*/
#include <stdio.h>
-
+#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
+st_get_program_binary_driver_sha1(struct gl_context *ctx, uint8_t *sha1)
+{
+ disk_cache_compute_key(ctx->Cache, NULL, 0, sha1);
+}
+
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
-write_tgsi_to_cache(struct blob *blob, struct pipe_shader_state *tgsi,
- struct st_context *st, unsigned char *sha1,
- unsigned num_tokens)
+copy_blob_to_driver_cache_blob(struct blob *blob, struct gl_program *prog)
{
- blob_write_uint32(blob, num_tokens);
- blob_write_bytes(blob, tgsi->tokens,
- num_tokens * sizeof(struct tgsi_token));
+ 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;
+}
- disk_cache_put(st->ctx->Cache, sha1, blob->data, blob->size);
+static void
+write_tgsi_to_cache(struct blob *blob, const struct tgsi_token *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);
}
-/**
- * Store tgsi 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,
- struct pipe_shader_state *out_state,
- unsigned num_tokens)
+static void
+write_nir_to_cache(struct blob *blob, struct gl_program *prog)
{
- if (!st->ctx->Cache)
- return;
+ struct st_program *stp = (struct st_program *)prog;
- /* Exit early when we are dealing with a ff shader with no source file to
- * generate a source from.
- */
- if (*prog->sh.data->sha1 == 0)
+ 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);
+}
+
+static void
+st_serialise_ir_program(struct gl_context *ctx, struct gl_program *prog,
+ bool nir)
+{
+ if (prog->driver_cache_blob)
return;
- unsigned char *sha1;
- struct blob *blob = blob_create(NULL);
+ 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;
- sha1 = stvp->sha1;
+ 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,
+ 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->result_to_output,
+ 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, st, sha1, num_tokens);
- break;
}
- case MESA_SHADER_TESS_CTRL: {
- struct st_tessctrl_program *stcp = (struct st_tessctrl_program *) prog;
- sha1 = stcp->sha1;
- write_stream_out_to_cache(blob, out_state);
- write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
- break;
- }
- case MESA_SHADER_TESS_EVAL: {
- struct st_tesseval_program *step = (struct st_tesseval_program *) prog;
- sha1 = step->sha1;
+ 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_stream_out_to_cache(blob, out_state);
- write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
- break;
- }
- case MESA_SHADER_GEOMETRY: {
- struct st_geometry_program *stgp = (struct st_geometry_program *) prog;
- sha1 = stgp->sha1;
+ if (nir)
+ write_nir_to_cache(&blob, prog);
+ else
+ write_tgsi_to_cache(&blob, stp->state.tokens, prog);
- write_stream_out_to_cache(blob, out_state);
- write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
- break;
- }
- case MESA_SHADER_FRAGMENT: {
- struct st_fragment_program *stfp = (struct st_fragment_program *) prog;
- sha1 = stfp->sha1;
+ blob_finish(&blob);
+}
- write_tgsi_to_cache(blob, &stfp->tgsi, st, sha1, num_tokens);
- break;
- }
- case MESA_SHADER_COMPUTE: {
- struct st_compute_program *stcp = (struct st_compute_program *) prog;
- sha1 = stcp->sha1;
+/**
+ * Store TGSI or NIR and any other required state in on-disk shader cache.
+ */
+void
+st_store_ir_in_disk_cache(struct st_context *st, struct gl_program *prog,
+ bool nir)
+{
+ if (!st->ctx->Cache)
+ return;
- write_tgsi_to_cache(blob, out_state, st, sha1, num_tokens);
- break;
- }
- default:
- unreachable("Unsupported stage");
- }
+ /* Exit early when we are dealing with a ff shader with no source file to
+ * generate a source from.
