r5xx: Moar LIT.

Still not working, but getting closer.

diff --git a/src/mesa/drivers/dri/r300/r500_fragprog.c b/src/mesa/drivers/dri/r300/r500_fragprog.c
index 81976573c6b881a2ec527273e68e9fc7b62fc54c..f55c8560c752197081f68a2e79a156d29a12950b 100644 (file)
--- a/src/mesa/drivers/dri/r300/r500_fragprog.c
+++ b/src/mesa/drivers/dri/r300/r500_fragprog.c
@@ -102,6 +102,10 @@
 /* Writemasks */
 #define R500_WRITEMASK_G 0x2
 #define R500_WRITEMASK_A 0x8
+#define R500_WRITEMASK_AR 0x9
+#define R500_WRITEMASK_AG 0xA
+#define R500_WRITEMASK_ARG 0xB
+#define R500_WRITEMASK_AB 0xC
 #define R500_WRITEMASK_ARGB 0xF
 
 /* 1/(2pi), needed for quick modulus in trig insts
@@ -678,20 +682,19 @@ static GLboolean parse_program(struct r500_fragment_program *fp)
                                        | R500_ALU_RGBA_ADDRD(dest);
                                break;
                        case OPCODE_LIT:
-                               /* I think I've got a pretty good path through this.
-                                * MAX temp1, tmp, [0, 0, 0, -128];
-                                * MIN temp1.w, temp1.w, [128];
-                                * POW temp1.z, temp1.y, temp1.w; (3 insts)
-                                * MOV result.xyzw, [1, temp1.y, temp1.z, 1]; */
+                               /* To be honest, I have no idea how I came up with the following.
+                                * All I know is that it's based on the r3xx stuff, and was
+                                * concieved with the help of NyQuil. Mmm, MyQuil. */
+
+                               /* First instruction */
                                src[0] = make_src(fp, fpi->SrcReg[0]);
                                src[1] = emit_const4fv(fp, LIT);
                                fp->inst[counter].inst0 = R500_INST_TYPE_ALU | R500_INST_TEX_SEM_WAIT
-                                       | (R500_WRITEMASK_ARGB << 11);
+                                       | (R500_WRITEMASK_ARG << 11);
                                fp->inst[counter].inst1 = R500_RGB_ADDR0(src[0]) | R500_RGB_ADDR1(src[1]);
                                fp->inst[counter].inst2 = R500_ALPHA_ADDR0(src[0]) | R500_ALPHA_ADDR1(src[1]);
                                fp->inst[counter].inst3 = R500_ALU_RGB_SEL_A_SRC0
                                        | MAKE_SWIZ_RGB_A(make_rgb_swizzle(fpi->SrcReg[0]))
-                                       | R500_ALU_RGB_SEL_B_SRC1
                                        | MAKE_SWIZ_RGB_B(R500_SWIZ_RGB_ZERO);
                                fp->inst[counter].inst4 = R500_ALPHA_OP_MAX
                                        | R500_ALPHA_ADDRD(get_temp(fp, 0))
@@ -700,81 +703,80 @@ static GLboolean parse_program(struct r500_fragment_program *fp)
                                fp->inst[counter].inst5 = R500_ALU_RGBA_OP_MAX
                                        | R500_ALU_RGBA_ADDRD(get_temp(fp, 0));
                                counter++;
-                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_A << 11);
+                               /* Second instruction */
+                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_AB << 11);
                                fp->inst[counter].inst1 = R500_RGB_ADDR0(get_temp(fp, 0));
                                fp->inst[counter].inst2 = R500_ALPHA_ADDR0(get_temp(fp, 0)) | R500_ALPHA_ADDR1(src[1]);
+                               /* Select [z, z, z, y] */
+                               temp_swiz = 2 | (2 << 3) | (2 << 6);
                                fp->inst[counter].inst3 = R500_ALU_RGB_SEL_A_SRC0
-                                       | MAKE_SWIZ_RGB_A(R500_SWIZ_RGB_RGB)
+                                       | MAKE_SWIZ_RGB_A(temp_swiz)
                                        | R500_ALU_RGB_SEL_B_SRC0
                                        | MAKE_SWIZ_RGB_B(R500_SWIZ_RGB_RGB);
-                               fp->inst[counter].inst4 = R500_ALPHA_OP_MAX
+                               fp->inst[counter].inst4 = R500_ALPHA_OP_LN2
                                        | R500_ALPHA_ADDRD(get_temp(fp, 0))
