static void
vc4_screen_destroy(struct pipe_screen *pscreen)
{
vc4_bufmgr_destroy(pscreen);
ralloc_free(pscreen);
}
bool
brwCreateContext(int api,
const struct gl_config *mesaVis,
__DRIcontext *driContextPriv,
unsigned major_version,
unsigned minor_version,
uint32_t flags,
unsigned *error,
void *sharedContextPrivate)
{
__DRIscreen *sPriv = driContextPriv->driScreenPriv;
struct intel_screen *screen = sPriv->driverPrivate;
struct dd_function_table functions;
struct brw_context *brw = rzalloc(NULL, struct brw_context);
if (!brw) {
printf("%s: failed to alloc context\n", __FUNCTION__);
*error = __DRI_CTX_ERROR_NO_MEMORY;
return false;
}
/* brwInitVtbl needs to know the chipset generation so that it can set the
* right pointers.
*/
brw->gen = screen->gen;
brwInitVtbl( brw );
brwInitDriverFunctions(screen, &functions);
struct gl_context *ctx = &brw->ctx;
if (!intelInitContext( brw, api, major_version, minor_version,
mesaVis, driContextPriv,
sharedContextPrivate, &functions,
error)) {
ralloc_free(brw);
return false;
}
brw_initialize_context_constants(brw);
/* Reinitialize the context point state. It depends on ctx->Const values. */
_mesa_init_point(ctx);
if (brw->gen >= 6) {
/* Create a new hardware context. Using a hardware context means that
* our GPU state will be saved/restored on context switch, allowing us
* to assume that the GPU is in the same state we left it in.
*
* This is required for transform feedback buffer offsets, query objects,
* and also allows us to reduce how much state we have to emit.
*/
brw->hw_ctx = drm_intel_gem_context_create(brw->bufmgr);
if (!brw->hw_ctx) {
fprintf(stderr, "Gen6+ requires Kernel 3.6 or later.\n");
ralloc_free(brw);
return false;
}
}
brw_init_surface_formats(brw);
/* Initialize swrast, tnl driver tables: */
TNLcontext *tnl = TNL_CONTEXT(ctx);
if (tnl)
tnl->Driver.RunPipeline = _tnl_run_pipeline;
ctx->DriverFlags.NewTransformFeedback = BRW_NEW_TRANSFORM_FEEDBACK;
ctx->DriverFlags.NewRasterizerDiscard = BRW_NEW_RASTERIZER_DISCARD;
ctx->DriverFlags.NewUniformBuffer = BRW_NEW_UNIFORM_BUFFER;
if (brw->is_g4x || brw->gen >= 5) {
brw->CMD_VF_STATISTICS = GM45_3DSTATE_VF_STATISTICS;
brw->CMD_PIPELINE_SELECT = CMD_PIPELINE_SELECT_GM45;
brw->has_surface_tile_offset = true;
if (brw->gen < 6)
brw->has_compr4 = true;
brw->has_aa_line_parameters = true;
brw->has_pln = true;
} else {
brw->CMD_VF_STATISTICS = GEN4_3DSTATE_VF_STATISTICS;
brw->CMD_PIPELINE_SELECT = CMD_PIPELINE_SELECT_965;
}
/* WM maximum threads is number of EUs times number of threads per EU. */
assert(brw->gen <= 7);
if (brw->is_haswell) {
if (brw->gt == 1) {
brw->max_wm_threads = 102;
brw->max_vs_threads = 70;
brw->urb.size = 128;
brw->urb.max_vs_entries = 640;
brw->urb.max_gs_entries = 256;
} else if (brw->gt == 2) {
brw->max_wm_threads = 204;
brw->max_vs_threads = 280;
brw->urb.size = 256;
brw->urb.max_vs_entries = 1664;
brw->urb.max_gs_entries = 640;
} else if (brw->gt == 3) {
brw->max_wm_threads = 408;
brw->max_vs_threads = 280;
brw->urb.size = 512;
brw->urb.max_vs_entries = 1664;
brw->urb.max_gs_entries = 640;
}
} else if (brw->gen == 7) {
if (brw->gt == 1) {
brw->max_wm_threads = 48;
brw->max_vs_threads = 36;
brw->max_gs_threads = 36;
brw->urb.size = 128;
brw->urb.max_vs_entries = 512;
brw->urb.max_gs_entries = 192;
} else if (brw->gt == 2) {
brw->max_wm_threads = 172;
brw->max_vs_threads = 128;
brw->max_gs_threads = 128;
brw->urb.size = 256;
brw->urb.max_vs_entries = 704;
brw->urb.max_gs_entries = 320;
} else {
assert(!"Unknown gen7 device.");
}
} else if (brw->gen == 6) {
if (brw->gt == 2) {
brw->max_wm_threads = 80;
brw->max_vs_threads = 60;
brw->max_gs_threads = 60;
brw->urb.size = 64; /* volume 5c.5 section 5.1 */
brw->urb.max_vs_entries = 256; /* volume 2a (see 3DSTATE_URB) */
brw->urb.max_gs_entries = 256;
} else {
brw->max_wm_threads = 40;
brw->max_vs_threads = 24;
brw->max_gs_threads = 21; /* conservative; 24 if rendering disabled */
brw->urb.size = 32; /* volume 5c.5 section 5.1 */
brw->urb.max_vs_entries = 256; /* volume 2a (see 3DSTATE_URB) */
brw->urb.max_gs_entries = 256;
}
brw->urb.gen6_gs_previously_active = false;
} else if (brw->gen == 5) {
brw->urb.size = 1024;
brw->max_vs_threads = 72;
brw->max_gs_threads = 32;
brw->max_wm_threads = 12 * 6;
} else if (brw->is_g4x) {
brw->urb.size = 384;
brw->max_vs_threads = 32;
brw->max_gs_threads = 2;
brw->max_wm_threads = 10 * 5;
} else if (brw->gen < 6) {
brw->urb.size = 256;
brw->max_vs_threads = 16;
brw->max_gs_threads = 2;
brw->max_wm_threads = 8 * 4;
brw->has_negative_rhw_bug = true;
}
if (brw->gen <= 7) {
brw->needs_unlit_centroid_workaround = true;
}
brw->prim_restart.in_progress = false;
brw->prim_restart.enable_cut_index = false;
brw_init_state( brw );
brw->curbe.last_buf = calloc(1, 4096);
brw->curbe.next_buf = calloc(1, 4096);
brw->state.dirty.mesa = ~0;
brw->state.dirty.brw = ~0;
brw->emit_state_always = 0;
brw->batch.need_workaround_flush = true;
ctx->VertexProgram._MaintainTnlProgram = true;
ctx->FragmentProgram._MaintainTexEnvProgram = true;
brw_draw_init( brw );
brw->precompile = driQueryOptionb(&brw->optionCache, "shader_precompile");
ctx->Const.ContextFlags = 0;
if ((flags & __DRI_CTX_FLAG_FORWARD_COMPATIBLE) != 0)
ctx->Const.ContextFlags |= GL_CONTEXT_FLAG_FORWARD_COMPATIBLE_BIT;
if ((flags & __DRI_CTX_FLAG_DEBUG) != 0) {
ctx->Const.ContextFlags |= GL_CONTEXT_FLAG_DEBUG_BIT;
/* Turn on some extra GL_ARB_debug_output generation. */
brw->perf_debug = true;
}
brw_fs_alloc_reg_sets(brw);
if (INTEL_DEBUG & DEBUG_SHADER_TIME)
brw_init_shader_time(brw);
_mesa_compute_version(ctx);
_mesa_initialize_dispatch_tables(ctx);
_mesa_initialize_vbo_vtxfmt(ctx);
return true;
}
static bool
brw_codegen_gs_prog(struct brw_context *brw,
struct brw_program *gp,
struct brw_gs_prog_key *key)
{
struct brw_compiler *compiler = brw->screen->compiler;
const struct gen_device_info *devinfo = &brw->screen->devinfo;
struct brw_stage_state *stage_state = &brw->gs.base;
struct brw_gs_prog_data prog_data;
bool start_busy = false;
double start_time = 0;
memset(&prog_data, 0, sizeof(prog_data));
assign_gs_binding_table_offsets(devinfo, &gp->program, &prog_data);
/* 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.
*
* Note: param_count needs to be num_uniform_components * 4, since we add
* padding around uniform values below vec4 size, so the worst case is that
* every uniform is a float which gets padded to the size of a vec4.
