From: Nicolai Hähnle <[email protected]>
---
src/gallium/drivers/radeonsi/si_compute.c | 63 ++--
src/gallium/drivers/radeonsi/si_debug.c | 74 +++--
src/gallium/drivers/radeonsi/si_pipe.c | 2 +-
src/gallium/drivers/radeonsi/si_pipe.h | 2 +-
src/gallium/drivers/radeonsi/si_shader.c | 291 +++++++++---------
src/gallium/drivers/radeonsi/si_shader.h | 19 +-
.../drivers/radeonsi/si_shader_internal.h | 3 +-
.../drivers/radeonsi/si_shader_tgsi_setup.c | 14 +-
.../drivers/radeonsi/si_state_shaders.c | 39 +--
9 files changed, 270 insertions(+), 237 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_compute.c
b/src/gallium/drivers/radeonsi/si_compute.c
index 2899ee146d4..e4ef138db33 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -21,20 +21,21 @@
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
*
*/
#include "tgsi/tgsi_parse.h"
#include "util/u_async_debug.h"
#include "util/u_memory.h"
#include "util/u_upload_mgr.h"
+#include "ac_rtld.h"
#include "amd_kernel_code_t.h"
#include "si_build_pm4.h"
#include "si_compute.h"
#define COMPUTE_DBG(sscreen, fmt, args...) \
do { \
if ((sscreen->debug_flags & DBG(COMPUTE))) fprintf(stderr, fmt,
##args); \
} while (0);
struct dispatch_packet {
@@ -54,22 +55,40 @@ struct dispatch_packet {
uint64_t reserved2;
};
static const amd_kernel_code_t *si_compute_get_code_object(
const struct si_compute *program,
uint64_t symbol_offset)
{
if (!program->use_code_object_v2) {
return NULL;
}
- return (const amd_kernel_code_t*)
- (program->shader.binary.code + symbol_offset);
+
+ struct ac_rtld_binary rtld;
+ if (!ac_rtld_open(&rtld, 1, &program->shader.binary.elf_buffer,
+ &program->shader.binary.elf_size))
+ return NULL;
+
+ const amd_kernel_code_t *result = NULL;
+ const char *text;
+ size_t size;
+ if (!ac_rtld_get_section_by_name(&rtld, ".text", &text, &size))
+ goto out;
+
+ if (symbol_offset + sizeof(amd_kernel_code_t) > size)
+ goto out;
+
+ result = (const amd_kernel_code_t*)(text + symbol_offset);
+
+out:
+ ac_rtld_close(&rtld);
+ return result;
}
static void code_object_to_config(const amd_kernel_code_t *code_object,
struct ac_shader_config *out_config) {
uint32_t rsrc1 = code_object->compute_pgm_resource_registers;
uint32_t rsrc2 = code_object->compute_pgm_resource_registers >> 32;
out_config->num_sgprs = code_object->wavefront_sgpr_count;
out_config->num_vgprs = code_object->workitem_vgpr_count;
out_config->float_mode = G_00B028_FLOAT_MODE(rsrc1);
@@ -137,21 +156,21 @@ static void si_create_compute_state_async(void *job, int
thread_index)
mtx_lock(&sscreen->shader_cache_mutex);
if (ir_binary &&
si_shader_cache_load_shader(sscreen, ir_binary, shader)) {
mtx_unlock(&sscreen->shader_cache_mutex);
si_shader_dump_stats_for_shader_db(shader, debug);
si_shader_dump(sscreen, shader, debug, PIPE_SHADER_COMPUTE,
stderr, true);
- if (!si_shader_binary_upload(sscreen, shader))
+ if (!si_shader_binary_upload(sscreen, shader, 0))
program->shader.compilation_failed = true;
} else {
mtx_unlock(&sscreen->shader_cache_mutex);
if (!si_shader_create(sscreen, compiler, &program->shader,
debug)) {
program->shader.compilation_failed = true;
if (program->ir_type == PIPE_SHADER_IR_TGSI)
FREE(program->ir.tgsi);
program->shader.selector = NULL;
@@ -229,39 +248,37 @@ static void *si_create_compute_state(
si_schedule_initial_compile(sctx, PIPE_SHADER_COMPUTE,
&program->ready,
&program->compiler_ctx_state,
program,
si_create_compute_state_async);
} else {
const struct pipe_llvm_program_header *header;
const char *code;
header = cso->prog;
code = cso->prog + sizeof(struct pipe_llvm_program_header);
- ac_elf_read(code, header->num_bytes, &program->shader.binary);
- if (program->use_code_object_v2) {
- const amd_kernel_code_t *code_object =
- si_compute_get_code_object(program, 0);
- code_object_to_config(code_object,
&program->shader.config);
- if (program->shader.binary.reloc_count != 0) {
- fprintf(stderr, "Error: %d unsupported
relocations\n",
- program->shader.binary.reloc_count);
- FREE(program);
- return NULL;
- }
- } else {
- ac_shader_binary_read_config(&program->shader.binary,
- &program->shader.config, 0, false);
+ program->shader.binary.elf_size = header->num_bytes;
+ program->shader.binary.elf_buffer = malloc(header->num_bytes);
+ if (!program->shader.binary.elf_buffer) {
+ FREE(program);
+ return NULL;
}
+ memcpy((void *)program->shader.binary.elf_buffer, code,
header->num_bytes);
+
+ const amd_kernel_code_t *code_object =
+ si_compute_get_code_object(program, 0);
+ code_object_to_config(code_object, &program->shader.config);
+
si_shader_dump(sctx->screen, &program->shader, &sctx->debug,
PIPE_SHADER_COMPUTE, stderr, true);
- if (!si_shader_binary_upload(sctx->screen, &program->shader)) {
+ if (!si_shader_binary_upload(sctx->screen, &program->shader,
0)) {
fprintf(stderr, "LLVM failed to upload shader\n");
+ free((void *)program->shader.binary.elf_buffer);
FREE(program);
return NULL;
}
}
return program;
}
static void si_bind_compute_state(struct pipe_context *ctx, void *state)
{
@@ -386,23 +403,21 @@ static bool si_setup_compute_scratch_buffer(struct
si_context *sctx,
PIPE_USAGE_DEFAULT,
scratch_needed, 256);
if (!sctx->compute_scratch_buffer)
return false;
}
if (sctx->compute_scratch_buffer != shader->scratch_bo &&
scratch_needed) {
uint64_t scratch_va = sctx->compute_scratch_buffer->gpu_address;
- si_shader_apply_scratch_relocs(shader, scratch_va);
-
- if (!si_shader_binary_upload(sctx->screen, shader))
+ if (!si_shader_binary_upload(sctx->screen, shader, scratch_va))
return false;
si_resource_reference(&shader->scratch_bo,
sctx->compute_scratch_buffer);
}
return true;
}
static bool si_switch_compute_shader(struct si_context *sctx,
@@ -419,25 +434,21 @@ static bool si_switch_compute_shader(struct si_context
*sctx,
if (sctx->cs_shader_state.emitted_program == program &&
sctx->cs_shader_state.offset == offset)
return true;
if (program->ir_type != PIPE_SHADER_IR_NATIVE) {
config = &shader->config;
} else {
unsigned lds_blocks;
config = &inline_config;
- if (code_object) {
- code_object_to_config(code_object, config);
- } else {
- ac_shader_binary_read_config(&shader->binary, config,
offset, false);
- }
+ code_object_to_config(code_object, config);
lds_blocks = config->lds_size;
/* XXX: We are over allocating LDS. For SI, the shader reports
* LDS in blocks of 256 bytes, so if there are 4 bytes lds
* allocated in the shader and 4 bytes allocated by the state
* tracker, then we will set LDS_SIZE to 512 bytes rather than
256.
