Revert "r600g: precompute some of the hw state"

This reverts commit de0b76cab22caa9fc7260f80acb8f151ccced6c5, its pre-computes the texture state wrong,

you can't just use an array of levels, since you can have FBOs to depth texture slices inside a level as well
it would get really messy quickly. Probably need to split commits like this up into pieces for each piece
of state, so we can revert bits easier in case of regressions.

This also break 5 piglit tests, and valgrind starts to warn about invalid read/writes after this.

1 files changed, 151 insertions, 75 deletions

diff --git a/src/gallium/drivers/r600/r600_state.c b/src/gallium/drivers/r600/r600_state.c
index 6049e1328ec..b5db848e35e 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -34,16 +34,6 @@
 #include "r600d.h"
 #include "r600_state_inlines.h"
 
-static struct radeon_state *r600_blend(struct r600_context *rctx, const struct pipe_blend_state *state);
-static struct radeon_state *r600_viewport(struct r600_context *rctx, const struct pipe_viewport_state *state);
-static struct radeon_state *r600_ucp(struct r600_context *rctx, const struct pipe_clip_state *state);
-static struct radeon_state *r600_sampler(struct r600_context *rctx,
-				const struct pipe_sampler_state *state,
-				unsigned id);
-static struct radeon_state *r600_resource(struct pipe_context *ctx,
-					const struct pipe_sampler_view *view,
-					unsigned id);
-
 static void *r600_create_blend_state(struct pipe_context *ctx,
 					const struct pipe_blend_state *state)
 {
@@ -96,7 +86,6 @@ static struct pipe_sampler_view *r600_create_sampler_view(struct pipe_context *c
 	rstate->state.sampler_view.texture = texture;
 	rstate->state.sampler_view.reference.count = 1;
 	rstate->state.sampler_view.context = ctx;
-	rstate->rstate = r600_resource(ctx, &rstate->state.sampler_view, 0);
 	return &rstate->state.sampler_view;
 }
 
@@ -240,9 +229,6 @@ static void r600_bind_ps_sampler(struct pipe_context *ctx,
 	for (i = 0; i < count; i++) {
 		rstate = (struct r600_context_state *)states[i];
 		rctx->ps_sampler[i] = r600_context_state_incref(rstate);
-		if (rstate) {
-			radeon_state_convert(rstate->rstate, R600_STATE_SAMPLER, i, R600_SHADER_PS);
-		}
 	}
 	rctx->ps_nsampler = count;
 }
@@ -260,9 +246,6 @@ static void r600_bind_vs_sampler(struct pipe_context *ctx,
 	for (i = 0; i < count; i++) {
 		rstate = (struct r600_context_state *)states[i];
 		rctx->vs_sampler[i] = r600_context_state_incref(rstate);
-		if (rstate) {
-			radeon_state_convert(rstate->rstate, R600_STATE_SAMPLER, i, R600_SHADER_VS);
-		}
 	}
 	rctx->vs_nsampler = count;
 }
@@ -354,9 +337,6 @@ static void r600_set_ps_sampler_view(struct pipe_context *ctx,
 	for (i = 0; i < count; i++) {
 		rstate = (struct r600_context_state *)views[i];
 		rctx->ps_sampler_view[i] = r600_context_state_incref(rstate);
-		if (rstate) {
-			radeon_state_convert(rstate->rstate, R600_STATE_RESOURCE, i, R600_SHADER_PS);
-		}
 	}
 	rctx->ps_nsampler_view = count;
 }
@@ -375,9 +355,6 @@ static void r600_set_vs_sampler_view(struct pipe_context *ctx,
 	for (i = 0; i < count; i++) {
 		rstate = (struct r600_context_state *)views[i];
 		rctx->vs_sampler_view[i] = r600_context_state_incref(rstate);
-		if (rstate) {
-			radeon_state_convert(rstate->rstate, R600_STATE_RESOURCE, i, R600_SHADER_VS);
-		}
 	}
 	rctx->vs_nsampler_view = count;
 }
@@ -386,19 +363,10 @@ static void r600_set_framebuffer_state(struct pipe_context *ctx,
 					const struct pipe_framebuffer_state *state)
 {
 	struct r600_context *rctx = r600_context(ctx);
-	struct r600_resource_texture *rtexture;
 	struct r600_context_state *rstate;
 
