[Mesa-dev] [PATCH] gallium: add double opcodes and TGSI execution (v2.1)
Glenn Kennard
glenn.kennard at gmail.com
Tue Dec 23 17:48:48 PST 2014
On Tue, 23 Dec 2014 22:50:30 +0100, Dave Airlie <airlied at gmail.com> wrote:
> This patch adds support for a set of double opcodes
> to TGSI. It is an update of work done originally
> by Michal Krol on the gallium-double-opcodes branch.
>
> The opcodes have a hint where they came from in the
> header file.
>
> v2: add unsigned/int <-> double
> v2.1: update docs.
> This is based on code by Michael Krol <michal at vmware.com>
>
> Signed-off-by: Dave Airlie <airlied at redhat.com>
> ---
> src/gallium/auxiliary/tgsi/tgsi_exec.c | 743
> ++++++++++++++++++++++++++++-
> src/gallium/auxiliary/tgsi/tgsi_info.c | 24 +-
> src/gallium/docs/source/tgsi.rst | 76 ++-
> src/gallium/include/pipe/p_shader_tokens.h | 26 +-
> 4 files changed, 850 insertions(+), 19 deletions(-)
>
> diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c
> b/src/gallium/auxiliary/tgsi/tgsi_exec.c
> index 834568b..6af4730 100644
> --- a/src/gallium/auxiliary/tgsi/tgsi_exec.c
> +++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c
> @@ -72,6 +72,16 @@
> #define TILE_BOTTOM_LEFT 2
> #define TILE_BOTTOM_RIGHT 3
> +union tgsi_double_channel {
> + double d[TGSI_QUAD_SIZE];
> + unsigned u[TGSI_QUAD_SIZE][2];
> +};
> +
> +struct tgsi_double_vector {
> + union tgsi_double_channel xy;
> + union tgsi_double_channel zw;
> +};
> +
> static void
> micro_abs(union tgsi_exec_channel *dst,
> const union tgsi_exec_channel *src)
> @@ -147,6 +157,55 @@ micro_cos(union tgsi_exec_channel *dst,
> }
> static void
> +micro_d2f(union tgsi_exec_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->f[0] = (float)src->d[0];
> + dst->f[1] = (float)src->d[1];
> + dst->f[2] = (float)src->d[2];
> + dst->f[3] = (float)src->d[3];
> +}
> +
> +static void
> +micro_d2i(union tgsi_exec_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->i[0] = (int)src->d[0];
> + dst->i[1] = (int)src->d[1];
> + dst->i[2] = (int)src->d[2];
> + dst->i[3] = (int)src->d[3];
> +}
> +
> +static void
> +micro_d2u(union tgsi_exec_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->u[0] = (unsigned)src->d[0];
> + dst->u[1] = (unsigned)src->d[1];
> + dst->u[2] = (unsigned)src->d[2];
> + dst->u[3] = (unsigned)src->d[3];
> +}
> +static void
> +micro_dabs(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src->d[0] >= 0.0 ? src->d[0] : -src->d[0];
> + dst->d[1] = src->d[1] >= 0.0 ? src->d[1] : -src->d[1];
> + dst->d[2] = src->d[2] >= 0.0 ? src->d[2] : -src->d[2];
> + dst->d[3] = src->d[3] >= 0.0 ? src->d[3] : -src->d[3];
> +}
> +
> +static void
> +micro_dadd(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src[0].d[0] + src[1].d[0];
> + dst->d[1] = src[0].d[1] + src[1].d[1];
> + dst->d[2] = src[0].d[2] + src[1].d[2];
> + dst->d[3] = src[0].d[3] + src[1].d[3];
> +}
> +
> +static void
> micro_ddx(union tgsi_exec_channel *dst,
> const union tgsi_exec_channel *src)
> {
> @@ -167,6 +226,159 @@ micro_ddy(union tgsi_exec_channel *dst,
> }
> static void
> +micro_ddiv(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src[0].d[0] / src[1].d[0];
> + dst->d[1] = src[0].d[1] / src[1].d[1];
> + dst->d[2] = src[0].d[2] / src[1].d[2];
> + dst->d[3] = src[0].d[3] / src[1].d[3];
> +}
> +
> +static void
> +micro_dmul(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src[0].d[0] * src[1].d[0];
> + dst->d[1] = src[0].d[1] * src[1].d[1];
> + dst->d[2] = src[0].d[2] * src[1].d[2];
> + dst->d[3] = src[0].d[3] * src[1].d[3];
> +}
> +
> +static void
> +micro_dmax(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src[0].d[0] > src[1].d[0] ? src[0].d[0] : src[1].d[0];
> + dst->d[1] = src[0].d[1] > src[1].d[1] ? src[0].d[1] : src[1].d[1];
