gallium: PIPE_COMPUTE_CAP_IR_TARGET - allow drivers to specify a processor v2

This target string now contains four values instead of three.  The old
processor field (which was really being interpreted as arch) has been split
into two fields: processor and arch.  This allows drivers to pass a
more a more detailed description of the hardware to compiler frontends.

v2:
  - Adapt to libclc changes

Reviewed-by: Francisco Jerez <currojerez@riseup.net>

src/gallium/docs/source/screen.rst

@@ -236,10 +236,10 @@ PIPE_COMPUTE_CAP_*
 Compute-specific capabilities. They can be queried using
 pipe_screen::get_compute_param.
 
-* ``PIPE_COMPUTE_CAP_IR_TARGET``: A description of the target as a target
+* ``PIPE_COMPUTE_CAP_IR_TARGET``: A description of the target of the form
-  triple specification of the form ``processor-manufacturer-os`` that will
+  ``processor-arch-manufacturer-os`` that will be passed on to the compiler.
-  be passed on to the compiler.  This CAP is only relevant for drivers
+  This CAP is only relevant for drivers that specify PIPE_SHADER_IR_LLVM for
-  that specify PIPE_SHADER_IR_LLVM for their preferred IR.
+  their preferred IR.
   Value type: null-terminated string.
 * ``PIPE_COMPUTE_CAP_GRID_DIMENSION``: Number of supported dimensions
   for grid and block coordinates.  Value type: ``uint64_t``.

src/gallium/drivers/r600/r600_llvm.c

@@ -550,69 +550,6 @@ LLVMModuleRef r600_tgsi_llvm(
 	return ctx->gallivm.module;
 }
 
-const char * r600_llvm_gpu_string(enum radeon_family family)
-{
-	const char * gpu_family;
-
-	switch (family) {
-	case CHIP_R600:
-	case CHIP_RV610:
-	case CHIP_RV630:
-	case CHIP_RV620:
-	case CHIP_RV635:
-	case CHIP_RV670:
-	case CHIP_RS780:
-	case CHIP_RS880:
-		gpu_family = "r600";
-		break;
-	case CHIP_RV710:
-		gpu_family = "rv710";
-		break;
-	case CHIP_RV730:
-		gpu_family = "rv730";
-		break;
-	case CHIP_RV740:
-	case CHIP_RV770:
-		gpu_family = "rv770";
-		break;
-	case CHIP_PALM:
-	case CHIP_CEDAR:
-		gpu_family = "cedar";
-		break;
-	case CHIP_SUMO:
-	case CHIP_SUMO2:
-	case CHIP_REDWOOD:
-		gpu_family = "redwood";
-		break;
-	case CHIP_JUNIPER:
-		gpu_family = "juniper";
-		break;
-	case CHIP_HEMLOCK:
-	case CHIP_CYPRESS:
-		gpu_family = "cypress";
-		break;
-	case CHIP_BARTS:
-		gpu_family = "barts";
-		break;
-	case CHIP_TURKS:
-		gpu_family = "turks";
-		break;
-	case CHIP_CAICOS:
-		gpu_family = "caicos";
-		break;
-	case CHIP_CAYMAN:
-        case CHIP_ARUBA:
-		gpu_family = "cayman";
-		break;
-	default:
-		gpu_family = "";
-		fprintf(stderr, "Chip not supported by r600 llvm "
-			"backend, please file a bug at " PACKAGE_BUGREPORT "\n");
-		break;
-	}
-	return gpu_family;
-}
-
 unsigned r600_llvm_compile(
 	LLVMModuleRef mod,
 	unsigned char ** inst_bytes,

src/gallium/drivers/r600/r600_llvm.h

@@ -15,8 +15,6 @@ LLVMModuleRef r600_tgsi_llvm(
 	struct radeon_llvm_context * ctx,
 	const struct tgsi_token * tokens);
 