*/
int param_count = gp->program.nir->num_uniforms / 4;
prog_data.base.base.param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.base.pull_param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.base.image_param =
rzalloc_array(NULL, struct brw_image_param,
gp->program.info.num_images);
prog_data.base.base.nr_params = param_count;
prog_data.base.base.nr_image_params = gp->program.info.num_images;
brw_nir_setup_glsl_uniforms(gp->program.nir, &gp->program,
&prog_data.base.base,
compiler->scalar_stage[MESA_SHADER_GEOMETRY]);
brw_nir_analyze_ubo_ranges(compiler, gp->program.nir,
prog_data.base.base.ubo_ranges);
uint64_t outputs_written = gp->program.info.outputs_written;
brw_compute_vue_map(devinfo,
&prog_data.base.vue_map, outputs_written,
gp->program.info.separate_shader);
int st_index = -1;
if (INTEL_DEBUG & DEBUG_SHADER_TIME)
st_index = brw_get_shader_time_index(brw, &gp->program, ST_GS, true);
if (unlikely(brw->perf_debug)) {
start_busy = brw->batch.last_bo && brw_bo_busy(brw->batch.last_bo);
start_time = get_time();
}
void *mem_ctx = ralloc_context(NULL);
unsigned program_size;
char *error_str;
const unsigned *program =
brw_compile_gs(brw->screen->compiler, brw, mem_ctx, key,
&prog_data, gp->program.nir, &gp->program,
st_index, &program_size, &error_str);
if (program == NULL) {
ralloc_strcat(&gp->program.sh.data->InfoLog, error_str);
_mesa_problem(NULL, "Failed to compile geometry shader: %s\n", error_str);
ralloc_free(mem_ctx);
return false;
}
if (unlikely(brw->perf_debug)) {
if (gp->compiled_once) {
brw_gs_debug_recompile(brw, &gp->program, key);
}
if (start_busy && !brw_bo_busy(brw->batch.last_bo)) {
perf_debug("GS compile took %.03f ms and stalled the GPU\n",
(get_time() - start_time) * 1000);
}
gp->compiled_once = true;
}
/* Scratch space is used for register spilling */
brw_alloc_stage_scratch(brw, stage_state,
prog_data.base.base.total_scratch,
devinfo->max_gs_threads);
brw_upload_cache(&brw->cache, BRW_CACHE_GS_PROG,
key, sizeof(*key),
program, program_size,
&prog_data, sizeof(prog_data),
&stage_state->prog_offset, &brw->gs.base.prog_data);
ralloc_free(mem_ctx);
return true;
}
~has_recursion_visitor()
{
hash_table_dtor(this->function_hash);
ralloc_free(this->mem_ctx);
}
/**
* All Mesa program -> GPU code generation goes through this function.
* Depending on the instructions used (i.e. flow control instructions)
* we'll use one of two code generators.
*/
bool do_wm_prog(struct brw_context *brw,
struct gl_shader_program *prog,
struct brw_fragment_program *fp,
struct brw_wm_prog_key *key)
{
struct brw_wm_compile *c;
const GLuint *program;
struct gl_shader *fs = NULL;
GLuint program_size;
if (prog)
fs = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
c = rzalloc(NULL, struct brw_wm_compile);
/* 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;
if (fs) {
param_count = fs->num_uniform_components;
} else {
param_count = fp->program.Base.Parameters->NumParameters * 4;
}
/* The backend also sometimes adds params for texture size. */
param_count += 2 * BRW_MAX_TEX_UNIT;
c->prog_data.param = rzalloc_array(NULL, const float *, param_count);
c->prog_data.pull_param = rzalloc_array(NULL, const float *, param_count);
memcpy(&c->key, key, sizeof(*key));
c->prog_data.barycentric_interp_modes =
brw_compute_barycentric_interp_modes(brw, c->key.flat_shade,
&fp->program);
program = brw_wm_fs_emit(brw, c, &fp->program, prog, &program_size);
if (program == NULL)
return false;
/* Scratch space is used for register spilling */
if (c->last_scratch) {
perf_debug("Fragment shader triggered register spilling. "
"Try reducing the number of live scalar values to "
"improve performance.\n");
c->prog_data.total_scratch = brw_get_scratch_size(c->last_scratch);
brw_get_scratch_bo(brw, &brw->wm.scratch_bo,
c->prog_data.total_scratch * brw->max_wm_threads);
}
if (unlikely(INTEL_DEBUG & DEBUG_WM))
fprintf(stderr, "\n");
brw_upload_cache(&brw->cache, BRW_WM_PROG,
&c->key, sizeof(c->key),
program, program_size,
&c->prog_data, sizeof(c->prog_data),
&brw->wm.prog_offset, &brw->wm.prog_data);
ralloc_free(c);
return true;
}
static void DeleteShader(struct gl_context *ctx, struct gl_shader *shader)
{
ralloc_free(shader);
}
static bool
constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
{
nir_const_value src[NIR_MAX_VEC_COMPONENTS];
if (!instr->dest.dest.is_ssa)
return false;
/* In the case that any outputs/inputs have unsized types, then we need to
* guess the bit-size. In this case, the validator ensures that all
* bit-sizes match so we can just take the bit-size from first
* output/input with an unsized type. If all the outputs/inputs are sized
* then we don't need to guess the bit-size at all because the code we
* generate for constant opcodes in this case already knows the sizes of
* the types involved and does not need the provided bit-size for anything
* (although it still requires to receive a valid bit-size).
*/
unsigned bit_size = 0;
if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type))
bit_size = instr->dest.dest.ssa.bit_size;
for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
if (!instr->src[i].src.is_ssa)
return false;
if (bit_size == 0 &&
!nir_alu_type_get_type_size(nir_op_infos[instr->op].input_sizes[i])) {
bit_size = instr->src[i].src.ssa->bit_size;
}
nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
if (src_instr->type != nir_instr_type_load_const)
return false;
nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
j++) {
switch(load_const->def.bit_size) {
case 64:
src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
break;
case 32:
src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
break;
case 16:
src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
break;
case 8:
src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
break;
default:
unreachable("Invalid bit size");
}
}
/* We shouldn't have any source modifiers in the optimization loop. */
assert(!instr->src[i].abs && !instr->src[i].negate);
}
if (bit_size == 0)
bit_size = 32;
/* We shouldn't have any saturate modifiers in the optimization loop. */
assert(!instr->dest.saturate);
nir_const_value dest =
nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
bit_size, src);
nir_load_const_instr *new_instr =
nir_load_const_instr_create(mem_ctx,
instr->dest.dest.ssa.num_components,
instr->dest.dest.ssa.bit_size);
new_instr->value = dest;
nir_instr_insert_before(&instr->instr, &new_instr->instr);
nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
nir_src_for_ssa(&new_instr->def));
nir_instr_remove(&instr->instr);
ralloc_free(instr);
return true;
}
brw_blorp_const_color_program::~brw_blorp_const_color_program()
{
ralloc_free(mem_ctx);
}
~get_sampler_name()
{
ralloc_free(this->mem_ctx);
}
~string_buffer()
{
ralloc_free(m_Ptr);
}
static bool
do_gs_prog(struct brw_context *brw,
struct gl_shader_program *prog,
struct brw_geometry_program *gp,
struct brw_gs_prog_key *key)
{
struct brw_stage_state *stage_state = &brw->gs.base;
struct brw_gs_compile c;
memset(&c, 0, sizeof(c));
c.key = *key;
c.gp = gp;
c.prog_data.include_primitive_id =
(gp->program.Base.InputsRead & VARYING_BIT_PRIMITIVE_ID) != 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.
*
* Note: param_count needs to be num_uniform_components * 4, since we add
* padding around uniform values below vec4 size, so the worst case is that
* every uniform is a float which gets padded to the size of a vec4.
*/
struct gl_shader *gs = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
int param_count = gs->num_uniform_components * 4;
/* We also upload clip plane data as uniforms */
param_count += MAX_CLIP_PLANES * 4;
c.prog_data.base.param = rzalloc_array(NULL, const float *, param_count);
c.prog_data.base.pull_param = rzalloc_array(NULL, const float *, param_count);
if (gp->program.OutputType == GL_POINTS) {
/* When the output type is points, the geometry shader may output data
* to multiple streams, and EndPrimitive() has no effect. So we
* configure the hardware to interpret the control data as stream ID.
*/
c.prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID;
/* However, StreamID is not yet supported, so we output zero bits of
* control data per vertex.
*/
c.control_data_bits_per_vertex = 0;
} else {
/* When the output type is triangle_strip or line_strip, EndPrimitive()
* may be used to terminate the current strip and start a new one
* (similar to primitive restart), and outputting data to multiple
* streams is not supported. So we configure the hardware to interpret
* the control data as EndPrimitive information (a.k.a. "cut bits").
*/
c.prog_data.control_data_format = GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT;
/* We only need to output control data if the shader actually calls
* EndPrimitive().
*/
c.control_data_bits_per_vertex = gp->program.UsesEndPrimitive ? 1 : 0;
}
c.control_data_header_size_bits =
gp->program.VerticesOut * c.control_data_bits_per_vertex;
/* 1 HWORD = 32 bytes = 256 bits */
c.prog_data.control_data_header_size_hwords =
ALIGN(c.control_data_header_size_bits, 256) / 256;
GLbitfield64 outputs_written = gp->program.Base.OutputsWritten;
/* 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 (c.key.base.userclip_active) {
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
}
brw_compute_vue_map(brw, &c.prog_data.base.vue_map, outputs_written);
/* Compute the output vertex size.
*
* From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
* Size (p168):
*
* [0,62] indicating [1,63] 16B units
*
* Specifies the size of each vertex stored in the GS output entry
* (following any Control Header data) as a number of 128-bit units
* (minus one).
*
* Programming Restrictions: The vertex size must be programmed as a
* multiple of 32B units with the following exception: Rendering is
* disabled (as per SOL stage state) and the vertex size output by the
* GS thread is 16B.
*
* If rendering is enabled (as per SOL state) the vertex size must be
* programmed as a multiple of 32B units. In other words, the only time
* software can program a vertex size with an odd number of 16B units
* is when rendering is disabled.
