static void
read_tgsi_from_cache(struct blob_reader *blob_reader,
const struct tgsi_token **tokens,
unsigned *num_tokens)
{
*num_tokens = blob_read_uint32(blob_reader);
unsigned tokens_size = *num_tokens * sizeof(struct tgsi_token);
*tokens = (const struct tgsi_token*) MALLOC(tokens_size);
blob_copy_bytes(blob_reader, (uint8_t *) *tokens, tokens_size);
}
/* Test that we detect overrun. */
static void
test_overrun(void)
{
struct blob blob;
struct blob_reader reader;
uint32_t value = 0xdeadbeef;
blob_init(&blob);
blob_write_uint32(&blob, value);
blob_reader_init(&reader, blob.data, blob.size);
expect_equal(value, blob_read_uint32(&reader), "read before overrun");
expect_equal(false, reader.overrun, "overrun flag not set");
expect_equal(0, blob_read_uint32(&reader), "read at overrun");
expect_equal(true, reader.overrun, "overrun flag set");
blob_finish(&blob);
}
/* Test that we detect overrun. */
static void
test_overrun(void)
{
void *ctx =ralloc_context(NULL);
struct blob *blob;
struct blob_reader reader;
uint32_t value = 0xdeadbeef;
blob = blob_create(ctx);
blob_write_uint32(blob, value);
blob_reader_init(&reader, blob->data, blob->size);
expect_equal(value, blob_read_uint32(&reader), "read before overrun");
expect_equal(false, reader.overrun, "overrun flag not set");
expect_equal(0, blob_read_uint32(&reader), "read at overrun");
expect_equal(true, reader.overrun, "overrun flag set");
ralloc_free(ctx);
}
static nir_constant *
read_constant(read_ctx *ctx, nir_variable *nvar)
{
nir_constant *c = ralloc(nvar, nir_constant);
blob_copy_bytes(ctx->blob, (uint8_t *)c->values, sizeof(c->values));
c->num_elements = blob_read_uint32(ctx->blob);
c->elements = ralloc_array(nvar, nir_constant *, c->num_elements);
for (unsigned i = 0; i < c->num_elements; i++)
c->elements[i] = read_constant(ctx, nvar);
return c;
}
static nir_variable *
read_variable(read_ctx *ctx)
{
nir_variable *var = rzalloc(ctx->nir, nir_variable);
read_add_object(ctx, var);
var->type = decode_type_from_blob(ctx->blob);
bool has_name = blob_read_uint32(ctx->blob);
if (has_name) {
const char *name = blob_read_string(ctx->blob);
var->name = ralloc_strdup(var, name);
} else {
var->name = NULL;
}
blob_copy_bytes(ctx->blob, (uint8_t *) &var->data, sizeof(var->data));
var->num_state_slots = blob_read_uint32(ctx->blob);
var->state_slots = ralloc_array(var, nir_state_slot, var->num_state_slots);
blob_copy_bytes(ctx->blob, (uint8_t *) var->state_slots,
var->num_state_slots * sizeof(nir_state_slot));
bool has_const_initializer = blob_read_uint32(ctx->blob);
if (has_const_initializer)
var->constant_initializer = read_constant(ctx, var);
else
var->constant_initializer = NULL;
bool has_interface_type = blob_read_uint32(ctx->blob);
if (has_interface_type)
var->interface_type = decode_type_from_blob(ctx->blob);
else
var->interface_type = NULL;
var->num_members = blob_read_uint32(ctx->blob);
if (var->num_members > 0) {
var->members = ralloc_array(var, struct nir_variable_data,
var->num_members);
blob_copy_bytes(ctx->blob, (uint8_t *) var->members,
var->num_members * sizeof(*var->members));
}
/* Test at least one call of each blob_write_foo and blob_read_foo function,
* verifying that we read out everything we wrote, that every bytes is
* consumed, and that the overrun bit is not set.
