use DP3 instead.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
dp2_emit /* emit */
};
-/* TGSI_OPCODE_DP2A */
-static void
-dp2a_fetch_args(
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
-{
- dp_fetch_args(bld_base, emit_data, 2);
- emit_data->args[5] = lp_build_emit_fetch(bld_base, emit_data->inst,
- 2, TGSI_CHAN_X);
-}
-
-static void
-dp2a_emit(
- const struct lp_build_tgsi_action * action,
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
-{
- LLVMValueRef tmp;
- tmp = lp_build_emit_llvm(bld_base, TGSI_OPCODE_DP2, emit_data);
- emit_data->output[emit_data->chan] = lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_ADD,
- emit_data->args[5], tmp);
-}
-
-static struct lp_build_tgsi_action dp2a_action = {
- dp2a_fetch_args, /* fetch_args */
- dp2a_emit /* emit */
-};
-
/* TGSI_OPCODE_DP3 */
static void
dp3_fetch_args(
bld_base->op_actions[TGSI_OPCODE_DP2] = dp2_action;
bld_base->op_actions[TGSI_OPCODE_DP3] = dp3_action;
bld_base->op_actions[TGSI_OPCODE_DP4] = dp4_action;
- bld_base->op_actions[TGSI_OPCODE_DP2A] = dp2a_action;
bld_base->op_actions[TGSI_OPCODE_DPH] = dph_action;
bld_base->op_actions[TGSI_OPCODE_DST] = dst_action;
bld_base->op_actions[TGSI_OPCODE_EXP] = exp_action;
dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2);
break;
- case TGSI_OPCODE_DP2A:
- return FALSE;
-
case TGSI_OPCODE_FRC:
src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL);
tmp0 = lp_build_floor(&bld->bld_base.base, src0);
ttn_move_dest_masked(b, dest, nir_imm_float(b, 1.0), TGSI_WRITEMASK_W);
}
-static void
-ttn_dp2a(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
-{
- ttn_move_dest(b, dest,
- ttn_channel(b, nir_fadd(b, nir_fdot2(b, src[0], src[1]),
- src[2]),
- X));
-}
-
static void
ttn_dp2(nir_builder *b, nir_op op, nir_alu_dest dest, nir_ssa_def **src)
{
[TGSI_OPCODE_MAD] = nir_op_ffma,
[TGSI_OPCODE_LRP] = 0,
[TGSI_OPCODE_SQRT] = nir_op_fsqrt,
- [TGSI_OPCODE_DP2A] = 0,
[TGSI_OPCODE_FRC] = nir_op_ffract,
[TGSI_OPCODE_FLR] = nir_op_ffloor,
[TGSI_OPCODE_ROUND] = nir_op_fround_even,
ttn_dp4(b, op_trans[tgsi_op], dest, src);
break;
- case TGSI_OPCODE_DP2A:
- ttn_dp2a(b, op_trans[tgsi_op], dest, src);
- break;
-
case TGSI_OPCODE_DPH:
ttn_dph(b, op_trans[tgsi_op], dest, src);
break;
}
}
-static void
-exec_dp2a(struct tgsi_exec_machine *mach,
- const struct tgsi_full_instruction *inst)
-{
- unsigned int chan;
- union tgsi_exec_channel arg[3];
-
- fetch_source(mach, &arg[0], &inst->Src[0], TGSI_CHAN_X, TGSI_EXEC_DATA_FLOAT);
- fetch_source(mach, &arg[1], &inst->Src[1], TGSI_CHAN_X, TGSI_EXEC_DATA_FLOAT);
- micro_mul(&arg[2], &arg[0], &arg[1]);
-
- fetch_source(mach, &arg[0], &inst->Src[0], TGSI_CHAN_Y, TGSI_EXEC_DATA_FLOAT);
- fetch_source(mach, &arg[1], &inst->Src[1], TGSI_CHAN_Y, TGSI_EXEC_DATA_FLOAT);
- micro_mad(&arg[0], &arg[0], &arg[1], &arg[2]);
-
- fetch_source(mach, &arg[1], &inst->Src[2], TGSI_CHAN_X, TGSI_EXEC_DATA_FLOAT);
- micro_add(&arg[0], &arg[0], &arg[1]);
-
- for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
- if (inst->Dst[0].Register.WriteMask & (1 << chan)) {
- store_dest(mach, &arg[0], &inst->Dst[0], inst, chan, TGSI_EXEC_DATA_FLOAT);
- }