*/
if (sctx->chip_class <= SI) {
lds_blocks += align(program->local_size, 256) >> 8;
} else {
diff --git a/src/gallium/drivers/radeonsi/si_debug.c
b/src/gallium/drivers/radeonsi/si_debug.c
index c40dcd0b5d6..ae7ac2f0442 100644
--- a/src/gallium/drivers/radeonsi/si_debug.c
+++ b/src/gallium/drivers/radeonsi/si_debug.c
@@ -26,20 +26,21 @@
#include "si_compute.h"
#include "sid.h"
#include "gfx9d.h"
#include "sid_tables.h"
#include "driver_ddebug/dd_util.h"
#include "util/u_dump.h"
#include "util/u_log.h"
#include "util/u_memory.h"
#include "util/u_string.h"
#include "ac_debug.h"
+#include "ac_rtld.h"
static void si_dump_bo_list(struct si_context *sctx,
const struct radeon_saved_cs *saved, FILE *f);
DEBUG_GET_ONCE_OPTION(replace_shaders, "RADEON_REPLACE_SHADERS", NULL)
/**
* Store a linearized copy of all chunks of \p cs together with the buffer
* list in \p saved.
*/
@@ -195,29 +196,30 @@ static void si_dump_compute_shader(struct si_context *ctx,
chunk->ctx = ctx;
chunk->processor = PIPE_SHADER_COMPUTE;
chunk->shader = &state->program->shader;
si_compute_reference(&chunk->program, state->program);
u_log_chunk(log, &si_log_chunk_type_shader, chunk);
}
/**
* Shader compiles can be overridden with arbitrary ELF objects by setting
* the environment variable
RADEON_REPLACE_SHADERS=num1:filename1[;num2:filename2]
+ *
+ * TODO: key this off some hash
*/
-bool si_replace_shader(unsigned num, struct ac_shader_binary *binary)
+bool si_replace_shader(unsigned num, struct si_shader_binary *binary)
{
const char *p = debug_get_option_replace_shaders();
const char *semicolon;
char *copy = NULL;
FILE *f;
long filesize, nread;
- char *buf = NULL;
bool replaced = false;
if (!p)
return false;
while (*p) {
unsigned long i;
char *endp;
i = strtoul(p, &endp, 0);
@@ -259,37 +261,39 @@ bool si_replace_shader(unsigned num, struct
ac_shader_binary *binary)
if (fseek(f, 0, SEEK_END) != 0)
goto file_error;
filesize = ftell(f);
if (filesize < 0)
goto file_error;
if (fseek(f, 0, SEEK_SET) != 0)
goto file_error;
- buf = MALLOC(filesize);
- if (!buf) {
+ binary->elf_buffer = MALLOC(filesize);
+ if (!binary->elf_buffer) {
fprintf(stderr, "out of memory\n");
goto out_close;
}
- nread = fread(buf, 1, filesize, f);
- if (nread != filesize)
+ nread = fread((void*)binary->elf_buffer, 1, filesize, f);
+ if (nread != filesize) {
+ FREE((void*)binary->elf_buffer);
+ binary->elf_buffer = NULL;
goto file_error;
+ }
- ac_elf_read(buf, filesize, binary);
+ binary->elf_size = nread;
replaced = true;
out_close:
fclose(f);
out_free:
- FREE(buf);
free(copy);
return replaced;
file_error:
perror("radeonsi: reading shader");
goto out_close;
}
/* Parsed IBs are difficult to read without colors. Use "less -R file" to
* read them, or use "aha -b -f file" to convert them to html.
@@ -888,60 +892,79 @@ static void si_dump_compute_descriptors(struct si_context
*sctx,
}
struct si_shader_inst {
const char *text; /* start of disassembly for this instruction */
unsigned textlen;
unsigned size; /* instruction size = 4 or 8 */
uint64_t addr; /* instruction address */
};
/**
- * Split a disassembly string into instructions and add them to the array
- * pointed to by \p instructions.
+ * Open the given \p binary as \p rtld_binary and split the contained
+ * disassembly string into instructions and add them to the array
+ * pointed to by \p instructions, which must be sufficiently large.
*
* Labels are considered to be part of the following instruction.
+ *
+ * The caller must keep \p rtld_binary alive as long as \p instructions are
+ * used and then close it afterwards.