> + dst->d[2] = src[0].d[2] > src[1].d[2] ? src[0].d[2] : src[1].d[2];
> + dst->d[3] = src[0].d[3] > src[1].d[3] ? src[0].d[3] : src[1].d[3];
> +}
> +
> +static void
> +micro_dmin(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src[0].d[0] < src[1].d[0] ? src[0].d[0] : src[1].d[0];
> + dst->d[1] = src[0].d[1] < src[1].d[1] ? src[0].d[1] : src[1].d[1];
> + dst->d[2] = src[0].d[2] < src[1].d[2] ? src[0].d[2] : src[1].d[2];
> + dst->d[3] = src[0].d[3] < src[1].d[3] ? src[0].d[3] : src[1].d[3];
> +}
> +
> +static void
> +micro_dneg(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = -src->d[0];
> + dst->d[1] = -src->d[1];
> + dst->d[2] = -src->d[2];
> + dst->d[3] = -src->d[3];
> +}
> +
> +static void
> +micro_dslt(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->u[0][0] = src[0].d[0] < src[1].d[0] ? ~0U : 0U;
> + dst->u[1][0] = src[0].d[1] < src[1].d[1] ? ~0U : 0U;
> + dst->u[2][0] = src[0].d[2] < src[1].d[2] ? ~0U : 0U;
> + dst->u[3][0] = src[0].d[3] < src[1].d[3] ? ~0U : 0U;
> +}
> +
> +static void
> +micro_dsne(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->u[0][0] = src[0].d[0] != src[1].d[0] ? ~0U : 0U;
> + dst->u[1][0] = src[0].d[1] != src[1].d[1] ? ~0U : 0U;
> + dst->u[2][0] = src[0].d[2] != src[1].d[2] ? ~0U : 0U;
> + dst->u[3][0] = src[0].d[3] != src[1].d[3] ? ~0U : 0U;
> +}
> +
> +static void
> +micro_dsge(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + // fprintf(stderr, "%f %f\n", src[0].d[0], src[1].d[0]);
leftover debug print
> + dst->u[0][0] = src[0].d[0] >= src[1].d[0] ? ~0U : 0U;
> + dst->u[1][0] = src[0].d[1] >= src[1].d[1] ? ~0U : 0U;
> + dst->u[2][0] = src[0].d[2] >= src[1].d[2] ? ~0U : 0U;
> + dst->u[3][0] = src[0].d[3] >= src[1].d[3] ? ~0U : 0U;
> +}
> +
> +static void
> +micro_dseq(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->u[0][0] = src[0].d[0] == src[1].d[0] ? ~0U : 0U;
> + dst->u[1][0] = src[0].d[1] == src[1].d[1] ? ~0U : 0U;
> + dst->u[2][0] = src[0].d[2] == src[1].d[2] ? ~0U : 0U;
> + dst->u[3][0] = src[0].d[3] == src[1].d[3] ? ~0U : 0U;
> +}
> +
> +static void
> +micro_drcp(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = 1.0 / src->d[0];
> + dst->d[1] = 1.0 / src->d[1];
> + dst->d[2] = 1.0 / src->d[2];
> + dst->d[3] = 1.0 / src->d[3];
> +}
> +
> +static void
> +micro_dsqrt(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = sqrt(src->d[0]);
> + dst->d[1] = sqrt(src->d[1]);
> + dst->d[2] = sqrt(src->d[2]);
> + dst->d[3] = sqrt(src->d[3]);
> +}
> +
> +static void
> +micro_dmad(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src[0].d[0] * src[1].d[0] + src[2].d[0];
> + dst->d[1] = src[0].d[1] * src[1].d[1] + src[2].d[1];
> + dst->d[2] = src[0].d[2] * src[1].d[2] + src[2].d[2];
> + dst->d[3] = src[0].d[3] * src[1].d[3] + src[2].d[3];
> +}
> +
> +static void
> +micro_dfrac(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = src->d[0] - floor(src->d[0]);
> + dst->d[1] = src->d[1] - floor(src->d[1]);
> + dst->d[2] = src->d[2] - floor(src->d[2]);
> + dst->d[3] = src->d[3] - floor(src->d[3]);
> +}
> +
> +static void
> +micro_dldexp(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src0,
> + union tgsi_exec_channel *src1)
> +{
> + dst->d[0] = ldexp(src0->d[0], src1->i[1]);
> + dst->d[1] = ldexp(src0->d[1], src1->i[1]);
> + dst->d[2] = ldexp(src0->d[2], src1->i[2]);
> + dst->d[3] = ldexp(src0->d[3], src1->i[3]);
> +}
> +
> +static void
> +micro_dfracexp(union tgsi_double_channel *dst,
> + union tgsi_exec_channel *dst_exp,
> + const union tgsi_double_channel *src)
> +{
> + dst->d[0] = frexp(src->d[0], &dst_exp->i[0]);
> + dst->d[1] = frexp(src->d[1], &dst_exp->i[1]);
> + dst->d[2] = frexp(src->d[2], &dst_exp->i[2]);
> + dst->d[3] = frexp(src->d[3], &dst_exp->i[3]);
> +}
> +
> +static void
> micro_exp2(union tgsi_exec_channel *dst,
> const union tgsi_exec_channel *src)
