r600g: add initial vertex translate support.

4 files changed, 239 insertions, 3 deletions

diff --git a/src/gallium/drivers/r600/r600_context.c b/src/gallium/drivers/r600/r600_context.c
index 091751e93a..cff15c57d2 100644
--- a/src/gallium/drivers/r600/r600_context.c
+++ b/src/gallium/drivers/r600/r600_context.c
@@ -62,6 +62,9 @@ static void r600_destroy_context(struct pipe_context *context)
 	u_upload_destroy(rctx->upload_vb);
 	u_upload_destroy(rctx->upload_ib);
 
+	if (rctx->tran.translate_cache)
+		translate_cache_destroy(rctx->tran.translate_cache);
+
 	radeon_ctx_fini(rctx->ctx);
 	FREE(rctx);
 }
@@ -153,6 +156,10 @@ struct pipe_context *r600_create_context(struct pipe_screen *screen, void *priv)
 		goto out_free;
 	}
 
+	rctx->tran.translate_cache = translate_cache_create();
+	if (rctx->tran.translate_cache == NULL)
+		goto out_free;
+
 	rctx->vs_constant = (struct radeon_state *)calloc(R600_MAX_CONSTANT, sizeof(struct radeon_state));
 	if (!rctx->vs_constant) {
 		goto out_free;
diff --git a/src/gallium/drivers/r600/r600_context.h b/src/gallium/drivers/r600/r600_context.h
index d104531d36..ae32194318 100644
--- a/src/gallium/drivers/r600/r600_context.h
+++ b/src/gallium/drivers/r600/r600_context.h
@@ -31,6 +31,7 @@
 #include <tgsi/tgsi_util.h>
 #include <util/u_blitter.h>
 #include <util/u_double_list.h>
+#include "translate/translate_cache.h"
 #include "radeon.h"
 #include "r600_shader.h"
 
@@ -115,7 +116,11 @@ struct r600_vertex_element
 {
 	unsigned			refcount;
 	unsigned			count;
-	struct pipe_vertex_element	elements[32];
+	struct pipe_vertex_element	elements[PIPE_MAX_ATTRIBS];
+
+	enum pipe_format hw_format[PIPE_MAX_ATTRIBS];
+	unsigned hw_format_size[PIPE_MAX_ATTRIBS];
+	boolean incompatible_layout;
 };
 
 struct r600_draw {
@@ -132,6 +137,18 @@ struct r600_draw {
 	unsigned		index_bias;
 };
 
+struct r600_translate_context {
+	/* Translate cache for incompatible vertex offset/stride/format fallback. */
+	struct translate_cache *translate_cache;
+
+	/* The vertex buffer slot containing the translated buffer. */
+	unsigned vb_slot;
+
+	/* Saved and new vertex element state. */
+	void *saved_velems, *new_velems;
+};
+
+
 struct r600_context_hw_states {
 	struct radeon_state	rasterizer;
 	struct radeon_state	scissor;
@@ -247,6 +264,10 @@ struct r600_context {
 	struct u_upload_mgr *upload_vb;
 	struct u_upload_mgr *upload_ib;
 	bool any_user_vbs;
+	unsigned vb_max_index;
+
+	/* For translating vertex buffers having incompatible vertex layout. */
+	struct r600_translate_context tran;
 };
 
 /* Convenience cast wrapper. */
diff --git a/src/gallium/drivers/r600/r600_draw.c b/src/gallium/drivers/r600/r600_draw.c
index c41156f15f..ed1b20f7b9 100644
--- a/src/gallium/drivers/r600/r600_draw.c
+++ b/src/gallium/drivers/r600/r600_draw.c
@@ -32,12 +32,173 @@
 #include <util/u_inlines.h>
 #include <util/u_memory.h>
 #include <util/u_index_modify.h>
+#include "translate/translate_cache.h"
+#include "translate/translate.h"
 #include "radeon.h"
 #include "r600_screen.h"
 #include "r600_context.h"
 #include "r600_resource.h"
 #include "r600_state_inlines.h"
 