*/
-static void si_add_split_disasm(const char *disasm,
+static void si_add_split_disasm(struct ac_rtld_binary *rtld_binary,
+ struct si_shader_binary *binary,
uint64_t *addr,
unsigned *num,
struct si_shader_inst *instructions)
{
- const char *semicolon;
+ if (!ac_rtld_open(rtld_binary, 1, &binary->elf_buffer,
&binary->elf_size))
+ return;
+
+ const char *disasm;
+ uint64_t nbytes;
+ if (!ac_rtld_get_section_by_name(rtld_binary, ".AMDGPU.disasm",
+ &disasm, &nbytes))
+ return;
+
+ const char *end = disasm + nbytes;
+ while (disasm < end) {
+ const char *semicolon = memchr(disasm, ';', end - disasm);
+ if (!semicolon)
+ break;
- while ((semicolon = strchr(disasm, ';'))) {
struct si_shader_inst *inst = &instructions[(*num)++];
- const char *end = util_strchrnul(semicolon, '\n');
+ const char *inst_end = memchr(semicolon + 1, '\n', end -
semicolon - 1);
+ if (!inst_end)
+ inst_end = end;
inst->text = disasm;
- inst->textlen = end - disasm;
+ inst->textlen = inst_end - disasm;
inst->addr = *addr;
/* More than 16 chars after ";" means the instruction is 8
bytes long. */
- inst->size = end - semicolon > 16 ? 8 : 4;
+ inst->size = inst_end - semicolon > 16 ? 8 : 4;
*addr += inst->size;
- if (!(*end))
+ if (inst_end == end)
break;
- disasm = end + 1;
+ disasm = inst_end + 1;
}
}
/* If the shader is being executed, print its asm instructions, and annotate
* those that are being executed right now with information about waves that
* execute them. This is most useful during a GPU hang.
*/
static void si_print_annotated_shader(struct si_shader *shader,
struct ac_wave_info *waves,
unsigned num_waves,
FILE *f)
{
- if (!shader || !shader->binary.disasm_string)
+ if (!shader)
return;
uint64_t start_addr = shader->bo->gpu_address;
uint64_t end_addr = start_addr + shader->bo->b.b.width0;
unsigned i;
/* See if any wave executes the shader. */
for (i = 0; i < num_waves; i++) {
if (start_addr <= waves[i].pc && waves[i].pc <= end_addr)
break;
@@ -951,39 +974,40 @@ static void si_print_annotated_shader(struct si_shader
*shader,
/* Remember the first found wave. The waves are sorted according to PC.
*/
waves = &waves[i];
num_waves -= i;
/* Get the list of instructions.
* Buffer size / 4 is the upper bound of the instruction count.
*/
unsigned num_inst = 0;
uint64_t inst_addr = start_addr;
+ struct ac_rtld_binary rtld_binaries[5] = {};
struct si_shader_inst *instructions =
calloc(shader->bo->b.b.width0 / 4, sizeof(struct
si_shader_inst));
if (shader->prolog) {
- si_add_split_disasm(shader->prolog->binary.disasm_string,
+ si_add_split_disasm(&rtld_binaries[0], &shader->prolog->binary,
&inst_addr, &num_inst, instructions);
}
if (shader->previous_stage) {
-
si_add_split_disasm(shader->previous_stage->binary.disasm_string,
+ si_add_split_disasm(&rtld_binaries[1],
&shader->previous_stage->binary,
&inst_addr, &num_inst, instructions);
}
if (shader->prolog2) {
- si_add_split_disasm(shader->prolog2->binary.disasm_string,
+ si_add_split_disasm(&rtld_binaries[2], &shader->prolog2->binary,
&inst_addr, &num_inst, instructions);
}
- si_add_split_disasm(shader->binary.disasm_string,
+ si_add_split_disasm(&rtld_binaries[3], &shader->binary,
&inst_addr, &num_inst, instructions);
if (shader->epilog) {
- si_add_split_disasm(shader->epilog->binary.disasm_string,
+ si_add_split_disasm(&rtld_binaries[4], &shader->epilog->binary,
&inst_addr, &num_inst, instructions);
}
fprintf(f, COLOR_YELLOW "%s - annotated disassembly:" COLOR_RESET "\n",
si_get_shader_name(shader, shader->selector->type));
/* Print instructions with annotations. */
for (i = 0; i < num_inst; i++) {
struct si_shader_inst *inst = &instructions[i];
@@ -1007,20 +1031,22 @@ static void si_print_annotated_shader(struct si_shader
*shader,
}
waves->matched = true;
waves = &waves[1];
num_waves--;
}
}
fprintf(f, "\n\n");
free(instructions);
+ for (unsigned i = 0; i < ARRAY_SIZE(rtld_binaries); ++i)
+ ac_rtld_close(&rtld_binaries[i]);
}
static void si_dump_annotated_shaders(struct si_context *sctx, FILE *f)
{
struct ac_wave_info waves[AC_MAX_WAVES_PER_CHIP];
unsigned num_waves = ac_get_wave_info(waves);
fprintf(f, COLOR_CYAN "The number of active waves = %u" COLOR_RESET
"\n\n", num_waves);
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c
b/src/gallium/drivers/radeonsi/si_pipe.c
index aaf5138a3a2..795aef97d73 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -723,21 +723,21 @@ static void si_destroy_screen(struct pipe_screen* pscreen)
for (i = 0; i < ARRAY_SIZE(sscreen->compiler_lowp); i++)
si_destroy_compiler(&sscreen->compiler_lowp[i]);
/* Free shader parts. */
for (i = 0; i < ARRAY_SIZE(parts); i++) {
while (parts[i]) {
struct si_shader_part *part = parts[i];
parts[i] = part->next;
- ac_shader_binary_clean(&part->binary);
+ si_shader_binary_clean(&part->binary);
FREE(part);
}
}
mtx_destroy(&sscreen->shader_parts_mutex);