-                                       | R500_ALPHA_SEL_A_SRC0 | R500_ALPHA_SWIZ_A_A
-                                       | R500_ALPHA_SEL_B_SRC1 | R500_ALPHA_SWIZ_B_A | R500_ALPHA_MOD_B_NEG;
-                               fp->inst[counter].inst5 = R500_ALU_RGBA_OP_MAX
+                                       | R500_ALPHA_SEL_A_SRC0 | R500_ALPHA_SWIZ_A_G;
+                               fp->inst[counter].inst5 = R500_ALU_RGBA_OP_MIN
                                        | R500_ALU_RGBA_ADDRD(get_temp(fp, 0));
                                counter++;
-                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_ARGB << 11);
+                               /* Third instruction */
+                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_AG << 11);
                                fp->inst[counter].inst1 = R500_RGB_ADDR0(get_temp(fp, 0));
                                fp->inst[counter].inst2 = R500_ALPHA_ADDR0(get_temp(fp, 0));
-                               fp->inst[counter].inst3 = R500_ALU_RGB_SEL_A_SRC0
-                                       | MAKE_SWIZ_RGB_A(R500_SWIZ_RGB_RGB);
-                               fp->inst[counter].inst4 = R500_ALPHA_OP_LN2
-                                       | R500_ALPHA_ADDRD(get_temp(fp, 1))
-                                       | R500_ALPHA_SEL_A_SRC0 | R500_ALPHA_SWIZ_A_G;
-                               fp->inst[counter].inst5 = R500_ALU_RGBA_OP_SOP
-                                       | R500_ALU_RGBA_ADDRD(get_temp(fp, 1));
-                               counter++;
-                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_ARGB << 11);
-                               fp->inst[counter].inst1 = R500_RGB_ADDR0(get_temp(fp, 0))
-                                       | R500_RGB_ADDR1(get_temp(fp, 1));
-                               fp->inst[counter].inst2 = R500_ALPHA_ADDR0(get_temp(fp, 0))
-                                       | R500_ALPHA_ADDR1(get_temp(fp, 1));
-                               /* Select [w, w, w, w] */
-                               temp_swiz = 3 | (3 << 3) | (3 << 6);
+                               /* Select [x, x, x, z] */
+                               temp_swiz = 0;
                                fp->inst[counter].inst3 = R500_ALU_RGB_SEL_A_SRC0
                                        | MAKE_SWIZ_RGB_A(temp_swiz)
-                                       | R500_ALU_RGB_SEL_B_SRC1
-                                       | MAKE_SWIZ_RGB_B(R500_SWIZ_RGB_RGB);
+                                       | R500_ALU_RGB_SEL_B_SRC0
+                                       | MAKE_SWIZ_RGB_B(R500_SWIZ_RGB_ONE);
                                fp->inst[counter].inst4 = R500_ALPHA_OP_MAD
                                        | R500_ALPHA_ADDRD(get_temp(fp, 1))
-                                       | R500_ALPHA_SEL_A_SRC0 | R500_ALPHA_SWIZ_A_A
-                                       | R500_ALPHA_SEL_B_SRC1 | R500_ALPHA_SWIZ_B_A;
+                                       | R500_ALPHA_SEL_A_SRC0 | R500_ALPHA_SWIZ_A_B
+                                       | R500_ALPHA_SEL_B_SRC0 | R500_ALPHA_SWIZ_B_A;
                                fp->inst[counter].inst5 = R500_ALU_RGBA_OP_MAD
                                        | R500_ALU_RGBA_ADDRD(get_temp(fp, 1))
                                        | MAKE_SWIZ_RGBA_C(R500_SWIZ_RGB_ZERO)
-                                       | MAKE_SWIZ_ALPHA_C(R500_SWIZZLE_ZERO);
+                                       | R500_ALU_RGBA_A_SWIZ_0;
                                counter++;
-                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_G << 11);
-                               fp->inst[counter].inst1 = R500_RGB_ADDR0(get_temp(fp, 1));
-                               fp->inst[counter].inst2 = R500_ALPHA_ADDR0(get_temp(fp, 1));
-                               fp->inst[counter].inst3 = R500_ALU_RGB_SEL_A_SRC0
-                                       | MAKE_SWIZ_RGB_A(R500_SWIZ_RGB_RGB);
+                               /* Fourth instruction */
