[Mesa-dev] [PATCH 25/25] radeonsi: implement binary shaders & shader cache in memory
Nicolai Hähnle
nhaehnle at gmail.com
Tue Feb 16 17:02:00 UTC 2016
On 15.02.2016 18:59, Marek Olšák wrote:
> From: Marek Olšák <marek.olsak at amd.com>
>
> ---
> src/gallium/drivers/radeonsi/si_pipe.c | 5 +-
> src/gallium/drivers/radeonsi/si_pipe.h | 16 ++
> src/gallium/drivers/radeonsi/si_shader.h | 4 +-
> src/gallium/drivers/radeonsi/si_state.h | 2 +
> src/gallium/drivers/radeonsi/si_state_shaders.c | 234 +++++++++++++++++++++++-
> 5 files changed, 254 insertions(+), 7 deletions(-)
>
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
> index 75d4775..a576237 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.c
> +++ b/src/gallium/drivers/radeonsi/si_pipe.c
> @@ -563,7 +563,7 @@ static void si_destroy_screen(struct pipe_screen* pscreen)
> }
> }
> pipe_mutex_destroy(sscreen->shader_parts_mutex);
> -
> + si_destroy_shader_cache(sscreen);
> r600_destroy_common_screen(&sscreen->b);
> }
>
> @@ -611,7 +611,8 @@ struct pipe_screen *radeonsi_screen_create(struct radeon_winsys *ws)
> sscreen->b.b.resource_create = r600_resource_create_common;
>
> if (!r600_common_screen_init(&sscreen->b, ws) ||
> - !si_init_gs_info(sscreen)) {
> + !si_init_gs_info(sscreen) ||
> + !si_init_shader_cache(sscreen)) {
> FREE(sscreen);
> return NULL;
> }
> diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
> index 1ac7bc4..ef860a5 100644
> --- a/src/gallium/drivers/radeonsi/si_pipe.h
> +++ b/src/gallium/drivers/radeonsi/si_pipe.h
> @@ -80,6 +80,7 @@
> #define SI_MAX_BORDER_COLORS 4096
>
> struct si_compute;
> +struct hash_table;
>
> struct si_screen {
> struct r600_common_screen b;
> @@ -94,6 +95,21 @@ struct si_screen {
> struct si_shader_part *tcs_epilogs;
> struct si_shader_part *ps_prologs;
> struct si_shader_part *ps_epilogs;
> +
> + /* Shader cache in memory.
> + *
> + * Design & limitations:
> + * - The shader cache is per screen (= per process), never saved to
> + * disk, and skips redundant shader compilations from TGSI to bytecode.
> + * - It can only be used with one-variant-per-shader support, in which
> + * case only the main (typically middle) part of shaders is cached.
> + * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
> + * variants of VS and TES are cached, so LS and ES aren't.
> + * - GS and CS aren't cached, but it's certainly possible to cache
> + * those as well.
> + */
> + pipe_mutex shader_cache_mutex;
> + struct hash_table *shader_cache;
> };
>
> struct si_blend_color {
> diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
> index 48e048d..7e46871 100644
> --- a/src/gallium/drivers/radeonsi/si_shader.h
> +++ b/src/gallium/drivers/radeonsi/si_shader.h
> @@ -362,8 +362,10 @@ struct si_shader {
> struct r600_resource *bo;
> struct r600_resource *scratch_bo;
> union si_shader_key key;
> - struct radeon_shader_binary binary;
> bool is_binary_shared;
> +
> + /* The following data is all that's needed for binary shaders. */
> + struct radeon_shader_binary binary;
> struct si_shader_config config;
> struct si_shader_info info;
> };
> diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
> index f64c4d4..40792cb 100644
> --- a/src/gallium/drivers/radeonsi/si_state.h
> +++ b/src/gallium/drivers/radeonsi/si_state.h
> @@ -280,6 +280,8 @@ si_create_sampler_view_custom(struct pipe_context *ctx,
> /* si_state_shader.c */
> bool si_update_shaders(struct si_context *sctx);
> void si_init_shader_functions(struct si_context *sctx);
> +bool si_init_shader_cache(struct si_screen *sscreen);
> +void si_destroy_shader_cache(struct si_screen *sscreen);
>
> /* si_state_draw.c */
> void si_emit_cache_flush(struct si_context *sctx, struct r600_atom *atom);
> diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
> index c62cbb7..bc3e5be 100644
> --- a/src/gallium/drivers/radeonsi/si_state_shaders.c
> +++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
> @@ -32,10 +32,217 @@
>
> #include "tgsi/tgsi_parse.h"
> #include "tgsi/tgsi_ureg.h"
> +#include "util/hash_table.h"
> +#include "util/u_hash.h"
> #include "util/u_memory.h"
> #include "util/u_prim.h"
> #include "util/u_simple_shaders.h"
>
> +/* SHADER_CACHE */
> +
> +/**
> + * Return the TGSI binary in a buffer. The first 4 bytes contain its size as
> + * integer.