si_destroy_shader_cache(sscreen);
si_destroy_perfcounters(sscreen);
si_gpu_load_kill_thread(sscreen);
mtx_destroy(&sscreen->gpu_load_mutex);
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h
b/src/gallium/drivers/radeonsi/si_pipe.h
index 695827c9dd7..c6a2e7179e8 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -1249,21 +1249,21 @@ void si_save_cs(struct radeon_winsys *ws, struct
radeon_cmdbuf *cs,
struct radeon_saved_cs *saved, bool get_buffer_list);
void si_clear_saved_cs(struct radeon_saved_cs *saved);
void si_destroy_saved_cs(struct si_saved_cs *scs);
void si_auto_log_cs(void *data, struct u_log_context *log);
void si_log_hw_flush(struct si_context *sctx);
void si_log_draw_state(struct si_context *sctx, struct u_log_context *log);
void si_log_compute_state(struct si_context *sctx, struct u_log_context *log);
void si_init_debug_functions(struct si_context *sctx);
void si_check_vm_faults(struct si_context *sctx,
struct radeon_saved_cs *saved, enum ring_type ring);
-bool si_replace_shader(unsigned num, struct ac_shader_binary *binary);
+bool si_replace_shader(unsigned num, struct si_shader_binary *binary);
/* si_dma.c */
void si_init_dma_functions(struct si_context *sctx);
/* si_dma_cs.c */
void si_dma_emit_timestamp(struct si_context *sctx, struct si_resource *dst,
uint64_t offset);
void si_sdma_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
uint64_t offset, uint64_t size, unsigned clear_value);
void si_need_dma_space(struct si_context *ctx, unsigned num_dw,
diff --git a/src/gallium/drivers/radeonsi/si_shader.c
b/src/gallium/drivers/radeonsi/si_shader.c
index c457ca12b9a..38658942007 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -21,22 +21,24 @@
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
#include "util/u_memory.h"
#include "util/u_string.h"
#include "tgsi/tgsi_build.h"
#include "tgsi/tgsi_util.h"
#include "tgsi/tgsi_dump.h"
+#include "ac_binary.h"
#include "ac_exp_param.h"
#include "ac_shader_util.h"
+#include "ac_rtld.h"
#include "ac_llvm_util.h"
#include "si_shader_internal.h"
#include "si_pipe.h"
#include "sid.h"
#include "compiler/nir/nir.h"
static const char scratch_rsrc_dword0_symbol[] =
"SCRATCH_RSRC_DWORD0";
@@ -4962,182 +4964,169 @@ static void si_llvm_emit_polygon_stipple(struct
si_shader_context *ctx,
/* The stipple pattern is 32x32, each row has 32 bits. */
offset = LLVMBuildMul(builder, address[1],
LLVMConstInt(ctx->i32, 4, 0), "");
row = buffer_load_const(ctx, desc, offset);
row = ac_to_integer(&ctx->ac, row);
bit = LLVMBuildLShr(builder, row, address[0], "");
bit = LLVMBuildTrunc(builder, bit, ctx->i1, "");
ac_build_kill_if_false(&ctx->ac, bit);
}
-void si_shader_apply_scratch_relocs(struct si_shader *shader,
- uint64_t scratch_va)
-{
- unsigned i;
- uint32_t scratch_rsrc_dword0 = scratch_va;
- uint32_t scratch_rsrc_dword1 =
- S_008F04_BASE_ADDRESS_HI(scratch_va >> 32);
-
- /* Enable scratch coalescing. */
- scratch_rsrc_dword1 |= S_008F04_SWIZZLE_ENABLE(1);
-
- for (i = 0 ; i < shader->binary.reloc_count; i++) {
- const struct ac_shader_reloc *reloc =
- &shader->binary.relocs[i];
- if (!strcmp(scratch_rsrc_dword0_symbol, reloc->name)) {
- util_memcpy_cpu_to_le32(shader->binary.code +
reloc->offset,
- &scratch_rsrc_dword0, 4);
- } else if (!strcmp(scratch_rsrc_dword1_symbol, reloc->name)) {
- util_memcpy_cpu_to_le32(shader->binary.code +
reloc->offset,
- &scratch_rsrc_dword1, 4);
- }
- }
-}
-
/* For the UMR disassembler. */
#define DEBUGGER_END_OF_CODE_MARKER 0xbf9f0000 /* invalid instruction */
#define DEBUGGER_NUM_MARKERS 5
+static bool si_shader_binary_open(const struct si_shader *shader,
+ struct ac_rtld_binary *rtld)
+{
+ const char *part_elfs[5];
+ uint64_t part_sizes[5];
+ unsigned num_parts = 0;
+
+#define add_part(shader_or_part) \
+ if (shader_or_part) { \
+ part_elfs[num_parts] = (shader_or_part)->binary.elf_buffer; \
+ part_sizes[num_parts] = (shader_or_part)->binary.elf_size; \
+ num_parts++; \
+ }
+
+ add_part(shader->prolog);
+ add_part(shader->previous_stage);
+ add_part(shader->prolog2);
+ add_part(shader);
+ add_part(shader->epilog);
+
+#undef add_part
+
+ return ac_rtld_open(rtld, num_parts, part_elfs, part_sizes);
+}
+
static unsigned si_get_shader_binary_size(const struct si_shader *shader)
{
- unsigned size = shader->binary.code_size;
-
- if (shader->prolog)
- size += shader->prolog->binary.code_size;
- if (shader->previous_stage)
- size += shader->previous_stage->binary.code_size;
- if (shader->prolog2)
- size += shader->prolog2->binary.code_size;
- if (shader->epilog)
- size += shader->epilog->binary.code_size;
- return size + DEBUGGER_NUM_MARKERS * 4;
-}
-
-bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader
*shader)
-{
- const struct ac_shader_binary *prolog =
- shader->prolog ? &shader->prolog->binary : NULL;
- const struct ac_shader_binary *previous_stage =