+                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_AR << 11);
+                               fp->inst[counter].inst2 = R500_ALPHA_ADDR0(get_temp(fp, 0));
+                               fp->inst[counter].inst3 = MAKE_SWIZ_RGB_A(R500_SWIZ_RGB_ONE)
+                                       | MAKE_SWIZ_RGB_B(R500_SWIZ_RGB_ONE);
                                fp->inst[counter].inst4 = R500_ALPHA_OP_EX2
                                        | R500_ALPHA_ADDRD(get_temp(fp, 0))
                                        | R500_ALPHA_SEL_A_SRC0 | R500_ALPHA_SWIZ_A_A;
-                               fp->inst[counter].inst5 = R500_ALU_RGBA_OP_SOP
-                                       | R500_ALU_RGBA_ADDRD(get_temp(fp, 0));
+                               fp->inst[counter].inst5 = R500_ALU_RGBA_OP_MAD
+                                       | R500_ALU_RGBA_ADDRD(get_temp(fp, 0))
+                                       | MAKE_SWIZ_RGBA_C(R500_SWIZ_RGB_ZERO)
+                                       | MAKE_SWIZ_ALPHA_C(R500_SWIZZLE_ZERO);
                                counter++;
-                               emit_alu(fp, counter, fpi);
+                               /* Fifth instruction */
+                               fp->inst[counter].inst0 = R500_INST_TYPE_ALU | (R500_WRITEMASK_AB << 11);
                                fp->inst[counter].inst1 = R500_RGB_ADDR0(get_temp(fp, 0));
-                               fp->inst[counter].inst2 = R500_ALPHA_ADDR0(get_temp(fp, 0));
-                               /* Select [1, y, z, 1] */
-                               temp_swiz = R500_SWIZZLE_ONE | (2 << 3) | (3 << 6);
+                               /* Select [w, w, w] */
+                               temp_swiz = 3 | (3 << 3) | (3 << 6);
                                fp->inst[counter].inst3 = R500_ALU_RGB_SEL_A_SRC0
-                                       | MAKE_SWIZ_RGB_A(temp_swiz)
+                                       | MAKE_SWIZ_RGB_A(R500_SWIZ_RGB_ZERO)
                                        | R500_ALU_RGB_SEL_B_SRC0
-                                       | MAKE_SWIZ_RGB_B(temp_swiz)
-                                       | R500_ALU_RGB_OMOD_DISABLE;
-                               fp->inst[counter].inst4 |= R500_ALPHA_OP_CMP
-                                       | R500_ALPHA_ADDRD(dest)
-                                       | R500_ALPHA_SEL_A_SRC0 | R500_ALPHA_SWIZ_A_1
-                                       | R500_ALPHA_SEL_B_SRC0 | R500_ALPHA_SWIZ_B_1
-                                       | R500_ALPHA_OMOD_DISABLE;
+                                       | MAKE_SWIZ_RGB_B(temp_swiz);
+                               fp->inst[counter].inst4 |= R500_ALPHA_OP_MAD
+                                       | R500_ALPHA_ADDRD(get_temp(fp, 0))
+                                       | R500_ALPHA_SWIZ_A_1
+                                       | R500_ALPHA_SWIZ_B_1;
+                               /* Select [-y, -y, -y] */
+                               temp_swiz = 1 | (1 << 3) | (1 << 6);
                                fp->inst[counter].inst5 = R500_ALU_RGBA_OP_CMP
-                                       | R500_ALU_RGBA_ADDRD(dest)
-                                       | MAKE_SWIZ_RGBA_C(R500_SWIZ_RGB_ZERO)
+                                       | R500_ALU_RGBA_ADDRD(get_temp(fp, 0))
+                                       | MAKE_SWIZ_RGBA_C(temp_swiz)
+                                       | R500_ALU_RGBA_MOD_C_NEG
                                        | MAKE_SWIZ_ALPHA_C(R500_SWIZZLE_ZERO);
+                               counter++;
+                               /* Final instruction */
+                               emit_alu(fp, counter, fpi);
+                               fpi->SrcReg[0].Index = get_temp(fp, 0);
+                               fpi->SrcReg[0].Swizzle = 1672;
+                               emit_mov(fp, counter, fpi->SrcReg[0], dest);
                                break;
                        case OPCODE_LRP:
                                /* src0 * src1 + INV(src0) * src2