> {
> @@ -201,6 +413,16 @@ micro_exp2(union tgsi_exec_channel *dst,
> }
> static void
> +micro_f2d(union tgsi_double_channel *dst,
> + const union tgsi_exec_channel *src)
> +{
> + dst->d[0] = (double)src->f[0];
> + dst->d[1] = (double)src->f[1];
> + dst->d[2] = (double)src->f[2];
> + dst->d[3] = (double)src->f[3];
> +}
> +
> +static void
> micro_flr(union tgsi_exec_channel *dst,
> const union tgsi_exec_channel *src)
> {
> @@ -221,6 +443,16 @@ micro_frc(union tgsi_exec_channel *dst,
> }
> static void
> +micro_i2d(union tgsi_double_channel *dst,
> + const union tgsi_exec_channel *src)
> +{
> + dst->d[0] = (double)src->i[0];
> + dst->d[1] = (double)src->i[1];
> + dst->d[2] = (double)src->i[2];
> + dst->d[3] = (double)src->i[3];
> +}
> +
> +static void
> micro_iabs(union tgsi_exec_channel *dst,
> const union tgsi_exec_channel *src)
> {
> @@ -449,11 +681,21 @@ micro_trunc(union tgsi_exec_channel *dst,
> dst->f[3] = (float)(int)src->f[3];
> }
> +static void
> +micro_u2d(union tgsi_double_channel *dst,
> + const union tgsi_exec_channel *src)
> +{
> + dst->d[0] = (double)src->u[0];
> + dst->d[1] = (double)src->u[1];
> + dst->d[2] = (double)src->u[2];
> + dst->d[3] = (double)src->u[3];
> +}
> enum tgsi_exec_datatype {
> TGSI_EXEC_DATA_FLOAT,
> TGSI_EXEC_DATA_INT,
> - TGSI_EXEC_DATA_UINT
> + TGSI_EXEC_DATA_UINT,
> + TGSI_EXEC_DATA_DOUBLE,
One comma too many
> };
> /*
> @@ -1090,11 +1332,11 @@ fetch_src_file_channel(const struct
> tgsi_exec_machine *mach,
> }
> static void
> -fetch_source(const struct tgsi_exec_machine *mach,
> - union tgsi_exec_channel *chan,
> - const struct tgsi_full_src_register *reg,
> - const uint chan_index,
> - enum tgsi_exec_datatype src_datatype)
> +fetch_source_d(const struct tgsi_exec_machine *mach,
> + union tgsi_exec_channel *chan,
> + const struct tgsi_full_src_register *reg,
> + const uint chan_index,
> + enum tgsi_exec_datatype src_datatype, bool dtype)
> {
> union tgsi_exec_channel index;
> union tgsi_exec_channel index2D;
> @@ -1238,6 +1480,9 @@ fetch_source(const struct tgsi_exec_machine *mach,
> &index2D,
> chan);
> + if (dtype)
> + return;
> +
Should probably note that modifiers are handled for doubles by caller.
> if (reg->Register.Absolute) {
> if (src_datatype == TGSI_EXEC_DATA_FLOAT) {
> micro_abs(chan, chan);
> @@ -1256,12 +1501,22 @@ fetch_source(const struct tgsi_exec_machine
> *mach,
> }
> static void
> -store_dest(struct tgsi_exec_machine *mach,
> - const union tgsi_exec_channel *chan,
> - const struct tgsi_full_dst_register *reg,
> - const struct tgsi_full_instruction *inst,
> - uint chan_index,
> - enum tgsi_exec_datatype dst_datatype)
> +fetch_source(const struct tgsi_exec_machine *mach,
> + union tgsi_exec_channel *chan,
> + const struct tgsi_full_src_register *reg,
> + const uint chan_index,
> + enum tgsi_exec_datatype src_datatype)
> +{
> + fetch_source_d(mach, chan, reg, chan_index, src_datatype, false);
> +}
> +
> +static void
> +store_dest_optsat(struct tgsi_exec_machine *mach,
> + const union tgsi_exec_channel *chan,
> + const struct tgsi_full_dst_register *reg,
> + const struct tgsi_full_instruction *inst,
> + uint chan_index,
> + enum tgsi_exec_datatype dst_datatype, bool sat)
> {
> uint i;
> union tgsi_exec_channel null;
> @@ -1471,6 +1726,14 @@ store_dest(struct tgsi_exec_machine *mach,
> }
> }
> + if (!sat) {
> + /* doubles path */
> + for (i = 0; i < TGSI_QUAD_SIZE; i++)
> + if (execmask & (1 << i))
> + dst->i[i] = chan->i[i];
> + return;
> + }
> +
> switch (inst->Instruction.Saturate) {
> case TGSI_SAT_NONE:
> for (i = 0; i < TGSI_QUAD_SIZE; i++)
> @@ -1505,8 +1768,20 @@ store_dest(struct tgsi_exec_machine *mach,
> default:
> assert( 0 );
> }
> +
> }
> +static void
> +store_dest(struct tgsi_exec_machine *mach,
> + const union tgsi_exec_channel *chan,