- shader->previous_stage ? &shader->previous_stage->binary : NULL;
- const struct ac_shader_binary *prolog2 =
- shader->prolog2 ? &shader->prolog2->binary : NULL;
- const struct ac_shader_binary *epilog =
- shader->epilog ? &shader->epilog->binary : NULL;
- const struct ac_shader_binary *mainb = &shader->binary;
- unsigned bo_size = si_get_shader_binary_size(shader) +
- (!epilog ? mainb->rodata_size : 0);
- unsigned char *ptr;
-
- assert(!prolog || !prolog->rodata_size);
- assert(!previous_stage || !previous_stage->rodata_size);
- assert(!prolog2 || !prolog2->rodata_size);
- assert((!prolog && !previous_stage && !prolog2 && !epilog) ||
- !mainb->rodata_size);
- assert(!epilog || !epilog->rodata_size);
+ struct ac_rtld_binary rtld;
+ si_shader_binary_open(shader, &rtld);
+ return rtld.rx_size;
+}
+
+
+static bool si_get_external_symbol(void *data, const char *name, uint64_t
*value)
+{
+ uint64_t *scratch_va = data;
+
+ if (!strcmp(scratch_rsrc_dword0_symbol, name)) {
+ *value = (uint32_t)*scratch_va;
+ return true;
+ }
+ if (!strcmp(scratch_rsrc_dword1_symbol, name)) {
+ /* Enable scratch coalescing. */
+ *value = S_008F04_BASE_ADDRESS_HI(*scratch_va >> 32) |
+ S_008F04_SWIZZLE_ENABLE(1);
+ if (HAVE_LLVM < 0x0800) {
+ /* Old LLVM created an R_ABS32_HI relocation for
+ * this symbol. */
+ *value <<= 32;
+ }
+ return true;
+ }
+
+ return false;
+}
+
+bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader
*shader,
+ uint64_t scratch_va)
+{
+ struct ac_rtld_binary binary;
+ if (!si_shader_binary_open(shader, &binary))
+ return false;
si_resource_reference(&shader->bo, NULL);
shader->bo = si_aligned_buffer_create(&sscreen->b,
sscreen->cpdma_prefetch_writes_memory ?
0 : SI_RESOURCE_FLAG_READ_ONLY,
PIPE_USAGE_IMMUTABLE,
- align(bo_size,
SI_CPDMA_ALIGNMENT),
+ align(binary.rx_size,
SI_CPDMA_ALIGNMENT),
256);
if (!shader->bo)
return false;
/* Upload. */
- ptr = sscreen->ws->buffer_map(shader->bo->buf, NULL,
+ struct ac_rtld_upload_info u;
+ u.binary = &binary;
+ u.get_external_symbol = si_get_external_symbol;
+ u.cb_data = &scratch_va;
+ u.rx_va = shader->bo->gpu_address;
+ u.rx_ptr = sscreen->ws->buffer_map(shader->bo->buf, NULL,
PIPE_TRANSFER_READ_WRITE |
PIPE_TRANSFER_UNSYNCHRONIZED |
RADEON_TRANSFER_TEMPORARY);
+ if (!u.rx_ptr)
+ return false;
- /* Don't use util_memcpy_cpu_to_le32. LLVM binaries are
- * endian-independent. */
- if (prolog) {
- memcpy(ptr, prolog->code, prolog->code_size);
- ptr += prolog->code_size;
- }
- if (previous_stage) {
- memcpy(ptr, previous_stage->code, previous_stage->code_size);
- ptr += previous_stage->code_size;
- }
- if (prolog2) {
- memcpy(ptr, prolog2->code, prolog2->code_size);
- ptr += prolog2->code_size;
- }
-
- memcpy(ptr, mainb->code, mainb->code_size);
- ptr += mainb->code_size;
-
- if (epilog) {
- memcpy(ptr, epilog->code, epilog->code_size);
- ptr += epilog->code_size;
- } else if (mainb->rodata_size > 0) {
- memcpy(ptr, mainb->rodata, mainb->rodata_size);
- ptr += mainb->rodata_size;
- }
-
- /* Add end-of-code markers for the UMR disassembler. */
- uint32_t *ptr32 = (uint32_t*)ptr;
- for (unsigned i = 0; i < DEBUGGER_NUM_MARKERS; i++)
- ptr32[i] = DEBUGGER_END_OF_CODE_MARKER;
+ bool ok = ac_rtld_upload(&u);
sscreen->ws->buffer_unmap(shader->bo->buf);
- return true;
+ return ok;
}
-static void si_shader_dump_disassembly(const struct ac_shader_binary *binary,
+static void si_shader_dump_disassembly(const struct si_shader_binary *binary,
struct pipe_debug_callback *debug,
const char *name, FILE *file)
{
- char *line, *p;
- unsigned i, count;
+ struct ac_rtld_binary rtld_binary;
- if (binary->disasm_string) {
- fprintf(file, "Shader %s disassembly:\n", name);
- fprintf(file, "%s", binary->disasm_string);
+ if (!ac_rtld_open(&rtld_binary, 1, &binary->elf_buffer,
&binary->elf_size))
+ return;
- if (debug && debug->debug_message) {
- /* Very long debug messages are cut off, so send the
- * disassembly one line at a time. This causes more
- * overhead, but on the plus side it simplifies
- * parsing of resulting logs.
- */
- pipe_debug_message(debug, SHADER_INFO,
- "Shader Disassembly Begin");
+ const char *disasm;
+ uint64_t nbytes;
- line = binary->disasm_string;
- while (*line) {
- p = util_strchrnul(line, '\n');
- count = p - line;
+ if (!ac_rtld_get_section_by_name(&rtld_binary, ".AMDGPU.disasm",
&disasm, &nbytes))
+ goto out;
- if (count) {
- pipe_debug_message(debug, SHADER_INFO,
- "%.*s", count, line);
- }
+ fprintf(file, "Shader %s disassembly:\n", name);
+ if (nbytes > INT_MAX) {
+ fprintf(file, "too long\n");
+ goto out;
+ }
- if (!*p)
- break;
- line = p + 1;
+ fprintf(file, "%*s", (int)nbytes, disasm);
+
+ if (debug && debug->debug_message) {
+ /* Very long debug messages are cut off, so send the
+ * disassembly one line at a time. This causes more
+ * overhead, but on the plus side it simplifies
+ * parsing of resulting logs.
