r600g/compute: Enable PIPE_SHADER_IR_NATIVE for compute shaders v2

v2:
  - Drop dependency on LLVM >= 3.5.1

8 files changed, 180 insertions, 97 deletions

diff --git a/src/gallium/drivers/r600/evergreen_compute.c b/src/gallium/drivers/r600/evergreen_compute.c
index 38b78c7dfcb..7a17d1ee089 100644
--- a/src/gallium/drivers/r600/evergreen_compute.c
+++ b/src/gallium/drivers/r600/evergreen_compute.c
@@ -49,6 +49,7 @@
 #ifdef HAVE_OPENCL
 #include "radeon_llvm_util.h"
 #endif
+#include "radeon_elf_util.h"
 #include <inttypes.h>
 
 /**
@@ -198,18 +199,42 @@ void *evergreen_create_compute_state(
 {
 	struct r600_context *ctx = (struct r600_context *)ctx_;
 	struct r600_pipe_compute *shader = CALLOC_STRUCT(r600_pipe_compute);
-
-#ifdef HAVE_OPENCL
 	const struct pipe_llvm_program_header * header;
-	const unsigned char * code;
-	unsigned i;
-
-	shader->llvm_ctx = LLVMContextCreate();
+	const char *code;
+	void *p;
+	boolean use_kill;
 
 	COMPUTE_DBG(ctx->screen, "*** evergreen_create_compute_state\n");
-
 	header = cso->prog;
 	code = cso->prog + sizeof(struct pipe_llvm_program_header);
+#if HAVE_LLVM < 0x0306
+#ifdef HAVE_OPENCL
+        (void)use_kill;
+	(void)p;
+	shader->llvm_ctx = LLVMContextCreate();
+	shader->num_kernels = radeon_llvm_get_num_kernels(shader->llvm_ctx,
+				code, header->num_bytes);
+	shader->kernels = CALLOC(sizeof(struct r600_kernel),
+				shader->num_kernels);
+	{
+		unsigned i;
+		for (i = 0; i < shader->num_kernels; i++) {
+			struct r600_kernel *kernel = &shader->kernels[i];
+			kernel->llvm_module = radeon_llvm_get_kernel_module(
+				shader->llvm_ctx, i, code, header->num_bytes);
+		}
+	}
+#endif
+#else
+	memset(&shader->binary, 0, sizeof(shader->binary));
+	radeon_elf_read(code, header->num_bytes, &shader->binary, true);
+	r600_create_shader(&shader->bc, &shader->binary, &use_kill);
+
+	shader->code_bo = r600_compute_buffer_alloc_vram(ctx->screen,
+							shader->bc.ndw * 4);
+	p = r600_buffer_map_sync_with_rings(&ctx->b, shader->code_bo, PIPE_TRANSFER_WRITE);
+	memcpy(p, shader->bc.bytecode, shader->bc.ndw * 4);
+	ctx->b.ws->buffer_unmap(shader->code_bo->cs_buf);
 #endif
 
 	shader->ctx = (struct r600_context*)ctx;
@@ -217,17 +242,6 @@ void *evergreen_create_compute_state(
 	shader->private_size = cso->req_private_mem;
 	shader->input_size = cso->req_input_mem;
 
-#ifdef HAVE_OPENCL 
-	shader->num_kernels = radeon_llvm_get_num_kernels(shader->llvm_ctx, code,
-							header->num_bytes);
-	shader->kernels = CALLOC(sizeof(struct r600_kernel), shader->num_kernels);
-
-	for (i = 0; i < shader->num_kernels; i++) {
-		struct r600_kernel *kernel = &shader->kernels[i];
-		kernel->llvm_module = radeon_llvm_get_kernel_module(shader->llvm_ctx, i,
-							code, header->num_bytes);
-	}
-#endif
 	return shader;
 }
 
@@ -238,14 +252,6 @@ void evergreen_delete_compute_state(struct pipe_context *ctx, void* state)
 	if (!shader)
 		return;
 
-	FREE(shader->kernels);
-
-#ifdef HAVE_OPENCL
-	if (shader->llvm_ctx){
-		LLVMContextDispose(shader->llvm_ctx);
-	}
-#endif
-
 	FREE(shader);
 }
 
@@ -347,7 +353,13 @@ static void evergreen_emit_direct_dispatch(
 	unsigned wave_divisor = (16 * num_pipes);
 	int group_size = 1;
 	int grid_size = 1;
-	unsigned lds_size = shader->local_size / 4 + shader->active_kernel->bc.nlds_dw;
+	unsigned lds_size = shader->local_size / 4 +
+#if HAVE_LLVM < 0x0306
+		shader->active_kernel->bc.nlds_dw;
+#else
+		shader->bc.nlds_dw;
+#endif
+
 