> + const struct tgsi_full_dst_register *reg,
> + const struct tgsi_full_instruction *inst,
> + uint chan_index,
> + enum tgsi_exec_datatype dst_datatype)
> +{
> + store_dest_optsat(mach, chan, reg, inst, chan_index,
> + dst_datatype, true);
> +}
> #define FETCH(VAL,INDEX,CHAN)\
> fetch_source(mach, VAL, &inst->Src[INDEX], CHAN,
> TGSI_EXEC_DATA_FLOAT)
> @@ -2980,6 +3255,358 @@ exec_endswitch(struct tgsi_exec_machine *mach)
> UPDATE_EXEC_MASK(mach);
> }
> +typedef void (* micro_dop)(union tgsi_double_channel *dst,
> + const union tgsi_double_channel *src);
> +
> +static void
> +fetch_double_channel(struct tgsi_exec_machine *mach,
> + union tgsi_double_channel *chan,
> + const struct tgsi_full_src_register *reg,
> + uint chan_0,
> + uint chan_1)
> +{
> + union tgsi_exec_channel src[2];
> + uint i;
> +
> + /* XXX: That won't work for operand modifiers.
> + */
Will work due to return in fetch_source_d for last function arg=true?
> + fetch_source_d(mach, &src[0], reg, chan_0, TGSI_EXEC_DATA_UINT,
> true);
> + fetch_source_d(mach, &src[1], reg, chan_1, TGSI_EXEC_DATA_UINT,
> true);
> +
> + for (i = 0; i < TGSI_QUAD_SIZE; i++) {
> + chan->u[i][0] = src[0].u[i];
> + chan->u[i][1] = src[1].u[i];
> + }
> + if (reg->Register.Absolute) {
> + micro_dabs(chan, chan);
> + }
> + if (reg->Register.Negate) {
> + micro_dneg(chan, chan);
> + }
> +
> +}
> +
> +static void
> +store_double_channel(struct tgsi_exec_machine *mach,
> + const union tgsi_double_channel *chan,
> + const struct tgsi_full_dst_register *reg,
> + const struct tgsi_full_instruction *inst,
> + uint chan_0,
> + uint chan_1)
> +{
> + union tgsi_exec_channel dst[2];
> + uint i;
> + union tgsi_double_channel temp;
> + const uint execmask = mach->ExecMask;
> + switch (inst->Instruction.Saturate) {
> + case TGSI_SAT_NONE:
> + for (i = 0; i < TGSI_QUAD_SIZE; i++)
> + if (execmask & (1 << i)) {
> + dst[0].u[i] = chan->u[i][0];
> + dst[1].u[i] = chan->u[i][1];
> + }
> + break;
> +
> + case TGSI_SAT_ZERO_ONE:
> + for (i = 0; i < TGSI_QUAD_SIZE; i++)
> + if (execmask & (1 << i)) {
> + if (chan->d[i] < 0.0f)
> + temp.d[i] = 0.0;
> + else if (chan->d[i] > 1.0f)
> + temp.d[i] = 1.0;
> + else
> + temp.d[i] = chan->d[i];
> +
> + dst[0].u[i] = temp.u[i][0];
> + dst[1].u[i] = temp.u[i][1];
> + }
> + break;
> +
> + case TGSI_SAT_MINUS_PLUS_ONE:
> + for (i = 0; i < TGSI_QUAD_SIZE; i++)
> + if (execmask & (1 << i)) {
> + if (chan->d[i] < -1.0)
> + temp.d[i] = -1.0;
> + else if (chan->d[i] > 1.0)
> + temp.d[i] = 1.0;
> + else
> + temp.d[i] = chan->d[i];
> +
> + dst[0].u[i] = temp.u[i][0];
> + dst[1].u[i] = temp.u[i][1];
> + }
> + break;
> +
> + default:
> + assert( 0 );
> + }
> +
> + /* XXX: That won't work for saturate modifiers.
> + */
Saturate handled above, and early break out before saturate application in
store_dest_optsat(), so comment outdated?
> + store_dest_optsat(mach, &dst[0], reg, inst, chan_0,
> TGSI_EXEC_DATA_UINT, false);
> + if (chan_1 != -1)
> + store_dest_optsat(mach, &dst[1], reg, inst, chan_1,
> TGSI_EXEC_DATA_UINT, false);
> +}
> +
> +static void
> +exec_double_unary(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst,
> + micro_dop op)
> +{
> + union tgsi_double_channel src;
> + union tgsi_double_channel dst;
> +
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_XY) ==
> TGSI_WRITEMASK_XY) {
> + fetch_double_channel(mach, &src, &inst->Src[0], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + op(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + }
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_ZW) ==
> TGSI_WRITEMASK_ZW) {
> + fetch_double_channel(mach, &src, &inst->Src[0], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + op(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + }
assert on any other write mask?