+ */
+ pipe_debug_message(debug, SHADER_INFO,
+ "Shader Disassembly Begin");
+
+ uint64_t line = 0;
+ while (line < nbytes) {
+ int count = nbytes - line;
+ const char *nl = memchr(disasm + line, '\n', nbytes -
line);
+ if (nl)
+ count = nl - disasm;
+
+ if (count) {
+ pipe_debug_message(debug, SHADER_INFO,
+ "%.*s", count, disasm +
line);
}
- pipe_debug_message(debug, SHADER_INFO,
- "Shader Disassembly End");
- }
- } else {
- fprintf(file, "Shader %s binary:\n", name);
- for (i = 0; i < binary->code_size; i += 4) {
- fprintf(file, "@0x%x: %02x%02x%02x%02x\n", i,
- binary->code[i + 3], binary->code[i + 2],
- binary->code[i + 1], binary->code[i]);
+ line += count + 1;
}
+
+ pipe_debug_message(debug, SHADER_INFO,
+ "Shader Disassembly End");
}
+
+out:
+ ac_rtld_close(&rtld_binary);
}
static void si_calculate_max_simd_waves(struct si_shader *shader)
{
struct si_screen *sscreen = shader->selector->screen;
struct ac_shader_config *conf = &shader->config;
unsigned num_inputs = shader->selector->info.num_inputs;
unsigned lds_increment = sscreen->info.chip_class >= CIK ? 512 : 256;
unsigned lds_per_wave = 0;
unsigned max_simd_waves;
@@ -5313,90 +5302,86 @@ void si_shader_dump(struct si_screen *sscreen, const
struct si_shader *shader,
if (shader->epilog)
si_shader_dump_disassembly(&shader->epilog->binary,
debug, "epilog", file);
fprintf(file, "\n");
}
si_shader_dump_stats(sscreen, shader, processor, file,
check_debug_option);
}
+bool si_shader_binary_read_config(struct si_shader_binary *binary,
+ struct ac_shader_config *conf)
+{
+ struct ac_rtld_binary rtld;
+ if (!ac_rtld_open(&rtld, 1, &binary->elf_buffer, &binary->elf_size))
+ return false;
+
+ bool ok = ac_rtld_read_config(&rtld, conf);
+
+ ac_rtld_close(&rtld);
+ return ok;
+}
+
static int si_compile_llvm(struct si_screen *sscreen,
- struct ac_shader_binary *binary,
+ struct si_shader_binary *binary,
struct ac_shader_config *conf,
struct ac_llvm_compiler *compiler,
LLVMModuleRef mod,
struct pipe_debug_callback *debug,
unsigned processor,
const char *name,
bool less_optimized)
{
- int r = 0;
unsigned count = p_atomic_inc_return(&sscreen->num_compilations);
if (si_can_dump_shader(sscreen, processor)) {
fprintf(stderr, "radeonsi: Compiling shader %d\n", count);
if (!(sscreen->debug_flags & (DBG(NO_IR) | DBG(PREOPT_IR)))) {
fprintf(stderr, "%s LLVM IR:\n\n", name);
ac_dump_module(mod);
fprintf(stderr, "\n");
}
}
if (sscreen->record_llvm_ir) {
char *ir = LLVMPrintModuleToString(mod);
binary->llvm_ir_string = strdup(ir);
LLVMDisposeMessage(ir);
}
if (!si_replace_shader(count, binary)) {
- r = si_llvm_compile(mod, binary, compiler, debug,
- less_optimized);
+ unsigned r = si_llvm_compile(mod, binary, compiler, debug,
+ less_optimized);
if (r)
return r;
}
- ac_shader_binary_read_config(binary, conf, 0, false);
+ if (!si_shader_binary_read_config(binary, conf))
+ return -1;
/* Enable 64-bit and 16-bit denormals, because there is no performance
* cost.
*
* If denormals are enabled, all floating-point output modifiers are
* ignored.
*
* Don't enable denormals for 32-bit floats, because:
* - Floating-point output modifiers would be ignored by the hw.
* - Some opcodes don't support denormals, such as v_mad_f32. We would
* have to stop using those.
* - SI & CI would be very slow.
*/
conf->float_mode |= V_00B028_FP_64_DENORMS;
- FREE(binary->config);
- FREE(binary->global_symbol_offsets);
- binary->config = NULL;
- binary->global_symbol_offsets = NULL;
-
- /* Some shaders can't have rodata because their binaries can be
- * concatenated.
