1 files changed, 189 insertions, 91 deletions

diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index 40874152274..63c54200d4e 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -36,7 +36,6 @@
 #include <util/u_pack_color.h>
 #include <util/u_memory.h>
 #include <util/u_inlines.h>
-#include <util/u_upload_mgr.h>
 #include <util/u_framebuffer.h>
 #include <pipebuffer/pb_buffer.h>
 #include "r600.h"
@@ -46,14 +45,169 @@
 #include "r600_pipe.h"
 #include "r600_state_inlines.h"
 
+void r600_polygon_offset_update(struct r600_pipe_context *rctx)
+{
+	struct r600_pipe_state state;
+
+	state.id = R600_PIPE_STATE_POLYGON_OFFSET;
+	state.nregs = 0;
+	if (rctx->rasterizer && rctx->framebuffer.zsbuf) {
+		float offset_units = rctx->rasterizer->offset_units;
+		unsigned offset_db_fmt_cntl = 0, depth;
+
+		switch (rctx->framebuffer.zsbuf->texture->format) {
+		case PIPE_FORMAT_Z24X8_UNORM:
+		case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
+			depth = -24;
+			offset_units *= 2.0f;
+			break;
+		case PIPE_FORMAT_Z32_FLOAT:
+			depth = -23;
+			offset_units *= 1.0f;
+			offset_db_fmt_cntl |= S_028DF8_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
+			break;
+		case PIPE_FORMAT_Z16_UNORM:
+			depth = -16;
+			offset_units *= 4.0f;
+			break;
+		default:
+			return;
+		}
+		/* FIXME some of those reg can be computed with cso */
+		offset_db_fmt_cntl |= S_028DF8_POLY_OFFSET_NEG_NUM_DB_BITS(depth);
+		r600_pipe_state_add_reg(&state,
+				R_028E00_PA_SU_POLY_OFFSET_FRONT_SCALE,
+				fui(rctx->rasterizer->offset_scale), 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(&state,
+				R_028E04_PA_SU_POLY_OFFSET_FRONT_OFFSET,
+				fui(offset_units), 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(&state,
+				R_028E08_PA_SU_POLY_OFFSET_BACK_SCALE,
+				fui(rctx->rasterizer->offset_scale), 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(&state,
+				R_028E0C_PA_SU_POLY_OFFSET_BACK_OFFSET,
+				fui(offset_units), 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(&state,
+				R_028DF8_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
+				offset_db_fmt_cntl, 0xFFFFFFFF, NULL);
+		r600_context_pipe_state_set(&rctx->ctx, &state);
+	}
+}
+
+/* FIXME optimize away spi update when it's not needed */
+static void r600_spi_update(struct r600_pipe_context *rctx)
+{
+	struct r600_pipe_shader *shader = rctx->ps_shader;
+	struct r600_pipe_state rstate;
+	struct r600_shader *rshader = &shader->shader;
+	unsigned i, tmp;
+
+	rstate.nregs = 0;
+	for (i = 0; i < rshader->ninput; i++) {
+		tmp = S_028644_SEMANTIC(r600_find_vs_semantic_index(&rctx->vs_shader->shader, rshader, i));
+		if (rshader->input[i].centroid)
+			tmp |= S_028644_SEL_CENTROID(1);
+		if (rshader->input[i].interpolate == TGSI_INTERPOLATE_LINEAR)
+			tmp |= S_028644_SEL_LINEAR(1);
+
+		if (rshader->input[i].name == TGSI_SEMANTIC_COLOR ||
+		    rshader->input[i].name == TGSI_SEMANTIC_BCOLOR ||
+		    rshader->input[i].name == TGSI_SEMANTIC_POSITION) {
+			tmp |= S_028644_FLAT_SHADE(rctx->flatshade);
+		}
+		if (rshader->input[i].name == TGSI_SEMANTIC_GENERIC &&
+			rctx->sprite_coord_enable & (1 << rshader->input[i].sid)) {
+			tmp |= S_028644_PT_SPRITE_TEX(1);
+		}
+		r600_pipe_state_add_reg(&rstate, R_028644_SPI_PS_INPUT_CNTL_0 + i * 4, tmp, 0xFFFFFFFF, NULL);
+	}
+	r600_context_pipe_state_set(&rctx->ctx, &rstate);
+}
+
+void r600_vertex_buffer_update(struct r600_pipe_context *rctx)
+{
+	struct r600_pipe_state *rstate;
+	struct r600_resource *rbuffer;
+	struct pipe_vertex_buffer *vertex_buffer;
+	unsigned i, offset;
+
+	/* we don't update until we know vertex elements */
+	if (rctx->vertex_elements == NULL || !rctx->nvertex_buffer)
+		return;
+
+	/* delete previous translated vertex elements */
+	if (rctx->tran.new_velems) {
+		r600_end_vertex_translate(rctx);
+	}
+
+	if (rctx->vertex_elements->incompatible_layout) {
+		/* translate rebind new vertex elements so
+		 * return once translated
+		 */
+		r600_begin_vertex_translate(rctx);
+		return;
+	}
+
+	if (rctx->any_user_vbs) {
+		r600_upload_user_buffers(rctx);
+		rctx->any_user_vbs = FALSE;
+	}
+
+	if (rctx->vertex_elements->vbuffer_need_offset) {
+		/* one resource per vertex elements */
+		rctx->nvs_resource = rctx->vertex_elements->count;
+	} else {
+		/* bind vertex buffer once */
+		rctx->nvs_resource = rctx->nvertex_buffer;
+	}
+
+	for (i = 0 ; i < rctx->nvs_resource; i++) {
+		rstate = &rctx->vs_resource[i];
+		rstate->id = R600_PIPE_STATE_RESOURCE;
+		rstate->nregs = 0;
+
+		if (rctx->vertex_elements->vbuffer_need_offset) {
+			/* one resource per vertex elements */
+			unsigned vbuffer_index;
+			vbuffer_index = rctx->vertex_elements->elements[i].vertex_buffer_index;
+			vertex_buffer = &rctx->vertex_buffer[vbuffer_index];
+			rbuffer = (struct r600_resource*)vertex_buffer->buffer;
+			offset = rctx->vertex_elements->vbuffer_offset[i] +
+				vertex_buffer->buffer_offset +
+				r600_bo_offset(rbuffer->bo);
+		} else {
+			/* bind vertex buffer once */
+			vertex_buffer = &rctx->vertex_buffer[i];
+			rbuffer = (struct r600_resource*)vertex_buffer->buffer;
+			offset = vertex_buffer->buffer_offset +
+				r600_bo_offset(rbuffer->bo);
+		}
+
+		r600_pipe_state_add_reg(rstate, R_038000_RESOURCE0_WORD0,
+					offset, 0xFFFFFFFF, rbuffer->bo);
+		r600_pipe_state_add_reg(rstate, R_038004_RESOURCE0_WORD1,
+					rbuffer->bo_size - offset - 1, 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(rstate, R_038008_RESOURCE0_WORD2,
+					S_038008_STRIDE(vertex_buffer->stride),
+					0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(rstate, R_03800C_RESOURCE0_WORD3,
+					0x00000000, 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(rstate, R_038010_RESOURCE0_WORD4,
+					0x00000000, 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(rstate, R_038014_RESOURCE0_WORD5,
+					0x00000000, 0xFFFFFFFF, NULL);
+		r600_pipe_state_add_reg(rstate, R_038018_RESOURCE0_WORD6,
+					0xC0000000, 0xFFFFFFFF, NULL);
+		r600_context_pipe_state_set_fs_resource(&rctx->ctx, rstate, i);
+	}
+}
+
 static void r600_draw_common(struct r600_drawl *draw)
 {
 	struct r600_pipe_context *rctx = (struct r600_pipe_context *)draw->ctx;
-	struct r600_pipe_state *rstate;
 	struct r600_resource *rbuffer;
-	unsigned i, j, offset, prim;
+	unsigned prim;
 	u32 vgt_dma_index_type, vgt_draw_initiator, mask;
-	struct pipe_vertex_buffer *vertex_buffer;
 	struct r600_draw rdraw;
 	struct r600_pipe_state vgt;
 