 	/* Calculate group_size/grid_size */
 	for (i = 0; i < 3; i++) {
@@ -520,19 +532,34 @@ void evergreen_emit_cs_shader(
 	struct r600_cs_shader_state *state =
 					(struct r600_cs_shader_state*)atom;
 	struct r600_pipe_compute *shader = state->shader;
-	struct r600_kernel *kernel = &shader->kernels[state->kernel_index];
 	struct radeon_winsys_cs *cs = rctx->b.rings.gfx.cs;
+	uint64_t va;
+	struct r600_resource *code_bo;
+	unsigned ngpr, nstack;
+
+#if HAVE_LLVM < 0x0306
+	struct r600_kernel *kernel = &shader->kernels[state->kernel_index];
+	code_bo = kernel->code_bo;
+	va = kernel->code_bo->gpu_address;
+	ngpr = kernel->bc.ngpr;
+	nstack = kernel->bc.nstack;
+#else
+	code_bo = shader->code_bo;
+	va = shader->code_bo->gpu_address + state->pc;
+	ngpr = shader->bc.ngpr;
+	nstack = shader->bc.nstack;
+#endif
 
 	r600_write_compute_context_reg_seq(cs, R_0288D0_SQ_PGM_START_LS, 3);
-	radeon_emit(cs, kernel->code_bo->gpu_address >> 8); /* R_0288D0_SQ_PGM_START_LS */
+	radeon_emit(cs, va >> 8); /* R_0288D0_SQ_PGM_START_LS */
 	radeon_emit(cs,           /* R_0288D4_SQ_PGM_RESOURCES_LS */
-			S_0288D4_NUM_GPRS(kernel->bc.ngpr)
-			| S_0288D4_STACK_SIZE(kernel->bc.nstack));
+			S_0288D4_NUM_GPRS(ngpr)
+			| S_0288D4_STACK_SIZE(nstack));
 	radeon_emit(cs, 0);	/* R_0288D8_SQ_PGM_RESOURCES_LS_2 */
 
 	radeon_emit(cs, PKT3C(PKT3_NOP, 0, 0));
 	radeon_emit(cs, r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx,
-					      kernel->code_bo, RADEON_USAGE_READ,
+					      code_bo, RADEON_USAGE_READ,
 					      RADEON_PRIO_SHADER_DATA));
 }
 
@@ -542,46 +569,54 @@ static void evergreen_launch_grid(
 		uint32_t pc, const void *input)
 {
 	struct r600_context *ctx = (struct r600_context *)ctx_;
-
 	struct r600_pipe_compute *shader = ctx->cs_shader_state.shader;
-	struct r600_kernel *kernel = &shader->kernels[pc];
-
-	COMPUTE_DBG(ctx->screen, "*** evergreen_launch_grid: pc = %u\n", pc);
+	boolean use_kill;
 