 	rstate = r600_context_state(rctx, pipe_framebuffer_type, state);
 	r600_bind_state(ctx, rstate);
-	for (int i = 0; i < state->nr_cbufs; i++) {
-		rtexture = (struct r600_resource_texture*)state->cbufs[i]->texture;
-		r600_texture_cb(ctx, rtexture, i, state->cbufs[i]->level);
-	}
-	if (state->zsbuf) {
-		rtexture = (struct r600_resource_texture*)state->zsbuf->texture;
-		r600_texture_db(ctx, rtexture, state->zsbuf->level);
-	}
 }
 
 static void r600_set_polygon_stipple(struct pipe_context *ctx,
@@ -597,7 +565,6 @@ struct r600_context_state *r600_context_state(struct r600_context *rctx, unsigne
 		break;
 	case pipe_viewport_type:
 		rstate->state.viewport = (*states).viewport;
-		rstate->rstate = r600_viewport(rctx, &rstate->state.viewport);
 		break;
 	case pipe_depth_type:
 		rstate->state.depth = (*states).depth;
@@ -613,7 +580,6 @@ struct r600_context_state *r600_context_state(struct r600_context *rctx, unsigne
 		break;
 	case pipe_clip_type:
 		rstate->state.clip = (*states).clip;
-		rstate->rstate = r600_ucp(rctx, &rstate->state.clip);
 		break;
 	case pipe_stencil_type:
 		rstate->state.stencil = (*states).stencil;
@@ -626,7 +592,6 @@ struct r600_context_state *r600_context_state(struct r600_context *rctx, unsigne
 		break;
 	case pipe_blend_type:
 		rstate->state.blend = (*states).blend;
-		rstate->rstate = r600_blend(rctx, &rstate->state.blend);
 		break;
 	case pipe_stencil_ref_type:
 		rstate->state.stencil_ref = (*states).stencil_ref;
@@ -641,7 +606,6 @@ struct r600_context_state *r600_context_state(struct r600_context *rctx, unsigne
 		break;
 	case pipe_sampler_type:
 		rstate->state.sampler = (*states).sampler;
-		rstate->rstate = r600_sampler(rctx, &rstate->state.sampler, 0);
 		break;
 	default:
 		R600_ERR("invalid type %d\n", rstate->type);
@@ -651,10 +615,11 @@ struct r600_context_state *r600_context_state(struct r600_context *rctx, unsigne
 	return rstate;
 }
 
-static struct radeon_state *r600_blend(struct r600_context *rctx, const struct pipe_blend_state *state)
+static struct radeon_state *r600_blend(struct r600_context *rctx)
 {
 	struct r600_screen *rscreen = rctx->screen;
 	struct radeon_state *rstate;
+	const struct pipe_blend_state *state = &rctx->blend->state.blend;
 	int i;
 
 	rstate = radeon_state(rscreen->rw, R600_STATE_BLEND, 0);
@@ -710,28 +675,129 @@ static struct radeon_state *r600_blend(struct r600_context *rctx, const struct p
 	return rstate;
 }
 
-static struct radeon_state *r600_ucp(struct r600_context *rctx, const struct pipe_clip_state *state)
+static struct radeon_state *r600_ucp(struct r600_context *rctx, int clip)
 {
 	struct r600_screen *rscreen = rctx->screen;
 	struct radeon_state *rstate;
+	const struct pipe_clip_state *state = &rctx->clip->state.clip;
 
-	rstate = radeon_state(rscreen->rw, R600_STATE_UCP, 0);
+	rstate = radeon_state(rscreen->rw, R600_STATE_CLIP, clip);
 	if (rstate == NULL)
 		return NULL;
 
