#include "st_debug.h"
#include "st_program.h"
#include "st_shader_cache.h"
+#include "st_util.h"
#include "compiler/glsl/program.h"
+#include "compiler/nir/nir.h"
+#include "compiler/nir/nir_serialize.h"
#include "pipe/p_shader_tokens.h"
#include "program/ir_to_mesa.h"
+#include "tgsi/tgsi_parse.h"
#include "util/u_memory.h"
void
static void
write_stream_out_to_cache(struct blob *blob,
- struct pipe_shader_state *tgsi)
+ struct pipe_shader_state *state)
{
- blob_write_bytes(blob, &tgsi->stream_output,
- sizeof(tgsi->stream_output));
+ blob_write_uint32(blob, state->stream_output.num_outputs);
+ if (state->stream_output.num_outputs) {
+ blob_write_bytes(blob, &state->stream_output.stride,
+ sizeof(state->stream_output.stride));
+ blob_write_bytes(blob, &state->stream_output.output,
+ sizeof(state->stream_output.output));
+ }
+}
+
+static void
+copy_blob_to_driver_cache_blob(struct blob *blob, struct gl_program *prog)
+{
+ prog->driver_cache_blob = ralloc_size(NULL, blob->size);
+ memcpy(prog->driver_cache_blob, blob->data, blob->size);
+ prog->driver_cache_blob_size = blob->size;
}
static void
write_tgsi_to_cache(struct blob *blob, const struct tgsi_token *tokens,
- struct gl_program *prog, unsigned num_tokens)
+ struct gl_program *prog)
{
+ unsigned num_tokens = tgsi_num_tokens(tokens);
+
blob_write_uint32(blob, num_tokens);
blob_write_bytes(blob, tokens, num_tokens * sizeof(struct tgsi_token));
+ copy_blob_to_driver_cache_blob(blob, prog);
+}
- prog->driver_cache_blob = ralloc_size(NULL, blob->size);
- memcpy(prog->driver_cache_blob, blob->data, blob->size);
- prog->driver_cache_blob_size = blob->size;
+static void
+write_nir_to_cache(struct blob *blob, struct gl_program *prog)
+{
+ struct st_program *stp = (struct st_program *)prog;
+
+ st_serialize_nir(stp);
+
+ blob_write_intptr(blob, stp->serialized_nir_size);
+ blob_write_bytes(blob, stp->serialized_nir, stp->serialized_nir_size);
+
+ copy_blob_to_driver_cache_blob(blob, prog);
}
-void
-st_serialise_tgsi_program(struct gl_context *ctx, struct gl_program *prog)
+static void
+st_serialise_ir_program(struct gl_context *ctx, struct gl_program *prog,
+ bool nir)
{
+ if (prog->driver_cache_blob)
+ return;
+
+ struct st_program *stp = (struct st_program *)prog;
struct blob blob;
blob_init(&blob);
- switch (prog->info.stage) {
- case MESA_SHADER_VERTEX: {
- struct st_vertex_program *stvp = (struct st_vertex_program *) prog;
+ if (prog->info.stage == MESA_SHADER_VERTEX) {
+ struct st_vertex_program *stvp = (struct st_vertex_program *)stp;
blob_write_uint32(&blob, stvp->num_inputs);
blob_write_bytes(&blob, stvp->index_to_input,
sizeof(stvp->index_to_input));
+ blob_write_bytes(&blob, stvp->input_to_index,
+ sizeof(stvp->input_to_index));
blob_write_bytes(&blob, stvp->result_to_output,
sizeof(stvp->result_to_output));
-
- write_stream_out_to_cache(&blob, &stvp->tgsi);
- write_tgsi_to_cache(&blob, stvp->tgsi.tokens, prog,
- stvp->num_tgsi_tokens);
- break;
}
- case MESA_SHADER_TESS_CTRL:
- case MESA_SHADER_TESS_EVAL:
- case MESA_SHADER_GEOMETRY: {
- struct st_common_program *stcp = (struct st_common_program *) prog;
-
- write_stream_out_to_cache(&blob, &stcp->tgsi);
- write_tgsi_to_cache(&blob, stcp->tgsi.tokens, prog,
- stcp->num_tgsi_tokens);
- break;
- }
- case MESA_SHADER_FRAGMENT: {
- struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
- write_tgsi_to_cache(&blob, stfp->tgsi.tokens, prog,
- stfp->num_tgsi_tokens);
- break;
- }
- case MESA_SHADER_COMPUTE: {
- struct st_compute_program *stcp = (struct st_compute_program *) prog;
+ if (prog->info.stage == MESA_SHADER_VERTEX ||
+ prog->info.stage == MESA_SHADER_TESS_EVAL ||
+ prog->info.stage == MESA_SHADER_GEOMETRY)
+ write_stream_out_to_cache(&blob, &stp->state);
- write_tgsi_to_cache(&blob, stcp->tgsi.prog, prog,
- stcp->num_tgsi_tokens);
- break;
- }
- default:
- unreachable("Unsupported stage");
- }
+ if (nir)
+ write_nir_to_cache(&blob, prog);
+ else
+ write_tgsi_to_cache(&blob, stp->state.tokens, prog);
blob_finish(&blob);
}
/**
- * Store tgsi and any other required state in on-disk shader cache.
+ * Store TGSI or NIR and any other required state in on-disk shader cache.