-const char * r600_llvm_gpu_string(enum radeon_family family);
-
 unsigned r600_llvm_compile(
 	LLVMModuleRef mod,
 	unsigned char ** inst_bytes,

src/gallium/drivers/r600/r600_pipe.c

@@ -764,18 +764,84 @@ static int r600_get_video_param(struct pipe_screen *screen,
 	}
 }
 
+const char * r600_llvm_gpu_string(enum radeon_family family)
+{
+	const char * gpu_family;
+
+	switch (family) {
+	case CHIP_R600:
+	case CHIP_RV610:
+	case CHIP_RV630:
+	case CHIP_RV620:
+	case CHIP_RV635:
+	case CHIP_RV670:
+	case CHIP_RS780:
+	case CHIP_RS880:
+		gpu_family = "r600";
+		break;
+	case CHIP_RV710:
+		gpu_family = "rv710";
+		break;
+	case CHIP_RV730:
+		gpu_family = "rv730";
+		break;
+	case CHIP_RV740:
+	case CHIP_RV770:
+		gpu_family = "rv770";
+		break;
+	case CHIP_PALM:
+	case CHIP_CEDAR:
+		gpu_family = "cedar";
+		break;
+	case CHIP_SUMO:
+	case CHIP_SUMO2:
+	case CHIP_REDWOOD:
+		gpu_family = "redwood";
+		break;
+	case CHIP_JUNIPER:
+		gpu_family = "juniper";
+		break;
+	case CHIP_HEMLOCK:
+	case CHIP_CYPRESS:
+		gpu_family = "cypress";
+		break;
+	case CHIP_BARTS:
+		gpu_family = "barts";
+		break;
+	case CHIP_TURKS:
+		gpu_family = "turks";
+		break;
+	case CHIP_CAICOS:
+		gpu_family = "caicos";
+		break;
+	case CHIP_CAYMAN:
+        case CHIP_ARUBA:
+		gpu_family = "cayman";
+		break;
+	default:
+		gpu_family = "";
+		fprintf(stderr, "Chip not supported by r600 llvm "
+			"backend, please file a bug at " PACKAGE_BUGREPORT "\n");
+		break;
+	}
+	return gpu_family;
+}
+
+
 static int r600_get_compute_param(struct pipe_screen *screen,
         enum pipe_compute_cap param,
         void *ret)
 {
+	struct r600_screen *rscreen = (struct r600_screen *)screen;
 	//TODO: select these params by asic
 	switch (param) {
-	case PIPE_COMPUTE_CAP_IR_TARGET:
+	case PIPE_COMPUTE_CAP_IR_TARGET: {
+		const char *gpu = r600_llvm_gpu_string(rscreen->family);
 		if (ret) {
-			strcpy(ret, "r600--");
+			sprintf(ret, "%s-r600--", gpu);
 		}
-		return 7 * sizeof(char);
+		return (8 + strlen(gpu)) * sizeof(char);
-
+	}
 	case PIPE_COMPUTE_CAP_GRID_DIMENSION:
 		if (ret) {
 			uint64_t * grid_dimension = ret;

src/gallium/drivers/r600/r600_pipe.h

@@ -746,6 +746,8 @@ boolean r600_rings_is_buffer_referenced(struct r600_context *ctx,
 void *r600_buffer_mmap_sync_with_rings(struct r600_context *ctx,
 					struct r600_resource *resource,
 					unsigned usage);
+const char * r600_llvm_gpu_string(enum radeon_family family);
+
 
 /* r600_query.c */
 void r600_init_query_functions(struct r600_context *rctx);

src/gallium/drivers/radeon/LLVM_REVISION.txt

@@ -1 +1 @@
-@178505
+@178928

src/gallium/drivers/radeonsi/radeonsi_pipe.c

@@ -274,6 +274,17 @@ static const char* r600_get_vendor(struct pipe_screen* pscreen)
 	return "X.Org";
 }
 