> +}
> +
> +static void
> +exec_double_binary(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst,
> + micro_dop op,
> + enum tgsi_exec_datatype dst_datatype)
> +{
> + union tgsi_double_channel src[2];
> + union tgsi_double_channel dst;
> + int first_dest_chan, second_dest_chan;
> + int wmask;
> +
> + wmask = inst->Dst[0].Register.WriteMask;
> + if (wmask & TGSI_WRITEMASK_XY) {
> + first_dest_chan = TGSI_CHAN_X;
> + second_dest_chan = TGSI_CHAN_Y;
> + if (dst_datatype == TGSI_EXEC_DATA_UINT) {
> + first_dest_chan = (wmask & TGSI_WRITEMASK_X) ? TGSI_CHAN_X :
> TGSI_CHAN_Y;
> + second_dest_chan = -1;
passing -1 to store_double_channel() taking uint, think some paths try to
use that as array index without checks and will blow up.
> + }
> +
> + fetch_double_channel(mach, &src[0], &inst->Src[0], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + fetch_double_channel(mach, &src[1], &inst->Src[1], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + op(&dst, src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> first_dest_chan, second_dest_chan);
> + }
> +
> + if (wmask & TGSI_WRITEMASK_ZW) {
> + first_dest_chan = TGSI_CHAN_Z;
> + second_dest_chan = TGSI_CHAN_W;
> + if (dst_datatype == TGSI_EXEC_DATA_UINT) {
> + first_dest_chan = (wmask & TGSI_WRITEMASK_Z) ? TGSI_CHAN_Z :
> TGSI_CHAN_W;
> + second_dest_chan = -1;
> + }
> +
> + fetch_double_channel(mach, &src[0], &inst->Src[0], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + fetch_double_channel(mach, &src[1], &inst->Src[1], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + op(&dst, src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> first_dest_chan, second_dest_chan);
> + }
> +}
> +
> +static void
> +exec_double_trinary(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst,
> + micro_dop op)
> +{
> + union tgsi_double_channel src[3];
> + union tgsi_double_channel dst;
> +
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_XY) ==
> TGSI_WRITEMASK_XY) {
> + fetch_double_channel(mach, &src[0], &inst->Src[0], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + fetch_double_channel(mach, &src[1], &inst->Src[1], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + fetch_double_channel(mach, &src[2], &inst->Src[2], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + op(&dst, src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + }
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_ZW) ==
> TGSI_WRITEMASK_ZW) {
> + fetch_double_channel(mach, &src[0], &inst->Src[0], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + fetch_double_channel(mach, &src[1], &inst->Src[1], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + fetch_double_channel(mach, &src[2], &inst->Src[2], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + op(&dst, src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + }
> +}
> +
> +static void
> +exec_f2d(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_exec_channel src;
> + union tgsi_double_channel dst;
> +
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_XY) ==
> TGSI_WRITEMASK_XY) {
> + fetch_source(mach, &src, &inst->Src[0], TGSI_CHAN_X,
> TGSI_EXEC_DATA_FLOAT);
> + micro_f2d(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + }
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_ZW) ==
> TGSI_WRITEMASK_ZW) {
> + fetch_source(mach, &src, &inst->Src[0], TGSI_CHAN_Y,
> TGSI_EXEC_DATA_FLOAT);
> + micro_f2d(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + }
> +}
> +
> +static void
> +exec_d2f(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_double_channel src;
> + union tgsi_exec_channel dst;
