X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fr600%2Fevergreen_compute.c;h=292b5e32afd3675d8c5d7580d227ce65e14b9594;hb=1da538f85bc327f4ae5e1a5b90c15b99f8cf48df;hp=3bb3895bf7348b5f73341dc3eb2a11a9e2b6b373;hpb=92af184690995d3b16731518f7becfaac3538edb;p=mesa.git

diff --git a/src/gallium/drivers/r600/evergreen_compute.c b/src/gallium/drivers/r600/evergreen_compute.c
index 3bb3895bf73..292b5e32afd 100644
--- a/src/gallium/drivers/r600/evergreen_compute.c
+++ b/src/gallium/drivers/r600/evergreen_compute.c
@@ -30,7 +30,7 @@
 #include "pipe/p_state.h"
 #include "pipe/p_context.h"
 #include "util/u_blitter.h"
-#include "util/u_double_list.h"
+#include "util/list.h"
 #include "util/u_transfer.h"
 #include "util/u_surface.h"
 #include "util/u_pack_color.h"
@@ -38,19 +38,16 @@
 #include "util/u_inlines.h"
 #include "util/u_framebuffer.h"
 #include "pipebuffer/pb_buffer.h"
-#include "r600.h"
 #include "evergreend.h"
-#include "r600_resource.h"
 #include "r600_shader.h"
 #include "r600_pipe.h"
 #include "r600_formats.h"
 #include "evergreen_compute.h"
-#include "r600_hw_context_priv.h"
 #include "evergreen_compute_internal.h"
 #include "compute_memory_pool.h"
-#ifdef HAVE_OPENCL
-#include "llvm_wrapper.h"
-#endif
+#include "sb/sb_public.h"
+#include "radeon/radeon_elf_util.h"
+#include <inttypes.h>
 
 /**
 RAT0 is for global binding write
@@ -83,127 +80,228 @@ writable images will consume TEX slots, VTX slots too because of linear indexing
 
 */
 
-const struct u_resource_vtbl r600_global_buffer_vtbl =
+struct r600_resource *r600_compute_buffer_alloc_vram(struct r600_screen *screen,
+						     unsigned size)
 {
-	u_default_resource_get_handle, /* get_handle */
-	r600_compute_global_buffer_destroy, /* resource_destroy */
-	r600_compute_global_get_transfer, /* get_transfer */
-	r600_compute_global_transfer_destroy, /* transfer_destroy */
-	r600_compute_global_transfer_map, /* transfer_map */
-	r600_compute_global_transfer_flush_region,/* transfer_flush_region */
-	r600_compute_global_transfer_unmap, /* transfer_unmap */
-	r600_compute_global_transfer_inline_write /* transfer_inline_write */
-};
+	struct pipe_resource *buffer = NULL;
+	assert(size);
 
+	buffer = pipe_buffer_create((struct pipe_screen*) screen,
+				    PIPE_BIND_CUSTOM,
+				    PIPE_USAGE_IMMUTABLE,
+				    size);
 
-void *evergreen_create_compute_state(
-	struct pipe_context *ctx_,
-	const const struct pipe_compute_state *cso)
+	return (struct r600_resource *)buffer;
+}
+
+
+static void evergreen_set_rat(struct r600_pipe_compute *pipe,
+			      unsigned id,
+			      struct r600_resource *bo,
+			      int start,
+			      int size)
 {
-	struct r600_context *ctx = (struct r600_context *)ctx_;
+	struct pipe_surface rat_templ;
+	struct r600_surface *surf = NULL;
+	struct r600_context *rctx = NULL;
+
+	assert(id < 12);
+	assert((size & 3) == 0);
+	assert((start & 0xFF) == 0);
+
+	rctx = pipe->ctx;
+
+	COMPUTE_DBG(rctx->screen, "bind rat: %i \n", id);
+
+	/* Create the RAT surface */
+	memset(&rat_templ, 0, sizeof(rat_templ));
+	rat_templ.format = PIPE_FORMAT_R32_UINT;
+	rat_templ.u.tex.level = 0;
+	rat_templ.u.tex.first_layer = 0;
+	rat_templ.u.tex.last_layer = 0;
+
+	/* Add the RAT the list of color buffers */
+	pipe->ctx->framebuffer.state.cbufs[id] = pipe->ctx->b.b.create_surface(
+		(struct pipe_context *)pipe->ctx,
+		(struct pipe_resource *)bo, &rat_templ);
+
+	/* Update the number of color buffers */
+	pipe->ctx->framebuffer.state.nr_cbufs =
+		MAX2(id + 1, pipe->ctx->framebuffer.state.nr_cbufs);
+
+	/* Update the cb_target_mask
+	 * XXX: I think this is a potential spot for bugs once we start doing
+	 * GL interop.  cb_target_mask may be modified in the 3D sections
+	 * of this driver. */
+	pipe->ctx->compute_cb_target_mask |= (0xf << (id * 4));
+
+	surf = (struct r600_surface*)pipe->ctx->framebuffer.state.cbufs[id];
+	evergreen_init_color_surface_rat(rctx, surf);
+}
 
-#ifdef HAVE_OPENCL
-	const struct pipe_llvm_program_header * header;
-	const unsigned char * code;
+static void evergreen_cs_set_vertex_buffer(struct r600_context *rctx,
+					   unsigned vb_index,
+					   unsigned offset,
+					   struct pipe_resource *buffer)
+{
+	struct r600_vertexbuf_state *state = &rctx->cs_vertex_buffer_state;
+	struct pipe_vertex_buffer *vb = &state->vb[vb_index];
+	vb->stride = 1;
+	vb->buffer_offset = offset;
+	vb->buffer = buffer;
+	vb->user_buffer = NULL;
+
+	/* The vertex instructions in the compute shaders use the texture cache,
+	 * so we need to invalidate it. */
+	rctx->b.flags |= R600_CONTEXT_INV_VERTEX_CACHE;
+	state->enabled_mask |= 1 << vb_index;
+	state->dirty_mask |= 1 << vb_index;
+	r600_mark_atom_dirty(rctx, &state->atom);
+}
 
-	header = cso->prog;
-	code = cso->prog + sizeof(struct pipe_llvm_program_header);
-#endif
+static void evergreen_cs_set_constant_buffer(struct r600_context *rctx,
+					     unsigned cb_index,
+					     unsigned offset,
+					     unsigned size,
+					     struct pipe_resource *buffer)
+{
+	struct pipe_constant_buffer cb;
+	cb.buffer_size = size;
+	cb.buffer_offset = offset;
+	cb.buffer = buffer;
+	cb.user_buffer = NULL;
 
-	if (!ctx->screen->screen.get_param(&ctx->screen->screen,
-							PIPE_CAP_COMPUTE)) {
-		fprintf(stderr, "Compute is not supported\n");
-		return NULL;
-	}
-	struct r600_pipe_compute *shader =	CALLOC_STRUCT(r600_pipe_compute);
+	rctx->b.b.set_constant_buffer(&rctx->b.b, PIPE_SHADER_COMPUTE, cb_index, &cb);
+}
 
-	shader->ctx = (struct r600_context*)ctx;
-	shader->resources = (struct evergreen_compute_resource*)
-			CALLOC(sizeof(struct evergreen_compute_resource),
-			get_compute_resource_num());
-	shader->local_size = cso->req_local_mem; ///TODO: assert it
-	shader->private_size = cso->req_private_mem;
-	shader->input_size = cso->req_input_mem;
+/* We need to define these R600 registers here, because we can't include
+ * evergreend.h and r600d.h.
+ */
+#define R_028868_SQ_PGM_RESOURCES_VS                 0x028868
+#define R_028850_SQ_PGM_RESOURCES_PS                 0x028850
 
-#ifdef HAVE_OPENCL 
-	shader->mod = llvm_parse_bitcode(code, header->num_bytes);
+#ifdef HAVE_OPENCL
 
-	r600_compute_shader_create(ctx_, shader->mod, &shader->bc);
-#endif
-	return shader;
+static void r600_shader_binary_read_config(const struct radeon_shader_binary *binary,
+					   struct r600_bytecode *bc,
+					   uint64_t symbol_offset,
+					   boolean *use_kill)
+{
+       unsigned i;
+       const unsigned char *config =
+               radeon_shader_binary_config_start(binary, symbol_offset);
+
+       for (i = 0; i < binary->config_size_per_symbol; i+= 8) {
+               unsigned reg =
+                       util_le32_to_cpu(*(uint32_t*)(config + i));
+               unsigned value =
+                       util_le32_to_cpu(*(uint32_t*)(config + i + 4));
+               switch (reg) {
+               /* R600 / R700 */
+               case R_028850_SQ_PGM_RESOURCES_PS:
+               case R_028868_SQ_PGM_RESOURCES_VS:
+               /* Evergreen / Northern Islands */
+               case R_028844_SQ_PGM_RESOURCES_PS:
+               case R_028860_SQ_PGM_RESOURCES_VS:
+               case R_0288D4_SQ_PGM_RESOURCES_LS:
+                       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);
+                       break;
+               case R_0288E8_SQ_LDS_ALLOC:
+                       bc->nlds_dw = value;
+                       break;
+               }
+       }
 }
 