*
* Note: B=bytes in the above text.
*
* It doesn't seem worth the extra trouble to optimize the case where the
* vertex size is 16B (especially since this would require special-casing
* the GEN assembly that writes to the URB). So we just set the vertex
* size to a multiple of 32B (2 vec4's) in all cases.
*
* The maximum output vertex size is 62*16 = 992 bytes (31 hwords). We
* budget that as follows:
*
* 512 bytes for varyings (a varying component is 4 bytes and
* gl_MaxGeometryOutputComponents = 128)
* 16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
* bytes)
* 16 bytes overhead for gl_Position (we allocate it a slot in the VUE
* even if it's not used)
* 32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
* whenever clip planes are enabled, even if the shader doesn't
* write to gl_ClipDistance)
* 16 bytes overhead since the VUE size must be a multiple of 32 bytes
* (see below)--this causes up to 1 VUE slot to be wasted
* 400 bytes available for varying packing overhead
*
* Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
* per interpolation type, so this is plenty.
*
*/
unsigned output_vertex_size_bytes = c.prog_data.base.vue_map.num_slots * 16;
assert(output_vertex_size_bytes <= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES);
c.prog_data.output_vertex_size_hwords =
ALIGN(output_vertex_size_bytes, 32) / 32;
/* Compute URB entry size. The maximum allowed URB entry size is 32k.
* That divides up as follows:
*
* 64 bytes for the control data header (cut indices or StreamID bits)
* 4096 bytes for varyings (a varying component is 4 bytes and
* gl_MaxGeometryTotalOutputComponents = 1024)
* 4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
* bytes/vertex and gl_MaxGeometryOutputVertices is 256)
* 4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
* even if it's not used)
* 8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
* whenever clip planes are enabled, even if the shader doesn't
* write to gl_ClipDistance)
* 4096 bytes overhead since the VUE size must be a multiple of 32
* bytes (see above)--this causes up to 1 VUE slot to be wasted
* 8128 bytes available for varying packing overhead
*
* Worst-case varying packing overhead is 3/4 of a varying slot per
* interpolation type, which works out to 3072 bytes, so this would allow
* us to accommodate 2 interpolation types without any danger of running
* out of URB space.
*
* In practice, the risk of running out of URB space is very small, since
* the above figures are all worst-case, and most of them scale with the
* number of output vertices. So we'll just calculate the amount of space
* we need, and if it's too large, fail to compile.
*/
unsigned output_size_bytes =
c.prog_data.output_vertex_size_hwords * 32 * gp->program.VerticesOut;
output_size_bytes += 32 * c.prog_data.control_data_header_size_hwords;
assert(output_size_bytes >= 1);
if (output_size_bytes > GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES)
return false;
/* URB entry sizes are stored as a multiple of 64 bytes. */
c.prog_data.base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
c.prog_data.output_topology = prim_to_hw_prim[gp->program.OutputType];
brw_compute_vue_map(brw, &c.input_vue_map, c.key.input_varyings);
/* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
* need to program a URB read length of ceiling(num_slots / 2).
*/
c.prog_data.base.urb_read_length = (c.input_vue_map.num_slots + 1) / 2;
void *mem_ctx = ralloc_context(NULL);
unsigned program_size;
const unsigned *program =
brw_gs_emit(brw, prog, &c, mem_ctx, &program_size);
if (program == NULL) {
ralloc_free(mem_ctx);
return false;
}
/* Scratch space is used for register spilling */
if (c.base.last_scratch) {
perf_debug("Geometry shader triggered register spilling. "
"Try reducing the number of live vec4 values to "
"improve performance.\n");
c.prog_data.base.total_scratch
= brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
brw_get_scratch_bo(brw, &stage_state->scratch_bo,
c.prog_data.base.total_scratch * brw->max_gs_threads);
}
brw_upload_cache(&brw->cache, BRW_GS_PROG,
&c.key, sizeof(c.key),
program, program_size,
&c.prog_data, sizeof(c.prog_data),
&stage_state->prog_offset, &brw->gs.prog_data);
ralloc_free(mem_ctx);
return true;
}
void
brw_wm_clear_compile(struct brw_context *brw,
struct brw_wm_compile *c)
{
ralloc_free(c);
}
/**
* All Mesa program -> GPU code generation goes through this function.
* Depending on the instructions used (i.e. flow control instructions)
* we'll use one of two code generators.
*/
bool do_wm_prog(struct brw_context *brw,
struct gl_shader_program *prog,
struct brw_fragment_program *fp,
struct brw_wm_prog_key *key)
{
struct gl_context *ctx = &brw->ctx;
void *mem_ctx = ralloc_context(NULL);
struct brw_wm_prog_data prog_data;
const GLuint *program;
struct gl_shader *fs = NULL;
GLuint program_size;
if (prog)
fs = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
memset(&prog_data, 0, sizeof(prog_data));
/* 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;
if (fs) {
param_count = fs->num_uniform_components;
} else {
param_count = fp->program.Base.Parameters->NumParameters * 4;
}
/* The backend also sometimes adds params for texture size. */
param_count += 2 * ctx->Const.Program[MESA_SHADER_FRAGMENT].MaxTextureImageUnits;
prog_data.base.param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.pull_param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.nr_params = param_count;
prog_data.barycentric_interp_modes =
brw_compute_barycentric_interp_modes(brw, key->flat_shade,
key->persample_shading,
&fp->program);
program = brw_wm_fs_emit(brw, mem_ctx, key, &prog_data,
&fp->program, prog, &program_size);
if (program == NULL) {
ralloc_free(mem_ctx);
return false;
}
if (prog_data.total_scratch) {
brw_get_scratch_bo(brw, &brw->wm.base.scratch_bo,
prog_data.total_scratch * brw->max_wm_threads);
}
if (unlikely(INTEL_DEBUG & DEBUG_WM))
fprintf(stderr, "\n");
brw_upload_cache(&brw->cache, BRW_WM_PROG,
key, sizeof(struct brw_wm_prog_key),
program, program_size,
&prog_data, sizeof(prog_data),
&brw->wm.base.prog_offset, &brw->wm.prog_data);
ralloc_free(mem_ctx);
return true;
}
static void compile_gs_prog( struct brw_context *brw,
struct brw_gs_prog_key *key )
{
struct brw_gs_compile c;
const GLuint *program;
void *mem_ctx;
GLuint program_size;
memset(&c, 0, sizeof(c));
c.key = *key;
c.vue_map = brw->vs.prog_data->base.vue_map;
c.nr_regs = (c.vue_map.num_slots + 1)/2;
mem_ctx = ralloc_context(NULL);
/* Begin the compilation:
*/
brw_init_compile(brw, &c.func, mem_ctx);
c.func.single_program_flow = 1;
/* For some reason the thread is spawned with only 4 channels
* unmasked.
*/
brw_set_mask_control(&c.func, BRW_MASK_DISABLE);
if (brw->gen >= 6) {
unsigned num_verts;
bool check_edge_flag;
/* On Sandybridge, we use the GS for implementing transform feedback
* (called "Stream Out" in the PRM).
*/
switch (key->primitive) {
case _3DPRIM_POINTLIST:
num_verts = 1;
check_edge_flag = false;
break;
case _3DPRIM_LINELIST:
case _3DPRIM_LINESTRIP:
case _3DPRIM_LINELOOP:
num_verts = 2;
check_edge_flag = false;
break;
case _3DPRIM_TRILIST:
case _3DPRIM_TRIFAN:
case _3DPRIM_TRISTRIP:
case _3DPRIM_RECTLIST:
num_verts = 3;
check_edge_flag = false;
break;
case _3DPRIM_QUADLIST:
case _3DPRIM_QUADSTRIP:
case _3DPRIM_POLYGON:
num_verts = 3;
check_edge_flag = true;
break;
default:
assert(!"Unexpected primitive type in Gen6 SOL program.");
return;
}
gen6_sol_program(&c, key, num_verts, check_edge_flag);
} else {
/* On Gen4-5, we use the GS to decompose certain types of primitives.
* Note that primitives which don't require a GS program have already
* been weeded out by now.
*/
switch (key->primitive) {
case _3DPRIM_QUADLIST:
brw_gs_quads( &c, key );
break;
case _3DPRIM_QUADSTRIP:
brw_gs_quad_strip( &c, key );
break;
case _3DPRIM_LINELOOP:
brw_gs_lines( &c );
break;
default:
ralloc_free(mem_ctx);
return;
}
}
/* get the program
*/
program = brw_get_program(&c.func, &program_size);
if (unlikely(INTEL_DEBUG & DEBUG_GS)) {
int i;
printf("gs:\n");
for (i = 0; i < program_size / sizeof(struct brw_instruction); i++)
brw_disasm(stdout, &((struct brw_instruction *)program)[i],
brw->gen);
printf("\n");
}
brw_upload_cache(&brw->cache, BRW_GS_PROG,
&c.key, sizeof(c.key),
program, program_size,
&c.prog_data, sizeof(c.prog_data),
&brw->gs.prog_offset, &brw->gs.prog_data);
ralloc_free(mem_ctx);
}
static void
setup_glsl_msaa_blit_shader(struct gl_context *ctx,
struct blit_state *blit,
const struct gl_framebuffer *drawFb,
struct gl_renderbuffer *src_rb,
GLenum target)
{
const char *vs_source;
char *fs_source;
void *mem_ctx;
enum blit_msaa_shader shader_index;
bool dst_is_msaa = false;
GLenum src_datatype;
const char *vec4_prefix;
const char *sampler_array_suffix = "";
char *name;
const char *texcoord_type = "vec2";
int samples;
int shader_offset = 0;
if (src_rb) {
samples = MAX2(src_rb->NumSamples, 1);
src_datatype = _mesa_get_format_datatype(src_rb->Format);
} else {
/* depth-or-color glCopyTexImage fallback path that passes a NULL rb and
* doesn't handle integer.