> + */
> +static void *si_get_tgsi_binary(struct si_shader_selector *sel)
> +{
> + unsigned tgsi_size = tgsi_num_tokens(sel->tokens) *
> + sizeof(struct tgsi_token);
> + unsigned size = 4 + tgsi_size + sizeof(sel->so);
> + char *result = (char*)MALLOC(size);
> +
> + if (!result)
> + return NULL;
> +
> + *((uint32_t*)result) = size;
> + memcpy(result + 4, sel->tokens, tgsi_size);
> + memcpy(result + 4 + tgsi_size, &sel->so, sizeof(sel->so));
> + return result;
> +}
> +
> +/** Copy "data" to "ptr" and return the next dword following copied data. */
> +static uint32_t *write_data(uint32_t *ptr, const void *data, unsigned size)
> +{
> + memcpy(ptr, data, size);
> + ptr += DIV_ROUND_UP(size, 4);
> + return ptr;
> +}
> +
> +/** Read data from "ptr". Return the next dword following the data. */
> +static uint32_t *read_data(uint32_t *ptr, void *data, unsigned size)
> +{
> + memcpy(data, ptr, size);
> + ptr += DIV_ROUND_UP(size, 4);
> + return ptr;
> +}
> +
> +/**
> + * Write the size as uint followed by the data. Return the next dword
> + * following the copied data.
> + */
> +static uint32_t *write_chunk(uint32_t *ptr, const void *data, unsigned size)
> +{
> + *ptr++ = size;
> + return write_data(ptr, data, size);
> +}
> +
> +/**
> + * Read the size as uint followed by the data. Return both via parameters.
> + * Return the next dword following the data.
> + */
> +static uint32_t *read_chunk(uint32_t *ptr, void **data, unsigned *size)
> +{
> + *size = *ptr++;
> + assert(*data == NULL);
> + *data = malloc(*size);
> + return read_data(ptr, *data, *size);
> +}
> +
> +/**
> + * Return the shader binary in a buffer. The first 4 bytes contain its size
> + * as integer.
> + */
> +static void *si_get_shader_binary(struct si_shader *shader)
> +{
> + /* There is always a size of data followed by the data itself. */
> + unsigned relocs_size = shader->binary.reloc_count *
> + sizeof(shader->binary.relocs[0]);
> + unsigned disasm_size = strlen(shader->binary.disasm_string) + 1;
> + unsigned size =
> + 4 + /* total size */
> + 4 + /* CRC32 of the data below */
> + align(sizeof(shader->config), 4) +
> + align(sizeof(shader->info), 4) +
> + 4 + align(shader->binary.code_size, 4) +
> + 4 + align(shader->binary.rodata_size, 4) +
> + 4 + align(relocs_size, 4) +
> + 4 + align(disasm_size, 4);
> + void *buffer = CALLOC(1, size);
> + uint32_t *ptr = (uint32_t*)buffer;
> +
> + if (!buffer)
> + return NULL;
> +
> + *ptr++ = size;
> + ptr++; /* CRC32 is calculated at the end. */
> +
> + ptr = write_data(ptr, &shader->config, sizeof(shader->config));
> + ptr = write_data(ptr, &shader->info, sizeof(shader->info));
> + ptr = write_chunk(ptr, shader->binary.code, shader->binary.code_size);
> + ptr = write_chunk(ptr, shader->binary.rodata, shader->binary.rodata_size);
> + ptr = write_chunk(ptr, shader->binary.relocs, relocs_size);
> + ptr = write_chunk(ptr, shader->binary.disasm_string, disasm_size);
Suggestion: assert((char *)ptr - (char *)buffer == size);
> +
> + /* Compute CRC32. */
> + ptr = (uint32_t*)buffer;
> + ptr++;
> + *ptr = util_hash_crc32(ptr + 1, size - 8);
> +
> + return buffer;
> +}
> +
> +static bool si_load_shader_binary(struct si_shader *shader, void *binary)
> +{
> + uint32_t *ptr = (uint32_t*)binary;
> + uint32_t size = *ptr++;
> + uint32_t crc32 = *ptr++;
> + unsigned chunk_size;
> +
> + if (util_hash_crc32(ptr, size - 8) != crc32) {
> + fprintf(stderr, "radeonsi: binary shader has invalid CRC32\n");
> + return false;
> + }
> +
> + ptr = read_data(ptr, &shader->config, sizeof(shader->config));
> + ptr = read_data(ptr, &shader->info, sizeof(shader->info));
> + ptr = read_chunk(ptr, (void**)&shader->binary.code,
> + &shader->binary.code_size);
> + ptr = read_chunk(ptr, (void**)&shader->binary.rodata,
> + &shader->binary.rodata_size);
> + ptr = read_chunk(ptr, (void**)&shader->binary.relocs, &chunk_size);
> + shader->binary.reloc_count = chunk_size / sizeof(shader->binary.relocs[0]);
> + ptr = read_chunk(ptr, (void**)&shader->binary.disasm_string, &chunk_size);
> +
> + return true;
> +}
> +
> +/**
> + * Insert a shader into the cache. It's assumed the shader is not in the cache.