> + int wm = inst->Dst[0].Register.WriteMask;
> + int i;
> + int bit;
> + for (i = 0; i < 2; i++) {
> + bit = ffs(wm);
> + if (bit) {
> + wm &= ~(1 << (bit - 1));
> + if (i == 0)
> + fetch_double_channel(mach, &src, &inst->Src[0],
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + else
> + fetch_double_channel(mach, &src, &inst->Src[0],
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + micro_d2f(&dst, &src);
> + store_dest(mach, &dst, &inst->Dst[0], inst, bit - 1,
> TGSI_EXEC_DATA_FLOAT);
> + }
> + }
> +}
> +
> +static void
> +exec_i2d(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_exec_channel src;
> + union tgsi_double_channel dst;
> +
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_XY) ==
> TGSI_WRITEMASK_XY) {
> + fetch_source(mach, &src, &inst->Src[0], TGSI_CHAN_X,
> TGSI_EXEC_DATA_INT);
> + micro_i2d(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + }
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_ZW) ==
> TGSI_WRITEMASK_ZW) {
> + fetch_source(mach, &src, &inst->Src[0], TGSI_CHAN_Y,
> TGSI_EXEC_DATA_INT);
> + micro_i2d(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + }
> +}
> +
> +static void
> +exec_d2i(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_double_channel src;
> + union tgsi_exec_channel dst;
> + int wm = inst->Dst[0].Register.WriteMask;
> + int i;
> + int bit;
> + for (i = 0; i < 2; i++) {
> + bit = ffs(wm);
> + if (bit) {
> + wm &= ~(1 << (bit - 1));
> + if (i == 0)
> + fetch_double_channel(mach, &src, &inst->Src[0],
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + else
> + fetch_double_channel(mach, &src, &inst->Src[0],
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + micro_d2i(&dst, &src);
> + store_dest(mach, &dst, &inst->Dst[0], inst, bit - 1,
> TGSI_EXEC_DATA_INT);
> + }
> + }
> +}
> +static void
> +exec_u2d(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_exec_channel src;
> + union tgsi_double_channel dst;
> +
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_XY) ==
> TGSI_WRITEMASK_XY) {
> + fetch_source(mach, &src, &inst->Src[0], TGSI_CHAN_X,
> TGSI_EXEC_DATA_UINT);
> + micro_u2d(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + }
> + if ((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_ZW) ==
> TGSI_WRITEMASK_ZW) {
> + fetch_source(mach, &src, &inst->Src[0], TGSI_CHAN_Y,
> TGSI_EXEC_DATA_UINT);
> + micro_u2d(&dst, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + }
> +}
> +
> +static void
> +exec_d2u(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_double_channel src;
> + union tgsi_exec_channel dst;
> + int wm = inst->Dst[0].Register.WriteMask;
> + int i;
> + int bit;
> + for (i = 0; i < 2; i++) {
> + bit = ffs(wm);
> + if (bit) {
> + wm &= ~(1 << (bit - 1));
> + if (i == 0)
> + fetch_double_channel(mach, &src, &inst->Src[0],
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + else
> + fetch_double_channel(mach, &src, &inst->Src[0],
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + micro_d2u(&dst, &src);
> + store_dest(mach, &dst, &inst->Dst[0], inst, bit - 1,
> TGSI_EXEC_DATA_UINT);
> + }
> + }
> +}
> +
> +static void
> +exec_dldexp(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_double_channel src0;
> + union tgsi_exec_channel src1;
> + union tgsi_double_channel dst;
> + int wmask;
> +
> + wmask = inst->Dst[0].Register.WriteMask;
> + if (wmask & TGSI_WRITEMASK_XY) {
> +
> + fetch_double_channel(mach, &src0, &inst->Src[0], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + fetch_source(mach, &src1, &inst->Src[1], TGSI_CHAN_X,
> TGSI_EXEC_DATA_INT);
> + micro_dldexp(&dst, &src0, &src1);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + }
> +
> + if (wmask & TGSI_WRITEMASK_ZW) {
> + fetch_double_channel(mach, &src0, &inst->Src[0], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + fetch_source(mach, &src1, &inst->Src[1], TGSI_CHAN_Z,
> TGSI_EXEC_DATA_INT);
> + micro_dldexp(&dst, &src0, &src1);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + }
> +}
> +
> +static void
> +exec_dfracexp(struct tgsi_exec_machine *mach,
> + const struct tgsi_full_instruction *inst)
> +{
> + union tgsi_double_channel src;
> + union tgsi_double_channel dst;
> + union tgsi_exec_channel dst_exp;
> +
> + if (((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_XY) ==
> TGSI_WRITEMASK_XY)) {
> + fetch_double_channel(mach, &src, &inst->Src[0], TGSI_CHAN_X,
> TGSI_CHAN_Y);
> + micro_dfracexp(&dst, &dst_exp, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_X, TGSI_CHAN_Y);
> + store_dest(mach, &dst_exp, &inst->Dst[1], inst,
> ffs(inst->Dst[1].Register.WriteMask) - 1, TGSI_EXEC_DATA_INT);
> + }
> + if (((inst->Dst[0].Register.WriteMask & TGSI_WRITEMASK_ZW) ==
> TGSI_WRITEMASK_ZW)) {
> + fetch_double_channel(mach, &src, &inst->Src[0], TGSI_CHAN_Z,
> TGSI_CHAN_W);
> + micro_dfracexp(&dst, &dst_exp, &src);
> + store_double_channel(mach, &dst, &inst->Dst[0], inst,
> TGSI_CHAN_Z, TGSI_CHAN_W);
> + store_dest(mach, &dst_exp, &inst->Dst[1], inst,
> ffs(inst->Dst[1].Register.WriteMask) - 1, TGSI_EXEC_DATA_INT);
> + }
> +}
> +
> +
> static void
> micro_i2f(union tgsi_exec_channel *dst,
> const union tgsi_exec_channel *src)
> @@ -4336,6 +4963,98 @@ exec_instruction(
> case TGSI_OPCODE_UMSB:
> exec_vector_unary(mach, inst, micro_umsb, TGSI_EXEC_DATA_INT,
> TGSI_EXEC_DATA_UINT);
> break;
> +
> + case TGSI_OPCODE_F2D:
> + exec_f2d(mach, inst);
> + break;
> +
> + case TGSI_OPCODE_D2F:
> + exec_d2f(mach, inst);
> + break;
> +
> + case TGSI_OPCODE_DABS:
> + exec_double_unary(mach, inst, micro_dabs);
> + break;
> +
> + case TGSI_OPCODE_DNEG:
> + exec_double_unary(mach, inst, micro_dneg);
> + break;
> +
> + case TGSI_OPCODE_DADD:
> + exec_double_binary(mach, inst, micro_dadd, TGSI_EXEC_DATA_DOUBLE);
> + break;
> +
> + case TGSI_OPCODE_DDIV:
> + exec_double_binary(mach, inst, micro_ddiv, TGSI_EXEC_DATA_DOUBLE);
> + break;
> +
> + case TGSI_OPCODE_DMUL:
> + exec_double_binary(mach, inst, micro_dmul, TGSI_EXEC_DATA_DOUBLE);
> + break;
> +
> + case TGSI_OPCODE_DMAX:
> + exec_double_binary(mach, inst, micro_dmax, TGSI_EXEC_DATA_DOUBLE);
> + break;
> +
> + case TGSI_OPCODE_DMIN:
> + exec_double_binary(mach, inst, micro_dmin, TGSI_EXEC_DATA_DOUBLE);
> + break;
> +
> + case TGSI_OPCODE_DSLT:
> + exec_double_binary(mach, inst, micro_dslt, TGSI_EXEC_DATA_UINT);
> + break;
> +
> + case TGSI_OPCODE_DSGE:
> + exec_double_binary(mach, inst, micro_dsge, TGSI_EXEC_DATA_UINT);
> + break;
> +
> + case TGSI_OPCODE_DSEQ:
> + exec_double_binary(mach, inst, micro_dseq, TGSI_EXEC_DATA_UINT);
> + break;
> +
> + case TGSI_OPCODE_DSNE:
> + exec_double_binary(mach, inst, micro_dsne, TGSI_EXEC_DATA_UINT);
> + break;
> +
> + case TGSI_OPCODE_DRCP:
> + exec_double_unary(mach, inst, micro_drcp);
> + break;
> +
> + case TGSI_OPCODE_DSQRT:
> + exec_double_unary(mach, inst, micro_dsqrt);
> + break;
> +
> + case TGSI_OPCODE_DMAD:
> + exec_double_trinary(mach, inst, micro_dmad);
> + break;
> +
> + case TGSI_OPCODE_DFRAC:
> + exec_double_unary(mach, inst, micro_dfrac);
> + break;
> +
> + case TGSI_OPCODE_DLDEXP:
> + exec_dldexp(mach, inst);
> + break;
> +
> + case TGSI_OPCODE_DFRACEXP:
> + exec_dfracexp(mach, inst);
> + break;
> +
> + case TGSI_OPCODE_I2D:
> + exec_i2d(mach, inst);
> + break;
> +