*/
samples = 1;
src_datatype = GL_UNSIGNED_NORMALIZED;
}
/* We expect only power of 2 samples in source multisample buffer. */
assert(samples > 0 && _mesa_is_pow_two(samples));
while (samples >> (shader_offset + 1)) {
shader_offset++;
}
/* Update the assert if we plan to support more than 16X MSAA. */
assert(shader_offset >= 0 && shader_offset <= 4);
if (drawFb->Visual.samples > 1) {
/* If you're calling meta_BlitFramebuffer with the destination
* multisampled, this is the only path that will work -- swrast and
* CopyTexImage won't work on it either.
*/
assert(ctx->Extensions.ARB_sample_shading);
dst_is_msaa = true;
/* We need shader invocation per sample, not per pixel */
_mesa_set_enable(ctx, GL_MULTISAMPLE, GL_TRUE);
_mesa_set_enable(ctx, GL_SAMPLE_SHADING, GL_TRUE);
_mesa_MinSampleShading(1.0);
}
switch (target) {
case GL_TEXTURE_2D_MULTISAMPLE:
case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
if (src_rb && (src_rb->_BaseFormat == GL_DEPTH_COMPONENT ||
src_rb->_BaseFormat == GL_DEPTH_STENCIL)) {
if (dst_is_msaa)
shader_index = BLIT_MSAA_SHADER_2D_MULTISAMPLE_DEPTH_COPY;
else
shader_index = BLIT_MSAA_SHADER_2D_MULTISAMPLE_DEPTH_RESOLVE;
} else {
if (dst_is_msaa)
shader_index = BLIT_MSAA_SHADER_2D_MULTISAMPLE_COPY;
else {
shader_index = BLIT_1X_MSAA_SHADER_2D_MULTISAMPLE_RESOLVE +
shader_offset;
}
}
if (target == GL_TEXTURE_2D_MULTISAMPLE_ARRAY) {
shader_index += (BLIT_1X_MSAA_SHADER_2D_MULTISAMPLE_ARRAY_RESOLVE -
BLIT_1X_MSAA_SHADER_2D_MULTISAMPLE_RESOLVE);
sampler_array_suffix = "Array";
texcoord_type = "vec3";
}
break;
default:
_mesa_problem(ctx, "Unkown texture target %s\n",
_mesa_enum_to_string(target));
shader_index = BLIT_2X_MSAA_SHADER_2D_MULTISAMPLE_RESOLVE;
}
/* We rely on the enum being sorted this way. */
STATIC_ASSERT(BLIT_1X_MSAA_SHADER_2D_MULTISAMPLE_RESOLVE_INT ==
BLIT_1X_MSAA_SHADER_2D_MULTISAMPLE_RESOLVE + 5);
STATIC_ASSERT(BLIT_1X_MSAA_SHADER_2D_MULTISAMPLE_RESOLVE_UINT ==
BLIT_1X_MSAA_SHADER_2D_MULTISAMPLE_RESOLVE + 10);
if (src_datatype == GL_INT) {
shader_index += 5;
vec4_prefix = "i";
} else if (src_datatype == GL_UNSIGNED_INT) {
shader_index += 10;
vec4_prefix = "u";
} else {
vec4_prefix = "";
}
if (blit->msaa_shaders[shader_index]) {
_mesa_UseProgram(blit->msaa_shaders[shader_index]);
return;
}
mem_ctx = ralloc_context(NULL);
if (shader_index == BLIT_MSAA_SHADER_2D_MULTISAMPLE_DEPTH_RESOLVE ||
shader_index == BLIT_MSAA_SHADER_2D_MULTISAMPLE_ARRAY_DEPTH_RESOLVE ||
shader_index == BLIT_MSAA_SHADER_2D_MULTISAMPLE_ARRAY_DEPTH_COPY ||
shader_index == BLIT_MSAA_SHADER_2D_MULTISAMPLE_DEPTH_COPY) {
char *sample_index;
const char *arb_sample_shading_extension_string;
if (dst_is_msaa) {
arb_sample_shading_extension_string = "#extension GL_ARB_sample_shading : enable";
sample_index = "gl_SampleID";
name = "depth MSAA copy";
} else {
/* Don't need that extension, since we're drawing to a single-sampled
* destination.
*/
arb_sample_shading_extension_string = "";
/* From the GL 4.3 spec:
*
* "If there is a multisample buffer (the value of SAMPLE_BUFFERS
* is one), then values are obtained from the depth samples in
* this buffer. It is recommended that the depth value of the
* centermost sample be used, though implementations may choose
* any function of the depth sample values at each pixel.
*
* We're slacking and instead of choosing centermost, we've got 0.
*/
sample_index = "0";
name = "depth MSAA resolve";
}
vs_source = ralloc_asprintf(mem_ctx,
"#version 130\n"
"in vec2 position;\n"
"in %s textureCoords;\n"
"out %s texCoords;\n"
"void main()\n"
"{\n"
" texCoords = textureCoords;\n"
" gl_Position = vec4(position, 0.0, 1.0);\n"
"}\n",
texcoord_type,
texcoord_type);
fs_source = ralloc_asprintf(mem_ctx,
"#version 130\n"
"#extension GL_ARB_texture_multisample : enable\n"
"%s\n"
"uniform sampler2DMS%s texSampler;\n"
"in %s texCoords;\n"
"out vec4 out_color;\n"
"\n"
"void main()\n"
"{\n"
" gl_FragDepth = texelFetch(texSampler, i%s(texCoords), %s).r;\n"
"}\n",
arb_sample_shading_extension_string,
sampler_array_suffix,
texcoord_type,
texcoord_type,
sample_index);
} else {
/* You can create 2D_MULTISAMPLE textures with 0 sample count (meaning 1
* sample). Yes, this is ridiculous.
*/
char *sample_resolve;
const char *arb_sample_shading_extension_string;
const char *merge_function;
name = ralloc_asprintf(mem_ctx, "%svec4 MSAA %s",
vec4_prefix,
dst_is_msaa ? "copy" : "resolve");
if (dst_is_msaa) {
arb_sample_shading_extension_string = "#extension GL_ARB_sample_shading : enable";
sample_resolve = ralloc_asprintf(mem_ctx, " out_color = texelFetch(texSampler, i%s(texCoords), gl_SampleID);", texcoord_type);
merge_function = "";
} else {
int i;
int step;
if (src_datatype == GL_INT || src_datatype == GL_UNSIGNED_INT) {
merge_function =
"gvec4 merge(gvec4 a, gvec4 b) { return (a >> gvec4(1)) + (b >> gvec4(1)) + (a & b & gvec4(1)); }\n";
} else {
/* The divide will happen at the end for floats. */
merge_function =
"vec4 merge(vec4 a, vec4 b) { return (a + b); }\n";
}
arb_sample_shading_extension_string = "";
/* We're assuming power of two samples for this resolution procedure.
*
* To avoid losing any floating point precision if the samples all
* happen to have the same value, we merge pairs of values at a time
* (so the floating point exponent just gets increased), rather than
* doing a naive sum and dividing.
*/
assert(_mesa_is_pow_two(samples));
/* Fetch each individual sample. */
sample_resolve = rzalloc_size(mem_ctx, 1);
for (i = 0; i < samples; i++) {
ralloc_asprintf_append(&sample_resolve,
" gvec4 sample_1_%d = texelFetch(texSampler, i%s(texCoords), %d);\n",
i, texcoord_type, i);
}
/* Now, merge each pair of samples, then merge each pair of those,
* etc.
*/
for (step = 2; step <= samples; step *= 2) {
for (i = 0; i < samples; i += step) {
ralloc_asprintf_append(&sample_resolve,
" gvec4 sample_%d_%d = merge(sample_%d_%d, sample_%d_%d);\n",
step, i,
step / 2, i,
step / 2, i + step / 2);
}
}
/* Scale the final result. */
if (src_datatype == GL_UNSIGNED_INT || src_datatype == GL_INT) {
ralloc_asprintf_append(&sample_resolve,
" out_color = sample_%d_0;\n",
samples);
} else {
ralloc_asprintf_append(&sample_resolve,
" gl_FragColor = sample_%d_0 / %f;\n",
samples, (float)samples);
}
}
vs_source = ralloc_asprintf(mem_ctx,
"#version 130\n"
"in vec2 position;\n"
"in %s textureCoords;\n"
"out %s texCoords;\n"
"void main()\n"
"{\n"
" texCoords = textureCoords;\n"
" gl_Position = vec4(position, 0.0, 1.0);\n"
"}\n",
texcoord_type,
texcoord_type);
fs_source = ralloc_asprintf(mem_ctx,
"#version 130\n"
"#extension GL_ARB_texture_multisample : enable\n"
"%s\n"
"#define gvec4 %svec4\n"
"uniform %ssampler2DMS%s texSampler;\n"
"in %s texCoords;\n"
"out gvec4 out_color;\n"
"\n"
"%s" /* merge_function */
"void main()\n"
"{\n"
"%s\n" /* sample_resolve */
"}\n",
arb_sample_shading_extension_string,
vec4_prefix,
vec4_prefix,
sampler_array_suffix,
texcoord_type,
merge_function,
sample_resolve);
}
_mesa_meta_compile_and_link_program(ctx, vs_source, fs_source, name,
&blit->msaa_shaders[shader_index]);
ralloc_free(mem_ctx);
}
void
program_resource_visitor::process(ir_variable *var)
{
const glsl_type *t = var->type;
const bool row_major =
var->data.matrix_layout == GLSL_MATRIX_LAYOUT_ROW_MAJOR;
/* false is always passed for the row_major parameter to the other
* processing functions because no information is available to do
* otherwise. See the warning in linker.h.