- */
- if (binary->rodata_size &&
- (processor == PIPE_SHADER_VERTEX ||
- processor == PIPE_SHADER_TESS_CTRL ||
- processor == PIPE_SHADER_TESS_EVAL ||
- processor == PIPE_SHADER_FRAGMENT)) {
- fprintf(stderr, "radeonsi: The shader can't have rodata.");
- return -EINVAL;
- }
-
- return r;
+ return 0;
}
static void si_llvm_build_ret(struct si_shader_context *ctx, LLVMValueRef ret)
{
if (LLVMGetTypeKind(LLVMTypeOf(ret)) == LLVMVoidTypeKind)
LLVMBuildRetVoid(ctx->ac.builder);
else
LLVMBuildRet(ctx->ac.builder, ret);
}
@@ -5567,21 +5552,25 @@ si_generate_gs_copy_shader(struct si_screen *sscreen,
bool ok = false;
if (si_compile_llvm(sscreen, &ctx.shader->binary,
&ctx.shader->config, ctx.compiler,
ctx.ac.module,
debug, PIPE_SHADER_GEOMETRY,
"GS Copy Shader", false) == 0) {
if (si_can_dump_shader(sscreen, PIPE_SHADER_GEOMETRY))
fprintf(stderr, "GS Copy Shader:\n");
si_shader_dump(sscreen, ctx.shader, debug,
PIPE_SHADER_GEOMETRY, stderr, true);
- ok = si_shader_binary_upload(sscreen, ctx.shader);
+
+ if (!ctx.shader->config.scratch_bytes_per_wave)
+ ok = si_shader_binary_upload(sscreen, ctx.shader, 0);
+ else
+ ok = true;
}
si_llvm_dispose(&ctx);
if (!ok) {
FREE(shader);
shader = NULL;
} else {
si_fix_resource_usage(sscreen, shader);
}
@@ -7899,30 +7888,30 @@ bool si_shader_create(struct si_screen *sscreen, struct
ac_llvm_compiler *compil
shader->epilog->config.num_vgprs);
}
si_calculate_max_simd_waves(shader);
}
si_fix_resource_usage(sscreen, shader);
si_shader_dump(sscreen, shader, debug, sel->info.processor,
stderr, true);
/* Upload. */
- if (!si_shader_binary_upload(sscreen, shader)) {
+ if (!si_shader_binary_upload(sscreen, shader, 0)) {
fprintf(stderr, "LLVM failed to upload shader\n");
return false;
}
return true;
}
void si_shader_destroy(struct si_shader *shader)
{
if (shader->scratch_bo)
si_resource_reference(&shader->scratch_bo, NULL);
si_resource_reference(&shader->bo, NULL);
if (!shader->is_binary_shared)
- ac_shader_binary_clean(&shader->binary);
+ si_shader_binary_clean(&shader->binary);
free(shader->shader_log);
}
diff --git a/src/gallium/drivers/radeonsi/si_shader.h
b/src/gallium/drivers/radeonsi/si_shader.h
index 26cfc432899..302de427c04 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -564,20 +564,27 @@ struct si_shader_info {
ubyte vs_output_param_offset[SI_MAX_VS_OUTPUTS];
ubyte num_input_sgprs;
ubyte num_input_vgprs;
signed char face_vgpr_index;
signed char ancillary_vgpr_index;
bool uses_instanceid;
ubyte nr_pos_exports;
ubyte nr_param_exports;
};
+struct si_shader_binary {
+ const char *elf_buffer;
+ uint64_t elf_size;
+
+ char *llvm_ir_string;
+};
+
struct si_shader {
struct si_compiler_ctx_state compiler_ctx_state;
struct si_shader_selector *selector;
struct si_shader_selector *previous_stage_sel; /* for refcounting
*/
struct si_shader *next_variant;
struct si_shader_part *prolog;
struct si_shader *previous_stage; /* for GFX9 */
struct si_shader_part *prolog2;
@@ -588,21 +595,21 @@ struct si_shader {
struct si_resource *scratch_bo;
struct si_shader_key key;
struct util_queue_fence ready;
bool compilation_failed;
bool is_monolithic;
bool is_optimized;
bool is_binary_shared;
bool is_gs_copy_shader;
/* The following data is all that's needed for binary shaders. */
- struct ac_shader_binary binary;
+ struct si_shader_binary binary;
struct ac_shader_config config;
struct si_shader_info info;
unsigned private_mem_vgprs;
unsigned max_simd_waves;
/* Shader key + LLVM IR + disassembly + statistics.
* Generated for debug contexts only.
*/
char *shader_log;
size_t shader_log_size;
@@ -647,52 +654,54 @@ struct si_shader {
} ctx_reg;
/*For save precompute registers value */
unsigned vgt_tf_param; /* VGT_TF_PARAM */
unsigned vgt_vertex_reuse_block_cntl; /* VGT_VERTEX_REUSE_BLOCK_CNTL */
};
struct si_shader_part {
struct si_shader_part *next;
union si_shader_part_key key;
- struct ac_shader_binary binary;
+ struct si_shader_binary binary;
struct ac_shader_config config;
};
/* si_shader.c */
struct si_shader *
si_generate_gs_copy_shader(struct si_screen *sscreen,
struct ac_llvm_compiler *compiler,
struct si_shader_selector *gs_selector,
struct pipe_debug_callback *debug);
int si_compile_tgsi_shader(struct si_screen *sscreen,
struct ac_llvm_compiler *compiler,
struct si_shader *shader,
struct pipe_debug_callback *debug);
bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler
*compiler,
struct si_shader *shader,
struct pipe_debug_callback *debug);
void si_shader_destroy(struct si_shader *shader);
unsigned si_shader_io_get_unique_index_patch(unsigned semantic_name, unsigned
index);
unsigned si_shader_io_get_unique_index(unsigned semantic_name, unsigned index,
unsigned is_varying);
-bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader
*shader);
+bool si_shader_binary_upload(struct si_screen *sscreen, struct si_shader
*shader,
+ uint64_t scratch_va);
void si_shader_dump(struct si_screen *sscreen, const struct si_shader *shader,
struct pipe_debug_callback *debug, unsigned processor,
FILE *f, bool check_debug_option);
void si_shader_dump_stats_for_shader_db(const struct si_shader *shader,
struct pipe_debug_callback *debug);
void si_multiwave_lds_size_workaround(struct si_screen *sscreen,
unsigned *lds_size);
-void si_shader_apply_scratch_relocs(struct si_shader *shader,
- uint64_t scratch_va);
const char *si_get_shader_name(const struct si_shader *shader, unsigned
processor);
+bool si_shader_binary_read_config(struct si_shader_binary *binary,
+ struct ac_shader_config *conf);
+void si_shader_binary_clean(struct si_shader_binary *binary);
/* si_shader_nir.c */
void si_nir_scan_shader(const struct nir_shader *nir,
struct tgsi_shader_info *info);
void si_nir_scan_tess_ctrl(const struct nir_shader *nir,
struct tgsi_tessctrl_info *out);
void si_lower_nir(struct si_shader_selector *sel);
/* Inline helpers. */
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h
b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 235c46ecf92..464439f5a6b 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -29,21 +29,20 @@
#include "gallivm/lp_bld_flow.h"
#include "gallivm/lp_bld_init.h"
#include "gallivm/lp_bld_tgsi.h"
#include "tgsi/tgsi_parse.h"
#include "ac_shader_abi.h"
#include <llvm-c/Core.h>
#include <llvm-c/TargetMachine.h>
struct pipe_debug_callback;
-struct ac_shader_binary;
#define RADEON_LLVM_MAX_INPUT_SLOTS 32
#define RADEON_LLVM_MAX_INPUTS 32 * 4
#define RADEON_LLVM_MAX_OUTPUTS 32 * 4
#define RADEON_LLVM_MAX_SYSTEM_VALUES 11
#define RADEON_LLVM_MAX_ADDRS 16
struct si_shader_context {
struct lp_build_tgsi_context bld_base;
@@ -211,21 +210,21 @@ si_shader_context(struct lp_build_tgsi_context *bld_base)
return (struct si_shader_context*)bld_base;
}
static inline struct si_shader_context *
si_shader_context_from_abi(struct ac_shader_abi *abi)
{
struct si_shader_context *ctx = NULL;
return container_of(abi, ctx, abi);
}
-unsigned si_llvm_compile(LLVMModuleRef M, struct ac_shader_binary *binary,
+unsigned si_llvm_compile(LLVMModuleRef M, struct si_shader_binary *binary,
struct ac_llvm_compiler *compiler,
struct pipe_debug_callback *debug,
bool less_optimized);
LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type);
LLVMValueRef bitcast(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type, LLVMValueRef value);
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
index 3ec919dd23b..3e27e9b0c38 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
@@ -73,48 +73,58 @@ static void si_diagnostic_handler(LLVMDiagnosticInfoRef di,
void *context)
}
LLVMDisposeMessage(description);
}
/**
* Compile an LLVM module to machine code.