- }
-}
-
static void
exec_dph(struct tgsi_exec_machine *mach,
const struct tgsi_full_instruction *inst)
exec_scalar_unary(mach, inst, micro_sqrt, TGSI_EXEC_DATA_FLOAT, TGSI_EXEC_DATA_FLOAT);
break;
- case TGSI_OPCODE_DP2A:
- exec_dp2a(mach, inst);
- break;
-
case TGSI_OPCODE_FRC:
exec_vector_unary(mach, inst, micro_frc, TGSI_EXEC_DATA_FLOAT, TGSI_EXEC_DATA_FLOAT);
break;
{ 1, 3, 0, 0, 0, 0, 0, COMP, "LRP", TGSI_OPCODE_LRP },
{ 1, 3, 0, 0, 0, 0, 0, COMP, "FMA", TGSI_OPCODE_FMA },
{ 1, 1, 0, 0, 0, 0, 0, REPL, "SQRT", TGSI_OPCODE_SQRT },
- { 1, 3, 0, 0, 0, 0, 0, REPL, "DP2A", TGSI_OPCODE_DP2A },
+ { 1, 3, 0, 0, 0, 0, 0, REPL, "", 21 }, /* removed */
{ 1, 1, 0, 0, 0, 0, 0, COMP, "F2U64", TGSI_OPCODE_F2U64 },
{ 1, 1, 0, 0, 0, 0, 0, COMP, "F2I64", TGSI_OPCODE_F2I64 },
{ 1, 1, 0, 0, 0, 0, 0, COMP, "FRC", TGSI_OPCODE_FRC },
* DP2 - 2-component Dot Product
* dst = src0.x \times src1.x + src0.y \times src1.y
*
- * DP2A - 2-component Dot Product And Add
- * dst = src0.x \times src1.x + src0.y \times src1.y + src2.x
- *
* NOTE: these are translated into sequence of MUL/MAD(/ADD) scalar
* operations, which is what you'd prefer for a ISA that is natively
* scalar. Probably a native vector ISA would at least already have
* DP4/DP3 instructions, but perhaps there is room for an alternative
- * translation for DPH/DP2/DP2A using vector instructions.
+ * translation for DPH/DP2 using vector instructions.
*
* ; needs: 1 tmp
* MUL tmpA.x, src0.x, src1.x
* } else if (DP4) {
* MAD tmpA.x, src0.w, src1.w, tmpA.x
* }
- * } else if (DP2A) {
- * ADD tmpA.x, src2.x, tmpA.x
* }
* ; fixup last instruction to replicate into dst
*/
#define DP3_GROW (NINST(2) + NINST(3) + NINST(3) - OINST(2))
#define DPH_GROW (NINST(2) + NINST(3) + NINST(3) + NINST(2) - OINST(2))
#define DP2_GROW (NINST(2) + NINST(3) - OINST(2))
-#define DP2A_GROW (NINST(2) + NINST(3) + NINST(2) - OINST(3))
#define DOTP_TMP 1
static void
transform_dotp(struct tgsi_transform_context *tctx,
struct tgsi_full_dst_register *dst = &inst->Dst[0];
struct tgsi_full_src_register *src0 = &inst->Src[0];
struct tgsi_full_src_register *src1 = &inst->Src[1];
- struct tgsi_full_src_register *src2 = &inst->Src[2]; /* only DP2A */
struct tgsi_full_instruction new_inst;
unsigned opcode = inst->Instruction.Opcode;
reg_src(&new_inst.Src[1], src1, SWIZ(W, W, W, W));
reg_src(&new_inst.Src[2], &ctx->tmp[A].src, SWIZ(X, X, X, X));
}
- } else if (opcode == TGSI_OPCODE_DP2A) {
- tctx->emit_instruction(tctx, &new_inst);
-
- /* ADD tmpA.x, src2.x, tmpA.x */
- new_inst = tgsi_default_full_instruction();
- new_inst.Instruction.Opcode = TGSI_OPCODE_ADD;
- new_inst.Instruction.NumDstRegs = 1;
- reg_dst(&new_inst.Dst[0], &ctx->tmp[A].dst, TGSI_WRITEMASK_X);
- new_inst.Instruction.NumSrcRegs = 2;
- reg_src(&new_inst.Src[0], src2, SWIZ(X, X, X, X));
- reg_src(&new_inst.Src[1], &ctx->tmp[A].src, SWIZ(X, X, X, X));
}
/* fixup last instruction to write to dst: */
goto skip;
transform_dotp(tctx, inst);
break;
- case TGSI_OPCODE_DP2A:
- if (!ctx->config->lower_DP2A)
- goto skip;
- transform_dotp(tctx, inst);
- break;
case TGSI_OPCODE_FLR:
if (!ctx->config->lower_FLR)
goto skip;
OPCS(DP3) ||