*/
/* Only strdup the name if we actually will need to modify it. */
if (var->data.from_named_ifc_block_array) {
/* lower_named_interface_blocks created this variable by lowering an
* interface block array to an array variable. For example if the
* original source code was:
*
* out Blk { vec4 bar } foo[3];
*
* Then the variable is now:
*
* out vec4 bar[3];
*
* We need to visit each array element using the names constructed like
* so:
*
* Blk[0].bar
* Blk[1].bar
* Blk[2].bar
*/
assert(t->is_array());
const glsl_type *ifc_type = var->get_interface_type();
char *name = ralloc_strdup(NULL, ifc_type->name);
size_t name_length = strlen(name);
for (unsigned i = 0; i < t->length; i++) {
size_t new_length = name_length;
ralloc_asprintf_rewrite_tail(&name, &new_length, "[%u].%s", i,
var->name);
/* Note: row_major is only meaningful for uniform blocks, and
* lowering is only applied to non-uniform interface blocks, so we
* can safely pass false for row_major.
*/
recursion(var->type, &name, new_length, row_major, NULL, false);
}
ralloc_free(name);
} else if (var->data.from_named_ifc_block_nonarray) {
/* lower_named_interface_blocks created this variable by lowering a
* named interface block (non-array) to an ordinary variable. For
* example if the original source code was:
*
* out Blk { vec4 bar } foo;
*
* Then the variable is now:
*
* out vec4 bar;
*
* We need to visit this variable using the name:
*
* Blk.bar
*/
const glsl_type *ifc_type = var->get_interface_type();
char *name = ralloc_asprintf(NULL, "%s.%s", ifc_type->name, var->name);
/* Note: row_major is only meaningful for uniform blocks, and lowering
* is only applied to non-uniform interface blocks, so we can safely
* pass false for row_major.
*/
recursion(var->type, &name, strlen(name), row_major, NULL, false);
ralloc_free(name);
} else if (t->without_array()->is_record()) {
char *name = ralloc_strdup(NULL, var->name);
recursion(var->type, &name, strlen(name), row_major, NULL, false);
ralloc_free(name);
} else if (t->is_interface()) {
char *name = ralloc_strdup(NULL, var->type->name);
recursion(var->type, &name, strlen(name), row_major, NULL, false);
ralloc_free(name);
} else if (t->is_array() && t->fields.array->is_interface()) {
char *name = ralloc_strdup(NULL, var->type->fields.array->name);
recursion(var->type, &name, strlen(name), row_major, NULL, false);
ralloc_free(name);
} else {
this->visit_field(t, var->name, row_major, NULL, false);
}
}
static void
setup_glsl_msaa_blit_scaled_shader(struct gl_context *ctx,
struct blit_state *blit,
struct gl_renderbuffer *src_rb,
GLenum target, GLenum filter)
{
GLint loc_src_width, loc_src_height;
int i, samples;
int shader_offset = 0;
void *mem_ctx = ralloc_context(NULL);
char *fs_source;
char *name, *sample_number;
const uint8_t *sample_map;
char *sample_map_str = rzalloc_size(mem_ctx, 1);
char *sample_map_expr = rzalloc_size(mem_ctx, 1);
char *texel_fetch_macro = rzalloc_size(mem_ctx, 1);
const char *sampler_array_suffix = "";
float y_scale;
enum blit_msaa_shader shader_index;
assert(src_rb);
samples = MAX2(src_rb->NumSamples, 1);
y_scale = samples * 0.5;
/* We expect only power of 2 samples in source multisample buffer. */
assert(samples > 0 && _mesa_is_pow_two(samples));
while (samples >> (shader_offset + 1)) {
shader_offset++;
}
/* Update the assert if we plan to support more than 8X MSAA. */
assert(shader_offset > 0 && shader_offset < 4);
assert(target == GL_TEXTURE_2D_MULTISAMPLE ||
target == GL_TEXTURE_2D_MULTISAMPLE_ARRAY);
shader_index = BLIT_2X_MSAA_SHADER_2D_MULTISAMPLE_SCALED_RESOLVE +
shader_offset - 1;
if (target == GL_TEXTURE_2D_MULTISAMPLE_ARRAY) {
shader_index += BLIT_2X_MSAA_SHADER_2D_MULTISAMPLE_ARRAY_SCALED_RESOLVE -
BLIT_2X_MSAA_SHADER_2D_MULTISAMPLE_SCALED_RESOLVE;
sampler_array_suffix = "Array";
}
if (blit->msaa_shaders[shader_index]) {
_mesa_UseProgram(blit->msaa_shaders[shader_index]);
/* Update the uniform values. */
loc_src_width =
_mesa_GetUniformLocation(blit->msaa_shaders[shader_index], "src_width");
loc_src_height =
_mesa_GetUniformLocation(blit->msaa_shaders[shader_index], "src_height");
_mesa_Uniform1f(loc_src_width, src_rb->Width);
_mesa_Uniform1f(loc_src_height, src_rb->Height);
return;
}
name = ralloc_asprintf(mem_ctx, "vec4 MSAA scaled resolve");
/* Below switch is used to setup the shader expression, which computes
* sample index and map it to to a sample number on hardware.
*/
switch(samples) {
case 2:
sample_number = "sample_map[int(2 * fract(coord.x))]";
sample_map = ctx->Const.SampleMap2x;
break;
case 4:
sample_number = "sample_map[int(2 * fract(coord.x) + 4 * fract(coord.y))]";
sample_map = ctx->Const.SampleMap4x;
break;
case 8:
sample_number = "sample_map[int(2 * fract(coord.x) + 8 * fract(coord.y))]";
sample_map = ctx->Const.SampleMap8x;
break;
default:
sample_number = NULL;
sample_map = NULL;
_mesa_problem(ctx, "Unsupported sample count %d\n", samples);
unreachable("Unsupported sample count");
}
/* Create sample map string. */
for (i = 0 ; i < samples - 1; i++) {
ralloc_asprintf_append(&sample_map_str, "%d, ", sample_map[i]);
}
ralloc_asprintf_append(&sample_map_str, "%d", sample_map[samples - 1]);
/* Create sample map expression using above string. */
ralloc_asprintf_append(&sample_map_expr,
" const int sample_map[%d] = int[%d](%s);\n",
samples, samples, sample_map_str);
if (target == GL_TEXTURE_2D_MULTISAMPLE) {
ralloc_asprintf_append(&texel_fetch_macro,
"#define TEXEL_FETCH(coord) texelFetch(texSampler, ivec2(coord), %s);\n",
sample_number);
} else {
ralloc_asprintf_append(&texel_fetch_macro,
"#define TEXEL_FETCH(coord) texelFetch(texSampler, ivec3(coord, layer), %s);\n",
sample_number);
}
static const char vs_source[] =
"#version 130\n"
"in vec2 position;\n"
"in vec3 textureCoords;\n"
"out vec2 texCoords;\n"
"flat out int layer;\n"
"void main()\n"
"{\n"
" texCoords = textureCoords.xy;\n"
" layer = int(textureCoords.z);\n"
" gl_Position = vec4(position, 0.0, 1.0);\n"
"}\n"
;
fs_source = ralloc_asprintf(mem_ctx,
"#version 130\n"
"#extension GL_ARB_texture_multisample : enable\n"
"uniform sampler2DMS%s texSampler;\n"
"uniform float src_width, src_height;\n"
"in vec2 texCoords;\n"
"flat in int layer;\n"
"out vec4 out_color;\n"
"\n"
"void main()\n"
"{\n"
"%s"
" vec2 interp;\n"
" const vec2 scale = vec2(2.0f, %ff);\n"
" const vec2 scale_inv = vec2(0.5f, %ff);\n"
" const vec2 s_0_offset = vec2(0.25f, %ff);\n"
" vec2 s_0_coord, s_1_coord, s_2_coord, s_3_coord;\n"
" vec4 s_0_color, s_1_color, s_2_color, s_3_color;\n"
" vec4 x_0_color, x_1_color;\n"
" vec2 tex_coord = texCoords - s_0_offset;\n"
"\n"
" tex_coord *= scale;\n"
" clamp(tex_coord.x, 0.0f, scale.x * src_width - 1.0f);\n"
" clamp(tex_coord.y, 0.0f, scale.y * src_height - 1.0f);\n"
" interp = fract(tex_coord);\n"
" tex_coord = ivec2(tex_coord) * scale_inv;\n"
"\n"
" /* Compute the sample coordinates used for filtering. */\n"
" s_0_coord = tex_coord;\n"
" s_1_coord = tex_coord + vec2(scale_inv.x, 0.0f);\n"
" s_2_coord = tex_coord + vec2(0.0f, scale_inv.y);\n"