-void evergreen_delete_compute_state(struct pipe_context *ctx, void* state)
+static unsigned r600_create_shader(struct r600_bytecode *bc,
+				   const struct radeon_shader_binary *binary,
+				   boolean *use_kill)
+
 {
-	struct r600_pipe_compute *shader = (struct r600_pipe_compute *)state;
+	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;
 
-	free(shader->resources);
-	free(shader);
+	r600_shader_binary_read_config(binary, bc, 0, use_kill);
+	return 0;
 }
 
-static void evergreen_bind_compute_state(struct pipe_context *ctx_, void *state)
+#endif
+
+static void r600_destroy_shader(struct r600_bytecode *bc)
 {
-	struct r600_context *ctx = (struct r600_context *)ctx_;
+	FREE(bc->bytecode);
+}
 
-	ctx->cs_shader = (struct r600_pipe_compute *)state;
+static void *evergreen_create_compute_state(struct pipe_context *ctx,
+					    const struct pipe_compute_state *cso)
+{
+	struct r600_context *rctx = (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 char *code;
+	void *p;
+	boolean use_kill;
 
-	assert(!ctx->cs_shader->shader_code_bo);
+	COMPUTE_DBG(rctx->screen, "*** evergreen_create_compute_state\n");
+	header = cso->prog;
+	code = cso->prog + sizeof(struct pipe_llvm_program_header);
+	radeon_shader_binary_init(&shader->binary);
+	radeon_elf_read(code, header->num_bytes, &shader->binary);
+	r600_create_shader(&shader->bc, &shader->binary, &use_kill);
+
+	/* Upload code + ROdata */
+	shader->code_bo = r600_compute_buffer_alloc_vram(rctx->screen,
+							shader->bc.ndw * 4);
+	p = r600_buffer_map_sync_with_rings(&rctx->b, shader->code_bo, PIPE_TRANSFER_WRITE);
+	//TODO: use util_memcpy_cpu_to_le32 ?
+	memcpy(p, shader->bc.bytecode, shader->bc.ndw * 4);
+	rctx->b.ws->buffer_unmap(shader->code_bo->buf);
+#endif
 
-	ctx->cs_shader->shader_code_bo =
-		r600_compute_buffer_alloc_vram(ctx->screen,
-					ctx->cs_shader->bc.ndw * 4);
+	shader->ctx = rctx;
+	shader->local_size = cso->req_local_mem;
+	shader->private_size = cso->req_private_mem;
+	shader->input_size = cso->req_input_mem;
 
-	void *p = ctx->ws->buffer_map(ctx->cs_shader->shader_code_bo->cs_buf,
-					ctx->cs, PIPE_TRANSFER_WRITE);
+	return shader;
+}
 
-	memcpy(p, ctx->cs_shader->bc.bytecode, ctx->cs_shader->bc.ndw * 4);
+static void evergreen_delete_compute_state(struct pipe_context *ctx, void *state)
+{
+	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct r600_pipe_compute *shader = state;
 
-	ctx->ws->buffer_unmap(ctx->cs_shader->shader_code_bo->cs_buf);
+	COMPUTE_DBG(rctx->screen, "*** evergreen_delete_compute_state\n");
 
-	evergreen_compute_init_config(ctx);
+	if (!shader)
+		return;
 
-	struct evergreen_compute_resource* res = get_empty_res(ctx->cs_shader,
-						COMPUTE_RESOURCE_SHADER, 0);
+	radeon_shader_binary_clean(&shader->binary);
+	r600_destroy_shader(&shader->bc);
 
-	if (ctx->chip_class < CAYMAN) {
-		evergreen_reg_set(res, R_008C0C_SQ_GPR_RESOURCE_MGMT_3,
-			S_008C0C_NUM_LS_GPRS(ctx->cs_shader->bc.ngpr));
-	}
+	/* TODO destroy shader->code_bo, shader->const_bo
+	 * we'll need something like r600_buffer_free */
+	FREE(shader);
+}
 
-	///maybe we can use it later
-	evergreen_reg_set(res, R_0286C8_SPI_THREAD_GROUPING, 0);
-	///maybe we can use it later
-	evergreen_reg_set(res, R_008C14_SQ_GLOBAL_GPR_RESOURCE_MGMT_2, 0);
-
-	evergreen_reg_set(res, R_0288D4_SQ_PGM_RESOURCES_LS,
-		S_0288D4_NUM_GPRS(ctx->cs_shader->bc.ngpr)
-		| S_0288D4_STACK_SIZE(ctx->cs_shader->bc.nstack));
-	evergreen_reg_set(res, R_0288D8_SQ_PGM_RESOURCES_LS_2, 0);
-
-	evergreen_reg_set(res, R_0288D0_SQ_PGM_START_LS, 0);
-	res->bo = ctx->cs_shader->shader_code_bo;
-	res->usage = RADEON_USAGE_READ;
-	res->coher_bo_size = ctx->cs_shader->bc.ndw*4;
-	res->flags = COMPUTE_RES_SH_FLUSH;
-
-	/* We can't always determine the
-	 * number of iterations in a loop before it's executed,
-	 * so we just need to set up the loop counter to give us the maximum
-	 * number of iterations possible.  Currently, loops in shader code
-	 * ignore the loop counter and use a break instruction to exit the
-	 * loop at the correct time.
-	 */
-	evergreen_set_loop_const(ctx->cs_shader,
-		0, /* index */
-		0xFFF, /* Maximum value of the loop counter (i.e. when the loop
-			* counter reaches this value, the program will break
-			* out of the loop. */
-		0x0,   /* Starting value of the loop counter. */
-		0x1);  /* Amount to increment the loop counter each iteration. */
+static void evergreen_bind_compute_state(struct pipe_context *ctx, void *state)
+{
+	struct r600_context *rctx = (struct r600_context *)ctx;
+
+	COMPUTE_DBG(rctx->screen, "*** evergreen_bind_compute_state\n");
+
+	rctx->cs_shader_state.shader = (struct r600_pipe_compute *)state;
 }
 
 /* The kernel parameters are stored a vtx buffer (ID=0), besides the explicit
- * kernel parameters there are inplicit parameters that need to be stored
+ * kernel parameters there are implicit parameters that need to be stored
  * in the vertex buffer as well.  Here is how these parameters are organized in
  * the buffer:
  *
@@ -213,420 +311,549 @@ static void evergreen_bind_compute_state(struct pipe_context *ctx_, void *state)
  *             (x,y,z)
  * DWORDS 9+ : Kernel parameters
  */
-void evergreen_compute_upload_input(
-	struct pipe_context *ctx_,
-	const uint *block_layout,
-	const uint *grid_layout,
-	const void *input)
+static void evergreen_compute_upload_input(struct pipe_context *ctx,
+					   const struct pipe_grid_info *info)
 {
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	int i;
-	unsigned kernel_parameters_offset_bytes = 36;
-	uint32_t * num_work_groups_start;
-	uint32_t * global_size_start;
-	uint32_t * local_size_start;
-	uint32_t * kernel_parameters_start;
-
-	if (ctx->cs_shader->input_size == 0) {
+	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct r600_pipe_compute *shader = rctx->cs_shader_state.shader;
+	unsigned i;
+	/* We need to reserve 9 dwords (36 bytes) for implicit kernel
+	 * parameters.
+	 */
+	unsigned input_size = shader->input_size + 36;
+	uint32_t *num_work_groups_start;
+	uint32_t *global_size_start;
+	uint32_t *local_size_start;
+	uint32_t *kernel_parameters_start;
+	struct pipe_box box;
+	struct pipe_transfer *transfer = NULL;
+
+	if (shader->input_size == 0) {
 		return;
 	}
 
-	if (!ctx->cs_shader->kernel_param) {
-		unsigned buffer_size = ctx->cs_shader->input_size;
-
+	if (!shader->kernel_param) {
 		/* Add space for the grid dimensions */
-		buffer_size += kernel_parameters_offset_bytes * sizeof(uint);
-		ctx->cs_shader->kernel_param =
-				r600_compute_buffer_alloc_vram(ctx->screen,
-						buffer_size);
+		shader->kernel_param = (struct r600_resource *)
+			pipe_buffer_create(ctx->screen, PIPE_BIND_CUSTOM,
+					PIPE_USAGE_IMMUTABLE, input_size);
 	}
 
