bool
brw_codegen_vs_prog(struct brw_context *brw,
struct gl_shader_program *prog,
struct brw_vertex_program *vp,
struct brw_vs_prog_key *key)
{
GLuint program_size;
const GLuint *program;
struct brw_vs_prog_data prog_data;
struct brw_stage_prog_data *stage_prog_data = &prog_data.base.base;
void *mem_ctx;
int i;
struct brw_shader *vs = NULL;
bool start_busy = false;
double start_time = 0;
if (prog)
vs = (struct brw_shader *) prog->_LinkedShaders[MESA_SHADER_VERTEX];
memset(&prog_data, 0, sizeof(prog_data));
/* Use ALT floating point mode for ARB programs so that 0^0 == 1. */
if (!prog)
stage_prog_data->use_alt_mode = true;
mem_ctx = ralloc_context(NULL);
brw_assign_common_binding_table_offsets(MESA_SHADER_VERTEX,
brw->intelScreen->devinfo,
prog, &vp->program.Base,
&prog_data.base.base, 0);
/* Allocate the references to the uniforms that will end up in the
* prog_data associated with the compiled program, and which will be freed
* by the state cache.
*/
int param_count = vp->program.Base.nir->num_uniforms;
if (!brw->intelScreen->compiler->scalar_vs)
param_count *= 4;
if (vs)
prog_data.base.base.nr_image_params = vs->base.NumImages;
/* vec4_visitor::setup_uniform_clipplane_values() also uploads user clip
* planes as uniforms.
*/
param_count += key->nr_userclip_plane_consts * 4;
stage_prog_data->param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
stage_prog_data->pull_param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
stage_prog_data->image_param =
rzalloc_array(NULL, struct brw_image_param,
stage_prog_data->nr_image_params);
stage_prog_data->nr_params = param_count;
if (prog) {
brw_nir_setup_glsl_uniforms(vp->program.Base.nir, prog, &vp->program.Base,
&prog_data.base.base,
brw->intelScreen->compiler->scalar_vs);
} else {
brw_nir_setup_arb_uniforms(vp->program.Base.nir, &vp->program.Base,
&prog_data.base.base);
}
GLbitfield64 outputs_written = vp->program.Base.OutputsWritten;
prog_data.inputs_read = vp->program.Base.InputsRead;
if (key->copy_edgeflag) {
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_EDGE);
prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
}
if (brw->gen < 6) {
/* Put dummy slots into the VUE for the SF to put the replaced
* point sprite coords in. We shouldn't need these dummy slots,
* which take up precious URB space, but it would mean that the SF
* doesn't get nice aligned pairs of input coords into output
* coords, which would be a pain to handle.
*/
for (i = 0; i < 8; i++) {
if (key->point_coord_replace & (1 << i))
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_TEX0 + i);
}
/* if back colors are written, allocate slots for front colors too */
if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC0))
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL0);
if (outputs_written & BITFIELD64_BIT(VARYING_SLOT_BFC1))
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_COL1);
}
/* In order for legacy clipping to work, we need to populate the clip
* distance varying slots whenever clipping is enabled, even if the vertex
* shader doesn't write to gl_ClipDistance.
*/
if (key->nr_userclip_plane_consts > 0) {
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
}
brw_compute_vue_map(brw->intelScreen->devinfo,
&prog_data.base.vue_map, outputs_written,
prog ? prog->SeparateShader : false);
if (0) {
_mesa_fprint_program_opt(stderr, &vp->program.Base, PROG_PRINT_DEBUG,
true);
}
if (unlikely(brw->perf_debug)) {
start_busy = (brw->batch.last_bo &&
drm_intel_bo_busy(brw->batch.last_bo));
start_time = get_time();
}
if (unlikely(INTEL_DEBUG & DEBUG_VS))
brw_dump_ir("vertex", prog, vs ? &vs->base : NULL, &vp->program.Base);
int st_index = -1;
if (INTEL_DEBUG & DEBUG_SHADER_TIME)
st_index = brw_get_shader_time_index(brw, prog, &vp->program.Base, ST_VS);
/* Emit GEN4 code.
*/
char *error_str;
program = brw_compile_vs(brw->intelScreen->compiler, brw, mem_ctx, key,
&prog_data, vp->program.Base.nir,
brw_select_clip_planes(&brw->ctx),
!_mesa_is_gles3(&brw->ctx),
st_index, &program_size, &error_str);
if (program == NULL) {
if (prog) {
prog->LinkStatus = false;
ralloc_strcat(&prog->InfoLog, error_str);
}
_mesa_problem(NULL, "Failed to compile vertex shader: %s\n", error_str);
ralloc_free(mem_ctx);
return false;
}
if (unlikely(brw->perf_debug) && vs) {
if (vs->compiled_once) {
brw_vs_debug_recompile(brw, prog, key);
}
if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
perf_debug("VS compile took %.03f ms and stalled the GPU\n",
(get_time() - start_time) * 1000);
}
vs->compiled_once = true;
}
/* Scratch space is used for register spilling */
if (prog_data.base.base.total_scratch) {
brw_get_scratch_bo(brw, &brw->vs.base.scratch_bo,
prog_data.base.base.total_scratch *
brw->max_vs_threads);
}
brw_upload_cache(&brw->cache, BRW_CACHE_VS_PROG,
key, sizeof(struct brw_vs_prog_key),
program, program_size,
&prog_data, sizeof(prog_data),
&brw->vs.base.prog_offset, &brw->vs.prog_data);
ralloc_free(mem_ctx);
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();
}
/**
* 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();
}
}