" s_3_coord = tex_coord + vec2(scale_inv.x, scale_inv.y);\n"
"\n"
" /* Fetch sample color values. */\n"
"%s"
" s_0_color = TEXEL_FETCH(s_0_coord)\n"
" s_1_color = TEXEL_FETCH(s_1_coord)\n"
" s_2_color = TEXEL_FETCH(s_2_coord)\n"
" s_3_color = TEXEL_FETCH(s_3_coord)\n"
"#undef TEXEL_FETCH\n"
"\n"
" /* Do bilinear filtering on sample colors. */\n"
" x_0_color = mix(s_0_color, s_1_color, interp.x);\n"
" x_1_color = mix(s_2_color, s_3_color, interp.x);\n"
" out_color = mix(x_0_color, x_1_color, interp.y);\n"
"}\n",
sampler_array_suffix,
sample_map_expr,
y_scale,
1.0f / y_scale,
1.0f / samples,
texel_fetch_macro);
_mesa_meta_compile_and_link_program(ctx, vs_source, fs_source, name,
&blit->msaa_shaders[shader_index]);
loc_src_width =
_mesa_GetUniformLocation(blit->msaa_shaders[shader_index], "src_width");
loc_src_height =
_mesa_GetUniformLocation(blit->msaa_shaders[shader_index], "src_height");
_mesa_Uniform1f(loc_src_width, src_rb->Width);
_mesa_Uniform1f(loc_src_height, src_rb->Height);
ralloc_free(mem_ctx);
}
static bool load_body(ALLEGRO_CONFIG *cfg, const char *fullname,
Body *body, char **primary_name)
{
const char *name;
char *type;
/* Just checking */
if (al_get_first_config_entry(cfg, fullname, NULL) == NULL)
{
log_err("Section %s not in file\n", fullname);
return false;
}
/* Fill in some common parameters */
if (!config_get_double(cfg, fullname, "Mass", &body->mass, false))
return false;
if (!config_get_double(cfg, fullname, "Gravitational parameter",
&body->grav_param, true))
{
body->grav_param = GRAV_CONST * body->mass;
}
if (!config_get_double(cfg, fullname, "Radius", &body->radius, false))
return false;
/* Figure out what kind of object it is */
config_get_string(cfg, fullname, "Type", &type, true);
if (type == NULL)
body->type = BODY_UNKNOWN;
else if (strcmp(type, "Star") == 0)
body->type = BODY_STAR;
else if (strcmp(type, "Planet") == 0)
body->type = BODY_PLANET;
else if (strcmp(type, "Moon") == 0)
body->type = BODY_PLANET;
else if (strcmp(type, "Comet") == 0)
body->type = BODY_COMET;
else
{
log_err("Unknown type: %s\n", type);
ralloc_free(type);
return false;
}
ralloc_free(type);
/* Does it have a primary or not?
* Full names are of the form of "Primary/Name"
* We search backwards to allow for things like "Sol/Earth/Moon" */
if ((name = strrchr(fullname, '/')) == NULL)
{
/* This is a body without a primary */
body->name = ralloc_strdup(body->ctx, fullname);
body->type = (body->type == BODY_UNKNOWN ? BODY_STAR : body->type);
body->primary = NULL;
*primary_name = NULL;
} else if (name == fullname) /* No primary name, eg: sec = "/Earth" */
{
log_err("Malformed name: %s", fullname);
return false;
} else
{
const char *c;
for (c = name - 1; c >= fullname && *c != '/'; c--);
c++;
body->name = ralloc_strdup(body->ctx, name + 1);
body->type = (body->type == BODY_UNKNOWN ? BODY_PLANET : body->type);
body->primary = NULL; /* Fill in later */
*primary_name = ralloc_strndup(body->ctx, c, name - c);
}
body->num_satellites = 0;
body->satellite = NULL;
/* Bodies without primaries can't orbit another body */
if (*primary_name == NULL)
return true;
if (!config_get_double(cfg, fullname, "Ecc", &body->orbit.Ecc, false) ||
!config_get_double(cfg, fullname, "SMa", &body->orbit.SMa, false) ||
!config_get_double(cfg, fullname, "Inc", &body->orbit.Inc, false) ||
!config_get_double(cfg, fullname, "LAN", &body->orbit.LAN, false) ||
!config_get_double(cfg, fullname, "APe", &body->orbit.APe, false) ||
!config_get_double(cfg, fullname, "MnA", &body->orbit.MnA, false))
{
log_err("Couldn't load orbital elements of %s\n", fullname);
return false;
}
return true;
}
~glslopt_ctx() {
ralloc_free (mem_ctx);
}
static SolarSystem *load_from_config(ALLEGRO_CONFIG *cfg)
{
SolarSystem *solsys;
char **primary_names;
const char *name;
long num_bodies;
ALLEGRO_CONFIG_SECTION *sec;
int i;
/* First pass: Determine the number of bodies in the file */
num_bodies = 0;
sec = NULL;
while ((name = get_next_config_section(cfg, &sec)) != NULL)
{
if (name[0] != '\0')
num_bodies++;
}
if (num_bodies == 0)
return NULL; /* Empty solarsystem */
solsys = ralloc_size(NULL, sizeof(SolarSystem) + num_bodies*sizeof(Body));
if (solsys == NULL)
return NULL;
solsys->num_bodies = num_bodies;
primary_names = ralloc_array(solsys, char *, num_bodies);
if (primary_names == NULL)
{
ralloc_free(solsys);
return NULL;
}
/* Second pass: Load all celestial bodies */
i = 0;
sec = NULL;
while ((name = get_next_config_section(cfg, &sec)) != NULL && i < num_bodies)
{
if (name[0] == '\0')
continue;
solsys->body[i].ctx = solsys;
if (!load_body(cfg, name, &solsys->body[i], &primary_names[i]))
{
log_err("Couldn't load body %s\n", name);
ralloc_free(solsys);
return NULL;
}
i++;
}
if (i < num_bodies)
log_err("Internal consistency error\n");
/* Third pass: Connect each satellite body to its primary */
for (i = 0; i < num_bodies; i++)
{
Body *body = &solsys->body[i];
char *primary_name = primary_names[i];
if (primary_name == NULL) /* Independent body */
continue;
/* Look for the primary */
body->primary = NULL;
for (int j = 0; j < num_bodies; j++)
{
Body *body2 = &solsys->body[j];
if (strcmp(primary_name, body2->name) == 0)
{
body->primary = body2;
break;
}
}
if (body->primary == NULL)
{
log_err("Couldn't find %s's primary: %s\n", body->name,
primary_name);
ralloc_free(solsys);
return NULL;
}
ralloc_free(primary_name);
primary_name = NULL; /* Won't ever be used again */
body->primary->num_satellites++;
body->primary->satellite = reralloc(solsys, body->primary->satellite,
Body *, body->primary->num_satellites);
if (body->primary->satellite == NULL)
{
log_err("Out of memory\n");
ralloc_free(solsys);
return NULL;
}
body->primary->satellite[body->primary->num_satellites - 1] = body;
body->orbit.epoch = 0;
body->orbit.period = M_TWO_PI *
sqrt(CUBE(body->orbit.SMa) / body->primary->grav_param);
body->orbit.plane_orientation = quat_euler(RAD(body->orbit.LAN),
RAD(body->orbit.Inc),
RAD(body->orbit.APe));
}
ralloc_free(primary_names);
return solsys;
}
/* If the user *does* call delete, that's OK, we will just
* ralloc_free in that case. */
static void operator delete(void *node)
{
ralloc_free(node);
}
loop_state::~loop_state()
{
hash_table_dtor(this->ht);
ralloc_free(this->mem_ctx);
}
static bool
do_vs_prog(struct brw_context *brw,
struct gl_shader_program *prog,
struct brw_vertex_program *vp,
struct brw_vs_prog_key *key)
{
struct gl_context *ctx = &brw->intel.ctx;
struct intel_context *intel = &brw->intel;
GLuint program_size;
const GLuint *program;
struct brw_vs_compile c;
void *mem_ctx;
int aux_size;
int i;
struct gl_shader *vs = NULL;
if (prog)
vs = prog->_LinkedShaders[MESA_SHADER_VERTEX];
memset(&c, 0, sizeof(c));
memcpy(&c.key, key, sizeof(*key));
mem_ctx = ralloc_context(NULL);
brw_init_compile(brw, &c.func, mem_ctx);
c.vp = vp;
/* 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;
if (vs) {
/* We add padding around uniform values below vec4 size, with the worst
* case being a float value that gets blown up to a vec4, so be
* conservative here.