+#if HAVE_LLVM < 0x0306
 #ifdef HAVE_OPENCL
-
-	if (!kernel->code_bo) {
-		void *p;
-		struct r600_bytecode *bc = &kernel->bc;
-		LLVMModuleRef mod = kernel->llvm_module;
-		boolean use_kill = false;
-		bool dump = (ctx->screen->b.debug_flags & DBG_CS) != 0;
-		unsigned use_sb = ctx->screen->b.debug_flags & DBG_SB_CS;
-		unsigned sb_disasm = use_sb ||
-			(ctx->screen->b.debug_flags & DBG_SB_DISASM);
-
-		r600_bytecode_init(bc, ctx->b.chip_class, ctx->b.family,
-			   ctx->screen->has_compressed_msaa_texturing);
-		bc->type = TGSI_PROCESSOR_COMPUTE;
-		bc->isa = ctx->isa;
-		r600_llvm_compile(mod, ctx->b.family, bc, &use_kill, dump);
-
-		if (dump && !sb_disasm) {
-			r600_bytecode_disasm(bc);
-		} else if ((dump && sb_disasm) || use_sb) {
-			if (r600_sb_bytecode_process(ctx, bc, NULL, dump, use_sb))
-				R600_ERR("r600_sb_bytecode_process failed!\n");
-		}
-
-		kernel->code_bo = r600_compute_buffer_alloc_vram(ctx->screen,
-							kernel->bc.ndw * 4);
-		p = r600_buffer_map_sync_with_rings(&ctx->b, kernel->code_bo, PIPE_TRANSFER_WRITE);
-		memcpy(p, kernel->bc.bytecode, kernel->bc.ndw * 4);
-		ctx->b.ws->buffer_unmap(kernel->code_bo->cs_buf);
-	}
+	struct r600_kernel *kernel = &shader->kernels[pc];
+	(void)use_kill;
+        if (!kernel->code_bo) {
+                void *p;
+                struct r600_bytecode *bc = &kernel->bc;
+                LLVMModuleRef mod = kernel->llvm_module;
+                boolean use_kill = false;
+                bool dump = (ctx->screen->b.debug_flags & DBG_CS) != 0;
+                unsigned use_sb = ctx->screen->b.debug_flags & DBG_SB_CS;
+                unsigned sb_disasm = use_sb ||
+                        (ctx->screen->b.debug_flags & DBG_SB_DISASM);
+
+                r600_bytecode_init(bc, ctx->b.chip_class, ctx->b.family,
+                           ctx->screen->has_compressed_msaa_texturing);
+                bc->type = TGSI_PROCESSOR_COMPUTE;
+                bc->isa = ctx->isa;
+                r600_llvm_compile(mod, ctx->b.family, bc, &use_kill, dump);
+
+                if (dump && !sb_disasm) {
+                        r600_bytecode_disasm(bc);
+                } else if ((dump && sb_disasm) || use_sb) {
+                        if (r600_sb_bytecode_process(ctx, bc, NULL, dump, use_sb))
+                                R600_ERR("r600_sb_bytecode_process failed!\n");
+                }
+
+                kernel->code_bo = r600_compute_buffer_alloc_vram(ctx->screen,
+                                                        kernel->bc.ndw * 4);
+                p = r600_buffer_map_sync_with_rings(&ctx->b, kernel->code_bo, PIPE_TRANSFER_WRITE);
+                memcpy(p, kernel->bc.bytecode, kernel->bc.ndw * 4);
+                ctx->b.ws->buffer_unmap(kernel->code_bo->cs_buf);
+        }
 #endif
 	shader->active_kernel = kernel;
 	ctx->cs_shader_state.kernel_index = pc;
+#else
+	ctx->cs_shader_state.pc = pc;
+	/* Get the config information for this kernel. */
+	r600_shader_binary_read_config(&shader->binary, &shader->bc, pc, &use_kill);
+#endif
+
+	COMPUTE_DBG(ctx->screen, "*** evergreen_launch_grid: pc = %u\n", pc);
+
+
 	evergreen_compute_upload_input(ctx_, block_layout, grid_layout, input);
 	compute_emit_cs(ctx, block_layout, grid_layout);
 }
diff --git a/src/gallium/drivers/r600/evergreen_compute_internal.h b/src/gallium/drivers/r600/evergreen_compute_internal.h
index 0929d8dcf27..95593dd8e13 100644
--- a/src/gallium/drivers/r600/evergreen_compute_internal.h
+++ b/src/gallium/drivers/r600/evergreen_compute_internal.h
@@ -27,6 +27,8 @@
 
 #include "r600_asm.h"
 
+#if HAVE_LLVM < 0x0306
+
 struct r600_kernel {
 	unsigned count;
 #ifdef HAVE_OPENCL
@@ -36,13 +38,21 @@ struct r600_kernel {
 	struct r600_bytecode bc;
 };
 
+#endif
+
 struct r600_pipe_compute {
 	struct r600_context *ctx;
 
+#if HAVE_LLVM < 0x0306
 	unsigned num_kernels;
 	struct r600_kernel *kernels;
-
 	struct r600_kernel *active_kernel;
+#endif
+
+	struct radeon_shader_binary binary;
+	struct r600_resource *code_bo;
+	struct r600_bytecode bc;
+
 	unsigned local_size;
 	unsigned private_size;
 	unsigned input_size;
diff --git a/src/gallium/drivers/r600/r600_llvm.c b/src/gallium/drivers/r600/r600_llvm.c
index 766141968a0..c19693a03e6 100644
--- a/src/gallium/drivers/r600/r600_llvm.c
+++ b/src/gallium/drivers/r600/r600_llvm.c
@@ -13,8 +13,9 @@
 #include "r600_opcodes.h"
 #include "r600_shader.h"
 #include "r600_pipe.h"
-#include "radeon/radeon_llvm.h"
-#include "radeon/radeon_llvm_emit.h"
+#include "radeon_llvm.h"
+#include "radeon_llvm_emit.h"
+#include "radeon_elf_util.h"
 
 #include <stdio.h>
 
@@ -818,31 +819,20 @@ LLVMModuleRef r600_tgsi_llvm(
 #define R_028868_SQ_PGM_RESOURCES_VS                 0x028868
 #define R_028850_SQ_PGM_RESOURCES_PS                 0x028850
 