-	num_work_groups_start = ctx->ws->buffer_map(
-			ctx->cs_shader->kernel_param->cs_buf,
-			ctx->cs, PIPE_TRANSFER_WRITE);
+	u_box_1d(0, input_size, &box);
+	num_work_groups_start = ctx->transfer_map(ctx,
+			(struct pipe_resource*)shader->kernel_param,
+			0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE,
+			&box, &transfer);
 	global_size_start = num_work_groups_start + (3 * (sizeof(uint) /4));
 	local_size_start = global_size_start + (3 * (sizeof(uint)) / 4);
 	kernel_parameters_start = local_size_start + (3 * (sizeof(uint)) / 4);
 
 	/* Copy the work group size */
-	memcpy(num_work_groups_start, grid_layout, 3 * sizeof(uint));
+	memcpy(num_work_groups_start, info->grid, 3 * sizeof(uint));
 
 	/* Copy the global size */
 	for (i = 0; i < 3; i++) {
-		global_size_start[i] = grid_layout[i] * block_layout[i];
+		global_size_start[i] = info->grid[i] * info->block[i];
 	}
 
 	/* Copy the local dimensions */
-	memcpy(local_size_start, block_layout, 3 * sizeof(uint));
+	memcpy(local_size_start, info->block, 3 * sizeof(uint));
 
 	/* Copy the kernel inputs */
-	memcpy(kernel_parameters_start, input, ctx->cs_shader->input_size);
+	memcpy(kernel_parameters_start, info->input, shader->input_size);
 
-	for (i = 0; i < (kernel_parameters_offset_bytes / 4) +
-					(ctx->cs_shader->input_size / 4); i++) {
-		COMPUTE_DBG("input %i : %i\n", i,
+	for (i = 0; i < (input_size / 4); i++) {
+		COMPUTE_DBG(rctx->screen, "input %i : %u\n", i,
 			((unsigned*)num_work_groups_start)[i]);
 	}
 
-	ctx->ws->buffer_unmap(ctx->cs_shader->kernel_param->cs_buf);
+	ctx->transfer_unmap(ctx, transfer);
 
-	///ID=0 is reserved for the parameters
-	evergreen_set_vtx_resource(ctx->cs_shader,
-		ctx->cs_shader->kernel_param, 0, 0, 0);
-	///ID=0 is reserved for parameters
-	evergreen_set_const_cache(ctx->cs_shader, 0,
-		ctx->cs_shader->kernel_param, ctx->cs_shader->input_size, 0);
+	/* ID=0 and ID=3 are reserved for the parameters.
+	 * LLVM will preferably use ID=0, but it does not work for dynamic
+	 * indices. */
+	evergreen_cs_set_vertex_buffer(rctx, 3, 0,
+			(struct pipe_resource*)shader->kernel_param);
+	evergreen_cs_set_constant_buffer(rctx, 0, 0, input_size,
+			(struct pipe_resource*)shader->kernel_param);
 }
 
-void evergreen_direct_dispatch(
-		struct pipe_context *ctx_,
-		const uint *block_layout, const uint *grid_layout)
+static void evergreen_emit_dispatch(struct r600_context *rctx,
+				    const struct pipe_grid_info *info)
 {
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-
 	int i;
+	struct radeon_winsys_cs *cs = rctx->b.gfx.cs;
+	struct r600_pipe_compute *shader = rctx->cs_shader_state.shader;
+	unsigned num_waves;
+	unsigned num_pipes = rctx->screen->b.info.r600_max_quad_pipes;
+	unsigned wave_divisor = (16 * num_pipes);
+	int group_size = 1;
+	int grid_size = 1;
+	unsigned lds_size = shader->local_size / 4 +
+		shader->bc.nlds_dw;
 
-	struct evergreen_compute_resource* res = get_empty_res(ctx->cs_shader,
-		COMPUTE_RESOURCE_DISPATCH, 0);
 
-	evergreen_reg_set(res, R_008958_VGT_PRIMITIVE_TYPE, V_008958_DI_PT_POINTLIST);
+	/* Calculate group_size/grid_size */
+	for (i = 0; i < 3; i++) {
+		group_size *= info->block[i];
+	}
 
-	evergreen_reg_set(res, R_00899C_VGT_COMPUTE_START_X, 0);
-	evergreen_reg_set(res, R_0089A0_VGT_COMPUTE_START_Y, 0);
-	evergreen_reg_set(res, R_0089A4_VGT_COMPUTE_START_Z, 0);
+	for (i = 0; i < 3; i++)	{
+		grid_size *= info->grid[i];
+	}
 
-	evergreen_reg_set(res, R_0286EC_SPI_COMPUTE_NUM_THREAD_X, block_layout[0]);
-	evergreen_reg_set(res, R_0286F0_SPI_COMPUTE_NUM_THREAD_Y, block_layout[1]);
-	evergreen_reg_set(res, R_0286F4_SPI_COMPUTE_NUM_THREAD_Z, block_layout[2]);
+	/* num_waves = ceil((tg_size.x * tg_size.y, tg_size.z) / (16 * num_pipes)) */
+	num_waves = (info->block[0] * info->block[1] * info->block[2] +
+			wave_divisor - 1) / wave_divisor;
 
-	int group_size = 1;
+	COMPUTE_DBG(rctx->screen, "Using %u pipes, "
+				"%u wavefronts per thread block, "
+				"allocating %u dwords lds.\n",
+				num_pipes, num_waves, lds_size);
 
-	int grid_size = 1;
+	radeon_set_config_reg(cs, R_008970_VGT_NUM_INDICES, group_size);
 
-	for (i = 0; i < 3; i++) {
-		group_size *= block_layout[i];
-	}
+	radeon_set_config_reg_seq(cs, R_00899C_VGT_COMPUTE_START_X, 3);
+	radeon_emit(cs, 0); /* R_00899C_VGT_COMPUTE_START_X */
+	radeon_emit(cs, 0); /* R_0089A0_VGT_COMPUTE_START_Y */
+	radeon_emit(cs, 0); /* R_0089A4_VGT_COMPUTE_START_Z */
 
-	for (i = 0; i < 3; i++)	{
-		grid_size *= grid_layout[i];
+	radeon_set_config_reg(cs, R_0089AC_VGT_COMPUTE_THREAD_GROUP_SIZE,
+								group_size);
+
+	radeon_compute_set_context_reg_seq(cs, R_0286EC_SPI_COMPUTE_NUM_THREAD_X, 3);
+	radeon_emit(cs, info->block[0]); /* R_0286EC_SPI_COMPUTE_NUM_THREAD_X */
+	radeon_emit(cs, info->block[1]); /* R_0286F0_SPI_COMPUTE_NUM_THREAD_Y */
+	radeon_emit(cs, info->block[2]); /* R_0286F4_SPI_COMPUTE_NUM_THREAD_Z */
+
+	if (rctx->b.chip_class < CAYMAN) {
+		assert(lds_size <= 8192);
+	} else {
+		/* Cayman appears to have a slightly smaller limit, see the
+		 * value of CM_R_0286FC_SPI_LDS_MGMT.NUM_LS_LDS */
+		assert(lds_size <= 8160);
 	}
 
-	evergreen_reg_set(res, R_008970_VGT_NUM_INDICES, group_size);
-	evergreen_reg_set(res, R_0089AC_VGT_COMPUTE_THREAD_GROUP_SIZE, group_size);
+	radeon_compute_set_context_reg(cs, R_0288E8_SQ_LDS_ALLOC,
+					lds_size | (num_waves << 14));
 
-	evergreen_emit_raw_value(res, PKT3C(PKT3_DISPATCH_DIRECT, 3, 0));
-	evergreen_emit_raw_value(res, grid_layout[0]);
-	evergreen_emit_raw_value(res, grid_layout[1]);
-	evergreen_emit_raw_value(res, grid_layout[2]);
-	///VGT_DISPATCH_INITIATOR = COMPUTE_SHADER_EN
-	evergreen_emit_raw_value(res, 1);
+	/* Dispatch packet */
+	radeon_emit(cs, PKT3C(PKT3_DISPATCH_DIRECT, 3, 0));
+	radeon_emit(cs, info->grid[0]);
+	radeon_emit(cs, info->grid[1]);
+	radeon_emit(cs, info->grid[2]);
+	/* VGT_DISPATCH_INITIATOR = COMPUTE_SHADER_EN */
+	radeon_emit(cs, 1);
 }
 