+const char *r600_get_llvm_processor_name(enum radeon_family family)
+{
+	switch (family) {
+		case CHIP_TAHITI: return "tahiti";
+		case CHIP_PITCAIRN: return "pitcairn";
+		case CHIP_VERDE: return "verde";
+		case CHIP_OLAND: return "oland";
+		default: return "";
+	}
+}
+
 static const char *r600_get_family_name(enum radeon_family family)
 {
 	switch(family) {

src/gallium/drivers/radeonsi/radeonsi_pipe.h

@@ -218,6 +218,7 @@ void r600_upload_index_buffer(struct r600_context *rctx,
 /* r600_pipe.c */
 void radeonsi_flush(struct pipe_context *ctx, struct pipe_fence_handle **fence,
 		    unsigned flags);
+const char *r600_get_llvm_processor_name(enum radeon_family family);
 
 /* r600_query.c */
 void r600_init_query_functions(struct r600_context *rctx);

src/gallium/drivers/radeonsi/radeonsi_shader.c

@@ -1165,7 +1165,9 @@ int si_pipe_shader_create(
 	if (dump) {
 		LLVMDumpModule(mod);
 	}
-	radeon_llvm_compile(mod, &inst_bytes, &inst_byte_count, "SI", dump);
+	radeon_llvm_compile(mod, &inst_bytes, &inst_byte_count,
+			r600_get_llvm_processor_name(rctx->screen->family)
+			, dump);
 	if (dump) {
 		fprintf(stderr, "SI CODE:\n");
 		for (i = 0; i < inst_byte_count; i+=4 ) {

src/gallium/state_trackers/clover/llvm/invocation.cpp

@@ -110,7 +110,8 @@ namespace {
 
    llvm::Module *
    compile(const std::string &source, const std::string &name,
-           const std::string &triple, const std::string &opts) {
+           const std::string &triple, const std::string &processor,
+           const std::string &opts) {
 
       clang::CompilerInstance c;
       clang::CompilerInvocation invocation;
@@ -175,6 +176,7 @@ namespace {
 
       c.getLangOpts().NoBuiltin = true;
       c.getTargetOpts().Triple = triple;
+      c.getTargetOpts().CPU = processor;
 #if HAVE_LLVM <= 0x0301
       c.getInvocation().setLangDefaults(clang::IK_OpenCL);
 #else
@@ -215,12 +217,14 @@ namespace {
 
    void
    link(llvm::Module *mod, const std::string &triple,
+        const std::string &processor,
         const std::vector<llvm::Function *> &kernels) {
 
       llvm::PassManager PM;
       llvm::PassManagerBuilder Builder;
       llvm::sys::Path libclc_path =
-                            llvm::sys::Path(LIBCLC_LIBEXECDIR + triple + ".bc");
+                            llvm::sys::Path(LIBCLC_LIBEXECDIR + processor +
+			                    "-" + triple + ".bc");
 
       // Link the kernel with libclc
 #if HAVE_LLVM < 0x0303
@@ -339,18 +343,22 @@ namespace {
 module
 clover::compile_program_llvm(const compat::string &source,
                              enum pipe_shader_ir ir,
-                             const compat::string &triple,
+                             const compat::string &target,
                              const compat::string &opts) {
 
    std::vector<llvm::Function *> kernels;
+   size_t processor_str_len = std::string(target.begin()).find_first_of("-");
+   std::string processor(target.begin(), 0, processor_str_len);
+   std::string triple(target.begin(), processor_str_len + 1,
+                      target.size() - processor_str_len - 1);
 
    // The input file name must have the .cl extension in order for the
    // CompilerInvocation class to recognize it as an OpenCL source file.
-   llvm::Module *mod = compile(source, "input.cl", triple, opts);
+   llvm::Module *mod = compile(source, "input.cl", triple, processor, opts);
 
    find_kernels(mod, kernels);
 
-   link(mod, triple, kernels);
+   link(mod, triple, processor, kernels);
 
    // Build the clover::module
    switch (ir) {