-unsigned r600_llvm_compile(
-	LLVMModuleRef mod,
-	enum radeon_family family,
-	struct r600_bytecode *bc,
-	boolean *use_kill,
-	unsigned dump)
+void r600_shader_binary_read_config(const struct radeon_shader_binary *binary,
+					struct r600_bytecode *bc,
+					uint64_t symbol_offset,
+					boolean *use_kill)
 {
-	unsigned r;
-	struct radeon_shader_binary binary;
-	const char * gpu_family = r600_get_llvm_processor_name(family);
 	unsigned i;
+	const unsigned char *config =
+		radeon_shader_binary_config_start(binary, symbol_offset);
 
-	memset(&binary, 0, sizeof(struct radeon_shader_binary));
-	r = radeon_llvm_compile(mod, &binary, gpu_family, dump);
-
-	assert(binary.code_size % 4 == 0);
-	bc->bytecode = CALLOC(1, binary.code_size);
-	memcpy(bc->bytecode, binary.code, binary.code_size);
-	bc->ndw = binary.code_size / 4;
-
-	for (i = 0; i < binary.config_size; i+= 8) {
+	for (i = 0; i < binary->config_size_per_symbol; i+= 8) {
 		unsigned reg =
-			util_le32_to_cpu(*(uint32_t*)(binary.config + i));
+			util_le32_to_cpu(*(uint32_t*)(config + i));
 		unsigned value =
-			util_le32_to_cpu(*(uint32_t*)(binary.config + i + 4));
+			util_le32_to_cpu(*(uint32_t*)(config + i + 4));
 		switch (reg) {
 		/* R600 / R700 */
 		case R_028850_SQ_PGM_RESOURCES_PS:
@@ -851,8 +841,8 @@ unsigned r600_llvm_compile(
 		case R_028844_SQ_PGM_RESOURCES_PS:
 		case R_028860_SQ_PGM_RESOURCES_VS:
 		case R_0288D4_SQ_PGM_RESOURCES_LS:
-			bc->ngpr = G_028844_NUM_GPRS(value);
-			bc->nstack = G_028844_STACK_SIZE(value);
+			bc->ngpr = MAX2(bc->ngpr, G_028844_NUM_GPRS(value));
+			bc->nstack = MAX2(bc->nstack, G_028844_STACK_SIZE(value));
 			break;
 		case R_02880C_DB_SHADER_CONTROL:
 			*use_kill = G_02880C_KILL_ENABLE(value);
@@ -863,6 +853,39 @@ unsigned r600_llvm_compile(
 		}
 	}
 
+}
+
+unsigned r600_create_shader(struct r600_bytecode *bc,
+		const struct radeon_shader_binary *binary,
+		boolean *use_kill)
+
+{
+	assert(binary->code_size % 4 == 0);
+	bc->bytecode = CALLOC(1, binary->code_size);
+	memcpy(bc->bytecode, binary->code, binary->code_size);
+	bc->ndw = binary->code_size / 4;
+
+	r600_shader_binary_read_config(binary, bc, 0, use_kill);
+
+	return 0;
+}
+
+unsigned r600_llvm_compile(
+	LLVMModuleRef mod,
+	enum radeon_family family,
+	struct r600_bytecode *bc,
+	boolean *use_kill,
+	unsigned dump)
+{
+	unsigned r;
+	struct radeon_shader_binary binary;
+	const char * gpu_family = r600_get_llvm_processor_name(family);
+
+	memset(&binary, 0, sizeof(struct radeon_shader_binary));
+	r = radeon_llvm_compile(mod, &binary, gpu_family, dump);
+
+	r = r600_create_shader(bc, &binary, use_kill);
+
 	FREE(binary.code);
 	FREE(binary.config);
 
diff --git a/src/gallium/drivers/r600/r600_llvm.h b/src/gallium/drivers/r600/r600_llvm.h
index 3840a5a2933..9b5304d9fcb 100644
--- a/src/gallium/drivers/r600/r600_llvm.h
+++ b/src/gallium/drivers/r600/r600_llvm.h
@@ -10,6 +10,7 @@
 struct r600_bytecode;
 struct r600_shader_ctx;
 struct radeon_llvm_context;
+struct radeon_shader_binary;
 enum radeon_family;
 
 LLVMModuleRef r600_tgsi_llvm(
@@ -23,6 +24,15 @@ unsigned r600_llvm_compile(
 	boolean *use_kill,
 	unsigned dump);
 