-static void compute_emit_cs(struct r600_context *ctx)
+static void compute_emit_cs(struct r600_context *rctx,
+			    const struct pipe_grid_info *info)
 {
-	struct radeon_winsys_cs *cs = ctx->cs;
-	int i;
+	struct radeon_winsys_cs *cs = rctx->b.gfx.cs;
+	unsigned i;
 
-	r600_emit_atom(ctx, &ctx->start_cs_cmd.atom);
+	/* make sure that the gfx ring is only one active */
+	if (radeon_emitted(rctx->b.dma.cs, 0)) {
+		rctx->b.dma.flush(rctx, RADEON_FLUSH_ASYNC, NULL);
+	}
 
-	struct r600_resource *onebo = NULL;
+	/* Initialize all the compute-related registers.
+	 *
+	 * See evergreen_init_atom_start_compute_cs() in this file for the list
+	 * of registers initialized by the start_compute_cs_cmd atom.
+	 */
+	r600_emit_command_buffer(cs, &rctx->start_compute_cs_cmd);
+
+	/* emit config state */
+	if (rctx->b.chip_class == EVERGREEN)
+		r600_emit_atom(rctx, &rctx->config_state.atom);
+
+	rctx->b.flags |= R600_CONTEXT_WAIT_3D_IDLE | R600_CONTEXT_FLUSH_AND_INV;
+	r600_flush_emit(rctx);
+
+	/* Emit colorbuffers. */
+	/* XXX support more than 8 colorbuffers (the offsets are not a multiple of 0x3C for CB8-11) */
+	for (i = 0; i < 8 && i < rctx->framebuffer.state.nr_cbufs; i++) {
+		struct r600_surface *cb = (struct r600_surface*)rctx->framebuffer.state.cbufs[i];
+		unsigned reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx,
+						       (struct r600_resource*)cb->base.texture,
+						       RADEON_USAGE_READWRITE,
+						       RADEON_PRIO_SHADER_RW_BUFFER);
+
+		radeon_compute_set_context_reg_seq(cs, R_028C60_CB_COLOR0_BASE + i * 0x3C, 7);
+		radeon_emit(cs, cb->cb_color_base);	/* R_028C60_CB_COLOR0_BASE */
+		radeon_emit(cs, cb->cb_color_pitch);	/* R_028C64_CB_COLOR0_PITCH */
+		radeon_emit(cs, cb->cb_color_slice);	/* R_028C68_CB_COLOR0_SLICE */
+		radeon_emit(cs, cb->cb_color_view);	/* R_028C6C_CB_COLOR0_VIEW */
+		radeon_emit(cs, cb->cb_color_info);	/* R_028C70_CB_COLOR0_INFO */
+		radeon_emit(cs, cb->cb_color_attrib);	/* R_028C74_CB_COLOR0_ATTRIB */
+		radeon_emit(cs, cb->cb_color_dim);		/* R_028C78_CB_COLOR0_DIM */
+
+		radeon_emit(cs, PKT3(PKT3_NOP, 0, 0)); /* R_028C60_CB_COLOR0_BASE */
+		radeon_emit(cs, reloc);
+
+		radeon_emit(cs, PKT3(PKT3_NOP, 0, 0)); /* R_028C74_CB_COLOR0_ATTRIB */
+		radeon_emit(cs, reloc);
+	}
+	for (; i < 8 ; i++)
+		radeon_compute_set_context_reg(cs, R_028C70_CB_COLOR0_INFO + i * 0x3C,
+					       S_028C70_FORMAT(V_028C70_COLOR_INVALID));
+	for (; i < 12; i++)
+		radeon_compute_set_context_reg(cs, R_028E50_CB_COLOR8_INFO + (i - 8) * 0x1C,
+					       S_028C70_FORMAT(V_028C70_COLOR_INVALID));
 
-	for (i = 0; i < get_compute_resource_num(); i++) {
-		if (ctx->cs_shader->resources[i].enabled) {
-			int j;
-			COMPUTE_DBG("resnum: %i, cdw: %i\n", i, cs->cdw);
+	/* Set CB_TARGET_MASK  XXX: Use cb_misc_state */
+	radeon_compute_set_context_reg(cs, R_028238_CB_TARGET_MASK,
+					rctx->compute_cb_target_mask);
 
-			for (j = 0; j < ctx->cs_shader->resources[i].cs_end; j++) {
-				if (ctx->cs_shader->resources[i].do_reloc[j]) {
-					assert(ctx->cs_shader->resources[i].bo);
-					evergreen_emit_ctx_reloc(ctx,
-						ctx->cs_shader->resources[i].bo,
-						ctx->cs_shader->resources[i].usage);
-				}
 
-				cs->buf[cs->cdw++] = ctx->cs_shader->resources[i].cs[j];
-			}
+	/* Emit vertex buffer state */
+	rctx->cs_vertex_buffer_state.atom.num_dw = 12 * util_bitcount(rctx->cs_vertex_buffer_state.dirty_mask);
+	r600_emit_atom(rctx, &rctx->cs_vertex_buffer_state.atom);
 
-			if (ctx->cs_shader->resources[i].bo) {
-				onebo = ctx->cs_shader->resources[i].bo;
-				evergreen_emit_ctx_reloc(ctx,
-					ctx->cs_shader->resources[i].bo,
-					ctx->cs_shader->resources[i].usage);
-
-				///special case for textures
-				if (ctx->cs_shader->resources[i].do_reloc
-					[ctx->cs_shader->resources[i].cs_end] == 2) {
-					evergreen_emit_ctx_reloc(ctx,
-						ctx->cs_shader->resources[i].bo,
-						ctx->cs_shader->resources[i].usage);
-				}
-
-				evergreen_set_buffer_sync(ctx, ctx->cs_shader->resources[i].bo,
-					ctx->cs_shader->resources[i].coher_bo_size,
-					ctx->cs_shader->resources[i].flags,
-					ctx->cs_shader->resources[i].usage);
-			}
-		}
-	}
+	/* Emit constant buffer state */
+	r600_emit_atom(rctx, &rctx->constbuf_state[PIPE_SHADER_COMPUTE].atom);
 
-#if 0
-	COMPUTE_DBG("cdw: %i\n", cs->cdw);
-	for (i = 0; i < cs->cdw; i++) {
-		COMPUTE_DBG("%4i : 0x%08X\n", i, ctx->cs->buf[i]);
-	}
-#endif
+	/* Emit sampler state */
+	r600_emit_atom(rctx, &rctx->samplers[PIPE_SHADER_COMPUTE].states.atom);
 
-	ctx->ws->cs_flush(ctx->cs, RADEON_FLUSH_ASYNC | RADEON_FLUSH_COMPUTE);
+	/* Emit sampler view (texture resource) state */
+	r600_emit_atom(rctx, &rctx->samplers[PIPE_SHADER_COMPUTE].views.atom);
 
-	ctx->pm4_dirty_cdwords = 0;
-	ctx->flags = 0;
+	/* Emit compute shader state */
+	r600_emit_atom(rctx, &rctx->cs_shader_state.atom);
 
-	COMPUTE_DBG("shader started\n");
+	/* Emit dispatch state and dispatch packet */
+	evergreen_emit_dispatch(rctx, info);
 
-	ctx->ws->buffer_wait(onebo->buf, 0);
+	/* XXX evergreen_flush_emit() hardcodes the CP_COHER_SIZE to 0xffffffff
+	 */
+	rctx->b.flags |= R600_CONTEXT_INV_CONST_CACHE |
+		      R600_CONTEXT_INV_VERTEX_CACHE |
+	              R600_CONTEXT_INV_TEX_CACHE;
+	r600_flush_emit(rctx);
+	rctx->b.flags = 0;
+
+	if (rctx->b.chip_class >= CAYMAN) {
+		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
+		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
+		/* DEALLOC_STATE prevents the GPU from hanging when a
+		 * SURFACE_SYNC packet is emitted some time after a DISPATCH_DIRECT
+		 * with any of the CB*_DEST_BASE_ENA or DB_DEST_BASE_ENA bits set.
+		 */
+		radeon_emit(cs, PKT3C(PKT3_DEALLOC_STATE, 0, 0));
+		radeon_emit(cs, 0);
+	}
 
-	COMPUTE_DBG("...\n");
+#if 0
+	COMPUTE_DBG(rctx->screen, "cdw: %i\n", cs->cdw);
+	for (i = 0; i < cs->cdw; i++) {
+		COMPUTE_DBG(rctx->screen, "%4i : 0x%08X\n", i, cs->buf[i]);
+	}
+#endif
 
-	r600_emit_atom(ctx, &ctx->start_cs_cmd.atom);
+}
 
-	ctx->streamout_start = TRUE;
-	ctx->streamout_append_bitmask = ~0;
 