-	for (int i = 0; i < state->nr; i++) {
-		rstate->states[i * 4 + 0] = fui(state->ucp[i][0]);
-		rstate->states[i * 4 + 1] = fui(state->ucp[i][1]);
-		rstate->states[i * 4 + 2] = fui(state->ucp[i][2]);
-		rstate->states[i * 4 + 3] = fui(state->ucp[i][3]);
+	rstate->states[R600_CLIP__PA_CL_UCP_X_0] = fui(state->ucp[clip][0]);
+	rstate->states[R600_CLIP__PA_CL_UCP_Y_0] = fui(state->ucp[clip][1]);
+	rstate->states[R600_CLIP__PA_CL_UCP_Z_0] = fui(state->ucp[clip][2]);
+	rstate->states[R600_CLIP__PA_CL_UCP_W_0] = fui(state->ucp[clip][3]);
+
+	if (radeon_state_pm4(rstate)) {
+		radeon_state_decref(rstate);
+		return NULL;
 	}
+	return rstate;
 
+}
+
+static struct radeon_state *r600_cb(struct r600_context *rctx, int cb)
+{
+	struct r600_screen *rscreen = rctx->screen;
+	struct r600_resource_texture *rtex;
+	struct r600_resource *rbuffer;
+	struct radeon_state *rstate;
+	const struct pipe_framebuffer_state *state = &rctx->framebuffer->state.framebuffer;
+	unsigned level = state->cbufs[cb]->level;
+	unsigned pitch, slice;
+	unsigned color_info;
+	unsigned format, swap, ntype;
+	const struct util_format_description *desc;
+
+	rstate = radeon_state(rscreen->rw, R600_STATE_CB0 + cb, 0);
+	if (rstate == NULL)
+		return NULL;
+	rtex = (struct r600_resource_texture*)state->cbufs[cb]->texture;
+	rbuffer = &rtex->resource;
+	rstate->bo[0] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
+	rstate->bo[1] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
+	rstate->bo[2] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
+	rstate->placement[0] = RADEON_GEM_DOMAIN_GTT;
+	rstate->placement[2] = RADEON_GEM_DOMAIN_GTT;
+	rstate->placement[4] = RADEON_GEM_DOMAIN_GTT;
+	rstate->nbo = 3;
+	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
+	slice = (rtex->pitch[level] / rtex->bpt) * state->cbufs[cb]->height / 64 - 1;
+
+	ntype = 0;
+	desc = util_format_description(rtex->resource.base.b.format);
+	if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
+		ntype = V_0280A0_NUMBER_SRGB;
+
+	format = r600_translate_colorformat(rtex->resource.base.b.format);
+	swap = r600_translate_colorswap(rtex->resource.base.b.format);
+
+	color_info = S_0280A0_FORMAT(format) |
+		S_0280A0_COMP_SWAP(swap) |
+		S_0280A0_BLEND_CLAMP(1) |
+		S_0280A0_SOURCE_FORMAT(1) |
+		S_0280A0_NUMBER_TYPE(ntype);
+
+	rstate->states[R600_CB0__CB_COLOR0_BASE] = rtex->offset[level] >> 8;
+	rstate->states[R600_CB0__CB_COLOR0_INFO] = color_info;
+	rstate->states[R600_CB0__CB_COLOR0_SIZE] = S_028060_PITCH_TILE_MAX(pitch) |
+						S_028060_SLICE_TILE_MAX(slice);
+	rstate->states[R600_CB0__CB_COLOR0_VIEW] = 0x00000000;
+	rstate->states[R600_CB0__CB_COLOR0_FRAG] = 0x00000000;
+	rstate->states[R600_CB0__CB_COLOR0_TILE] = 0x00000000;
+	rstate->states[R600_CB0__CB_COLOR0_MASK] = 0x00000000;
 	if (radeon_state_pm4(rstate)) {
 		radeon_state_decref(rstate);
 		return NULL;
 	}
 	return rstate;
+}
 