*/
param_count = vs->num_uniform_components * 4;
/* We also upload clip plane data as uniforms */
param_count += MAX_CLIP_PLANES * 4;
} else {
param_count = vp->program.Base.Parameters->NumParameters * 4;
}
c.prog_data.param = rzalloc_array(NULL, const float *, param_count);
c.prog_data.pull_param = rzalloc_array(NULL, const float *, param_count);
c.prog_data.outputs_written = vp->program.Base.OutputsWritten;
c.prog_data.inputs_read = vp->program.Base.InputsRead;
if (c.key.copy_edgeflag) {
c.prog_data.outputs_written |= BITFIELD64_BIT(VERT_RESULT_EDGE);
c.prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
}
/* 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 (c.key.point_coord_replace & (1 << i))
c.prog_data.outputs_written |= BITFIELD64_BIT(VERT_RESULT_TEX0 + i);
}
brw_compute_vue_map(&c);
if (0) {
_mesa_fprint_program_opt(stdout, &c.vp->program.Base, PROG_PRINT_DEBUG,
true);
}
/* Emit GEN4 code.
*/
if (prog) {
if (!brw_vs_emit(prog, &c)) {
ralloc_free(mem_ctx);
return false;
}
} else {
brw_old_vs_emit(&c);
}
if (c.prog_data.nr_pull_params)
c.prog_data.num_surfaces = 1;
if (c.vp->program.Base.SamplersUsed)
c.prog_data.num_surfaces = SURF_INDEX_VS_TEXTURE(BRW_MAX_TEX_UNIT);
if (prog &&
prog->_LinkedShaders[MESA_SHADER_VERTEX]->NumUniformBlocks) {
c.prog_data.num_surfaces =
SURF_INDEX_VS_UBO(prog->_LinkedShaders[MESA_SHADER_VERTEX]->NumUniformBlocks);
}
/* Scratch space is used for register spilling */
if (c.last_scratch) {
perf_debug("Vertex shader triggered register spilling. "
"Try reducing the number of live vec4 values to "
"improve performance.\n");
c.prog_data.total_scratch = brw_get_scratch_size(c.last_scratch);
brw_get_scratch_bo(intel, &brw->vs.scratch_bo,
c.prog_data.total_scratch * brw->max_vs_threads);
}
/* get the program
*/
program = brw_get_program(&c.func, &program_size);
/* We upload from &c.prog_data including the constant_map assuming
* they're packed together. It would be nice to have a
* compile-time assert macro here.
*/
assert(c.constant_map == (int8_t *)&c.prog_data +
sizeof(c.prog_data));
assert(ctx->Const.VertexProgram.MaxNativeParameters ==
ARRAY_SIZE(c.constant_map));
(void) ctx;
aux_size = sizeof(c.prog_data);
/* constant_map */
aux_size += c.vp->program.Base.Parameters->NumParameters;
brw_upload_cache(&brw->cache, BRW_VS_PROG,
&c.key, sizeof(c.key),
program, program_size,
&c.prog_data, aux_size,
&brw->vs.prog_offset, &brw->vs.prog_data);
ralloc_free(mem_ctx);
return true;
}
bool
brw_codegen_gs_prog(struct brw_context *brw,
struct gl_shader_program *prog,
struct brw_geometry_program *gp,
struct brw_gs_prog_key *key)
{
struct brw_compiler *compiler = brw->intelScreen->compiler;
struct brw_stage_state *stage_state = &brw->gs.base;
struct brw_gs_prog_data prog_data;
bool start_busy = false;
double start_time = 0;
memset(&prog_data, 0, sizeof(prog_data));
assign_gs_binding_table_offsets(brw->intelScreen->devinfo, prog,
&gp->program.Base, &prog_data);
/* 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.
*
* Note: param_count needs to be num_uniform_components * 4, since we add
* padding around uniform values below vec4 size, so the worst case is that
* every uniform is a float which gets padded to the size of a vec4.
*/
struct gl_linked_shader *gs = prog->_LinkedShaders[MESA_SHADER_GEOMETRY];
struct brw_shader *bgs = (struct brw_shader *) gs;
int param_count = gp->program.Base.nir->num_uniforms / 4;
prog_data.base.base.param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.base.pull_param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.base.image_param =
rzalloc_array(NULL, struct brw_image_param, gs->NumImages);
prog_data.base.base.nr_params = param_count;
prog_data.base.base.nr_image_params = gs->NumImages;
brw_nir_setup_glsl_uniforms(gp->program.Base.nir, prog, &gp->program.Base,
&prog_data.base.base,
compiler->scalar_stage[MESA_SHADER_GEOMETRY]);
GLbitfield64 outputs_written = gp->program.Base.OutputsWritten;
prog_data.base.cull_distance_mask =
((1 << gp->program.Base.CullDistanceArraySize) - 1) <<
gp->program.Base.ClipDistanceArraySize;
brw_compute_vue_map(brw->intelScreen->devinfo,
&prog_data.base.vue_map, outputs_written,
prog->SeparateShader);
if (unlikely(INTEL_DEBUG & DEBUG_GS))
brw_dump_ir("geometry", prog, gs, NULL);
int st_index = -1;
if (INTEL_DEBUG & DEBUG_SHADER_TIME)
st_index = brw_get_shader_time_index(brw, prog, NULL, ST_GS);
if (unlikely(brw->perf_debug)) {
start_busy = brw->batch.last_bo && drm_intel_bo_busy(brw->batch.last_bo);
start_time = get_time();
}
void *mem_ctx = ralloc_context(NULL);
unsigned program_size;
char *error_str;
const unsigned *program =
brw_compile_gs(brw->intelScreen->compiler, brw, mem_ctx, key,
&prog_data, gs->Program->nir, prog,
st_index, &program_size, &error_str);
if (program == NULL) {
ralloc_free(mem_ctx);
return false;
}
if (unlikely(brw->perf_debug)) {
if (bgs->compiled_once) {
brw_gs_debug_recompile(brw, prog, key);
}
if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
perf_debug("GS compile took %.03f ms and stalled the GPU\n",
(get_time() - start_time) * 1000);
}
bgs->compiled_once = true;
}
/* Scratch space is used for register spilling */
brw_alloc_stage_scratch(brw, stage_state,
prog_data.base.base.total_scratch,
brw->max_gs_threads);
brw_upload_cache(&brw->cache, BRW_CACHE_GS_PROG,
key, sizeof(*key),
program, program_size,
&prog_data, sizeof(prog_data),
&stage_state->prog_offset, &brw->gs.prog_data);
ralloc_free(mem_ctx);
return true;
}
extern GLboolean
_mesa_validate_program_pipeline(struct gl_context* ctx,
struct gl_pipeline_object *pipe,
GLboolean IsBound)
{
unsigned i;
pipe->Validated = GL_FALSE;
/* Release and reset the info log.
*/
if (pipe->InfoLog != NULL)
ralloc_free(pipe->InfoLog);
pipe->InfoLog = NULL;
/* Section 2.11.11 (Shader Execution), subheading "Validation," of the
* OpenGL 4.1 spec says:
*
* "[INVALID_OPERATION] is generated by any command that transfers
* vertices to the GL if:
*
* - A program object is active for at least one, but not all of
* the shader stages that were present when the program was
* linked."
*
* For each possible program stage, verify that the program bound to that
* stage has all of its stages active. In other words, if the program
* bound to the vertex stage also has a fragment shader, the fragment
* shader must also be bound to the fragment stage.
*/
for (i = 0; i < MESA_SHADER_STAGES; i++) {
if (!program_stages_all_active(pipe, pipe->CurrentProgram[i])) {
goto err;
}
}
/* Section 2.11.11 (Shader Execution), subheading "Validation," of the
* OpenGL 4.1 spec says:
*
* "[INVALID_OPERATION] is generated by any command that transfers
* vertices to the GL if:
*
* ...
*
* - One program object is active for at least two shader stages
* and a second program is active for a shader stage between two
* stages for which the first program was active."
*
* Without Tesselation, the only case where this can occur is the geometry
* shader between the fragment shader and vertex shader.
*/
if (pipe->CurrentProgram[MESA_SHADER_GEOMETRY]
&& pipe->CurrentProgram[MESA_SHADER_FRAGMENT]
&& pipe->CurrentProgram[MESA_SHADER_VERTEX]) {
if (pipe->CurrentProgram[MESA_SHADER_VERTEX]->Name == pipe->CurrentProgram[MESA_SHADER_FRAGMENT]->Name &&
pipe->CurrentProgram[MESA_SHADER_GEOMETRY]->Name != pipe->CurrentProgram[MESA_SHADER_VERTEX]->Name) {
pipe->InfoLog =
ralloc_asprintf(pipe,
"Program %d is active for geometry stage between "
"two stages for which another program %d is "
"active",
pipe->CurrentProgram[MESA_SHADER_GEOMETRY]->Name,
pipe->CurrentProgram[MESA_SHADER_VERTEX]->Name);
goto err;
}
}
/* Section 2.11.11 (Shader Execution), subheading "Validation," of the
* OpenGL 4.1 spec says:
*
* "[INVALID_OPERATION] is generated by any command that transfers
* vertices to the GL if:
*
* ...
*
* - There is an active program for tessellation control,
* tessellation evaluation, or geometry stages with corresponding
* executable shader, but there is no active program with
* executable vertex shader."
*/
if (!pipe->CurrentProgram[MESA_SHADER_VERTEX]
&& pipe->CurrentProgram[MESA_SHADER_GEOMETRY]) {
pipe->InfoLog = ralloc_strdup(pipe, "Program lacks a vertex shader");
goto err;
}
/* Section 2.11.11 (Shader Execution), subheading "Validation," of the
* OpenGL 4.1 spec says:
*
* "[INVALID_OPERATION] is generated by any command that transfers
* vertices to the GL if:
*
* ...
