Share routing code between immd/vb. Fix breakage of vertex programs in vb mode.

6 files changed, 128 insertions, 101 deletions

diff --git a/src/mesa/drivers/dri/r300/r300_context.h b/src/mesa/drivers/dri/r300/r300_context.h
index b6a80ae005..eb76a1ca05 100644
--- a/src/mesa/drivers/dri/r300/r300_context.h
+++ b/src/mesa/drivers/dri/r300/r300_context.h
@@ -116,14 +116,6 @@ struct r300_dma_region {
     int aos_size;       /* number of components (1-4) */
     int aos_format;     /* format of components */
     int aos_reg;        /* VAP register assignment */
-
-/* Left here so immediate mode still works unmodified */
-	int element_size;
-	int stride;
-	int offset;
-	int ncomponents;
-	int reg;
-	int format;
 };
 
 struct r300_dma {
diff --git a/src/mesa/drivers/dri/r300/r300_maos.c b/src/mesa/drivers/dri/r300/r300_maos.c
index a6c15e5220..27cb4afa1c 100644
--- a/src/mesa/drivers/dri/r300/r300_maos.c
+++ b/src/mesa/drivers/dri/r300/r300_maos.c
@@ -208,11 +208,11 @@ static void emit_vector(GLcontext * ctx,
 	}
 
 }
-
-/* Emit any changed arrays to new GART memory, re-emit a packet to
- * update the arrays.
+				
+/* Emit vertex data to GART memory (unless immediate mode)
+ * Route inputs to the vertex processor
  */
-void r300EmitArrays(GLcontext * ctx, GLuint inputs)
+void r300EmitArrays(GLcontext * ctx, GLboolean immd)
 {
 	r300ContextPtr rmesa = R300_CONTEXT(ctx);
 	r300ContextPtr r300 = rmesa;
@@ -223,88 +223,126 @@ void r300EmitArrays(GLcontext * ctx, GLuint inputs)
 	GLuint vic_1 = 0;	/* R300_VAP_INPUT_CNTL_1 */
 	GLuint aa_vap_reg = 0; /* VAP register assignment */
 	GLuint i;
+	GLuint inputs = 0;
+
+#define CONFIGURE_AOS(f, v, sz, cn) { \
+		if (RADEON_DEBUG & DEBUG_STATE) \
+			fprintf(stderr, "Enabling "#v "\n"); \
+		if (++nr >= R300_MAX_AOS_ARRAYS) { \
+			fprintf(stderr, "Aieee! AOS array count exceeded!\n"); \
+			exit(-1); \
+		} \
+		\
+		if (rmesa->current_vp == NULL) \
+			rmesa->state.aos[nr-1].aos_reg = aa_vap_reg++; \
+		rmesa->state.aos[nr-1].aos_format = f; \
+		if (immd) { \
+			rmesa->state.aos[nr-1].aos_size = 4; \
+			rmesa->state.aos[nr-1].aos_stride = 4; \
+			rmesa->state.aos[nr-1].aos_offset = 0; \
+		} else { \
+			emit_vector(ctx, \
+						&rmesa->state.aos[nr-1], \
+						v->data, \
+						sz, \
+						v->stride, \
+						cn); \
+		} \
+}
+
+	if (rmesa->current_vp != NULL) {
+		if (rmesa->current_vp->inputs[VERT_ATTRIB_POS] != -1) {
+			inputs |= _TNL_BIT_POS;
+			rmesa->state.aos[nr++].aos_reg = rmesa->current_vp->inputs[VERT_ATTRIB_POS];
+		}
+		if (rmesa->current_vp->inputs[VERT_ATTRIB_NORMAL] != -1) {
+			inputs |= _TNL_BIT_NORMAL;
+			rmesa->state.aos[nr++].aos_reg = rmesa->current_vp->inputs[VERT_ATTRIB_NORMAL];
+		}
+		if (rmesa->current_vp->inputs[VERT_ATTRIB_COLOR0] != -1) {
+			inputs |= _TNL_BIT_COLOR0;
+			rmesa->state.aos[nr++].aos_reg = rmesa->current_vp->inputs[VERT_ATTRIB_COLOR0];
+		}
+		if (rmesa->current_vp->inputs[VERT_ATTRIB_COLOR1] != -1) {
+			inputs |= _TNL_BIT_COLOR1;
+			rmesa->state.aos[nr++].aos_reg = rmesa->current_vp->inputs[VERT_ATTRIB_COLOR1];
+		}
+		if (rmesa->current_vp->inputs[VERT_ATTRIB_FOG] != -1) {
+			inputs |= _TNL_BIT_FOG;
+			rmesa->state.aos[nr++].aos_reg = rmesa->current_vp->inputs[VERT_ATTRIB_FOG];
+		}
+		for (i=0;i<ctx->Const.MaxTextureUnits;i++) {
+			if (rmesa->current_vp->inputs[VERT_ATTRIB_TEX0+i] != -1)
+				inputs |= _TNL_BIT_TEX0<<i;
+				rmesa->state.aos[nr++].aos_reg = rmesa->current_vp->inputs[VERT_ATTRIB_TEX0+i];
+		}
+		nr = 0;
+	} else {
+		inputs = TNL_CONTEXT(ctx)->render_inputs;
+	}
+	rmesa->state.render_inputs = inputs;
 