+ */
+ static const char zero[sizeof(prog->sh.data->sha1)] = {0};
+ if (memcmp(prog->sh.data->sha1, zero, sizeof(prog->sh.data->sha1)) == 0)
+ return;
+
+ st_serialise_ir_program(st->ctx, prog, nir);
if (st->ctx->_Shader->Flags & GLSL_CACHE_INFO) {
- char sha1_buf[41];
- _mesa_sha1_format(sha1_buf, sha1);
- fprintf(stderr, "putting %s tgsi_tokens in cache: %s\n",
- _mesa_shader_stage_to_string(prog->info.stage), sha1_buf);
+ fprintf(stderr, "putting %s state tracker IR in cache\n",
+ _mesa_shader_stage_to_string(prog->info.stage));
}
-
- ralloc_free(blob);
}
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);
}
-bool
-st_load_tgsi_from_disk_cache(struct gl_context *ctx,
- struct gl_shader_program *prog)
+static void
+st_deserialise_ir_program(struct gl_context *ctx,
+ struct gl_shader_program *shProg,
+ struct gl_program *prog, bool nir)
{
- if (!ctx->Cache)
- return false;
-
- unsigned char *stage_sha1[MESA_SHADER_STAGES];
- char sha1_buf[41];
-
- /* Compute and store sha1 for each stage. These will be reused by the
- * cache store pass if we fail to find the cached tgsi.
- */
- for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
- if (prog->_LinkedShaders[i] == NULL)
- continue;
-
- char *buf = ralloc_strdup(NULL, "tgsi_tokens ");
- _mesa_sha1_format(sha1_buf,
- prog->_LinkedShaders[i]->Program->sh.data->sha1);
- ralloc_strcat(&buf, sha1_buf);
-
- struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
- switch (glprog->info.stage) {
- case MESA_SHADER_VERTEX: {
- struct st_vertex_program *stvp = (struct st_vertex_program *) glprog;
- stage_sha1[i] = stvp->sha1;
- ralloc_strcat(&buf, " vs");
- _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
- break;
- }
- case MESA_SHADER_TESS_CTRL: {
- struct st_tessctrl_program *stcp =
- (struct st_tessctrl_program *) glprog;
- stage_sha1[i] = stcp->sha1;
- ralloc_strcat(&buf, " tcs");
- _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
- break;
- }
- case MESA_SHADER_TESS_EVAL: {
- struct st_tesseval_program *step =
- (struct st_tesseval_program *) glprog;
- stage_sha1[i] = step->sha1;
- ralloc_strcat(&buf, " tes");
- _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
- break;
- }
- case MESA_SHADER_GEOMETRY: {
- struct st_geometry_program *stgp =
- (struct st_geometry_program *) glprog;
- stage_sha1[i] = stgp->sha1;
- ralloc_strcat(&buf, " gs");
- _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
- break;
- }
- case MESA_SHADER_FRAGMENT: {
- struct st_fragment_program *stfp =
- (struct st_fragment_program *) glprog;
- stage_sha1[i] = stfp->sha1;
- ralloc_strcat(&buf, " fs");
- _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
- break;
- }
- case MESA_SHADER_COMPUTE: {
- struct st_compute_program *stcp =
- (struct st_compute_program *) glprog;
- stage_sha1[i] = stcp->sha1;
- ralloc_strcat(&buf, " cs");
- _mesa_sha1_compute(buf, strlen(buf), stage_sha1[i]);
- break;
- }
- default:
- unreachable("Unsupported stage");
- }
-
- ralloc_free(buf);
- }
-
- /* Now that we have created the sha1 keys that will be used for writting to
- * the tgsi cache fallback to the regular glsl to tgsi path if we didn't
- * load the GLSL IR from cache. We do this as glsl to tgsi can alter things
- * such as gl_program_parameter_list which holds things like uniforms.