+/**
+ * Emit function for r600_cs_shader_state atom
+ */
+void evergreen_emit_cs_shader(struct r600_context *rctx,
+			      struct r600_atom *atom)
+{
+	struct r600_cs_shader_state *state =
+					(struct r600_cs_shader_state*)atom;
+	struct r600_pipe_compute *shader = state->shader;
+	struct radeon_winsys_cs *cs = rctx->b.gfx.cs;
+	uint64_t va;
+	struct r600_resource *code_bo;
+	unsigned ngpr, nstack;
+
+	code_bo = shader->code_bo;
+	va = shader->code_bo->gpu_address + state->pc;
+	ngpr = shader->bc.ngpr;
+	nstack = shader->bc.nstack;
+
+	radeon_compute_set_context_reg_seq(cs, R_0288D0_SQ_PGM_START_LS, 3);
+	radeon_emit(cs, va >> 8); /* R_0288D0_SQ_PGM_START_LS */
+	radeon_emit(cs,           /* R_0288D4_SQ_PGM_RESOURCES_LS */
+			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, radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx,
+					      code_bo, RADEON_USAGE_READ,
+					      RADEON_PRIO_USER_SHADER));
 }
 
-static void evergreen_launch_grid(
-		struct pipe_context *ctx_,
-		const uint *block_layout, const uint *grid_layout,
-		uint32_t pc, const void *input)
+static void evergreen_launch_grid(struct pipe_context *ctx,
+				  const struct pipe_grid_info *info)
 {
-	COMPUTE_DBG("PC: %i\n", pc);
+	struct r600_context *rctx = (struct r600_context *)ctx;
+#ifdef HAVE_OPENCL
+	struct r600_pipe_compute *shader = rctx->cs_shader_state.shader;
+	boolean use_kill;
 
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	unsigned num_waves;
-	unsigned num_pipes = ctx->screen->info.r600_max_pipes;
-	unsigned wave_divisor = (16 * num_pipes);
+	rctx->cs_shader_state.pc = info->pc;
+	/* Get the config information for this kernel. */
+	r600_shader_binary_read_config(&shader->binary, &shader->bc,
+                                  info->pc, &use_kill);
+#endif
 
-	/* num_waves = ceil((tg_size.x * tg_size.y, tg_size.z) / (16 * num_pipes)) */
-	num_waves = (block_layout[0] * block_layout[1] * block_layout[2] +
-			wave_divisor - 1) / wave_divisor;
+	COMPUTE_DBG(rctx->screen, "*** evergreen_launch_grid: pc = %u\n", info->pc);
 
-	COMPUTE_DBG("Using %u pipes, there are %u wavefronts per thread block\n",
-							num_pipes, num_waves);
 
-	evergreen_set_lds(ctx->cs_shader, 0, 0, num_waves);
-	evergreen_compute_upload_input(ctx_, block_layout, grid_layout, input);
-	evergreen_direct_dispatch(ctx_, block_layout, grid_layout);
-	compute_emit_cs(ctx);
+	evergreen_compute_upload_input(ctx, info);
+	compute_emit_cs(rctx, info);
 }
 
-static void evergreen_set_compute_resources(struct pipe_context * ctx_,
-		unsigned start, unsigned count,
-		struct pipe_surface ** surfaces)
+static void evergreen_set_compute_resources(struct pipe_context *ctx,
+					    unsigned start, unsigned count,
+					    struct pipe_surface **surfaces)
 {
-	struct r600_context *ctx = (struct r600_context *)ctx_;
+	struct r600_context *rctx = (struct r600_context *)ctx;
 	struct r600_surface **resources = (struct r600_surface **)surfaces;
-	for (int i = 0; i < count; i++)	{
+
+	COMPUTE_DBG(rctx->screen, "*** evergreen_set_compute_resources: start = %u count = %u\n",
+			start, count);
+
+	for (unsigned i = 0; i < count; i++) {
+		/* The First four vertex buffers are reserved for parameters and
+		 * global buffers. */
+		unsigned vtx_id = 4 + i;
 		if (resources[i]) {
 			struct r600_resource_global *buffer =
-				(struct r600_resource_global*)resources[i]->base.texture;
+				(struct r600_resource_global*)
+				resources[i]->base.texture;
 			if (resources[i]->base.writable) {
 				assert(i+1 < 12);
-				struct r600_resource_global *buffer =
-					(struct r600_resource_global*)
-					resources[i]->base.texture;
 
-				evergreen_set_rat(ctx->cs_shader, i+1,
+				evergreen_set_rat(rctx->cs_shader_state.shader, i+1,
 				(struct r600_resource *)resources[i]->base.texture,
 				buffer->chunk->start_in_dw*4,
 				resources[i]->base.texture->width0);
 			}
 
-			evergreen_set_vtx_resource(ctx->cs_shader,
-				(struct r600_resource *)resources[i]->base.texture, i+2,
-				 buffer->chunk->start_in_dw*4, resources[i]->base.writable);
-		}
-	}
-
-}
-
-static void evergreen_set_cs_sampler_view(struct pipe_context *ctx_,
-		unsigned start_slot, unsigned count,
-		struct pipe_sampler_view **views)
-{
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	struct r600_pipe_sampler_view **resource =
-		(struct r600_pipe_sampler_view **)views;
-
-	for (int i = 0; i < count; i++)	{
-		if (resource[i]) {
-			assert(i+1 < 12);
-			///FETCH0 = VTX0 (param buffer),
-			//FETCH1 = VTX1 (global buffer pool), FETCH2... = TEX
-			evergreen_set_tex_resource(ctx->cs_shader, resource[i], i+2);
-		}
-	}
-}
-
-static void evergreen_bind_compute_sampler_states(
-	struct pipe_context *ctx_,
-	unsigned start_slot,
-	unsigned num_samplers,
-	void **samplers_)
-{
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	struct compute_sampler_state ** samplers =
-		(struct compute_sampler_state **)samplers_;
-
-	for (int i = 0; i < num_samplers; i++) {
-		if (samplers[i]) {
-			evergreen_set_sampler_resource(ctx->cs_shader, samplers[i], i);
+			evergreen_cs_set_vertex_buffer(rctx, vtx_id,
+					buffer->chunk->start_in_dw * 4,
+					resources[i]->base.texture);
 		}
 	}
 }
 
-static void evergreen_set_global_binding(
-	struct pipe_context *ctx_, unsigned first, unsigned n,
-	struct pipe_resource **resources,
-	uint32_t **handles)
+static void evergreen_set_global_binding(struct pipe_context *ctx,
+					 unsigned first, unsigned n,
+					 struct pipe_resource **resources,
+					 uint32_t **handles)
 {
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	struct compute_memory_pool *pool = ctx->screen->global_pool;
+	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct compute_memory_pool *pool = rctx->screen->global_pool;
 	struct r600_resource_global **buffers =
 		(struct r600_resource_global **)resources;
+	unsigned i;
+
+	COMPUTE_DBG(rctx->screen, "*** evergreen_set_global_binding first = %u n = %u\n",
+			first, n);
 
 	if (!resources) {
 		/* XXX: Unset */
 		return;
 	}
 
-	compute_memory_finalize_pending(pool, ctx_);
+	/* We mark these items for promotion to the pool if they
+	 * aren't already there */
+	for (i = first; i < first + n; i++) {
+		struct compute_memory_item *item = buffers[i]->chunk;
 
-	for (int i = 0; i < n; i++)
+		if (!is_item_in_pool(item))
+			buffers[i]->chunk->status |= ITEM_FOR_PROMOTING;
+	}
+
+	if (compute_memory_finalize_pending(pool, ctx) == -1) {
+		/* XXX: Unset */
+		return;
+	}
+
+	for (i = first; i < first + n; i++)
 	{
+		uint32_t buffer_offset;
+		uint32_t handle;
 		assert(resources[i]->target == PIPE_BUFFER);
 		assert(resources[i]->bind & PIPE_BIND_GLOBAL);
 
-		*(handles[i]) = buffers[i]->chunk->start_in_dw * 4;
+		buffer_offset = util_le32_to_cpu(*(handles[i]));
+		handle = buffer_offset + buffers[i]->chunk->start_in_dw * 4;
+
+		*(handles[i]) = util_cpu_to_le32(handle);
 	}
 
-	evergreen_set_rat(ctx->cs_shader, 0, pool->bo, 0, pool->size_in_dw * 4);
-	evergreen_set_vtx_resource(ctx->cs_shader, pool->bo, 1, 0, 1);
-}
+	/* globals for writing */
+	evergreen_set_rat(rctx->cs_shader_state.shader, 0, pool->bo, 0, pool->size_in_dw * 4);
+	/* globals for reading */
+	evergreen_cs_set_vertex_buffer(rctx, 1, 0,
+				(struct pipe_resource*)pool->bo);
 