+static struct radeon_state *r600_db(struct r600_context *rctx)
+{
+	struct r600_screen *rscreen = rctx->screen;
+	struct r600_resource_texture *rtex;
+	struct r600_resource *rbuffer;
+	struct radeon_state *rstate;
+	const struct pipe_framebuffer_state *state = &rctx->framebuffer->state.framebuffer;
+	unsigned level;
+	unsigned pitch, slice, format;
+
+	if (state->zsbuf == NULL)
+		return NULL;
+
+	rstate = radeon_state(rscreen->rw, R600_STATE_DB, 0);
+	if (rstate == NULL)
+		return NULL;
+
+	rtex = (struct r600_resource_texture*)state->zsbuf->texture;
+	rtex->tilled = 1;
+	rtex->array_mode = 2;
+	rtex->tile_type = 1;
+	rtex->depth = 1;
+	rbuffer = &rtex->resource;
+
+	rstate->bo[0] = radeon_bo_incref(rscreen->rw, rbuffer->bo);
+	rstate->nbo = 1;
+	rstate->placement[0] = RADEON_GEM_DOMAIN_VRAM;
+	level = state->zsbuf->level;
+	pitch = (rtex->pitch[level] / rtex->bpt) / 8 - 1;
+	slice = (rtex->pitch[level] / rtex->bpt) * state->zsbuf->height / 64 - 1;
+	format = r600_translate_dbformat(state->zsbuf->texture->format);
+	rstate->states[R600_DB__DB_DEPTH_BASE] = rtex->offset[level] >> 8;
+	rstate->states[R600_DB__DB_DEPTH_INFO] = S_028010_ARRAY_MODE(rtex->array_mode) |
+					S_028010_FORMAT(format);
+	rstate->states[R600_DB__DB_DEPTH_VIEW] = 0x00000000;
+	rstate->states[R600_DB__DB_PREFETCH_LIMIT] = (state->zsbuf->height / 8) -1;
+	rstate->states[R600_DB__DB_DEPTH_SIZE] = S_028000_PITCH_TILE_MAX(pitch) |
+						S_028000_SLICE_TILE_MAX(slice);
+	if (radeon_state_pm4(rstate)) {
+		radeon_state_decref(rstate);
+		return NULL;
+	}
+	return rstate;
 }
 
 static struct radeon_state *r600_rasterizer(struct r600_context *rctx)
@@ -888,8 +954,9 @@ static struct radeon_state *r600_scissor(struct r600_context *rctx)
 	return rstate;
 }
 
-static struct radeon_state *r600_viewport(struct r600_context *rctx, const struct pipe_viewport_state *state)
+static struct radeon_state *r600_viewport(struct r600_context *rctx)
 {
+	const struct pipe_viewport_state *state = &rctx->viewport->state.viewport;
 	struct r600_screen *rscreen = rctx->screen;
 	struct radeon_state *rstate;
 
@@ -1299,7 +1366,6 @@ static struct radeon_state *r600_cb_cntl(struct r600_context *rctx)
 int r600_context_hw_states(struct pipe_context *ctx)
 {
 	struct r600_context *rctx = r600_context(ctx);
-	struct r600_resource_texture *rtexture;
 	unsigned i;
 	int r;
 	int nr_cbufs = rctx->framebuffer->state.framebuffer.nr_cbufs;
@@ -1311,59 +1377,69 @@ int r600_context_hw_states(struct pipe_context *ctx)
 	/* free previous TODO determine what need to be updated, what
 	 * doesn't
 	 */
+	//radeon_state_decref(rctx->hw_states.config);
 	rctx->hw_states.cb_cntl = radeon_state_decref(rctx->hw_states.cb_cntl);
+	rctx->hw_states.db = radeon_state_decref(rctx->hw_states.db);
 	rctx->hw_states.rasterizer = radeon_state_decref(rctx->hw_states.rasterizer);
 	rctx->hw_states.scissor = radeon_state_decref(rctx->hw_states.scissor);
 	rctx->hw_states.dsa = radeon_state_decref(rctx->hw_states.dsa);
+	rctx->hw_states.blend = radeon_state_decref(rctx->hw_states.blend);
+	rctx->hw_states.viewport = radeon_state_decref(rctx->hw_states.viewport);
+	for (i = 0; i < 8; i++) {
+		rctx->hw_states.cb[i] = radeon_state_decref(rctx->hw_states.cb[i]);
+	}
+	for (i = 0; i < 6; i++) {
+		rctx->hw_states.ucp[i] = radeon_state_decref(rctx->hw_states.ucp[i]);
+	}
+	for (i = 0; i < rctx->hw_states.ps_nresource; i++) {
+		radeon_state_decref(rctx->hw_states.ps_resource[i]);
+		rctx->hw_states.ps_resource[i] = NULL;
+	}
+	rctx->hw_states.ps_nresource = 0;
+	for (i = 0; i < rctx->hw_states.ps_nsampler; i++) {
+		radeon_state_decref(rctx->hw_states.ps_sampler[i]);
+		rctx->hw_states.ps_sampler[i] = NULL;
+	}
+	rctx->hw_states.ps_nsampler = 0;
 