+static void r600_begin_vertex_translate(struct r600_context *rctx)
+{
+	struct pipe_context *pipe = &rctx->context;
+	struct translate_key key = {0};
+	struct translate_element *te;
+	unsigned tr_elem_index[PIPE_MAX_ATTRIBS] = {0};
+	struct translate *tr;
+	struct r600_vertex_element *ve = rctx->vertex_elements;
+	boolean vb_translated[PIPE_MAX_ATTRIBS] = {0};
+	void *vb_map[PIPE_MAX_ATTRIBS] = {0}, *out_map;
+	struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS] = {0}, *out_transfer;
+	struct pipe_resource *out_buffer;
+	unsigned i, num_verts;
+
+	/* Initialize the translate key, i.e. the recipe how vertices should be
+	 * translated. */
+	for (i = 0; i < ve->count; i++) {
+		struct pipe_vertex_buffer *vb =
+			&rctx->vertex_buffer[ve->elements[i].vertex_buffer_index];
+		enum pipe_format output_format = ve->hw_format[i];
+		unsigned output_format_size = ve->hw_format_size[i];
+
+		/* Check for support. */
+		if (ve->elements[i].src_format == ve->hw_format[i] &&
+		    (vb->buffer_offset + ve->elements[i].src_offset) % 4 == 0 &&
+		    vb->stride % 4 == 0) {
+			continue;
+		}
+
+		/* Workaround for translate: output floats instead of halfs. */
+		switch (output_format) {
+		case PIPE_FORMAT_R16_FLOAT:
+			output_format = PIPE_FORMAT_R32_FLOAT;
+			output_format_size = 4;
+			break;
+		case PIPE_FORMAT_R16G16_FLOAT:
+			output_format = PIPE_FORMAT_R32G32_FLOAT;
+			output_format_size = 8;
+			break;
+		case PIPE_FORMAT_R16G16B16_FLOAT:
+			output_format = PIPE_FORMAT_R32G32B32_FLOAT;
+			output_format_size = 12;
+			break;
+		case PIPE_FORMAT_R16G16B16A16_FLOAT:
+			output_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
+			output_format_size = 16;
+			break;
+		default:;
+		}
+
+		/* Add this vertex element. */
+		te = &key.element[key.nr_elements];
+		/*te->type;
+		  te->instance_divisor;*/
+		te->input_buffer = ve->elements[i].vertex_buffer_index;
+		te->input_format = ve->elements[i].src_format;
+		te->input_offset = vb->buffer_offset + ve->elements[i].src_offset;
+		te->output_format = output_format;
+		te->output_offset = key.output_stride;
+
+		key.output_stride += output_format_size;
+		vb_translated[ve->elements[i].vertex_buffer_index] = TRUE;
+		tr_elem_index[i] = key.nr_elements;
+		key.nr_elements++;
+	}
+
+	/* Get a translate object. */
+	tr = translate_cache_find(rctx->tran.translate_cache, &key);
+
+	/* Map buffers we want to translate. */
+	for (i = 0; i < rctx->nvertex_buffer; i++) {
+		if (vb_translated[i]) {
+			struct pipe_vertex_buffer *vb = &rctx->vertex_buffer[i];
+
+			vb_map[i] = pipe_buffer_map(pipe, vb->buffer,
+						    PIPE_TRANSFER_READ, &vb_transfer[i]);
+
+			tr->set_buffer(tr, i, vb_map[i], vb->stride, vb->max_index);