*/
static void
test_write_and_read_functions (void)
{
struct blob blob;
struct blob_reader reader;
ssize_t reserved;
size_t str_offset, uint_offset;
uint8_t reserve_buf[sizeof(reserve_test_str)];
blob_init(&blob);
/*** Test blob by writing one of every possible kind of value. */
blob_write_bytes(&blob, bytes_test_str, sizeof(bytes_test_str));
reserved = blob_reserve_bytes(&blob, sizeof(reserve_test_str));
blob_overwrite_bytes(&blob, reserved, reserve_test_str, sizeof(reserve_test_str));
/* Write a placeholder, (to be replaced later via overwrite_bytes) */
str_offset = blob.size;
blob_write_bytes(&blob, placeholder_str, sizeof(placeholder_str));
blob_write_uint32(&blob, uint32_test);
/* Write a placeholder, (to be replaced later via overwrite_uint32) */
uint_offset = blob.size;
blob_write_uint32(&blob, uint32_placeholder);
blob_write_uint64(&blob, uint64_test);
blob_write_intptr(&blob, (intptr_t) &blob);
blob_write_string(&blob, string_test_str);
/* Finally, overwrite our placeholders. */
blob_overwrite_bytes(&blob, str_offset, overwrite_test_str,
sizeof(overwrite_test_str));
blob_overwrite_uint32(&blob, uint_offset, uint32_overwrite);
/*** Now read each value and verify. */
blob_reader_init(&reader, blob.data, blob.size);
expect_equal_str(bytes_test_str,
blob_read_bytes(&reader, sizeof(bytes_test_str)),
"blob_write/read_bytes");
blob_copy_bytes(&reader, reserve_buf, sizeof(reserve_buf));
expect_equal_str(reserve_test_str, (char *) reserve_buf,
"blob_reserve_bytes/blob_copy_bytes");
expect_equal_str(overwrite_test_str,
blob_read_bytes(&reader, sizeof(overwrite_test_str)),
"blob_overwrite_bytes");
expect_equal(uint32_test, blob_read_uint32(&reader),
"blob_write/read_uint32");
expect_equal(uint32_overwrite, blob_read_uint32(&reader),
"blob_overwrite_uint32");
expect_equal(uint64_test, blob_read_uint64(&reader),
"blob_write/read_uint64");
expect_equal((intptr_t) &blob, blob_read_intptr(&reader),
"blob_write/read_intptr");
expect_equal_str(string_test_str, blob_read_string(&reader),
"blob_write/read_string");
expect_equal(reader.end - reader.data, reader.current - reader.data,
"read_consumes_all_bytes");
expect_equal(false, reader.overrun, "read_does_not_overrun");
blob_finish(&blob);
}
static void
st_deserialise_ir_program(struct gl_context *ctx,
struct gl_shader_program *shProg,
struct gl_program *prog, bool nir)
{
struct st_context *st = st_context(ctx);
size_t size = prog->driver_cache_blob_size;
uint8_t *buffer = (uint8_t *) prog->driver_cache_blob;
const struct nir_shader_compiler_options *options =
ctx->Const.ShaderCompilerOptions[prog->info.stage].NirOptions;
assert(prog->driver_cache_blob && prog->driver_cache_blob_size > 0);
struct blob_reader blob_reader;
blob_reader_init(&blob_reader, buffer, size);
switch (prog->info.stage) {
case MESA_SHADER_VERTEX: {
struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
st_release_vp_variants(st, stvp);
stvp->num_inputs = blob_read_uint32(&blob_reader);
blob_copy_bytes(&blob_reader, (uint8_t *) stvp->index_to_input,
sizeof(stvp->index_to_input));
blob_copy_bytes(&blob_reader, (uint8_t *) stvp->input_to_index,
sizeof(stvp->input_to_index));
blob_copy_bytes(&blob_reader, (uint8_t *) stvp->result_to_output,
sizeof(stvp->result_to_output));
read_stream_out_from_cache(&blob_reader, &stvp->tgsi);
if (nir) {
stvp->tgsi.type = PIPE_SHADER_IR_NIR;
stvp->shader_program = shProg;
stvp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
prog->nir = stvp->tgsi.ir.nir;
} else {
read_tgsi_from_cache(&blob_reader, &stvp->tgsi.tokens,
&stvp->num_tgsi_tokens);
}
if (st->vp == stvp)