@@ -76,41 +230,22 @@ static void r600_draw_common(struct r600_drawl *draw)
 	}
 	if (r600_conv_pipe_prim(draw->mode, &prim))
 		return;
-
-
-	/* rebuild vertex shader if input format changed */
-	if (r600_pipe_shader_update(&rctx->context, rctx->vs_shader))
+	if (unlikely(rctx->ps_shader == NULL)) {
+		R600_ERR("missing vertex shader\n");
 		return;
-	if (r600_pipe_shader_update(&rctx->context, rctx->ps_shader))
+	}
+	if (unlikely(rctx->vs_shader == NULL)) {
+		R600_ERR("missing vertex shader\n");
 		return;
-
-	for (i = 0 ; i < rctx->vertex_elements->count; i++) {
-		uint32_t word2, format;
-
-		rstate = &rctx->vs_resource[i];
-		rstate->id = R600_PIPE_STATE_RESOURCE;
-		rstate->nregs = 0;
-
-		j = rctx->vertex_elements->elements[i].vertex_buffer_index;
-		vertex_buffer = &rctx->vertex_buffer[j];
-		rbuffer = (struct r600_resource*)vertex_buffer->buffer;
-		offset = rctx->vertex_elements->elements[i].src_offset +
-			vertex_buffer->buffer_offset +
-			r600_bo_offset(rbuffer->bo);
-
-		format = r600_translate_vertex_data_type(rctx->vertex_elements->hw_format[i]);
-
-		word2 = format | S_038008_STRIDE(vertex_buffer->stride);
-
-		r600_pipe_state_add_reg(rstate, R_038000_RESOURCE0_WORD0, offset, 0xFFFFFFFF, rbuffer->bo);
-		r600_pipe_state_add_reg(rstate, R_038004_RESOURCE0_WORD1, rbuffer->size - offset - 1, 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(rstate, R_038008_RESOURCE0_WORD2, word2, 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(rstate, R_03800C_RESOURCE0_WORD3, 0x00000000, 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(rstate, R_038010_RESOURCE0_WORD4, 0x00000000, 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(rstate, R_038014_RESOURCE0_WORD5, 0x00000000, 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(rstate, R_038018_RESOURCE0_WORD6, 0xC0000000, 0xFFFFFFFF, NULL);
-		r600_context_pipe_state_set_fs_resource(&rctx->ctx, rstate, i);
 	}
+	/* there should be enough input */
+	if (rctx->vertex_elements->count < rctx->vs_shader->shader.bc.nresource) {
+		R600_ERR("%d resources provided, expecting %d\n",
+			rctx->vertex_elements->count, rctx->vs_shader->shader.bc.nresource);
+		return;
+	}
+
+	r600_spi_update(rctx);
 