*
* @returns 0 for success, 1 for failure
*/
-unsigned si_llvm_compile(LLVMModuleRef M, struct ac_shader_binary *binary,
+unsigned si_llvm_compile(LLVMModuleRef M, struct si_shader_binary *binary,
struct ac_llvm_compiler *compiler,
struct pipe_debug_callback *debug,
bool less_optimized)
{
struct ac_compiler_passes *passes =
less_optimized && compiler->low_opt_passes ?
compiler->low_opt_passes : compiler->passes;
struct si_llvm_diagnostics diag;
LLVMContextRef llvm_ctx;
diag.debug = debug;
diag.retval = 0;
/* Setup Diagnostic Handler*/
llvm_ctx = LLVMGetModuleContext(M);
LLVMContextSetDiagnosticHandler(llvm_ctx, si_diagnostic_handler, &diag);
/* Compile IR. */
- if (!ac_compile_module_to_binary(passes, M, binary))
+ if (!ac_compile_module_to_elf(passes, M, (char **)&binary->elf_buffer,
+ &binary->elf_size))
diag.retval = 1;
if (diag.retval != 0)
pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed");
return diag.retval;
}
+void si_shader_binary_clean(struct si_shader_binary *binary)
+{
+ free((void *)binary->elf_buffer);
+ binary->elf_buffer = NULL;
+
+ free(binary->llvm_ir_string);
+ binary->llvm_ir_string = NULL;
+}
+
LLVMTypeRef tgsi2llvmtype(struct lp_build_tgsi_context *bld_base,
enum tgsi_opcode_type type)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
switch (type) {
case TGSI_TYPE_UNSIGNED:
case TGSI_TYPE_SIGNED:
return ctx->ac.i32;
case TGSI_TYPE_UNSIGNED64:
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c
b/src/gallium/drivers/radeonsi/si_state_shaders.c
index b8a587b50eb..041ed1c0c62 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -120,83 +120,77 @@ static uint32_t *read_chunk(uint32_t *ptr, void **data,
unsigned *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 = shader->binary.disasm_string ?
- strlen(shader->binary.disasm_string) + 1 : 0;
unsigned llvm_ir_size = shader->binary.llvm_ir_string ?
strlen(shader->binary.llvm_ir_string) + 1 : 0;
+
+ /* Refuse to allocate overly large buffers and guard against integer
+ * overflow. */
+ if (shader->binary.elf_size > UINT_MAX / 4 ||
+ llvm_ir_size > UINT_MAX / 4)
+ return NULL;
+
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) +
+ 4 + align(shader->binary.elf_size, 4) +
4 + align(llvm_ir_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);
+ ptr = write_chunk(ptr, shader->binary.elf_buffer,
shader->binary.elf_size);
ptr = write_chunk(ptr, shader->binary.llvm_ir_string, llvm_ir_size);
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;
+ unsigned elf_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);
+ ptr = read_chunk(ptr, (void**)&shader->binary.elf_buffer,
+ &elf_size);
+ shader->binary.elf_size = elf_size;
ptr = read_chunk(ptr, (void**)&shader->binary.llvm_ir_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 false on failure, in which case the ir_binary should be freed.
@@ -3095,27 +3089,22 @@ static int si_update_scratch_buffer(struct si_context
*sctx,
/* This shader is already configured to use the current
* scratch buffer. */
if (shader->scratch_bo == sctx->scratch_buffer) {
si_shader_unlock(shader);
return 0;
}
assert(sctx->scratch_buffer);
- if (shader->previous_stage)
- si_shader_apply_scratch_relocs(shader->previous_stage,
scratch_va);
-
- si_shader_apply_scratch_relocs(shader, scratch_va);
-
/* Replace the shader bo with a new bo that has the relocs applied. */
- if (!si_shader_binary_upload(sctx->screen, shader)) {
+ if (!si_shader_binary_upload(sctx->screen, shader, scratch_va)) {
si_shader_unlock(shader);
return -1;
}
/* Update the shader state to use the new shader bo. */
si_shader_init_pm4_state(sctx->screen, shader);
si_resource_reference(&shader->scratch_bo, sctx->scratch_buffer);
si_shader_unlock(shader);
--
2.20.1
_______________________________________________
mesa-dev mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