+	/* constants for reading, LLVM puts them in text segment */
+	evergreen_cs_set_vertex_buffer(rctx, 2, 0,
+				(struct pipe_resource*)rctx->cs_shader_state.shader->code_bo);
+}
 
-void evergreen_compute_init_config(struct r600_context *ctx)
+/**
+ * This function initializes all the compute specific registers that need to
+ * be initialized for each compute command stream.  Registers that are common
+ * to both compute and 3D will be initialized at the beginning of each compute
+ * command stream by the start_cs_cmd atom.  However, since the SET_CONTEXT_REG
+ * packet requires that the shader type bit be set, we must initialize all
+ * context registers needed for compute in this function.  The registers
+ * initialized by the start_cs_cmd atom can be found in evergreen_state.c in the
+ * functions evergreen_init_atom_start_cs or cayman_init_atom_start_cs depending
+ * on the GPU family.
+ */
+void evergreen_init_atom_start_compute_cs(struct r600_context *rctx)
 {
-	struct evergreen_compute_resource* res =
-		get_empty_res(ctx->cs_shader, COMPUTE_RESOURCE_CONFIG, 0);
-
+	struct r600_command_buffer *cb = &rctx->start_compute_cs_cmd;
 	int num_threads;
 	int num_stack_entries;
-	int num_temp_gprs;
 
-	enum radeon_family family;
-	unsigned tmp;
+	/* since all required registers are initialized in the
+	 * start_compute_cs_cmd atom, we can EMIT_EARLY here.
+	 */
+	r600_init_command_buffer(cb, 256);
+	cb->pkt_flags = RADEON_CP_PACKET3_COMPUTE_MODE;
+
+	/* This must be first. */
+	r600_store_value(cb, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
+	r600_store_value(cb, 0x80000000);
+	r600_store_value(cb, 0x80000000);
 
-	family = ctx->family;
+	/* We're setting config registers here. */
+	r600_store_value(cb, PKT3(PKT3_EVENT_WRITE, 0, 0));
+	r600_store_value(cb, EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
 
-	switch (family) {
+	switch (rctx->b.family) {
 	case CHIP_CEDAR:
 	default:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 256;
 		break;
 	case CHIP_REDWOOD:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 256;
 		break;
 	case CHIP_JUNIPER:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 512;
 		break;
 	case CHIP_CYPRESS:
 	case CHIP_HEMLOCK:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 512;
 		break;
 	case CHIP_PALM:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 256;
 		break;
 	case CHIP_SUMO:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 256;
 		break;
 	case CHIP_SUMO2:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 512;
 		break;
 	case CHIP_BARTS:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 512;
 		break;
 	case CHIP_TURKS:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 256;
 		break;
 	case CHIP_CAICOS:
-		num_temp_gprs = 4;
 		num_threads = 128;
 		num_stack_entries = 256;
 		break;
 	}
 
-	tmp = 0x00000000;
-	switch (family) {
-	case CHIP_CEDAR:
-	case CHIP_PALM:
-	case CHIP_SUMO:
-	case CHIP_SUMO2:
-	case CHIP_CAICOS:
-		break;
-	default:
-		tmp |= S_008C00_VC_ENABLE(1);
-		break;
+	/* Config Registers */
+	if (rctx->b.chip_class < CAYMAN)
+		evergreen_init_common_regs(rctx, cb, rctx->b.chip_class, rctx->b.family,
+					   rctx->screen->b.info.drm_minor);
+	else
+		cayman_init_common_regs(cb, rctx->b.chip_class, rctx->b.family,
+					rctx->screen->b.info.drm_minor);
+
+	/* The primitive type always needs to be POINTLIST for compute. */
+	r600_store_config_reg(cb, R_008958_VGT_PRIMITIVE_TYPE,
+						V_008958_DI_PT_POINTLIST);
+
+	if (rctx->b.chip_class < CAYMAN) {
+
+		/* These registers control which simds can be used by each stage.
+		 * The default for these registers is 0xffffffff, which means
+		 * all simds are available for each stage.  It's possible we may
+		 * want to play around with these in the future, but for now
+		 * the default value is fine.
+		 *
+		 * R_008E20_SQ_STATIC_THREAD_MGMT1
+		 * R_008E24_SQ_STATIC_THREAD_MGMT2
+		 * R_008E28_SQ_STATIC_THREAD_MGMT3
+		 */
+
+		/* XXX: We may need to adjust the thread and stack resource
+		 * values for 3D/compute interop */
+
+		r600_store_config_reg_seq(cb, R_008C18_SQ_THREAD_RESOURCE_MGMT_1, 5);
+
+		/* R_008C18_SQ_THREAD_RESOURCE_MGMT_1
+		 * Set the number of threads used by the PS/VS/GS/ES stage to
+		 * 0.
+		 */
+		r600_store_value(cb, 0);
+
+		/* R_008C1C_SQ_THREAD_RESOURCE_MGMT_2
+		 * Set the number of threads used by the CS (aka LS) stage to
+		 * the maximum number of threads and set the number of threads
+		 * for the HS stage to 0. */
+		r600_store_value(cb, S_008C1C_NUM_LS_THREADS(num_threads));
+
+		/* R_008C20_SQ_STACK_RESOURCE_MGMT_1
+		 * Set the Control Flow stack entries to 0 for PS/VS stages */
+		r600_store_value(cb, 0);
+
+		/* R_008C24_SQ_STACK_RESOURCE_MGMT_2
+		 * Set the Control Flow stack entries to 0 for GS/ES stages */
+		r600_store_value(cb, 0);
+
+		/* R_008C28_SQ_STACK_RESOURCE_MGMT_3
+		 * Set the Contol Flow stack entries to 0 for the HS stage, and
+		 * set it to the maximum value for the CS (aka LS) stage. */
+		r600_store_value(cb,
+			S_008C28_NUM_LS_STACK_ENTRIES(num_stack_entries));
 	}
-	tmp |= S_008C00_EXPORT_SRC_C(1);
-	tmp |= S_008C00_CS_PRIO(0);
-	tmp |= S_008C00_LS_PRIO(0);
-	tmp |= S_008C00_HS_PRIO(0);
-	tmp |= S_008C00_PS_PRIO(0);
-	tmp |= S_008C00_VS_PRIO(0);
-	tmp |= S_008C00_GS_PRIO(0);
-	tmp |= S_008C00_ES_PRIO(0);
-
-	evergreen_reg_set(res, R_008C00_SQ_CONFIG, tmp);
-
-	evergreen_reg_set(res, R_008C04_SQ_GPR_RESOURCE_MGMT_1,
-				S_008C04_NUM_CLAUSE_TEMP_GPRS(num_temp_gprs));
-	if (ctx->chip_class < CAYMAN) {
-		evergreen_reg_set(res, R_008C08_SQ_GPR_RESOURCE_MGMT_2, 0);
-	}
-	evergreen_reg_set(res, R_008C10_SQ_GLOBAL_GPR_RESOURCE_MGMT_1, 0);
-	evergreen_reg_set(res, R_008C14_SQ_GLOBAL_GPR_RESOURCE_MGMT_2, 0);
-	evergreen_reg_set(res, R_008D8C_SQ_DYN_GPR_CNTL_PS_FLUSH_REQ, (1 << 8));
-
-	/* workaround for hw issues with dyn gpr - must set all limits to 240
-	 * instead of 0, 0x1e == 240/8 */
-	if (ctx->chip_class < CAYMAN) {
-		evergreen_reg_set(res, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1,
-				S_028838_PS_GPRS(0x1e) |
-				S_028838_VS_GPRS(0x1e) |
-				S_028838_GS_GPRS(0x1e) |
-				S_028838_ES_GPRS(0x1e) |
-				S_028838_HS_GPRS(0x1e) |
-				S_028838_LS_GPRS(0x1e));
+	/* Give the compute shader all the available LDS space.
+	 * NOTE: This only sets the maximum number of dwords that a compute
+	 * shader can allocate.  When a shader is executed, we still need to
+	 * allocate the appropriate amount of LDS dwords using the
+	 * CM_R_0288E8_SQ_LDS_ALLOC register.
+	 */
+	if (rctx->b.chip_class < CAYMAN) {
+		r600_store_config_reg(cb, R_008E2C_SQ_LDS_RESOURCE_MGMT,
+			S_008E2C_NUM_PS_LDS(0x0000) | S_008E2C_NUM_LS_LDS(8192));
 	} else {
-		evergreen_reg_set(res, 0x286f8,
+		r600_store_context_reg(cb, CM_R_0286FC_SPI_LDS_MGMT,
+			S_0286FC_NUM_PS_LDS(0) |
+			S_0286FC_NUM_LS_LDS(255)); /* 255 * 32 = 8160 dwords */
+	}
+
+	/* Context Registers */
+
+	if (rctx->b.chip_class < CAYMAN) {
+		/* workaround for hw issues with dyn gpr - must set all limits
+		 * to 240 instead of 0, 0x1e == 240 / 8
+		 */
+		r600_store_context_reg(cb, R_028838_SQ_DYN_GPR_RESOURCE_LIMIT_1,
 				S_028838_PS_GPRS(0x1e) |
 				S_028838_VS_GPRS(0x1e) |
 				S_028838_GS_GPRS(0x1e) |
@@ -635,97 +862,132 @@ void evergreen_compute_init_config(struct r600_context *ctx)
 				S_028838_LS_GPRS(0x1e));
 	}
 
