From: Marek Olšák Date: Wed, 15 Jan 2020 01:40:51 +0000 (-0500) Subject: radeonsi: merge si_compile_llvm and si_llvm_compile functions X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=c4daf2b485e18b7f5d87392244ce740e5c5a9fbd;p=mesa.git radeonsi: merge si_compile_llvm and si_llvm_compile functions Reviewed-by: Timothy Arceri Tested-by: Marge Bot Part-of: --- diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index 49538e823f4..766d25b396e 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -2086,74 +2086,6 @@ void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader, si_shader_dump_stats(sscreen, shader, file, check_debug_option); } -int si_compile_llvm(struct si_screen *sscreen, - struct si_shader_binary *binary, - struct ac_shader_config *conf, - struct ac_llvm_compiler *compiler, - LLVMModuleRef mod, - struct pipe_debug_callback *debug, - enum pipe_shader_type shader_type, - unsigned wave_size, - const char *name, - bool less_optimized) -{ - unsigned count = p_atomic_inc_return(&sscreen->num_compilations); - - if (si_can_dump_shader(sscreen, shader_type)) { - 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)) { - unsigned r = si_llvm_compile(mod, binary, compiler, debug, - less_optimized, wave_size); - if (r) - return r; - } - - struct ac_rtld_binary rtld; - if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){ - .info = &sscreen->info, - .shader_type = tgsi_processor_to_shader_stage(shader_type), - .wave_size = wave_size, - .num_parts = 1, - .elf_ptrs = &binary->elf_buffer, - .elf_sizes = &binary->elf_size })) - return -1; - - bool ok = ac_rtld_read_config(&rtld, conf); - ac_rtld_close(&rtld); - if (!ok) - 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. - * - GFX6 & GFX7 would be very slow. - */ - conf->float_mode |= V_00B028_FP_64_DENORMS; - - return 0; -} - static void si_dump_shader_key_vs(const struct si_shader_key *key, const struct si_vs_prolog_bits *prolog, const char *prefix, FILE *f) @@ -3130,8 +3062,7 @@ int si_compile_shader(struct si_screen *sscreen, /* Compile to bytecode. */ r = si_compile_llvm(sscreen, &shader->binary, &shader->config, compiler, - ctx.ac.module, debug, ctx.type, ctx.ac.wave_size, - si_get_shader_name(shader), + &ctx.ac, debug, ctx.type, si_get_shader_name(shader), si_should_optimize_less(compiler, shader->selector)); si_llvm_dispose(&ctx); if (r) { @@ -3268,8 +3199,7 @@ si_get_shader_part(struct si_screen *sscreen, si_llvm_optimize_module(&ctx); if (si_compile_llvm(sscreen, &result->binary, &result->config, compiler, - ctx.ac.module, debug, ctx.type, ctx.ac.wave_size, - name, false)) { + &ctx.ac, debug, ctx.type, name, false)) { FREE(result); result = NULL; goto out; diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h index e9b41e0ff16..80011edb3d5 100644 --- a/src/gallium/drivers/radeonsi/si_shader_internal.h +++ b/src/gallium/drivers/radeonsi/si_shader_internal.h @@ -204,10 +204,6 @@ si_shader_context_from_abi(struct ac_shader_abi *abi) return container_of(abi, ctx, abi); } -unsigned si_llvm_compile(LLVMModuleRef M, struct si_shader_binary *binary, - struct ac_llvm_compiler *compiler, - struct pipe_debug_callback *debug, - bool less_optimized, unsigned wave_size); void si_llvm_context_init(struct si_shader_context *ctx, struct si_screen *sscreen, struct ac_llvm_compiler *compiler, @@ -292,10 +288,9 @@ int si_compile_llvm(struct si_screen *sscreen, struct si_shader_binary *binary, struct ac_shader_config *conf, struct ac_llvm_compiler *compiler, - LLVMModuleRef mod, + struct ac_llvm_context *ac, struct pipe_debug_callback *debug, enum pipe_shader_type shader_type, - unsigned wave_size, const char *name, bool less_optimized); void si_fix_resource_usage(struct si_screen *sscreen, struct si_shader *shader); diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm.c b/src/gallium/drivers/radeonsi/si_shader_llvm.c index 841c97b00bf..539b2531ab7 100644 --- a/src/gallium/drivers/radeonsi/si_shader_llvm.c +++ b/src/gallium/drivers/radeonsi/si_shader_llvm.c @@ -24,6 +24,10 @@ #include "si_shader_internal.h" #include "si_pipe.h" +#include "ac_rtld.h" +#include "sid.h" + +#include "tgsi/tgsi_from_mesa.h" #include "util/u_memory.h" struct si_llvm_diagnostics { @@ -63,42 +67,86 @@ 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 si_shader_binary *binary, - struct ac_llvm_compiler *compiler, - struct pipe_debug_callback *debug, - bool less_optimized, unsigned wave_size) +int si_compile_llvm(struct si_screen *sscreen, + struct si_shader_binary *binary, + struct ac_shader_config *conf, + struct ac_llvm_compiler *compiler, + struct ac_llvm_context *ac, + struct pipe_debug_callback *debug, + enum pipe_shader_type shader_type, + const char *name, + bool less_optimized) { - struct ac_compiler_passes *passes = compiler->passes; + unsigned count = p_atomic_inc_return(&sscreen->num_compilations); + + if (si_can_dump_shader(sscreen, shader_type)) { + 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(ac->module); + fprintf(stderr, "\n"); + } + } - if (wave_size == 32) - passes = compiler->passes_wave32; - else if (less_optimized && compiler->low_opt_passes) - passes = compiler->low_opt_passes; + if (sscreen->record_llvm_ir) { + char *ir = LLVMPrintModuleToString(ac->module); + binary->llvm_ir_string = strdup(ir); + LLVMDisposeMessage(ir); + } - struct si_llvm_diagnostics diag; - LLVMContextRef llvm_ctx; + if (!si_replace_shader(count, binary)) { + struct ac_compiler_passes *passes = compiler->passes; - diag.debug = debug; - diag.retval = 0; + if (ac->wave_size == 32) + passes = compiler->passes_wave32; + else if (less_optimized && compiler->low_opt_passes) + passes = compiler->low_opt_passes; - /* Setup Diagnostic Handler*/ - llvm_ctx = LLVMGetModuleContext(M); + struct si_llvm_diagnostics diag = {debug}; + LLVMContextSetDiagnosticHandler(ac->context, si_diagnostic_handler, &diag); - LLVMContextSetDiagnosticHandler(llvm_ctx, si_diagnostic_handler, &diag); + if (!ac_compile_module_to_elf(passes, ac->module, + (char **)&binary->elf_buffer, + &binary->elf_size)) + diag.retval = 1; - /* Compile IR. */ - 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 compilation failed"); + return diag.retval; + } + } + + struct ac_rtld_binary rtld; + if (!ac_rtld_open(&rtld, (struct ac_rtld_open_info){ + .info = &sscreen->info, + .shader_type = tgsi_processor_to_shader_stage(shader_type), + .wave_size = ac->wave_size, + .num_parts = 1, + .elf_ptrs = &binary->elf_buffer, + .elf_sizes = &binary->elf_size })) + return -1; + + bool ok = ac_rtld_read_config(&rtld, conf); + ac_rtld_close(&rtld); + if (!ok) + 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. + * - GFX6 & GFX7 would be very slow. + */ + conf->float_mode |= V_00B028_FP_64_DENORMS; - if (diag.retval != 0) - pipe_debug_message(debug, SHADER_INFO, "LLVM compile failed"); - return diag.retval; + return 0; } void si_shader_binary_clean(struct si_shader_binary *binary) diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c index 69dc7080a6c..a11748325dd 100644 --- a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c +++ b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c @@ -623,9 +623,8 @@ 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, ctx.ac.wave_size, + &ctx.shader->config, ctx.compiler, &ctx.ac, + debug, PIPE_SHADER_GEOMETRY, "GS Copy Shader", false) == 0) { if (si_can_dump_shader(sscreen, PIPE_SHADER_GEOMETRY)) fprintf(stderr, "GS Copy Shader:\n");