 	mask = 0;
 	for (int i = 0; i < rctx->framebuffer.nr_cbufs; i++) {
@@ -126,46 +261,6 @@ static void r600_draw_common(struct r600_drawl *draw)
 	r600_pipe_state_add_reg(&vgt, R_028238_CB_TARGET_MASK, rctx->cb_target_mask & mask, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(&vgt, R_03CFF0_SQ_VTX_BASE_VTX_LOC, 0, 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(&vgt, R_03CFF4_SQ_VTX_START_INST_LOC, 0, 0xFFFFFFFF, NULL);
-	/* build late state */
-	if (rctx->rasterizer && rctx->framebuffer.zsbuf) {
-		float offset_units = rctx->rasterizer->offset_units;
-		unsigned offset_db_fmt_cntl = 0, depth;
-
-		switch (rctx->framebuffer.zsbuf->texture->format) {
-		case PIPE_FORMAT_Z24X8_UNORM:
-		case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
-			depth = -24;
-			offset_units *= 2.0f;
-			break;
-		case PIPE_FORMAT_Z32_FLOAT:
-			depth = -23;
-			offset_units *= 1.0f;
-			offset_db_fmt_cntl |= S_028DF8_POLY_OFFSET_DB_IS_FLOAT_FMT(1);
-			break;
-		case PIPE_FORMAT_Z16_UNORM:
-			depth = -16;
-			offset_units *= 4.0f;
-			break;
-		default:
-			return;
-		}
-		offset_db_fmt_cntl |= S_028DF8_POLY_OFFSET_NEG_NUM_DB_BITS(depth);
-		r600_pipe_state_add_reg(&vgt,
-				R_028E00_PA_SU_POLY_OFFSET_FRONT_SCALE,
-				fui(rctx->rasterizer->offset_scale), 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(&vgt,
-				R_028E04_PA_SU_POLY_OFFSET_FRONT_OFFSET,
-				fui(offset_units), 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(&vgt,
-				R_028E08_PA_SU_POLY_OFFSET_BACK_SCALE,
-				fui(rctx->rasterizer->offset_scale), 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(&vgt,
-				R_028E0C_PA_SU_POLY_OFFSET_BACK_OFFSET,
-				fui(offset_units), 0xFFFFFFFF, NULL);
-		r600_pipe_state_add_reg(&vgt,
-				R_028DF8_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
-				offset_db_fmt_cntl, 0xFFFFFFFF, NULL);
-	}
 	r600_context_pipe_state_set(&rctx->ctx, &vgt);
 
 	rdraw.vgt_num_indices = draw->count;
@@ -187,15 +282,6 @@ void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
 	struct r600_drawl draw;
 	boolean translate = FALSE;
 