-	if (ctx->chip_class < CAYMAN) {
-
-		evergreen_reg_set(res, R_008E20_SQ_STATIC_THREAD_MGMT1, 0xFFFFFFFF);
-		evergreen_reg_set(res, R_008E24_SQ_STATIC_THREAD_MGMT2, 0xFFFFFFFF);
-		evergreen_reg_set(res, R_008E20_SQ_STATIC_THREAD_MGMT1, 0xFFFFFFFF);
-		evergreen_reg_set(res, R_008E24_SQ_STATIC_THREAD_MGMT2, 0xFFFFFFFF);
-		evergreen_reg_set(res, R_008E28_SQ_STATIC_THREAD_MGMT3, 0xFFFFFFFF);
-		evergreen_reg_set(res, R_008C18_SQ_THREAD_RESOURCE_MGMT_1, 0);
-		tmp = S_008C1C_NUM_LS_THREADS(num_threads);
-		evergreen_reg_set(res, R_008C1C_SQ_THREAD_RESOURCE_MGMT_2, tmp);
-		evergreen_reg_set(res, R_008C20_SQ_STACK_RESOURCE_MGMT_1, 0);
-		evergreen_reg_set(res, R_008C24_SQ_STACK_RESOURCE_MGMT_2, 0);
-		tmp = S_008C28_NUM_LS_STACK_ENTRIES(num_stack_entries);
-		evergreen_reg_set(res, R_008C28_SQ_STACK_RESOURCE_MGMT_3, tmp);
-	}
-	evergreen_reg_set(res, R_0286CC_SPI_PS_IN_CONTROL_0, S_0286CC_LINEAR_GRADIENT_ENA(1));
-	evergreen_reg_set(res, R_0286D0_SPI_PS_IN_CONTROL_1, 0);
-	evergreen_reg_set(res, R_0286E4_SPI_PS_IN_CONTROL_2, 0);
-	evergreen_reg_set(res, R_0286D8_SPI_INPUT_Z, 0);
-	evergreen_reg_set(res, R_0286E0_SPI_BARYC_CNTL, 1 << 20);
-	tmp = S_0286E8_TID_IN_GROUP_ENA | S_0286E8_TGID_ENA | S_0286E8_DISABLE_INDEX_PACK;
-	evergreen_reg_set(res, R_0286E8_SPI_COMPUTE_INPUT_CNTL, tmp);
-	tmp = S_028A40_COMPUTE_MODE(1) | S_028A40_PARTIAL_THD_AT_EOI(1);
-	evergreen_reg_set(res, R_028A40_VGT_GS_MODE, tmp);
-	evergreen_reg_set(res, R_028B54_VGT_SHADER_STAGES_EN, 2/*CS_ON*/);
-	evergreen_reg_set(res, R_028800_DB_DEPTH_CONTROL, 0);
-	evergreen_reg_set(res, R_02880C_DB_SHADER_CONTROL, 0);
-	evergreen_reg_set(res, R_028000_DB_RENDER_CONTROL, S_028000_COLOR_DISABLE(1));
-	evergreen_reg_set(res, R_02800C_DB_RENDER_OVERRIDE, 0);
-	evergreen_reg_set(res, R_0286E8_SPI_COMPUTE_INPUT_CNTL,
+	/* XXX: Investigate setting bit 15, which is FAST_COMPUTE_MODE */
+	r600_store_context_reg(cb, R_028A40_VGT_GS_MODE,
+		S_028A40_COMPUTE_MODE(1) | S_028A40_PARTIAL_THD_AT_EOI(1));
+
+	r600_store_context_reg(cb, R_028B54_VGT_SHADER_STAGES_EN, 2/*CS_ON*/);
+
+	r600_store_context_reg(cb, R_0286E8_SPI_COMPUTE_INPUT_CNTL,
 						S_0286E8_TID_IN_GROUP_ENA
 						| S_0286E8_TGID_ENA
 						| S_0286E8_DISABLE_INDEX_PACK)
 						;
+
+	/* The LOOP_CONST registers are an optimizations for loops that allows
+	 * you to store the initial counter, increment value, and maximum
+	 * counter value in a register so that hardware can calculate the
+	 * correct number of iterations for the loop, so that you don't need
+	 * to have the loop counter in your shader code.  We don't currently use
+	 * this optimization, so we must keep track of the counter in the
+	 * shader and use a break instruction to exit loops.  However, the
+	 * hardware will still uses this register to determine when to exit a
+	 * loop, so we need to initialize the counter to 0, set the increment
+	 * value to 1 and the maximum counter value to the 4095 (0xfff) which
+	 * is the maximum value allowed.  This gives us a maximum of 4096
+	 * iterations for our loops, but hopefully our break instruction will
+	 * execute before some time before the 4096th iteration.
+	 */
+	eg_store_loop_const(cb, R_03A200_SQ_LOOP_CONST_0 + (160 * 4), 0x1000FFF);
 }
 
-void evergreen_init_compute_state_functions(struct r600_context *ctx)
+void evergreen_init_compute_state_functions(struct r600_context *rctx)
 {
-	ctx->context.create_compute_state = evergreen_create_compute_state;
-	ctx->context.delete_compute_state = evergreen_delete_compute_state;
-	ctx->context.bind_compute_state = evergreen_bind_compute_state;
-//	 ctx->context.create_sampler_view = evergreen_compute_create_sampler_view;
-	ctx->context.set_compute_resources = evergreen_set_compute_resources;
-	ctx->context.set_compute_sampler_views = evergreen_set_cs_sampler_view;
-	ctx->context.bind_compute_sampler_states = evergreen_bind_compute_sampler_states;
-	ctx->context.set_global_binding = evergreen_set_global_binding;
-	ctx->context.launch_grid = evergreen_launch_grid;
-}
+	rctx->b.b.create_compute_state = evergreen_create_compute_state;
+	rctx->b.b.delete_compute_state = evergreen_delete_compute_state;
+	rctx->b.b.bind_compute_state = evergreen_bind_compute_state;
+//	 rctx->context.create_sampler_view = evergreen_compute_create_sampler_view;
+	rctx->b.b.set_compute_resources = evergreen_set_compute_resources;
+	rctx->b.b.set_global_binding = evergreen_set_global_binding;
+	rctx->b.b.launch_grid = evergreen_launch_grid;
 
+}
 
-struct pipe_resource *r600_compute_global_buffer_create(
-	struct pipe_screen *screen,
-	const struct pipe_resource *templ)
+static void *r600_compute_global_transfer_map(struct pipe_context *ctx,
+					      struct pipe_resource *resource,
+					      unsigned level,
+					      unsigned usage,
+					      const struct pipe_box *box,
+					      struct pipe_transfer **ptransfer)
 {
-	assert(templ->target == PIPE_BUFFER);
-	assert(templ->bind & PIPE_BIND_GLOBAL);
-	assert(templ->array_size == 1 || templ->array_size == 0);
-	assert(templ->depth0 == 1 || templ->depth0 == 0);
-	assert(templ->height0 == 1 || templ->height0 == 0);
+	struct r600_context *rctx = (struct r600_context*)ctx;
+	struct compute_memory_pool *pool = rctx->screen->global_pool;
+	struct r600_resource_global* buffer =
+		(struct r600_resource_global*)resource;
 
-	struct r600_resource_global* result = (struct r600_resource_global*)
-		CALLOC(sizeof(struct r600_resource_global), 1);
-	struct r600_screen* rscreen = (struct r600_screen*)screen;
+	struct compute_memory_item *item = buffer->chunk;
+	struct pipe_resource *dst = NULL;
+	unsigned offset = box->x;
 
-	result->base.b.vtbl = &r600_global_buffer_vtbl;
-	result->base.b.b.screen = screen;
-	result->base.b.b = *templ;
-	pipe_reference_init(&result->base.b.b.reference, 1);
+	if (is_item_in_pool(item)) {
+		compute_memory_demote_item(pool, item, ctx);
+	}
+	else {
+		if (item->real_buffer == NULL) {
+			item->real_buffer =
+					r600_compute_buffer_alloc_vram(pool->screen, item->size_in_dw * 4);
+		}
+	}
 
