#include "tnl/tnl.h"
#include "compiler/radeon_compiler.h"
-#include "compiler/radeon_nqssadce.h"
+#include "radeon_mesa_to_rc.h"
#include "r300_context.h"
+#include "r300_fragprog_common.h"
#include "r300_state.h"
/**
}
}
- if (vp->code.constants.Count * 4 > VSF_MAX_FRAGMENT_LENGTH) {
- /* Should have checked this earlier... */
- fprintf(stderr, "%s:Params exhausted\n", __FUNCTION__);
- _mesa_exit(-1);
- }
-
for(i = 0; i < vp->code.constants.Count; ++i) {
const float * src = 0;
const struct rc_constant * constant = &vp->code.constants.Constants[i];
return outputs;
}
+
+static void t_inputs_outputs(struct r300_vertex_program_compiler * c)
+{
+ int i;
+ int cur_reg;
+ GLuint OutputsWritten, InputsRead;
+
+ OutputsWritten = c->Base.Program.OutputsWritten;
+ InputsRead = c->Base.Program.InputsRead;
+
+ cur_reg = -1;
+ for (i = 0; i < VERT_ATTRIB_MAX; i++) {
+ if (InputsRead & (1 << i))
+ c->code->inputs[i] = ++cur_reg;
+ else
+ c->code->inputs[i] = -1;
+ }
+
+ cur_reg = 0;
+ for (i = 0; i < VERT_RESULT_MAX; i++)
+ c->code->outputs[i] = -1;
+
+ assert(OutputsWritten & (1 << VERT_RESULT_HPOS));
+
+ if (OutputsWritten & (1 << VERT_RESULT_HPOS)) {
+ c->code->outputs[VERT_RESULT_HPOS] = cur_reg++;
+ }
+
+ if (OutputsWritten & (1 << VERT_RESULT_PSIZ)) {
+ c->code->outputs[VERT_RESULT_PSIZ] = cur_reg++;
+ }
+
+ /* If we're writing back facing colors we need to send
+ * four colors to make front/back face colors selection work.
+ * If the vertex program doesn't write all 4 colors, lets
+ * pretend it does by skipping output index reg so the colors
+ * get written into appropriate output vectors.
+ */
+ if (OutputsWritten & (1 << VERT_RESULT_COL0)) {
+ c->code->outputs[VERT_RESULT_COL0] = cur_reg++;
+ } else if (OutputsWritten & (1 << VERT_RESULT_BFC0) ||
+ OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+ cur_reg++;
+ }
+
+ if (OutputsWritten & (1 << VERT_RESULT_COL1)) {
+ c->code->outputs[VERT_RESULT_COL1] = cur_reg++;
+ } else if (OutputsWritten & (1 << VERT_RESULT_BFC0) ||
+ OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+ cur_reg++;
+ }
+
+ if (OutputsWritten & (1 << VERT_RESULT_BFC0)) {
+ c->code->outputs[VERT_RESULT_BFC0] = cur_reg++;
+ } else if (OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+ cur_reg++;
+ }
+
+ if (OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+ c->code->outputs[VERT_RESULT_BFC1] = cur_reg++;
+ } else if (OutputsWritten & (1 << VERT_RESULT_BFC0)) {
+ cur_reg++;
+ }
+
+ for (i = VERT_RESULT_TEX0; i <= VERT_RESULT_TEX7; i++) {
+ if (OutputsWritten & (1 << i)) {
+ c->code->outputs[i] = cur_reg++;
+ }
+ }
+
+ if (OutputsWritten & (1 << VERT_RESULT_FOGC)) {
+ c->code->outputs[VERT_RESULT_FOGC] = cur_reg++;
+ }
+}
+
+/**
+ * The NV_vertex_program spec mandates that all registers be
+ * initialized to zero. We do this here unconditionally.
+ *
+ * \note We rely on dead-code elimination in the compiler.
