r300g: implement TRUNC correctly

This fixes some integer division tests.

diff --git a/src/gallium/drivers/r300/compiler/radeon_opcodes.c b/src/gallium/drivers/r300/compiler/radeon_opcodes.c
index 9bcb3c9..916baa2 100644 (file)
--- a/src/gallium/drivers/r300/compiler/radeon_opcodes.c
+++ b/src/gallium/drivers/r300/compiler/radeon_opcodes.c
@@ -343,6 +343,13 @@ struct rc_opcode_info rc_opcodes[MAX_RC_OPCODE] = {
                .IsComponentwise = 1
        },
        {
+               .Opcode = RC_OPCODE_TRUNC,
+               .Name = "TRUNC",
+               .NumSrcRegs = 1,
+               .HasDstReg = 1,
+               .IsComponentwise = 1
+       },
+       {
                .Opcode = RC_OPCODE_XPD,
                .Name = "XPD",
                .NumSrcRegs = 2,

diff --git a/src/gallium/drivers/r300/compiler/radeon_opcodes.h b/src/gallium/drivers/r300/compiler/radeon_opcodes.h
index 9c4b456..0a70901 100644 (file)
--- a/src/gallium/drivers/r300/compiler/radeon_opcodes.h
+++ b/src/gallium/drivers/r300/compiler/radeon_opcodes.h
@@ -175,6 +175,9 @@ typedef enum {
        /** vec4 instruction: dst.c = src0.c */
        RC_OPCODE_SWZ,
 
+       /** vec4 instruction: dst.c = (abs(src0.c) - fract(abs(src0.c))) * sgn(src0.c) */
+       RC_OPCODE_TRUNC,
+
        /** special instruction, see ARB_fragment_program */
        RC_OPCODE_XPD,
 

diff --git a/src/gallium/drivers/r300/compiler/radeon_program_alu.c b/src/gallium/drivers/r300/compiler/radeon_program_alu.c
index b3da311..f4ee86d 100644 (file)
--- a/src/gallium/drivers/r300/compiler/radeon_program_alu.c
+++ b/src/gallium/drivers/r300/compiler/radeon_program_alu.c
@@ -321,6 +321,24 @@ static void transform_FLR(struct radeon_compiler* c,
        rc_remove_instruction(inst);
 }
 
+static void transform_TRUNC(struct radeon_compiler* c,
+       struct rc_instruction* inst)
+{
+       /* Definition of trunc:
+        *   trunc(x) = (abs(x) - fract(abs(x))) * sgn(x)
+        *
+        * The multiplication by sgn(x) can be simplified using CMP:
+        *   y * sgn(x) = (x < 0 ? -y : y)
+        */
+       struct rc_dst_register dst = try_to_reuse_dst(c, inst);
+       emit1(c, inst->Prev, RC_OPCODE_FRC, 0, dst, absolute(inst->U.I.SrcReg[0]));
+       emit2(c, inst->Prev, RC_OPCODE_ADD, 0, dst, absolute(inst->U.I.SrcReg[0]),
+             negate(srcreg(RC_FILE_TEMPORARY, dst.Index)));
+       emit3(c, inst->Prev, RC_OPCODE_CMP, &inst->U.I, inst->U.I.DstReg, inst->U.I.SrcReg[0],
+             negate(srcreg(RC_FILE_TEMPORARY, dst.Index)), srcreg(RC_FILE_TEMPORARY, dst.Index));
+       rc_remove_instruction(inst);
+}
+
 /**
  * Definition of LIT (from ARB_fragment_program):
  *
@@ -666,6 +684,7 @@ int radeonTransformALU(
        case RC_OPCODE_SSG: transform_SSG(c, inst); return 1;
        case RC_OPCODE_SUB: transform_SUB(c, inst); return 1;
        case RC_OPCODE_SWZ: transform_SWZ(c, inst); return 1;
+       case RC_OPCODE_TRUNC: transform_TRUNC(c, inst); return 1;
        case RC_OPCODE_XPD: transform_XPD(c, inst); return 1;
        default:
                return 0;
@@ -866,6 +885,17 @@ static void transform_r300_vertex_SSG(struct radeon_compiler* c,
        rc_remove_instruction(inst);
 }
 
+static void transform_vertex_TRUNC(struct radeon_compiler* c,
+       struct rc_instruction* inst)
+{
+       struct rc_instruction *next = inst->Next;
+
+       /* next->Prev is removed after each transformation and replaced
+        * by a new instruction. */
+       transform_TRUNC(c, next->Prev);
+       transform_r300_vertex_CMP(c, next->Prev);
+}
+
 /**
  * For use with rc_local_transform, this transforms non-native ALU
  * instructions of the r300 up to r500 vertex engine.
@@ -904,6 +934,7 @@ int r300_transform_vertex_alu(
        case RC_OPCODE_SSG: transform_r300_vertex_SSG(c, inst); return 1;
        case RC_OPCODE_SUB: transform_SUB(c, inst); return 1;
        case RC_OPCODE_SWZ: transform_SWZ(c, inst); return 1;
+       case RC_OPCODE_TRUNC: transform_vertex_TRUNC(c, inst); return 1;
        case RC_OPCODE_XPD: transform_XPD(c, inst); return 1;
        default:
                return 0;

diff --git a/src/gallium/drivers/r300/r300_tgsi_to_rc.c b/src/gallium/drivers/r300/r300_tgsi_to_rc.c
index 4cb08b5..a0587b4 100644 (file)
--- a/src/gallium/drivers/r300/r300_tgsi_to_rc.c
+++ b/src/gallium/drivers/r300/r300_tgsi_to_rc.c
@@ -114,7 +114,7 @@ static unsigned translate_opcode(unsigned opcode)
         case TGSI_OPCODE_CEIL: return RC_OPCODE_CEIL;
      /* case TGSI_OPCODE_I2F: return RC_OPCODE_I2F; */
      /* case TGSI_OPCODE_NOT: return RC_OPCODE_NOT; */
-        case TGSI_OPCODE_TRUNC: return RC_OPCODE_FLR;
+        case TGSI_OPCODE_TRUNC: return RC_OPCODE_TRUNC;
      /* case TGSI_OPCODE_SHL: return RC_OPCODE_SHL; */
      /* case TGSI_OPCODE_ISHR: return RC_OPCODE_SHR; */
      /* case TGSI_OPCODE_AND: return RC_OPCODE_AND; */