-/*FIXME: handle vertex program input */
 	if (inputs & _TNL_BIT_POS) {
-		if (RADEON_DEBUG & DEBUG_ALL)
-			fprintf(stderr, "[%d] _TNL_BIT_POS: sz=%d, st=%d, c=%d\n",
-				nr, VB->ObjPtr->size, VB->ObjPtr->stride, count);
-		emit_vector(ctx, &rmesa->state.aos[nr++],
-					(char *)VB->ObjPtr->data,
-					VB->ObjPtr->size,
-					VB->ObjPtr->stride, count);
+		CONFIGURE_AOS(	AOS_FORMAT_FLOAT,
+						VB->ObjPtr,
+						immd ? 4 : VB->ObjPtr->size,
+						count);
 
 		vic_1 |= R300_INPUT_CNTL_POS;
-		rmesa->state.aos[nr-1].aos_reg	  = aa_vap_reg++;
-		rmesa->state.aos[nr-1].aos_format = AOS_FORMAT_FLOAT;
 	}
 
 	if (inputs & _TNL_BIT_NORMAL) {
-		if (RADEON_DEBUG & DEBUG_ALL)
-			fprintf(stderr, "[%d] _TNL_BIT_NORMAL: sz=%d, st=%d, c=%d\n",
-				nr, VB->NormalPtr->size, VB->NormalPtr->stride, count);
-		emit_vector(ctx, &rmesa->state.aos[nr++],
-				(char *)VB->NormalPtr->data,
-				VB->NormalPtr->size,
-				VB->NormalPtr->stride, count);
+		CONFIGURE_AOS(	AOS_FORMAT_FLOAT,
+						VB->NormalPtr,
+						immd ? 4 : VB->NormalPtr->size,
+						count);
 
 		vic_1 |= R300_INPUT_CNTL_NORMAL;
-		rmesa->state.aos[nr-1].aos_reg	  = aa_vap_reg++;
-		rmesa->state.aos[nr-1].aos_format = AOS_FORMAT_FLOAT;
 	}
 
-	if (inputs & VERT_BIT_COLOR0) {
+	if (inputs & _TNL_BIT_COLOR0) {
 		int emitsize;
-		if (RADEON_DEBUG & DEBUG_ALL)
-			fprintf(stderr, "[%d] _TNL_BIT_COLOR0: sz=%d, st=%d, c=%d\n",
-				nr, VB->ColorPtr[0]->size, VB->ColorPtr[0]->stride, count);
-
-		if (VB->ColorPtr[0]->size == 4 &&
-		    (VB->ColorPtr[0]->stride != 0 ||
-		     VB->ColorPtr[0]->data[0][3] != 1.0)) {
-			emitsize = 4;
-		} else {
-			emitsize = 3;
+
+		if (!immd) {
+			if (VB->ColorPtr[0]->size == 4 &&
+			    (VB->ColorPtr[0]->stride != 0 ||
+			     VB->ColorPtr[0]->data[0][3] != 1.0)) {
+				emitsize = 4;
+			} else {
+				emitsize = 3;
+			}
 		}
 