+ */
+static void initialize_NV_registers(struct radeon_compiler * compiler)
+{
+ unsigned int reg;
+ struct rc_instruction * inst;
+
+ for(reg = 0; reg < 12; ++reg) {
+ inst = rc_insert_new_instruction(compiler, &compiler->Program.Instructions);
+ inst->U.I.Opcode = RC_OPCODE_MOV;
+ inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
+ inst->U.I.DstReg.Index = reg;
+ inst->U.I.SrcReg[0].File = RC_FILE_NONE;
+ inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_0000;
+ }
+
+ inst = rc_insert_new_instruction(compiler, &compiler->Program.Instructions);
+ inst->U.I.Opcode = RC_OPCODE_ARL;
+ inst->U.I.DstReg.File = RC_FILE_ADDRESS;
+ inst->U.I.DstReg.Index = 0;
+ inst->U.I.DstReg.WriteMask = WRITEMASK_X;
+ inst->U.I.SrcReg[0].File = RC_FILE_NONE;
+ inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_0000;
+}
+
static struct r300_vertex_program *build_program(GLcontext *ctx,
struct r300_vertex_program_key *wanted_key,
const struct gl_vertex_program *mesa_vp)
_mesa_memcpy(&vp->key, wanted_key, sizeof(vp->key));
rc_init(&compiler.Base);
- compiler.Base.Debug = (RADEON_DEBUG & DEBUG_VERTS) ? GL_TRUE : GL_FALSE;
+ compiler.Base.Debug = (RADEON_DEBUG & RADEON_VERTS) ? GL_TRUE : GL_FALSE;
compiler.code = &vp->code;
- compiler.state.FogAttr = vp->key.FogAttr;
- compiler.state.WPosAttr = vp->key.WPosAttr;
compiler.RequiredOutputs = compute_required_outputs(vp->Base, vp->key.FpReads);
- compiler.program = &vp->Base->Base;
+ compiler.SetHwInputOutput = &t_inputs_outputs;
if (compiler.Base.Debug) {
fprintf(stderr, "Initial vertex program:\n");
- _mesa_print_program(compiler.program);
- fflush(stdout);
+ _mesa_print_program(&vp->Base->Base);
+ fflush(stderr);
}
if (mesa_vp->IsPositionInvariant) {
- _mesa_insert_mvp_code(ctx, (struct gl_vertex_program *)compiler.program);
+ _mesa_insert_mvp_code(ctx, vp->Base);
+ }
+
+ radeon_mesa_to_rc_program(&compiler.Base, &vp->Base->Base);
+
+ if (mesa_vp->IsNVProgram)
+ initialize_NV_registers(&compiler.Base);
+
+ rc_move_output(&compiler.Base, VERT_RESULT_PSIZ, VERT_RESULT_PSIZ, WRITEMASK_X);
+
+ if (vp->key.WPosAttr != FRAG_ATTRIB_MAX) {
+ rc_copy_output(&compiler.Base,
+ VERT_RESULT_HPOS,
+ vp->key.WPosAttr - FRAG_ATTRIB_TEX0 + VERT_RESULT_TEX0);
+ }
+
+ if (vp->key.FogAttr != FRAG_ATTRIB_MAX) {
+ rc_move_output(&compiler.Base,
+ VERT_RESULT_FOGC,
+ vp->key.FogAttr - FRAG_ATTRIB_TEX0 + VERT_RESULT_TEX0, WRITEMASK_X);
}
r3xx_compile_vertex_program(&compiler);
+
+ if (vp->code.constants.Count > ctx->Const.VertexProgram.MaxParameters) {
+ rc_error(&compiler.Base, "Program exceeds constant buffer size limit\n");
+ }
+
vp->error = compiler.Base.Error;
+ vp->Base->Base.InputsRead = vp->code.InputsRead;
+ vp->Base->Base.OutputsWritten = vp->code.OutputsWritten;
+
rc_destroy(&compiler.Base);
return vp;
struct r300_vertex_program *vp;
vpc = (struct r300_vertex_program_cont *)ctx->VertexProgram._Current;
+
+ if (!r300->selected_fp) {
+ /* This can happen when GetProgramiv is called to check
+ * whether the program runs natively.
+ *
+ * To be honest, this is not a very good solution,
+ * but solving the problem of reporting good values
+ * for those queries is tough anyway considering that
+ * we recompile vertex programs based on the precise
+ * fragment program that is in use.
+ */
+ r300SelectAndTranslateFragmentShader(ctx);
+ }
+
wanted_key.FpReads = r300->selected_fp->InputsRead;
- wanted_key.FogAttr = r300->selected_fp->code.fog_attr;
- wanted_key.WPosAttr = r300->selected_fp->code.wpos_attr;
+ wanted_key.FogAttr = r300->selected_fp->fog_attr;
+ wanted_key.WPosAttr = r300->selected_fp->wpos_attr;
for (vp = vpc->progs; vp; vp = vp->next) {
if (_mesa_memcmp(&vp->key, &wanted_key, sizeof(wanted_key))
#define bump_vpu_count(ptr, new_count) do { \
drm_r300_cmd_header_t* _p=((drm_r300_cmd_header_t*)(ptr)); \
int _nc=(new_count)/4; \
- assert(_nc < 256); \
if(_nc>_p->vpu.count)_p->vpu.count=_nc; \
} while(0)
assert((code->length > 0) && (code->length % 4 == 0));
+ R300_STATECHANGE( r300, vap_flush );
+
switch ((dest >> 8) & 0xf) {
case 0:
R300_STATECHANGE(r300, vpi);
break;
default:
fprintf(stderr, "%s:%s don't know how to handle dest %04x\n", __FILE__, __FUNCTION__, dest);
- _mesa_exit(-1);
+ exit(-1);
}
}
((drm_r300_cmd_header_t *) rmesa->hw.vpi.cmd)->vpu.count = 0;
((drm_r300_cmd_header_t *) rmesa->hw.vps.cmd)->vpu.count = 0;
+ R300_STATECHANGE(rmesa, vap_flush);
R300_STATECHANGE(rmesa, vpp);
param_count = r300VertexProgUpdateParams(ctx, prog, (float *)&rmesa->hw.vpp.cmd[R300_VPP_PARAM_0]);
bump_vpu_count(rmesa->hw.vpp.cmd, param_count);