OPCS(DPH) ||
OPCS(DP2) ||
- OPCS(DP2A) ||
OPCS(FLR) ||
OPCS(CEIL) ||
OPCS(TRUNC) ||
newlen += DP2_GROW * OPCS(DP2);
numtmp = MAX2(numtmp, DOTP_TMP);
}
- if (OPCS(DP2A)) {
- newlen += DP2A_GROW * OPCS(DP2A);
- numtmp = MAX2(numtmp, DOTP_TMP);
- }
if (OPCS(FLR)) {
newlen += FLR_GROW * OPCS(FLR);
numtmp = MAX2(numtmp, FLR_TMP);
unsigned lower_DP3:1;
unsigned lower_DPH:1;
unsigned lower_DP2:1;
- unsigned lower_DP2A:1;
unsigned lower_FLR:1;
unsigned lower_CEIL:1;
unsigned lower_TRUNC:1;
OP12_TEX(TEX_LZ)
OP13(LRP)
OP11(SQRT)
-OP13(DP2A)
OP11(FRC)
OP12_TEX(TXF_LZ)
OP11(FLR)
read_mask = write_mask & TGSI_WRITEMASK_XYZ ? TGSI_WRITEMASK_X : 0;
break;
- case TGSI_OPCODE_DP2A:
- read_mask = src_idx == 2 ? TGSI_WRITEMASK_X : TGSI_WRITEMASK_XY;
- break;
-
case TGSI_OPCODE_DP2:
read_mask = TGSI_WRITEMASK_XY;
break;
dst.w = src0.w \times src1.w + src2.w
-.. opcode:: DP2A - 2-component Dot Product And Add
-
-.. math::
-
- dst.x = src0.x \times src1.x + src0.y \times src1.y + src2.x
-
- dst.y = src0.x \times src1.x + src0.y \times src1.y + src2.x
-
- dst.z = src0.x \times src1.x + src0.y \times src1.y + src2.x
-
- dst.w = src0.x \times src1.x + src0.y \times src1.y + src2.x
-
-
.. opcode:: FRC - Fraction
.. math::
.lower_EXP = true,
.lower_LOG = true,
.lower_DP2 = true,
- .lower_DP2A = true,
.lower_TRUNC = true,
.lower_XPD = true
};
case TGSI_OPCODE_SGE: return RC_OPCODE_SGE;
case TGSI_OPCODE_MAD: return RC_OPCODE_MAD;
case TGSI_OPCODE_LRP: return RC_OPCODE_LRP;
- /* case TGSI_OPCODE_DP2A: return RC_OPCODE_DP2A; */
case TGSI_OPCODE_FRC: return RC_OPCODE_FRC;
case TGSI_OPCODE_FLR: return RC_OPCODE_FLR;
case TGSI_OPCODE_ROUND: return RC_OPCODE_ROUND;
[TGSI_OPCODE_LRP] = { ALU_OP0_NOP, tgsi_lrp},
[TGSI_OPCODE_FMA] = { ALU_OP0_NOP, tgsi_unsupported},
[TGSI_OPCODE_SQRT] = { ALU_OP1_SQRT_IEEE, tgsi_trans_srcx_replicate},
- [TGSI_OPCODE_DP2A] = { ALU_OP0_NOP, tgsi_unsupported},
+ [21] = { ALU_OP0_NOP, tgsi_unsupported},
[22] = { ALU_OP0_NOP, tgsi_unsupported},
[23] = { ALU_OP0_NOP, tgsi_unsupported},
[TGSI_OPCODE_FRC] = { ALU_OP1_FRACT, tgsi_op2},
[TGSI_OPCODE_LRP] = { ALU_OP0_NOP, tgsi_lrp},
[TGSI_OPCODE_FMA] = { ALU_OP3_FMA, tgsi_op3},
[TGSI_OPCODE_SQRT] = { ALU_OP1_SQRT_IEEE, tgsi_trans_srcx_replicate},
- [TGSI_OPCODE_DP2A] = { ALU_OP0_NOP, tgsi_unsupported},
+ [21] = { ALU_OP0_NOP, tgsi_unsupported},
[22] = { ALU_OP0_NOP, tgsi_unsupported},
[23] = { ALU_OP0_NOP, tgsi_unsupported},
[TGSI_OPCODE_FRC] = { ALU_OP1_FRACT, tgsi_op2},
[TGSI_OPCODE_LRP] = { ALU_OP0_NOP, tgsi_lrp},
[TGSI_OPCODE_FMA] = { ALU_OP3_FMA, tgsi_op3},
[TGSI_OPCODE_SQRT] = { ALU_OP1_SQRT_IEEE, cayman_emit_float_instr},
- [TGSI_OPCODE_DP2A] = { ALU_OP0_NOP, tgsi_unsupported},
+ [21] = { ALU_OP0_NOP, tgsi_unsupported},
[22] = { ALU_OP0_NOP, tgsi_unsupported},
[23] = { ALU_OP0_NOP, tgsi_unsupported},
[TGSI_OPCODE_FRC] = { ALU_OP1_FRACT, tgsi_op2},
{
switch (opcode) {
case TGSI_OPCODE_ADD: return SVGA3DOP_ADD;
- case TGSI_OPCODE_DP2A: return SVGA3DOP_DP2ADD;
case TGSI_OPCODE_DP3: return SVGA3DOP_DP3;
case TGSI_OPCODE_DP4: return SVGA3DOP_DP4;
case TGSI_OPCODE_FRC: return SVGA3DOP_FRC;
}
-/**
- * Emit code for TGSI_OPCODE_DP2A instruction.