> + * Use si_shader_cache_load_shader before calling this.
> + *
> + * Returns true if the tgsi_binary should be deleted after this.
> + */
> +static bool si_shader_cache_insert_shader(struct si_screen *sscreen,
> + void *tgsi_binary,
> + struct si_shader *shader)
> +{
> + struct hash_entry *entry;
> + void *hw_binary = si_get_shader_binary(shader);
> +
> + if (!hw_binary)
> + return true;
> +
> + entry = _mesa_hash_table_insert(sscreen->shader_cache, tgsi_binary,
> + hw_binary);
> + return entry->key != tgsi_binary;
_mesa_hash_table_insert can return NULL, indicating that it didn't
insert anything.
If it is successful, entry->key == tgsi_binary will always hold, so the
second check is unnecessary.
Nicolai
> +}
> +
> +static bool si_shader_cache_load_shader(struct si_screen *sscreen,
> + void *tgsi_binary,
> + struct si_shader *shader)
> +{
> + struct hash_entry *entry =
> + _mesa_hash_table_search(sscreen->shader_cache, tgsi_binary);
> + if (!entry)
> + return false;
> +
> + return si_load_shader_binary(shader, entry->data);
> +}
> +
> +static uint32_t si_shader_cache_key_hash(const void *key)
> +{
> + /* The first dword is the key size. */
> + return util_hash_crc32(key, *(uint32_t*)key);
> +}
> +
> +static bool si_shader_cache_key_equals(const void *a, const void *b)
> +{
> + uint32_t *keya = (uint32_t*)a;
> + uint32_t *keyb = (uint32_t*)b;
> +
> + /* The first dword is the key size. */
> + if (*keya != *keyb)
> + return false;
> +
> + return memcmp(keya, keyb, *keya) == 0;
> +}
> +
> +static void si_destroy_shader_cache_entry(struct hash_entry *entry)
> +{
> + FREE((void*)entry->key);
> + FREE(entry->data);
> +}
> +
> +bool si_init_shader_cache(struct si_screen *sscreen)
> +{
> + pipe_mutex_init(sscreen->shader_cache_mutex);
> + sscreen->shader_cache =
> + _mesa_hash_table_create(NULL,
> + si_shader_cache_key_hash,
> + si_shader_cache_key_equals);
> + return sscreen->shader_cache != NULL;
> +}
> +
> +void si_destroy_shader_cache(struct si_screen *sscreen)
> +{
> + if (sscreen->shader_cache)
> + _mesa_hash_table_destroy(sscreen->shader_cache,
> + si_destroy_shader_cache_entry);
> + pipe_mutex_destroy(sscreen->shader_cache_mutex);
> +}
> +
> +/* SHADER STATES */
> +
> static void si_set_tesseval_regs(struct si_shader *shader,
> struct si_pm4_state *pm4)
> {
> @@ -936,17 +1143,36 @@ static void *si_create_shader_selector(struct pipe_context *ctx,
> if (sel->type != PIPE_SHADER_GEOMETRY &&
> !sscreen->use_monolithic_shaders) {
> struct si_shader *shader = CALLOC_STRUCT(si_shader);
> + void *tgsi_binary;
>
> if (!shader)
> goto error;
>
> shader->selector = sel;
>
> - if (si_compile_tgsi_shader(sscreen, sctx->tm, shader, false,
> - &sctx->b.debug) != 0) {
> - FREE(shader);
> - goto error;
> + tgsi_binary = si_get_tgsi_binary(sel);
> +
> + /* Try to load the shader from the shader cache. */
> + pipe_mutex_lock(sscreen->shader_cache_mutex);
> +
> + if (tgsi_binary &&
> + si_shader_cache_load_shader(sscreen, tgsi_binary, shader)) {
> + FREE(tgsi_binary);
> + } else {
> + /* Compile the shader if it hasn't been loaded from the cache. */
> + if (si_compile_tgsi_shader(sscreen, sctx->tm, shader, false,
> + &sctx->b.debug) != 0) {
> + FREE(shader);
> + FREE(tgsi_binary);
> + pipe_mutex_unlock(sscreen->shader_cache_mutex);
> + goto error;
> + }
> +
> + if (si_shader_cache_insert_shader(sscreen, tgsi_binary, shader))
> + FREE(tgsi_binary);
> }
> + pipe_mutex_unlock(sscreen->shader_cache_mutex);
> +
> sel->main_shader_part = shader;
> }
>
>
More information about the mesa-dev
mailing list