+		}
+	}
+
+	/* Create and map the output buffer. */
+	num_verts = rctx->vb_max_index + 1;
+
+	out_buffer = pipe_buffer_create(&rctx->screen->screen,
+					PIPE_BIND_VERTEX_BUFFER,
+					key.output_stride * num_verts);
+
+	out_map = pipe_buffer_map(pipe, out_buffer, PIPE_TRANSFER_WRITE,
+				  &out_transfer);
+
+	/* Translate. */
+	tr->run(tr, 0, num_verts, 0, out_map);
+
+	/* Unmap all buffers. */
+	for (i = 0; i < rctx->nvertex_buffer; i++) {
+		if (vb_translated[i]) {
+			pipe_buffer_unmap(pipe, rctx->vertex_buffer[i].buffer,
+					  vb_transfer[i]);
+		}
+	}
+
+	pipe_buffer_unmap(pipe, out_buffer, out_transfer);
+
+	/* Setup the new vertex buffer in the first free slot. */
+	for (i = 0; i < PIPE_MAX_ATTRIBS; i++) {
+		struct pipe_vertex_buffer *vb = &rctx->vertex_buffer[i];
+
+		if (!vb->buffer) {
+			pipe_resource_reference(&vb->buffer, out_buffer);
+			vb->buffer_offset = 0;
+			vb->max_index = num_verts - 1;
+			vb->stride = key.output_stride;
+			rctx->tran.vb_slot = i;
+			break;
+		}
+	}
+
+	/* Save and replace vertex elements. */
+	{
+		struct pipe_vertex_element new_velems[PIPE_MAX_ATTRIBS];
+
+		rctx->tran.saved_velems = rctx->vertex_elements;
+
+		for (i = 0; i < ve->count; i++) {
+			if (vb_translated[ve->elements[i].vertex_buffer_index]) {
+				te = &key.element[tr_elem_index[i]];
+				new_velems[i].instance_divisor = ve->elements[i].instance_divisor;
+				new_velems[i].src_format = te->output_format;
+				new_velems[i].src_offset = te->output_offset;
+				new_velems[i].vertex_buffer_index = rctx->tran.vb_slot;
+			} else {
+				memcpy(&new_velems[i], &ve->elements[i],
+				       sizeof(struct pipe_vertex_element));
+			}
+		}
+
+		rctx->tran.new_velems =
+			pipe->create_vertex_elements_state(pipe, ve->count, new_velems);
+		pipe->bind_vertex_elements_state(pipe, rctx->tran.new_velems);
+	}
+
+	pipe_resource_reference(&out_buffer, NULL);
+}
+
+static void r600_end_vertex_translate(struct r600_context *rctx)
+{
+	struct pipe_context *pipe = &rctx->context;
+
+	/* Restore vertex elements. */
+	pipe->bind_vertex_elements_state(pipe, rctx->tran.saved_velems);
+	pipe->delete_vertex_elements_state(pipe, rctx->tran.new_velems);
+
+	/* Delete the now-unused VBO. */
+	pipe_resource_reference(&rctx->vertex_buffer[rctx->tran.vb_slot].buffer,
+				NULL);
+}
+
+
 static void r600_translate_index_buffer(struct r600_context *r600,
 					struct pipe_resource **index_buffer,
 					unsigned *index_size,
@@ -117,7 +278,7 @@ static int r600_draw_common(struct r600_draw *draw)
 		rbuffer = (struct r600_resource*)vertex_buffer->buffer;
 		offset = rctx->vertex_elements->elements[i].src_offset + vertex_buffer->buffer_offset;
 		