 	/* build new states */
-	rctx->hw_states.blend = NULL;
-	rctx->hw_states.viewport = NULL;
-	rctx->hw_states.ucp = NULL;
 	rctx->hw_states.rasterizer = r600_rasterizer(rctx);
 	rctx->hw_states.scissor = r600_scissor(rctx);
 	rctx->hw_states.dsa = r600_dsa(rctx);
-	rctx->hw_states.cb_cntl = r600_cb_cntl(rctx);
-	if (rctx->viewport) {
-		rctx->hw_states.viewport = rctx->viewport->rstate;
-	}
-	if (rctx->blend) {
-		rctx->hw_states.blend = rctx->blend->rstate;
-	}
-	if (rctx->clip) {
-		rctx->hw_states.ucp = rctx->clip->rstate;
-	}
-	for (i = 0; i < rctx->framebuffer->state.framebuffer.nr_cbufs; i++) {
-		rtexture = (struct r600_resource_texture*)rctx->framebuffer->state.framebuffer.cbufs[i]->texture;
-		rctx->hw_states.cb[i] = rtexture->cb[i][rctx->framebuffer->state.framebuffer.cbufs[i]->level];
+	rctx->hw_states.blend = r600_blend(rctx);
+	rctx->hw_states.viewport = r600_viewport(rctx);
+	for (i = 0; i < nr_cbufs; i++) {
+		rctx->hw_states.cb[i] = r600_cb(rctx, i);
 	}
-	if (rctx->framebuffer->state.framebuffer.zsbuf) {
-		rtexture = (struct r600_resource_texture*)rctx->framebuffer->state.framebuffer.zsbuf->texture;
-		rctx->hw_states.db = rtexture->db[rctx->framebuffer->state.framebuffer.zsbuf->level];
+	for (i = 0; i < ucp_nclip; i++) {
+		rctx->hw_states.ucp[i] = r600_ucp(rctx, i);
 	}
-
+	rctx->hw_states.db = r600_db(rctx);
+	rctx->hw_states.cb_cntl = r600_cb_cntl(rctx);
 
 	for (i = 0; i < rctx->ps_nsampler; i++) {
 		if (rctx->ps_sampler[i]) {
-			rctx->hw_states.ps_sampler[i] = rctx->ps_sampler[i]->rstate;
-		} else {
-			rctx->hw_states.ps_sampler[i] = NULL;
+			rctx->hw_states.ps_sampler[i] = r600_sampler(rctx,
+							&rctx->ps_sampler[i]->state.sampler,
+							i);
 		}
 	}
 	rctx->hw_states.ps_nsampler = rctx->ps_nsampler;
 	for (i = 0; i < rctx->ps_nsampler_view; i++) {
 		if (rctx->ps_sampler_view[i]) {
-			rctx->hw_states.ps_resource[i] = rctx->ps_sampler_view[i]->rstate;
-		} else {
-			rctx->hw_states.ps_resource[i] = NULL;
+			rctx->hw_states.ps_resource[i] = r600_resource(ctx,
+							&rctx->ps_sampler_view[i]->state.sampler_view,
+							i);
 		}
 	}
 	rctx->hw_states.ps_nresource = rctx->ps_nsampler_view;
 
 	/* bind states */
-	r = radeon_draw_set(rctx->draw, rctx->hw_states.ucp);
-	if (r)
-		return r;
+	for (i = 0; i < ucp_nclip; i++) {
+		r = radeon_draw_set(rctx->draw, rctx->hw_states.ucp[i]);
+		if (r)
+			return r;
+	}
 	r = radeon_draw_set(rctx->draw, rctx->hw_states.db);
 	if (r)
 		return r;