nvfx: implement DP2 in vp and fp

This commit is contained in:
Luca Barbieri 2010-08-21 19:43:46 +02:00
parent 4aec8aa2e3
commit 32d2525d64
2 changed files with 10 additions and 0 deletions

View file

@ -573,6 +573,11 @@ nvfx_fragprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_fpc *fpc,
nvfx_fp_emit(fpc, arith(sat, DDY, dst, mask, src[0], none, none));
}
break;
case TGSI_OPCODE_DP2:
tmp = nvfx_src(temp(fpc));
nvfx_fp_emit(fpc, arith(0, MUL, tmp.reg, NVFX_FP_MASK_X | NVFX_FP_MASK_Y, src[0], src[1], none));
nvfx_fp_emit(fpc, arith(0, ADD, dst, mask, swz(tmp, X, X, X, X), swz(tmp, Y, Y, Y, Y), none));
break;
case TGSI_OPCODE_DP3:
nvfx_fp_emit(fpc, arith(sat, DP3, dst, mask, src[0], src[1], none));
break;

View file

@ -514,6 +514,11 @@ nvfx_vertprog_parse_instruction(struct nvfx_context* nvfx, struct nvfx_vpc *vpc,
case TGSI_OPCODE_COS:
nvfx_vp_emit(vpc, arith(SCA, COS, dst, mask, none, none, src[0]));
break;
case TGSI_OPCODE_DP2:
tmp = nvfx_src(temp(vpc));
nvfx_vp_emit(vpc, arith(VEC, MUL, tmp.reg, NVFX_VP_MASK_X | NVFX_VP_MASK_Y, src[0], src[1], none));
nvfx_vp_emit(vpc, arith(VEC, ADD, dst, mask, swz(tmp, X, X, X, X), swz(tmp, Y, Y, Y, Y), none));
break;
case TGSI_OPCODE_DP3:
nvfx_vp_emit(vpc, arith(VEC, DP3, dst, mask, src[0], src[1], none));
break;