-		emit_vector(ctx,
-			    &(rmesa->state.aos[nr++]),
-			    (char *)VB->ColorPtr[0]->data,
-			    emitsize, VB->ColorPtr[0]->stride, count);
-//			    emitsize, VB->ColorPtr[0]->stride, count);
+		CONFIGURE_AOS(	AOS_FORMAT_FLOAT_COLOR,
+						VB->ColorPtr[0],
+						immd ? 4 : emitsize,
+						count);
 
 		vic_1 |= R300_INPUT_CNTL_COLOR;
-		rmesa->state.aos[nr-1].aos_reg	  = aa_vap_reg++;
-		rmesa->state.aos[nr-1].aos_format = AOS_FORMAT_FLOAT_COLOR;
 	}
 
-	if (inputs & VERT_BIT_COLOR1) {
-		if (RADEON_DEBUG & DEBUG_ALL)
-			fprintf(stderr, "[%d] _TNL_BIT_COLOR1: sz=%d, st=%d, c=%d\n",
-				nr, VB->SecondaryColorPtr[0]->size, VB->SecondaryColorPtr[0]->stride, count);
-
-		emit_vector(ctx,
-			    &rmesa->state.aos[nr++],
-			    (char *)VB->SecondaryColorPtr[0]->data,
-			    VB->SecondaryColorPtr[0]->size, VB->SecondaryColorPtr[0]->stride, count);
-		rmesa->state.aos[nr-1].aos_reg	  = aa_vap_reg++;
-		rmesa->state.aos[nr-1].aos_format = AOS_FORMAT_FLOAT_COLOR;
+	if (inputs & _TNL_BIT_COLOR1) {
+		CONFIGURE_AOS(	AOS_FORMAT_FLOAT_COLOR,
+						VB->SecondaryColorPtr[0],
+						immd ? 4 : VB->SecondaryColorPtr[0]->size,
+						count);
+	}
+
+	if (inputs & _TNL_BIT_FOG) {
+		CONFIGURE_AOS(	AOS_FORMAT_FLOAT,
+						VB->FogCoordPtr,
+						immd ? 4 : VB->FogCoordPtr->size,
+						count);
 	}
 
 	for (i = 0; i < ctx->Const.MaxTextureUnits; i++) {
 		if (inputs & (_TNL_BIT_TEX0 << i)) {
-			if (RADEON_DEBUG & DEBUG_ALL)
-				fprintf(stderr, "[%d] _TNL_BIT_TEX%02d: sz=%d, st=%d, c=%d\n",
-					nr, i, VB->TexCoordPtr[i]->size, VB->TexCoordPtr[i]->stride, count);
-			emit_vector(ctx,
-				    &(rmesa->state.aos[nr++]),
-				    (char *)VB->TexCoordPtr[i]->data,
-				    VB->TexCoordPtr[i]->size,
-				    VB->TexCoordPtr[i]->stride, count);
+			CONFIGURE_AOS(	AOS_FORMAT_FLOAT,
+							VB->TexCoordPtr[i],
+							immd ? 4 : VB->TexCoordPtr[i]->size,
+							count);
 
 			vic_1 |= R300_INPUT_CNTL_TC0 << i;
-			rmesa->state.aos[nr-1].aos_reg	  = aa_vap_reg++;
-			rmesa->state.aos[nr-1].aos_format = AOS_FORMAT_FLOAT;
 		}
 	}
 
@@ -398,7 +436,7 @@ drm_radeon_cmd_header_t *cmd = NULL;
 	((drm_r300_cmd_header_t*)r300->hw.vir[1].cmd)->unchecked_state.count = (nr+1)>>1;
 
