offset = brw->curbe.clip_start * 16;
GLbitfield mask;
/* If any planes are going this way, send them all this way:
*/
for (i = 0; i < 6; i++) {
buf[offset + i * 4 + 0].f = fixed_plane[i][0];
buf[offset + i * 4 + 1].f = fixed_plane[i][1];
buf[offset + i * 4 + 2].f = fixed_plane[i][2];
buf[offset + i * 4 + 3].f = fixed_plane[i][3];
}
/* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
* clip-space:
*/
clip_planes = brw_select_clip_planes(ctx);
mask = ctx->Transform.ClipPlanesEnabled;
while (mask) {
const int j = u_bit_scan(&mask);
buf[offset + i * 4 + 0].f = clip_planes[j][0];
buf[offset + i * 4 + 1].f = clip_planes[j][1];
buf[offset + i * 4 + 2].f = clip_planes[j][2];
buf[offset + i * 4 + 3].f = clip_planes[j][3];
i++;
}
}
/* vertex shader constants */
if (brw->curbe.vs_size) {
_mesa_load_state_parameters(ctx, brw->vertex_program->Parameters);
GLuint offset = brw->curbe.vs_start * 16;
/* BRW_NEW_VS_PROG_DATA | _NEW_PROGRAM_CONSTANTS: copy uniform values */
for (i = 0; i < brw->vs.base.prog_data->nr_params; i++) {
buf[offset + i] = *brw->vs.base.prog_data->param[i];
}
}
if (0) {
for (i = 0; i < sz*16; i+=4)
fprintf(stderr, "curbe %d.%d: %f %f %f %f\n", i/8, i&4,
buf[i+0].f, buf[i+1].f, buf[i+2].f, buf[i+3].f);
}
/* 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:
/* BRW_NEW_URB_FENCE: From the gen4 PRM, volume 1, section 3.9.8
* (CONSTANT_BUFFER (CURBE Load)):
*
* "Modifying the CS URB allocation via URB_FENCE invalidates any
* previous CURBE entries. Therefore software must subsequently
* [re]issue a CONSTANT_BUFFER command before CURBE data can be used
* in the pipeline."
*/
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();
/* Work around a Broadwater/Crestline depth interpolator bug. The
* following sequence will cause GPU hangs:
*
* 1. Change state so that all depth related fields in CC_STATE are
* disabled, and in WM_STATE, only "PS Use Source Depth" is enabled.
* 2. Emit a CONSTANT_BUFFER packet.
* 3. Draw via 3DPRIMITIVE.
*
* The recommended workaround is to emit a non-pipelined state change after
* emitting CONSTANT_BUFFER, in order to drain the windowizer pipeline.
*
* We arbitrarily choose 3DSTATE_GLOBAL_DEPTH_CLAMP_OFFSET (as it's small),
* and always emit it when "PS Use Source Depth" is set. We could be more
* precise, but the additional complexity is probably not worth it.
*
* BRW_NEW_FRAGMENT_PROGRAM
*/
if (brw->gen == 4 && !brw->is_g4x &&
(brw->fragment_program->info.inputs_read & (1 << VARYING_SLOT_POS))) {
BEGIN_BATCH(2);
OUT_BATCH(_3DSTATE_GLOBAL_DEPTH_OFFSET_CLAMP << 16 | (2 - 2));
OUT_BATCH(0);
ADVANCE_BATCH();
}
}
/**
* Creates a streamed BO containing the push constants for the VS or GS on
* gen6+.
*
* Push constants are constant values (such as GLSL uniforms) that are
* pre-loaded into a shader stage's register space at thread spawn time.
*
* Not all GLSL uniforms will be uploaded as push constants: The hardware has
* a limitation of 32 or 64 EU registers (256 or 512 floats) per stage to be
* uploaded as push constants, while GL 4.4 requires at least 1024 components
* to be usable for the VS. Plus, currently we always use pull constants
* instead of push constants when doing variable-index array access.
*
* See brw_curbe.c for the equivalent gen4/5 code.
*/
void
gen6_upload_push_constants(struct brw_context *brw,
const struct gl_program *prog,
const struct brw_stage_prog_data *prog_data,
struct brw_stage_state *stage_state)
{
struct gl_context *ctx = &brw->ctx;
if (prog_data->nr_params == 0) {
stage_state->push_const_size = 0;
} else {
/* Updates the ParamaterValues[i] pointers for all parameters of the
* basic type of PROGRAM_STATE_VAR.
*/
/* XXX: Should this happen somewhere before to get our state flag set? */
if (prog)
_mesa_load_state_parameters(ctx, prog->Parameters);
int i;
const int size = prog_data->nr_params * sizeof(gl_constant_value);
gl_constant_value *param;
if (brw->gen >= 8 || brw->is_haswell) {
param = intel_upload_space(brw, size, 32,
&stage_state->push_const_bo,
&stage_state->push_const_offset);
} else {
param = brw_state_batch(brw, size, 32,
&stage_state->push_const_offset);
}
STATIC_ASSERT(sizeof(gl_constant_value) == sizeof(float));
/* _NEW_PROGRAM_CONSTANTS
*
* Also _NEW_TRANSFORM -- we may reference clip planes other than as a
* side effect of dereferencing uniforms, so _NEW_PROGRAM_CONSTANTS
* wouldn't be set for them.
*/
for (i = 0; i < prog_data->nr_params; i++) {
param[i] = *prog_data->param[i];
}
if (0) {
fprintf(stderr, "%s constants:\n",
_mesa_shader_stage_to_string(stage_state->stage));
for (i = 0; i < prog_data->nr_params; i++) {
if ((i & 7) == 0)
fprintf(stderr, "g%d: ",
prog_data->dispatch_grf_start_reg + i / 8);
fprintf(stderr, "%8f ", param[i].f);
if ((i & 7) == 7)
fprintf(stderr, "\n");
}
if ((i & 7) != 0)
fprintf(stderr, "\n");
fprintf(stderr, "\n");
}
stage_state->push_const_size = ALIGN(prog_data->nr_params, 8) / 8;
/* We can only push 32 registers of constants at a time. */
/* From the SNB PRM (vol2, part 1, section 3.2.1.4: 3DSTATE_CONSTANT_VS:
*
* "The sum of all four read length fields (each incremented to
* represent the actual read length) must be less than or equal to
* 32"
*
* From the IVB PRM (vol2, part 1, section 3.2.1.3: 3DSTATE_CONSTANT_VS:
*
* "The sum of all four read length fields must be less than or
* equal to the size of 64"
*
* The other shader stages all match the VS's limits.
*/
assert(stage_state->push_const_size <= 32);
}
stage_state->push_constants_dirty = true;
}