-		rctx->vtbl->vs_resource(rctx, i, rbuffer, offset, vertex_buffer->stride, rctx->vertex_elements->elements[i].src_format);
+		rctx->vtbl->vs_resource(rctx, i, rbuffer, offset, vertex_buffer->stride, rctx->vertex_elements->hw_format[i]);
 		radeon_draw_bind(&rctx->draw, vs_resource);
 	}
 	rctx->vs_nresource = rctx->vertex_elements->count;
@@ -144,9 +305,14 @@ void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	struct r600_context *rctx = r600_context(ctx);
 	struct r600_draw draw;
 	int r;
+	boolean translate = FALSE;
 
 	memset(&draw, 0, sizeof(draw));
 
+	if (rctx->vertex_elements->incompatible_layout) {
+		r600_begin_vertex_translate(rctx);
+		translate = TRUE;
+	}
 	if (rctx->any_user_vbs) {
 		r600_upload_user_buffers(rctx);
 		rctx->any_user_vbs = FALSE;
@@ -186,5 +352,8 @@ void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	if (r)
 	  fprintf(stderr,"draw common failed %d\n", r);
 
+	if (translate) {
+		r600_end_vertex_translate(rctx);
+	}
 	pipe_resource_reference(&draw.index_buffer, NULL);
 }
diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index 86f9825b52..1cadeea2fd 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -204,15 +204,41 @@ static void *r600_create_shader_state(struct pipe_context *ctx,
 	return rstate;
 }
 
+#define FORMAT_REPLACE(what, withwhat) \
+    case PIPE_FORMAT_##what: *format = PIPE_FORMAT_##withwhat; break
+
 static void *r600_create_vertex_elements(struct pipe_context *ctx,
 				unsigned count,
 				const struct pipe_vertex_element *elements)
 {
 	struct r600_vertex_element *v = CALLOC_STRUCT(r600_vertex_element);
+	unsigned i;
+	enum pipe_format *format;
 
 	assert(count < 32);
 	v->count = count;
 	memcpy(v->elements, elements, count * sizeof(struct pipe_vertex_element));
+
+	for (i = 0; i < count; i++) {
+		v->hw_format[i] = v->elements[i].src_format;
+		format = &v->hw_format[i];
+
+		switch (*format) {
+                    FORMAT_REPLACE(R64_FLOAT,           R32_FLOAT);
+                    FORMAT_REPLACE(R64G64_FLOAT,        R32G32_FLOAT);
+                    FORMAT_REPLACE(R64G64B64_FLOAT,     R32G32B32_FLOAT);
+                    FORMAT_REPLACE(R64G64B64A64_FLOAT,  R32G32B32A32_FLOAT);
+		default:;
+		}
+		v->incompatible_layout =
+			v->incompatible_layout ||
+			v->elements[i].src_format != v->hw_format[i] ||
+			v->elements[i].src_offset % 4 != 0;
+
+                v->hw_format_size[i] =
+			align(util_format_get_blocksize(v->hw_format[i]), 4);
+	}
+
 	v->refcount = 1;
 	return v;
 }
@@ -455,19 +481,32 @@ static void r600_set_vertex_buffers(struct pipe_context *ctx,
 	struct r600_context *rctx = r600_context(ctx);
 	unsigned i;
 	boolean any_user_buffers = FALSE;
-
+	unsigned max_index, vbo_max_index;
+	const struct pipe_vertex_buffer *vbo;
 	for (i = 0; i < rctx->nvertex_buffer; i++) {
 		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, NULL);
 	}
 	memcpy(rctx->vertex_buffer, buffers, sizeof(struct pipe_vertex_buffer) * count);
+	max_index = 0;
 	for (i = 0; i < count; i++) {
+		vbo = &buffers[i];
 		rctx->vertex_buffer[i].buffer = NULL;
 		if (r600_buffer_is_user_buffer(buffers[i].buffer))
 			any_user_buffers = TRUE;
 		pipe_resource_reference(&rctx->vertex_buffer[i].buffer, buffers[i].buffer);
+		if (vbo->max_index == ~0) {
+			if (!vbo->stride)
+				vbo_max_index = 1;
+			else
+				vbo_max_index = (vbo->buffer->width0 - vbo->buffer_offset) / vbo->stride;
+		} else
+			vbo_max_index = vbo->max_index;
+	
+		max_index = MIN2(vbo_max_index, max_index);
 	}
 	rctx->any_user_vbs = any_user_buffers;
 	rctx->nvertex_buffer = count;
+	rctx->vb_max_index = max_index;
 }
 
 static void r600_set_index_buffer(struct pipe_context *ctx,