 	/* Set up input_cntl */
-
+	/* I don't think this is needed for vertex buffers, but it doesn't hurt anything */
 	R300_STATECHANGE(r300, vic);
 	r300->hw.vic.cmd[R300_VIC_CNTL_0]=0x5555;  /* Hard coded value, no idea what it means */
 	r300->hw.vic.cmd[R300_VIC_CNTL_1]=vic_1;
diff --git a/src/mesa/drivers/dri/r300/r300_maos.h b/src/mesa/drivers/dri/r300/r300_maos.h
index 6f6cce2f46..ec827a39b3 100644
--- a/src/mesa/drivers/dri/r300/r300_maos.h
+++ b/src/mesa/drivers/dri/r300/r300_maos.h
@@ -40,7 +40,7 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 #include "r300_context.h"
 
-extern void r300EmitArrays(GLcontext * ctx, GLuint inputs);
+extern void r300EmitArrays(GLcontext * ctx, GLboolean immd);
 extern void r300ReleaseArrays(GLcontext * ctx);
 
 #endif
diff --git a/src/mesa/drivers/dri/r300/r300_render.c b/src/mesa/drivers/dri/r300/r300_render.c
index 8f9d892982..862fabbc14 100644
--- a/src/mesa/drivers/dri/r300/r300_render.c
+++ b/src/mesa/drivers/dri/r300/r300_render.c
@@ -485,9 +485,12 @@ static GLboolean r300_run_immediate_render(GLcontext *ctx,
    efloat(1.0);
    #endif
    
+/* Setup INPUT_ROUTE and INPUT_CNTL */
+	r300EmitArrays(ctx, GL_TRUE);
+
 /* Why do we need this for immediate mode?? Vertex processor needs it to know proper regs */
-//   r300EmitLOAD_VBPNTR(rmesa, 0);
-   
+//	r300EmitLOAD_VBPNTR(rmesa, 0);
+
    for(i=0; i < VB->PrimitiveCount; i++){
        GLuint prim = VB->Primitive[i].mode;
        GLuint start = VB->Primitive[i].start;
@@ -542,7 +545,7 @@ static GLboolean r300_run_vb_render(GLcontext *ctx,
 
    
 	r300ReleaseArrays(ctx);
-	r300EmitArrays(ctx, rmesa->state.render_inputs);
+	r300EmitArrays(ctx, GL_FALSE);
 	
 //	LOCK_HARDWARE(&(rmesa->radeon));
 
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index d0f49b75f0..94324b339d 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -1017,18 +1017,11 @@ void r300_setup_routing(GLcontext *ctx, GLboolean immediate)
 	#define CONFIGURE_AOS(v, o, r, f) \
 		{\
 		if (RADEON_DEBUG & DEBUG_STATE)fprintf(stderr, "Enabling "#r "\n"); \
-		if(immediate){ \
-			r300->state.aos[count].element_size=4; \
-			r300->state.aos[count].stride=4; \
-			r300->state.aos[count].ncomponents=4; \
-			} else { \
-			r300->state.aos[count].element_size=v->size; \
-			r300->state.aos[count].stride=v->size; \
-			r300->state.aos[count].ncomponents=v->size; \
-			} \
-		r300->state.aos[count].offset=o; \
-		r300->state.aos[count].reg=reg; \
-		r300->state.aos[count].format=(f); \
+		r300->state.aos[count].aos_size=4; \
+		r300->state.aos[count].aos_stride=4; \
+		r300->state.aos[count].aos_offset=o; \
+		r300->state.aos[count].aos_reg=reg; \
+		r300->state.aos[count].aos_format=(f); \
 		r300->state.vap_reg.r=reg; \
 		count++; \
 		reg++; \
@@ -1123,12 +1116,12 @@ void r300_setup_routing(GLcontext *ctx, GLboolean immediate)
 	/* setup INPUT_ROUTE */
 	R300_STATECHANGE(r300, vir[0]);
 	for(i=0;i+1<count;i+=2){
-		dw=(r300->state.aos[i].ncomponents-1)
-		| ((r300->state.aos[i].reg)<<8)
-		| (r300->state.aos[i].format<<14)
-		| (((r300->state.aos[i+1].ncomponents-1)
-		| ((r300->state.aos[i+1].reg)<<8)
-		| (r300->state.aos[i+1].format<<14))<<16);
+		dw=(r300->state.aos[i].aos_size-1)
+		| ((r300->state.aos[i].aos_reg)<<8)
+		| (r300->state.aos[i].aos_format<<14)
+		| (((r300->state.aos[i+1].aos_size-1)
+		| ((r300->state.aos[i+1].aos_reg)<<8)
+		| (r300->state.aos[i+1].aos_format<<14))<<16);
 