-	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;
-	}
 	memset(&draw, 0, sizeof(struct r600_drawl));
 	draw.ctx = ctx;
 	draw.mode = info->mode;
@@ -603,9 +689,9 @@ static struct pipe_sampler_view *r600_create_sampler_view(struct pipe_context *c
 				word4 | S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_NORM) |
 				S_038010_SRF_MODE_ALL(V_038010_SFR_MODE_NO_ZERO) |
 				S_038010_REQUEST_SIZE(1) |
-				S_038010_BASE_LEVEL(state->first_level), 0xFFFFFFFF, NULL);
+				S_038010_BASE_LEVEL(state->u.tex.first_level), 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_038014_RESOURCE0_WORD5,
-				S_038014_LAST_LEVEL(state->last_level) |
+				S_038014_LAST_LEVEL(state->u.tex.last_level) |
 				S_038014_BASE_ARRAY(0) |
 				S_038014_LAST_ARRAY(0), 0xFFFFFFFF, NULL);
 	r600_pipe_state_add_reg(rstate, R_038018_RESOURCE0_WORD6,
@@ -824,10 +910,11 @@ static void r600_cb(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 	struct r600_resource_texture *rtex;
 	struct r600_resource *rbuffer;
 	struct r600_surface *surf;
-	unsigned level = state->cbufs[cb]->level;
+	unsigned level = state->cbufs[cb]->u.tex.level;
 	unsigned pitch, slice;
 	unsigned color_info;
 	unsigned format, swap, ntype;
+	unsigned offset;
 	const struct util_format_description *desc;
 	struct r600_bo *bo[3];
 
@@ -838,6 +925,9 @@ static void r600_cb(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 	bo[1] = rbuffer->bo;
 	bo[2] = rbuffer->bo;
 
+	/* XXX quite sure for dx10+ hw don't need any offset hacks */
+	offset = r600_texture_get_offset((struct r600_resource_texture *)state->cbufs[cb]->texture,
+					 level, state->cbufs[cb]->u.tex.first_layer);
 	pitch = rtex->pitch_in_pixels[level] / 8 - 1;
 	slice = rtex->pitch_in_pixels[level] * surf->aligned_height / 64 - 1;
 	ntype = 0;
@@ -868,7 +958,7 @@ static void r600_cb(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 
 	r600_pipe_state_add_reg(rstate,
 				R_028040_CB_COLOR0_BASE + cb * 4,
-				(state->cbufs[cb]->offset + r600_bo_offset(bo[0])) >> 8, 0xFFFFFFFF, bo[0]);
+				(offset + r600_bo_offset(bo[0])) >> 8, 0xFFFFFFFF, bo[0]);
 	r600_pipe_state_add_reg(rstate,
 				R_0280A0_CB_COLOR0_INFO + cb * 4,
 				color_info, 0xFFFFFFFF, bo[0]);
@@ -899,11 +989,12 @@ static void r600_db(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 	struct r600_surface *surf;
 	unsigned level;
 	unsigned pitch, slice, format;
+	unsigned offset;
 
 	if (state->zsbuf == NULL)
 		return;
 
-	level = state->zsbuf->level;
+	level = state->zsbuf->u.tex.level;
 
 	surf = (struct r600_surface *)state->zsbuf;
 	rtex = (struct r600_resource_texture*)state->zsbuf->texture;
@@ -913,12 +1004,15 @@ static void r600_db(struct r600_pipe_context *rctx, struct r600_pipe_state *rsta
 	rtex->depth = 1;
 	rbuffer = &rtex->resource;
 
+	/* XXX quite sure for dx10+ hw don't need any offset hacks */
+	offset = r600_texture_get_offset((struct r600_resource_texture *)state->zsbuf->texture,
+					 level, state->zsbuf->u.tex.first_layer);
 	pitch = rtex->pitch_in_pixels[level] / 8 - 1;
 	slice = rtex->pitch_in_pixels[level] * surf->aligned_height / 64 - 1;
 	format = r600_translate_dbformat(state->zsbuf->texture->format);
 
 	r600_pipe_state_add_reg(rstate, R_02800C_DB_DEPTH_BASE,
-				(state->zsbuf->offset + r600_bo_offset(rbuffer->bo)) >> 8, 0xFFFFFFFF, rbuffer->bo);
+				(offset + r600_bo_offset(rbuffer->bo)) >> 8, 0xFFFFFFFF, rbuffer->bo);
 	r600_pipe_state_add_reg(rstate, R_028000_DB_DEPTH_SIZE,
 				S_028000_PITCH_TILE_MAX(pitch) | S_028000_SLICE_TILE_MAX(slice),
 				0xFFFFFFFF, NULL);
@@ -1026,6 +1120,10 @@ static void r600_set_framebuffer_state(struct pipe_context *ctx,
 	free(rctx->states[R600_PIPE_STATE_FRAMEBUFFER]);
 	rctx->states[R600_PIPE_STATE_FRAMEBUFFER] = rstate;
 	r600_context_pipe_state_set(&rctx->ctx, rstate);
+
+	if (state->zsbuf) {
+		r600_polygon_offset_update(rctx);
+	}
 }
 
 static void r600_set_constant_buffer(struct pipe_context *ctx, uint shader, uint index,