*/
void
-st_store_tgsi_in_disk_cache(struct st_context *st, struct gl_program *prog)
+st_store_ir_in_disk_cache(struct st_context *st, struct gl_program *prog,
+ bool nir)
{
if (!st->ctx->Cache)
return;
if (memcmp(prog->sh.data->sha1, zero, sizeof(prog->sh.data->sha1)) == 0)
return;
- st_serialise_tgsi_program(st->ctx, prog);
+ st_serialise_ir_program(st->ctx, prog, nir);
if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
- fprintf(stderr, "putting %s tgsi_tokens in cache\n",
+ fprintf(stderr, "putting %s state tracker IR in cache\n",
_mesa_shader_stage_to_string(prog->info.stage));
}
}
static void
read_stream_out_from_cache(struct blob_reader *blob_reader,
- struct pipe_shader_state *tgsi)
+ struct pipe_shader_state *state)
{
- blob_copy_bytes(blob_reader, (uint8_t *) &tgsi->stream_output,
- sizeof(tgsi->stream_output));
+ memset(&state->stream_output, 0, sizeof(state->stream_output));
+ state->stream_output.num_outputs = blob_read_uint32(blob_reader);
+ if (state->stream_output.num_outputs) {
+ blob_copy_bytes(blob_reader, &state->stream_output.stride,
+ sizeof(state->stream_output.stride));
+ blob_copy_bytes(blob_reader, &state->stream_output.output,
+ sizeof(state->stream_output.output));
+ }
}
static void
read_tgsi_from_cache(struct blob_reader *blob_reader,
const struct tgsi_token **tokens)
{
- uint32_t num_tokens = blob_read_uint32(blob_reader);
+ unsigned 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);
}
-void
-st_deserialise_tgsi_program(struct gl_context *ctx,
- struct gl_shader_program *shProg,
- struct gl_program *prog)
+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;
+ st_set_prog_affected_state_flags(prog);
+ _mesa_associate_uniform_storage(ctx, shProg, prog);
+
+ assert(prog->driver_cache_blob && prog->driver_cache_blob_size > 0);
+
+ struct st_program *stp = st_program(prog);
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);
+ st_release_variants(st, stp);
+ if (prog->info.stage == MESA_SHADER_VERTEX) {
+ struct st_vertex_program *stvp = (struct st_vertex_program *)stp;
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);
- read_tgsi_from_cache(&blob_reader, &stvp->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);
- read_tgsi_from_cache(&blob_reader, &sttcp->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);
- read_tgsi_from_cache(&blob_reader, &sttep->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);
- read_tgsi_from_cache(&blob_reader, &stgp->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);
-
- read_tgsi_from_cache(&blob_reader, &stfp->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);
-
- read_tgsi_from_cache(&blob_reader,
- (const struct tgsi_token**) &stcp->tgsi.prog);
- 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");
+ if (prog->info.stage == MESA_SHADER_VERTEX ||
+ prog->info.stage == MESA_SHADER_TESS_EVAL ||
+ prog->info.stage == MESA_SHADER_GEOMETRY)
+ read_stream_out_from_cache(&blob_reader, &stp->state);
+
+ if (nir) {
+ assert(prog->nir == NULL);
+ assert(stp->serialized_nir == NULL);
+
+ stp->state.type = PIPE_SHADER_IR_NIR;
+ stp->serialized_nir_size = blob_read_intptr(&blob_reader);
+ stp->serialized_nir = malloc(stp->serialized_nir_size);
+ blob_copy_bytes(&blob_reader, stp->serialized_nir, stp->serialized_nir_size);
+ stp->shader_program = shProg;
+ } else {
+ read_tgsi_from_cache(&blob_reader, &stp->state.tokens);
}
/* Make sure we don't try to read more data than we wrote. This should
}
}
- 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);
+ st_finalize_program(st, prog);
}
bool
-st_load_tgsi_from_disk_cache(struct gl_context *ctx,
- struct gl_shader_program *prog)
+st_load_ir_from_disk_cache(struct gl_context *ctx,
+ struct gl_shader_program *prog,
+ bool nir)
{
if (!ctx->Cache)
return false;
/* If we didn't load the GLSL metadata from cache then we could not have
- * loaded the tgsi either.
+ * loaded TGSI or NIR either.
*/
if (prog->data->LinkStatus != LINKING_SKIPPED)
return false;
continue;
struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
- st_deserialise_tgsi_program(ctx, prog, glprog);
+ st_deserialise_ir_program(ctx, prog, glprog, nir);
/* We don't need the cached blob anymore so free it */
ralloc_free(glprog->driver_cache_blob);
glprog->driver_cache_blob_size = 0;
if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
- fprintf(stderr, "%s tgsi_tokens retrieved from cache\n",
+ fprintf(stderr, "%s state tracker IR retrieved from cache\n",
_mesa_shader_stage_to_string(i));
}
}
return true;
}
+
+void
+st_serialise_tgsi_program(struct gl_context *ctx, struct gl_program *prog)
+{
+ st_serialise_ir_program(ctx, prog, false);
+}
+
+void
+st_serialise_tgsi_program_binary(struct gl_context *ctx,
+ struct gl_shader_program *shProg,
+ struct gl_program *prog)
+{
+ st_serialise_ir_program(ctx, prog, false);
+}
+
+void
+st_deserialise_tgsi_program(struct gl_context *ctx,
+ struct gl_shader_program *shProg,
+ struct gl_program *prog)
+{
+ st_deserialise_ir_program(ctx, shProg, prog, false);
+}
+
+void
+st_serialise_nir_program(struct gl_context *ctx, struct gl_program *prog)
+{
+ st_serialise_ir_program(ctx, prog, true);
+}
+
+void
+st_serialise_nir_program_binary(struct gl_context *ctx,
+ struct gl_shader_program *shProg,
+ struct gl_program *prog)
+{
+ st_serialise_ir_program(ctx, prog, true);
+}
+
+void
+st_deserialise_nir_program(struct gl_context *ctx,
+ struct gl_shader_program *shProg,
+ struct gl_program *prog)
+{
+ st_deserialise_ir_program(ctx, shProg, prog, true);
+}