+unsigned r600_create_shader(struct r600_bytecode *bc,
+		const struct radeon_shader_binary *binary,
+		boolean *use_kill);
+
+void r600_shader_binary_read_config(const struct radeon_shader_binary *binary,
+		struct r600_bytecode *bc,
+		uint64_t symbol_offset,
+		boolean *use_kill);
+
 #endif /* defined R600_USE_LLVM || defined HAVE_OPENCL */
 
 #endif /* R600_LLVM_H */
diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
index c86daa6c249..0b571e45e9b 100644
--- a/src/gallium/drivers/r600/r600_pipe.c
+++ b/src/gallium/drivers/r600/r600_pipe.c
@@ -472,7 +472,11 @@ static int r600_get_shader_param(struct pipe_screen* pscreen, unsigned shader, e
 		return 16;
         case PIPE_SHADER_CAP_PREFERRED_IR:
 		if (shader == PIPE_SHADER_COMPUTE) {
+#if HAVE_LLVM < 0x0306
 			return PIPE_SHADER_IR_LLVM;
+#else
+			return PIPE_SHADER_IR_NATIVE;
+#endif
 		} else {
 			return PIPE_SHADER_IR_TGSI;
 		}
diff --git a/src/gallium/drivers/r600/r600_pipe.h b/src/gallium/drivers/r600/r600_pipe.h
index fa9d34b0d71..40b0328ea20 100644
--- a/src/gallium/drivers/r600/r600_pipe.h
+++ b/src/gallium/drivers/r600/r600_pipe.h
@@ -146,6 +146,7 @@ struct r600_clip_state {
 struct r600_cs_shader_state {
 	struct r600_atom atom;
 	unsigned kernel_index;
+	unsigned pc;
 	struct r600_pipe_compute *shader;
 };
 
diff --git a/src/gallium/drivers/radeon/radeon_llvm_util.c b/src/gallium/drivers/radeon/radeon_llvm_util.c
index ec1155923fe..0dfd9ad4867 100644
--- a/src/gallium/drivers/radeon/radeon_llvm_util.c
+++ b/src/gallium/drivers/radeon/radeon_llvm_util.c
@@ -34,7 +34,7 @@
 #include <llvm-c/Transforms/PassManagerBuilder.h>
 
 LLVMModuleRef radeon_llvm_parse_bitcode(LLVMContextRef ctx,
-							const unsigned char * bitcode, unsigned bitcode_len)
+							const char * bitcode, unsigned bitcode_len)
 {
 	LLVMMemoryBufferRef buf;
 	LLVMModuleRef module;
@@ -47,7 +47,7 @@ LLVMModuleRef radeon_llvm_parse_bitcode(LLVMContextRef ctx,
 }
 
 unsigned radeon_llvm_get_num_kernels(LLVMContextRef ctx,
-				const unsigned char *bitcode, unsigned bitcode_len)
+				const char *bitcode, unsigned bitcode_len)
 {
 	LLVMModuleRef mod = radeon_llvm_parse_bitcode(ctx, bitcode, bitcode_len);
 	return LLVMGetNamedMetadataNumOperands(mod, "opencl.kernels");
@@ -88,7 +88,7 @@ static void radeon_llvm_optimize(LLVMModuleRef mod)
 }
 
 LLVMModuleRef radeon_llvm_get_kernel_module(LLVMContextRef ctx, unsigned index,
-		const unsigned char *bitcode, unsigned bitcode_len)
+		const char *bitcode, unsigned bitcode_len)
 {
 	LLVMModuleRef mod;
 	unsigned num_kernels;
diff --git a/src/gallium/drivers/radeon/radeon_llvm_util.h b/src/gallium/drivers/radeon/radeon_llvm_util.h
index 733c329e99e..cc1932aef47 100644
--- a/src/gallium/drivers/radeon/radeon_llvm_util.h
+++ b/src/gallium/drivers/radeon/radeon_llvm_util.h
@@ -30,10 +30,10 @@
 #include <llvm-c/Core.h>
 
 LLVMModuleRef radeon_llvm_parse_bitcode(LLVMContextRef ctx,
-			const unsigned char * bitcode, unsigned bitcode_len);
+			const char * bitcode, unsigned bitcode_len);
 unsigned radeon_llvm_get_num_kernels(LLVMContextRef ctx,
-			const unsigned char *bitcode, unsigned bitcode_len);
+			const char *bitcode, unsigned bitcode_len);
 LLVMModuleRef radeon_llvm_get_kernel_module(LLVMContextRef ctx, unsigned index,
-			const unsigned char *bitcode, unsigned bitcode_len);
+			const char *bitcode, unsigned bitcode_len);
 
 #endif