- */
-static boolean
-emit_dp2a(struct svga_shader_emitter_v10 *emit,
- const struct tgsi_full_instruction *inst)
-{
- /* dst.x = src0.x * src1.x + src0.y * src1.y + src2.x
- * dst.y = src0.x * src1.x + src0.y * src1.y + src2.x
- * dst.z = src0.x * src1.x + src0.y * src1.y + src2.x
- * dst.w = src0.x * src1.x + src0.y * src1.y + src2.x
- * Translate into
- * MAD tmp.x, s0.y, s1.y, s2.x
- * MAD tmp.x, s0.x, s1.x, tmp.x
- * MOV dst.xyzw, tmp.xxxx
- */
- unsigned tmp = get_temp_index(emit);
- struct tgsi_full_src_register tmp_src = make_src_temp_reg(tmp);
- struct tgsi_full_dst_register tmp_dst = make_dst_temp_reg(tmp);
-
- struct tgsi_full_src_register tmp_src_xxxx =
- scalar_src(&tmp_src, TGSI_SWIZZLE_X);
- struct tgsi_full_dst_register tmp_dst_x =
- writemask_dst(&tmp_dst, TGSI_WRITEMASK_X);
-
- struct tgsi_full_src_register src0_xxxx =
- scalar_src(&inst->Src[0], TGSI_SWIZZLE_X);
- struct tgsi_full_src_register src0_yyyy =
- scalar_src(&inst->Src[0], TGSI_SWIZZLE_Y);
- struct tgsi_full_src_register src1_xxxx =
- scalar_src(&inst->Src[1], TGSI_SWIZZLE_X);
- struct tgsi_full_src_register src1_yyyy =
- scalar_src(&inst->Src[1], TGSI_SWIZZLE_Y);
- struct tgsi_full_src_register src2_xxxx =
- scalar_src(&inst->Src[2], TGSI_SWIZZLE_X);
-
- emit_instruction_op3(emit, VGPU10_OPCODE_MAD, &tmp_dst_x, &src0_yyyy,
- &src1_yyyy, &src2_xxxx, FALSE);
- emit_instruction_op3(emit, VGPU10_OPCODE_MAD, &tmp_dst_x, &src0_xxxx,
- &src1_xxxx, &tmp_src_xxxx, FALSE);
- emit_instruction_op1(emit, VGPU10_OPCODE_MOV, &inst->Dst[0],
- &tmp_src_xxxx, inst->Instruction.Saturate);
-
- free_temp_indexes(emit);
-
- return TRUE;
-}
-
-
/**
* Emit code for TGSI_OPCODE_DPH instruction.
*/
return emit_cmp(emit, inst);
case TGSI_OPCODE_COS:
return emit_sincos(emit, inst);
- case TGSI_OPCODE_DP2A:
- return emit_dp2a(emit, inst);
case TGSI_OPCODE_DPH:
return emit_dph(emit, inst);
case TGSI_OPCODE_DST:
#define TGSI_OPCODE_LRP 18
#define TGSI_OPCODE_FMA 19
#define TGSI_OPCODE_SQRT 20
-#define TGSI_OPCODE_DP2A 21
+/* gap */
#define TGSI_OPCODE_F2U64 22
#define TGSI_OPCODE_F2I64 23
#define TGSI_OPCODE_FRC 24
src[2] = args[0];
ureg_insn(t->ureg, TGSI_OPCODE_CMP, dst, 1, src, 3, 0);
} else if (!strcmp(desc->name, "DOT2_ADD")) {
- /* note: DP2A is not implemented in most pipe drivers */
tmp[0] = get_temp(t, MAX_NUM_FRAGMENT_REGISTERS_ATI); /* re-purpose a1 */
src[0] = args[0];
src[1] = args[1];