-	int size_in_dw = (templ->width0+3) / 4;
+	dst = (struct pipe_resource*)item->real_buffer;
 
-	result->chunk = compute_memory_alloc(rscreen->global_pool, size_in_dw);
+	if (usage & PIPE_TRANSFER_READ)
+		buffer->chunk->status |= ITEM_MAPPED_FOR_READING;
 
-	if (result->chunk == NULL)
-	{
-		free(result);
-		return NULL;
-	}
+	COMPUTE_DBG(rctx->screen, "* r600_compute_global_transfer_map()\n"
+			"level = %u, usage = %u, box(x = %u, y = %u, z = %u "
+			"width = %u, height = %u, depth = %u)\n", level, usage,
+			box->x, box->y, box->z, box->width, box->height,
+			box->depth);
+	COMPUTE_DBG(rctx->screen, "Buffer id = %"PRIi64" offset = "
+		"%u (box.x)\n", item->id, box->x);
 
-	return &result->base.b.b;
+
+	assert(resource->target == PIPE_BUFFER);
+	assert(resource->bind & PIPE_BIND_GLOBAL);
+	assert(box->x >= 0);
+	assert(box->y == 0);
+	assert(box->z == 0);
+
+	///TODO: do it better, mapping is not possible if the pool is too big
+	return pipe_buffer_map_range(ctx, dst,
+			offset, box->width, usage, ptransfer);
 }
 
-void r600_compute_global_buffer_destroy(
-	struct pipe_screen *screen,
-	struct pipe_resource *res)
+static void r600_compute_global_transfer_unmap(struct pipe_context *ctx,
+					       struct pipe_transfer *transfer)
 {
+	/* struct r600_resource_global are not real resources, they just map
+	 * to an offset within the compute memory pool.  The function
+	 * r600_compute_global_transfer_map() maps the memory pool
+	 * resource rather than the struct r600_resource_global passed to
+	 * it as an argument and then initalizes ptransfer->resource with
+	 * the memory pool resource (via pipe_buffer_map_range).
+	 * When transfer_unmap is called it uses the memory pool's
+	 * vtable which calls r600_buffer_transfer_map() rather than
+	 * this function.
+	 */
+	assert (!"This function should not be called");
+}
+
+static void r600_compute_global_transfer_flush_region(struct pipe_context *ctx,
+						      struct pipe_transfer *transfer,
+						      const struct pipe_box *box)
+{
+	assert(0 && "TODO");
+}
+
+static void r600_compute_global_buffer_destroy(struct pipe_screen *screen,
+					       struct pipe_resource *res)
+{
+	struct r600_resource_global* buffer = NULL;
+	struct r600_screen* rscreen = NULL;
+
 	assert(res->target == PIPE_BUFFER);
 	assert(res->bind & PIPE_BIND_GLOBAL);
 
-	struct r600_resource_global* buffer = (struct r600_resource_global*)res;
-	struct r600_screen* rscreen = (struct r600_screen*)screen;
+	buffer = (struct r600_resource_global*)res;
+	rscreen = (struct r600_screen*)screen;
 
 	compute_memory_free(rscreen->global_pool, buffer->chunk->id);
 
@@ -733,101 +995,50 @@ void r600_compute_global_buffer_destroy(
 	free(res);
 }
 
-void* r600_compute_global_transfer_map(
-	struct pipe_context *ctx_,
-	struct pipe_transfer* transfer)
+static const struct u_resource_vtbl r600_global_buffer_vtbl =
 {
-	assert(transfer->resource->target == PIPE_BUFFER);
-	assert(transfer->resource->bind & PIPE_BIND_GLOBAL);
-	assert(transfer->box.x >= 0);
-	assert(transfer->box.y == 0);
-	assert(transfer->box.z == 0);
-
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	struct r600_resource_global* buffer =
-		(struct r600_resource_global*)transfer->resource;
-
-	uint32_t* map;
-	///TODO: do it better, mapping is not possible if the pool is too big
-
-	if (!(map = ctx->ws->buffer_map(buffer->chunk->pool->bo->cs_buf,
-						ctx->cs, transfer->usage))) {
-		return NULL;
-	}
-
-	COMPUTE_DBG("buffer start: %lli\n", buffer->chunk->start_in_dw);
-	return ((char*)(map + buffer->chunk->start_in_dw)) + transfer->box.x;
-}
+	u_default_resource_get_handle, /* get_handle */
+	r600_compute_global_buffer_destroy, /* resource_destroy */
+	r600_compute_global_transfer_map, /* transfer_map */
+	r600_compute_global_transfer_flush_region,/* transfer_flush_region */
+	r600_compute_global_transfer_unmap, /* transfer_unmap */
+};
 
-void r600_compute_global_transfer_unmap(
-	struct pipe_context *ctx_,
-	struct pipe_transfer* transfer)
+struct pipe_resource *r600_compute_global_buffer_create(struct pipe_screen *screen,
+							const struct pipe_resource *templ)
 {
-	assert(transfer->resource->target == PIPE_BUFFER);
-	assert(transfer->resource->bind & PIPE_BIND_GLOBAL);
+	struct r600_resource_global* result = NULL;
+	struct r600_screen* rscreen = NULL;
+	int size_in_dw = 0;
 
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	struct r600_resource_global* buffer =
-		(struct r600_resource_global*)transfer->resource;
+	assert(templ->target == PIPE_BUFFER);
+	assert(templ->bind & PIPE_BIND_GLOBAL);
+	assert(templ->array_size == 1 || templ->array_size == 0);
+	assert(templ->depth0 == 1 || templ->depth0 == 0);
+	assert(templ->height0 == 1 || templ->height0 == 0);
 
-	ctx->ws->buffer_unmap(buffer->chunk->pool->bo->cs_buf);
-}
+	result = (struct r600_resource_global*)
+	CALLOC(sizeof(struct r600_resource_global), 1);
+	rscreen = (struct r600_screen*)screen;
 
-struct pipe_transfer * r600_compute_global_get_transfer(
-	struct pipe_context *ctx_,
-	struct pipe_resource *resource,
-	unsigned level,
-	unsigned usage,
-	const struct pipe_box *box)
-{
-	struct r600_context *ctx = (struct r600_context *)ctx_;
-	struct compute_memory_pool *pool = ctx->screen->global_pool;
+	COMPUTE_DBG(rscreen, "*** r600_compute_global_buffer_create\n");
+	COMPUTE_DBG(rscreen, "width = %u array_size = %u\n", templ->width0,
+			templ->array_size);
 
-	compute_memory_finalize_pending(pool, ctx_);
+	result->base.b.vtbl = &r600_global_buffer_vtbl;
+	result->base.b.b = *templ;
+	result->base.b.b.screen = screen;
+	pipe_reference_init(&result->base.b.b.reference, 1);
 
-	assert(resource->target == PIPE_BUFFER);
-	struct r600_context *rctx = (struct r600_context*)ctx_;
-	struct pipe_transfer *transfer = util_slab_alloc(&rctx->pool_transfers);
-
-	transfer->resource = resource;
-	transfer->level = level;
-	transfer->usage = usage;
-	transfer->box = *box;
-	transfer->stride = 0;
-	transfer->layer_stride = 0;
-	transfer->data = NULL;
-
-	/* Note strides are zero, this is ok for buffers, but not for
-	* textures 2d & higher at least.
-	*/
-	return transfer;
-}
+	size_in_dw = (templ->width0+3) / 4;
 
-void r600_compute_global_transfer_destroy(
-	struct pipe_context *ctx_,
-	struct pipe_transfer *transfer)
-{
-	struct r600_context *rctx = (struct r600_context*)ctx_;
-	util_slab_free(&rctx->pool_transfers, transfer);
-}
+	result->chunk = compute_memory_alloc(rscreen->global_pool, size_in_dw);
 
-void r600_compute_global_transfer_flush_region(
-	struct pipe_context *ctx_,
-	struct pipe_transfer *transfer,
-	const struct pipe_box *box)
-{
-	assert(0 && "TODO");
-}
+	if (result->chunk == NULL)
+	{
+		free(result);
+		return NULL;
+	}
 
-void r600_compute_global_transfer_inline_write(
-	struct pipe_context *pipe,
-	struct pipe_resource *resource,
-	unsigned level,
-	unsigned usage,
-	const struct pipe_box *box,
-	const void *data,
-	unsigned stride,
-	unsigned layer_stride)
-{
-	assert(0 && "TODO");
+	return &result->base.b.b;
 }