- */
- if (prog->data->LinkStatus != linking_skipped)
- return false;
-
struct st_context *st = st_context(ctx);
- uint8_t *buffer = NULL;
- for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
- if (prog->_LinkedShaders[i] == NULL)
- continue;
-
- unsigned char *sha1 = stage_sha1[i];
- size_t size;
- buffer = (uint8_t *) disk_cache_get(ctx->Cache, sha1, &size);
- if (buffer) {
- struct blob_reader blob_reader;
- blob_reader_init(&blob_reader, buffer, size);
-
- struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
- switch (glprog->info.stage) {
- case MESA_SHADER_VERTEX: {
- struct st_vertex_program *stvp =
- (struct st_vertex_program *) glprog;
-
- 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->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_tessctrl_program *sttcp =
- (struct st_tessctrl_program *) glprog;
-
- st_release_basic_variants(st, sttcp->Base.Target,
- &sttcp->variants, &sttcp->tgsi);
+ size_t size = prog->driver_cache_blob_size;
+ uint8_t *buffer = (uint8_t *) prog->driver_cache_blob;
- read_stream_out_from_cache(&blob_reader, &sttcp->tgsi);
- read_tgsi_from_cache(&blob_reader, &sttcp->tgsi.tokens);
+ st_set_prog_affected_state_flags(prog);
+ _mesa_associate_uniform_storage(ctx, shProg, prog);
- if (st->tcp == sttcp)
- st->dirty |= sttcp->affected_states;
+ assert(prog->driver_cache_blob && prog->driver_cache_blob_size > 0);
- break;
- }
- case MESA_SHADER_TESS_EVAL: {
- struct st_tesseval_program *sttep =
- (struct st_tesseval_program *) glprog;
+ struct st_program *stp = st_program(prog);
+ struct blob_reader blob_reader;
+ blob_reader_init(&blob_reader, buffer, size);
- st_release_basic_variants(st, sttep->Base.Target,
- &sttep->variants, &sttep->tgsi);
+ st_release_variants(st, stp);
- 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_geometry_program *stgp =
- (struct st_geometry_program *) glprog;
-
- 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 *) glprog;
-
- 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 *) glprog;
-
- 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 (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));
+ }
- if (st->cp == stcp)
- st->dirty |= stcp->affected_states;
+ 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);
+ }
- 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 (blob_reader.current != blob_reader.end || blob_reader.overrun) {
- /* Something very bad has gone wrong discard the item from the
- * cache and rebuild/link from source.
- */
- 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");
+ }
+ }
- if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
- fprintf(stderr, "Error reading program from cache (invalid "
- "TGSI cache item)\n");
- }
+ st_finalize_program(st, prog);
+}
- disk_cache_remove(ctx->Cache, sha1);
+bool
+st_load_ir_from_disk_cache(struct gl_context *ctx,
+ struct gl_shader_program *prog,
+ bool nir)
+{
+ if (!ctx->Cache)
+ return false;
- goto fallback_recompile;
- }
+ /* If we didn't load the GLSL metadata from cache then we could not have
+ * loaded TGSI or NIR either.
+ */
+ if (prog->data->LinkStatus != LINKING_SKIPPED)
+ return false;
- if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
- _mesa_sha1_format(sha1_buf, sha1);
- fprintf(stderr, "%s tgsi_tokens retrieved from cache: %s\n",
- _mesa_shader_stage_to_string(i), sha1_buf);
- }
+ for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
+ if (prog->_LinkedShaders[i] == NULL)
+ continue;
- st_set_prog_affected_state_flags(glprog);
- _mesa_associate_uniform_storage(ctx, prog, glprog->Parameters);
+ struct gl_program *glprog = prog->_LinkedShaders[i]->Program;
+ st_deserialise_ir_program(ctx, prog, glprog, nir);
- free(buffer);
- } else {
- /* Failed to find a matching cached shader so fallback to recompile.
- */
- if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
- fprintf(stderr, "TGSI cache item not found falling back to "
- "compile.\n");
- }
+ /* We don't need the cached blob anymore so free it */
+ ralloc_free(glprog->driver_cache_blob);
+ glprog->driver_cache_blob = NULL;
+ glprog->driver_cache_blob_size = 0;
- goto fallback_recompile;
+ if (ctx->_Shader->Flags & GLSL_CACHE_INFO) {
+ fprintf(stderr, "%s state tracker IR retrieved from cache\n",
+ _mesa_shader_stage_to_string(i));
}
}
return true;
+}
-fallback_recompile:
- free(buffer);
+void
+st_serialise_tgsi_program(struct gl_context *ctx, struct gl_program *prog)
+{
+ st_serialise_ir_program(ctx, prog, false);
+}
- for (unsigned i = 0; i < prog->NumShaders; i++) {
- _mesa_glsl_compile_shader(ctx, prog->Shaders[i], false, false, true);
- }
+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);
+}
- prog->data->cache_fallback = true;
- _mesa_glsl_link_shader(ctx, prog);
+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);
+}
- return true;
+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);
}