*
* - There is no current program object specified by UseProgram,
* there is a current program pipeline object, and the current
* program for any shader stage has been relinked since being
* applied to the pipeline object via UseProgramStages with the
* PROGRAM_SEPARABLE parameter set to FALSE.
*/
for (i = 0; i < MESA_SHADER_STAGES; i++) {
if (pipe->CurrentProgram[i] && !pipe->CurrentProgram[i]->SeparateShader) {
pipe->InfoLog = ralloc_asprintf(pipe,
"Program %d was relinked without "
"PROGRAM_SEPARABLE state",
pipe->CurrentProgram[i]->Name);
goto err;
}
}
/* Section 2.11.11 (Shader Execution), subheading "Validation," of the
* OpenGL 4.1 spec says:
*
* "[INVALID_OPERATION] is generated by any command that transfers
* vertices to the GL if:
*
* ...
*
* - Any two active samplers in the current program object are of
* different types, but refer to the same texture image unit.
*
* - The number of active samplers in the program exceeds the
* maximum number of texture image units allowed."
*/
if (!_mesa_sampler_uniforms_pipeline_are_valid(pipe))
goto err;
pipe->Validated = GL_TRUE;
return GL_TRUE;
err:
if (IsBound)
_mesa_error(ctx, GL_INVALID_OPERATION,
"glValidateProgramPipeline failed to validate the pipeline");
return GL_FALSE;
}
static bool
do_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_compile c;
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 gl_shader *vs = NULL;
if (prog)
vs = prog->_LinkedShaders[MESA_SHADER_VERTEX];
memset(&c, 0, sizeof(c));
memcpy(&c.key, key, sizeof(*key));
memset(&prog_data, 0, sizeof(prog_data));
mem_ctx = ralloc_context(NULL);
c.vp = vp;
/* 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;
if (vs) {
/* We add padding around uniform values below vec4 size, with the worst
* case being a float value that gets blown up to a vec4, so be
* conservative here.
*/
param_count = vs->num_uniform_components * 4;
} else {
param_count = vp->program.Base.Parameters->NumParameters * 4;
}
/* vec4_visitor::setup_uniform_clipplane_values() also uploads user clip
* planes as uniforms.
*/
param_count += c.key.base.nr_userclip_plane_consts * 4;
stage_prog_data->param = rzalloc_array(NULL, const float *, param_count);
stage_prog_data->pull_param = rzalloc_array(NULL, const float *, param_count);
/* Setting nr_params here NOT to the size of the param and pull_param
* arrays, but to the number of uniform components vec4_visitor
* needs. vec4_visitor::setup_uniforms() will set it back to a proper value.
*/
stage_prog_data->nr_params = ALIGN(param_count, 4) / 4;
if (vs) {
stage_prog_data->nr_params += vs->num_samplers;
}
GLbitfield64 outputs_written = vp->program.Base.OutputsWritten;
prog_data.inputs_read = vp->program.Base.InputsRead;
if (c.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 (c.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 (c.key.base.userclip_active) {
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST0);
outputs_written |= BITFIELD64_BIT(VARYING_SLOT_CLIP_DIST1);
}
brw_compute_vue_map(brw, &prog_data.base.vue_map, outputs_written);
if (0) {
_mesa_fprint_program_opt(stderr, &c.vp->program.Base, PROG_PRINT_DEBUG,
true);
}
/* Emit GEN4 code.
*/
program = brw_vs_emit(brw, prog, &c, &prog_data, mem_ctx, &program_size);
if (program == NULL) {
ralloc_free(mem_ctx);
return false;
}
/* Scratch space is used for register spilling */
if (c.base.last_scratch) {
perf_debug("Vertex shader triggered register spilling. "
"Try reducing the number of live vec4 values to "
"improve performance.\n");
prog_data.base.total_scratch
= brw_get_scratch_size(c.base.last_scratch*REG_SIZE);
brw_get_scratch_bo(brw, &brw->vs.base.scratch_bo,
prog_data.base.total_scratch * brw->max_vs_threads);
}
brw_upload_cache(&brw->cache, BRW_VS_PROG,
&c.key, sizeof(c.key),
program, program_size,
&prog_data, sizeof(prog_data),
&brw->vs.base.prog_offset, &brw->vs.prog_data);
ralloc_free(mem_ctx);
return true;
}
output_read_remover::~output_read_remover()
{
hash_table_dtor(replacements);
ralloc_free(mem_ctx);
}
glsl_symbol_table::~glsl_symbol_table()
{
_mesa_symbol_table_dtor(table);
ralloc_free(mem_ctx);
}
/**
* All Mesa program -> GPU code generation goes through this function.
* Depending on the instructions used (i.e. flow control instructions)
* we'll use one of two code generators.
*/
bool
brw_codegen_wm_prog(struct brw_context *brw,
struct gl_shader_program *prog,
struct brw_fragment_program *fp,
struct brw_wm_prog_key *key)
{
struct gl_context *ctx = &brw->ctx;
void *mem_ctx = ralloc_context(NULL);
struct brw_wm_prog_data prog_data;
const GLuint *program;
struct gl_shader *fs = NULL;
GLuint program_size;
if (prog)
fs = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
memset(&prog_data, 0, sizeof(prog_data));
/* key->alpha_test_func means simulating alpha testing via discards,
* so the shader definitely kills pixels.
*/
prog_data.uses_kill = fp->program.UsesKill || key->alpha_test_func;
prog_data.uses_omask =
fp->program.Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK);
prog_data.computed_depth_mode = computed_depth_mode(&fp->program);
/* Use ALT floating point mode for ARB programs so that 0^0 == 1. */
if (!prog)
prog_data.base.use_alt_mode = true;
/* 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;
if (fs) {
param_count = fs->num_uniform_components;
} else {
param_count = fp->program.Base.Parameters->NumParameters * 4;
}
/* The backend also sometimes adds params for texture size. */
param_count += 2 * ctx->Const.Program[MESA_SHADER_FRAGMENT].MaxTextureImageUnits;
prog_data.base.param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.pull_param =
rzalloc_array(NULL, const gl_constant_value *, param_count);
prog_data.base.nr_params = param_count;
prog_data.barycentric_interp_modes =
brw_compute_barycentric_interp_modes(brw, key->flat_shade,
key->persample_shading,
&fp->program);
program = brw_wm_fs_emit(brw, mem_ctx, key, &prog_data,
&fp->program, prog, &program_size);
if (program == NULL) {
ralloc_free(mem_ctx);
return false;
}
if (prog_data.base.total_scratch) {
brw_get_scratch_bo(brw, &brw->wm.base.scratch_bo,
prog_data.base.total_scratch * brw->max_wm_threads);
}
if (unlikely(INTEL_DEBUG & DEBUG_WM))
fprintf(stderr, "\n");
brw_upload_cache(&brw->cache, BRW_CACHE_FS_PROG,
key, sizeof(struct brw_wm_prog_key),
program, program_size,
&prog_data, sizeof(prog_data),
&brw->wm.base.prog_offset, &brw->wm.prog_data);
ralloc_free(mem_ctx);
return true;
}
static void compile_clip_prog( struct brw_context *brw,
struct brw_clip_prog_key *key )
{
struct brw_clip_compile c;
const GLuint *program;
void *mem_ctx;
GLuint program_size;
memset(&c, 0, sizeof(c));
mem_ctx = ralloc_context(NULL);
/* Begin the compilation:
*/
brw_init_codegen(&brw->screen->devinfo, &c.func, mem_ctx);
c.func.single_program_flow = 1;
c.key = *key;
c.vue_map = brw->vue_map_geom_out;
/* nr_regs is the number of registers filled by reading data from the VUE.
* This program accesses the entire VUE, so nr_regs needs to be the size of
* the VUE (measured in pairs, since two slots are stored in each
* register).
*/
c.nr_regs = (c.vue_map.num_slots + 1)/2;
c.prog_data.clip_mode = c.key.clip_mode; /* XXX */
/* For some reason the thread is spawned with only 4 channels
* unmasked.
*/
brw_set_default_mask_control(&c.func, BRW_MASK_DISABLE);
/* Would ideally have the option of producing a program which could
* do all three:
*/
switch (key->primitive) {
case GL_TRIANGLES:
if (key->do_unfilled)
brw_emit_unfilled_clip( &c );
else
brw_emit_tri_clip( &c );
break;
case GL_LINES:
brw_emit_line_clip( &c );
break;
case GL_POINTS:
brw_emit_point_clip( &c );
break;
default:
unreachable("not reached");
}
brw_compact_instructions(&c.func, 0, 0, NULL);
/* get the program
*/
program = brw_get_program(&c.func, &program_size);
if (unlikely(INTEL_DEBUG & DEBUG_CLIP)) {
fprintf(stderr, "clip:\n");
brw_disassemble(&brw->screen->devinfo, c.func.store,
0, program_size, stderr);
fprintf(stderr, "\n");
}
brw_upload_cache(&brw->cache,
BRW_CACHE_CLIP_PROG,
&c.key, sizeof(c.key),
program, program_size,
&c.prog_data, sizeof(c.prog_data),
&brw->clip.prog_offset, &brw->clip.prog_data);
ralloc_free(mem_ctx);
}