 		if(i+2==count){
 			dw|=(1<<(13+16));
@@ -1136,9 +1129,9 @@ void r300_setup_routing(GLcontext *ctx, GLboolean immediate)
 		r300->hw.vir[0].cmd[R300_VIR_CNTL_0+(i>>1)]=dw;
 		}
 	if(count & 1){
-		dw=(r300->state.aos[count-1].ncomponents-1)
-		| (r300->state.aos[count-1].format<<14)
-		| ((r300->state.aos[count-1].reg)<<8)
+		dw=(r300->state.aos[count-1].aos_size-1)
+		| (r300->state.aos[count-1].aos_format<<14)
+		| ((r300->state.aos[count-1].aos_reg)<<8)
 		| (1<<13);
 		r300->hw.vir[0].cmd[R300_VIR_CNTL_0+(count>>1)]=dw;
 		//fprintf(stderr, "vir0 dw=%08x\n", dw);
@@ -1163,13 +1156,13 @@ void r300_setup_routing(GLcontext *ctx, GLboolean immediate)
 
 	for(i=0;i+1<count;i+=2){
 		/* do i first.. */
-		mask=(1<<(r300->state.aos[i].ncomponents*3))-1;
+		mask=(1<<(r300->state.aos[i].aos_size*3))-1;
 		dw=(ALL_COMPONENTS & mask)
 		| (ALL_DEFAULT & ~mask)
 		| R300_INPUT_ROUTE_ENABLE;
 
 		/* i+1 */
-		mask=(1<<(r300->state.aos[i+1].ncomponents*3))-1;
+		mask=(1<<(r300->state.aos[i+1].aos_size*3))-1;
 		dw|=(
 		(ALL_COMPONENTS & mask)
 		| (ALL_DEFAULT & ~mask)
@@ -1179,7 +1172,7 @@ void r300_setup_routing(GLcontext *ctx, GLboolean immediate)
 		r300->hw.vir[1].cmd[R300_VIR_CNTL_0+(i>>1)]=dw;
 		}
 	if(count & 1){
-		mask=(1<<(r300->state.aos[count-1].ncomponents*3))-1;
+		mask=(1<<(r300->state.aos[count-1].aos_size*3))-1;
 		dw=(ALL_COMPONENTS & mask)
 		| (ALL_DEFAULT & ~mask)
 		| R300_INPUT_ROUTE_ENABLE;
@@ -1833,7 +1826,8 @@ void r300ResetHwState(r300ContextPtr r300)
 
 	r300UpdateTextureState(ctx);
 	
-	r300_setup_routing(ctx, GL_TRUE);
+//	r300_setup_routing(ctx, GL_TRUE);
+	r300EmitArrays(ctx, GL_TRUE); /* Just do the routing */
 	r300_setup_textures(ctx);
 	r300_setup_rs_unit(ctx);
 
diff --git a/src/mesa/drivers/dri/r300/r300_texstate.c b/src/mesa/drivers/dri/r300/r300_texstate.c
index 8499ef89ab..0946680671 100644
--- a/src/mesa/drivers/dri/r300/r300_texstate.c
+++ b/src/mesa/drivers/dri/r300/r300_texstate.c
@@ -119,7 +119,7 @@ static const struct {
 	    {0xffffff16,  0},
 	    {0xffffff17, 0},
 	    };
-#elif
+#else
 	    {R300_EASY_TX_FORMAT(Y, Z, W, X, W8Z8Y8X8), 0},
 	    {0xffffff01, 0},
 	    {0xffffff02, 0},