st->dirty |= ST_NEW_VERTEX_PROGRAM(st, stvp);
break;
}
case MESA_SHADER_TESS_CTRL: {
struct st_common_program *sttcp = st_common_program(prog);
st_release_basic_variants(st, sttcp->Base.Target,
&sttcp->variants, &sttcp->tgsi);
read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
if (nir) {
sttcp->tgsi.type = PIPE_SHADER_IR_NIR;
sttcp->shader_program = shProg;
sttcp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
prog->nir = sttcp->tgsi.ir.nir;
} else {
read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens,
&sttcp->num_tgsi_tokens);
}
if (st->tcp == sttcp)
st->dirty |= sttcp->affected_states;
break;
}
case MESA_SHADER_TESS_EVAL: {
struct st_common_program *sttep = st_common_program(prog);
st_release_basic_variants(st, sttep->Base.Target,
&sttep->variants, &sttep->tgsi);
read_stream_out_from_cache(&blob_reader, &sttep->tgsi);
if (nir) {
sttep->tgsi.type = PIPE_SHADER_IR_NIR;
sttep->shader_program = shProg;
sttep->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
prog->nir = sttep->tgsi.ir.nir;
} else {
read_tgsi_from_cache(&blob_reader, &sttep->tgsi.tokens,
&sttep->num_tgsi_tokens);
}
if (st->tep == sttep)
st->dirty |= sttep->affected_states;
break;
}
case MESA_SHADER_GEOMETRY: {
struct st_common_program *stgp = st_common_program(prog);
st_release_basic_variants(st, stgp->Base.Target, &stgp->variants,
&stgp->tgsi);
read_stream_out_from_cache(&blob_reader, &stgp->tgsi);
if (nir) {
stgp->tgsi.type = PIPE_SHADER_IR_NIR;
stgp->shader_program = shProg;
stgp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
prog->nir = stgp->tgsi.ir.nir;
} else {
read_tgsi_from_cache(&blob_reader, &stgp->tgsi.tokens,
&stgp->num_tgsi_tokens);
}
if (st->gp == stgp)
st->dirty |= stgp->affected_states;
break;
}
case MESA_SHADER_FRAGMENT: {
struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
st_release_fp_variants(st, stfp);
if (nir) {
stfp->tgsi.type = PIPE_SHADER_IR_NIR;
stfp->shader_program = shProg;
stfp->tgsi.ir.nir = nir_deserialize(NULL, options, &blob_reader);
prog->nir = stfp->tgsi.ir.nir;
} else {
read_tgsi_from_cache(&blob_reader, &stfp->tgsi.tokens,
&stfp->num_tgsi_tokens);
}
if (st->fp == stfp)
st->dirty |= stfp->affected_states;
break;
}
case MESA_SHADER_COMPUTE: {
struct st_compute_program *stcp = (struct st_compute_program *) prog;
st_release_cp_variants(st, stcp);
if (nir) {
stcp->tgsi.ir_type = PIPE_SHADER_IR_NIR;
stcp->shader_program = shProg;
stcp->tgsi.prog = nir_deserialize(NULL, options, &blob_reader);
prog->nir = (nir_shader *) stcp->tgsi.prog;
} else {
read_tgsi_from_cache(&blob_reader,
(const struct tgsi_token**) &stcp->tgsi.prog,
&stcp->num_tgsi_tokens);
}
stcp->tgsi.req_local_mem = stcp->Base.info.cs.shared_size;
stcp->tgsi.req_private_mem = 0;
stcp->tgsi.req_input_mem = 0;
if (st->cp == stcp)
st->dirty |= stcp->affected_states;
break;
}
default:
unreachable("Unsupported stage");
}
/* Make sure we don't try to read more data than we wrote. This should
* never happen in release builds but its useful to have this check to
* catch development bugs.
*/
if (blob_reader.current != blob_reader.end || blob_reader.overrun) {
assert(!"Invalid TGSI shader disk cache item!");
if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
fprintf(stderr, "Error reading program from cache (invalid "
"TGSI cache item)\n");
}
}
st_set_prog_affected_state_flags(prog);
_mesa_associate_uniform_storage(ctx, shProg, prog, false);
/* Create Gallium shaders now instead of on demand. */
if (ST_DEBUG & DEBUG_PRECOMPILE ||
st->shader_has_one_variant[prog->info.stage])
st_precompile_shader_variant(st, prog);
}