> + case TGSI_OPCODE_D2I:
> + exec_d2i(mach, inst);
> + break;
> +
> + case TGSI_OPCODE_U2D:
> + exec_u2d(mach, inst);
> + break;
> +
> + case TGSI_OPCODE_D2U:
> + exec_d2u(mach, inst);
> + break;
> default:
> assert( 0 );
> }
> diff --git a/src/gallium/auxiliary/tgsi/tgsi_info.c
> b/src/gallium/auxiliary/tgsi/tgsi_info.c
> index c90d24c..05c5cb5 100644
> --- a/src/gallium/auxiliary/tgsi/tgsi_info.c
> +++ b/src/gallium/auxiliary/tgsi/tgsi_info.c
> @@ -231,10 +231,32 @@ static const struct tgsi_opcode_info
> opcode_info[TGSI_OPCODE_LAST] =
> { 1, 1, 0, 0, 0, 0, COMP, "LSB", TGSI_OPCODE_LSB },
> { 1, 1, 0, 0, 0, 0, COMP, "IMSB", TGSI_OPCODE_IMSB },
> { 1, 1, 0, 0, 0, 0, COMP, "UMSB", TGSI_OPCODE_UMSB },
> -
> { 1, 1, 0, 0, 0, 0, OTHR, "INTERP_CENTROID",
> TGSI_OPCODE_INTERP_CENTROID },
> { 1, 2, 0, 0, 0, 0, OTHR, "INTERP_SAMPLE", TGSI_OPCODE_INTERP_SAMPLE
> },
> { 1, 2, 0, 0, 0, 0, OTHR, "INTERP_OFFSET", TGSI_OPCODE_INTERP_OFFSET
> },
> + { 1, 1, 0, 0, 0, 0, COMP, "F2D", TGSI_OPCODE_F2D },
> + { 1, 1, 0, 0, 0, 0, COMP, "D2F", TGSI_OPCODE_D2F },
> + { 1, 1, 0, 0, 0, 0, COMP, "DABS", TGSI_OPCODE_DABS },
> + { 1, 1, 0, 0, 0, 0, COMP, "DNEG", TGSI_OPCODE_DNEG },
> + { 1, 2, 0, 0, 0, 0, COMP, "DADD", TGSI_OPCODE_DADD },
> + { 1, 2, 0, 0, 0, 0, COMP, "DDIV", TGSI_OPCODE_DDIV },
> + { 1, 2, 0, 0, 0, 0, COMP, "DMUL", TGSI_OPCODE_DMUL },
> + { 1, 2, 0, 0, 0, 0, COMP, "DMAX", TGSI_OPCODE_DMAX },
> + { 1, 2, 0, 0, 0, 0, COMP, "DMIN", TGSI_OPCODE_DMIN },
> + { 1, 2, 0, 0, 0, 0, COMP, "DSLT", TGSI_OPCODE_DSLT },
> + { 1, 2, 0, 0, 0, 0, COMP, "DSGE", TGSI_OPCODE_DSGE },
> + { 1, 2, 0, 0, 0, 0, COMP, "DSEQ", TGSI_OPCODE_DSEQ },
> + { 1, 2, 0, 0, 0, 0, COMP, "DSNE", TGSI_OPCODE_DSNE },
> + { 1, 1, 0, 0, 0, 0, COMP, "DRCP", TGSI_OPCODE_DRCP },
> + { 1, 1, 0, 0 ,0, 0, COMP, "DSQRT", TGSI_OPCODE_DSQRT },
> + { 1, 3, 0, 0 ,0, 0, COMP, "DMAD", TGSI_OPCODE_DMAD },
> + { 1, 1, 0, 0, 0, 0, COMP, "DFRAC", TGSI_OPCODE_DFRAC},
> + { 1, 2, 0, 0, 0, 0, COMP, "DLDEXP", TGSI_OPCODE_DLDEXP},
> + { 2, 1, 0, 0, 0, 0, COMP, "DFRACEXP", TGSI_OPCODE_DFRACEXP},
> + { 1, 1, 0, 0, 0, 0, COMP, "D2I", TGSI_OPCODE_D2I },
> + { 1, 1, 0, 0, 0, 0, COMP, "I2D", TGSI_OPCODE_I2D },
> + { 1, 1, 0, 0, 0, 0, COMP, "D2U", TGSI_OPCODE_D2U },
> + { 1, 1, 0, 0, 0, 0, COMP, "U2D", TGSI_OPCODE_U2D },
> };
> const struct tgsi_opcode_info *
> diff --git a/src/gallium/docs/source/tgsi.rst
> b/src/gallium/docs/source/tgsi.rst
> index ff322e8..20b0102 100644
> --- a/src/gallium/docs/source/tgsi.rst
> +++ b/src/gallium/docs/source/tgsi.rst
> @@ -1810,7 +1810,10 @@ Double ISA
> The double-precision opcodes reinterpret four-component vectors into
> two-component vectors with doubled precision in each component.
> -Support for these opcodes is XXX undecided. :T
> +.. opcode:: DABS - Absolute
> +
> + dst.xy = |src0.xy|
> + dst.zw = |src0.zw|
> .. opcode:: DADD - Add
> @@ -1833,17 +1836,33 @@ Support for these opcodes is XXX undecided. :T
> .. math::
> - dst.xy = src0.xy == src1.xy ? 1.0F : 0.0F
> + dst.x = src0.xy == src1.xy ? \sim 0 : 0
> +
> + dst.z = src0.zw == src1.zw ? \sim 0 : 0
> +
> +.. opcode:: DSNE - Set on Equal
> +
> +.. math::
> +
> + dst.x = src0.xy != src1.xy ? \sim 0 : 0
> - dst.zw = src0.zw == src1.zw ? 1.0F : 0.0F
> + dst.z = src0.zw != src1.zw ? \sim 0 : 0
> .. opcode:: DSLT - Set on Less than
> .. math::
> - dst.xy = src0.xy < src1.xy ? 1.0F : 0.0F
> + dst.x = src0.xy < src1.xy ? \sim 0 : 0
> +
> + dst.z = src0.zw < src1.zw ? \sim 0 : 0
> +
> +.. opcode:: DSGE - Set on Greater equal
> +
> +.. math::
> - dst.zw = src0.zw < src1.zw ? 1.0F : 0.0F
> + dst.x = src0.xy >= src1.xy ? \sim 0 : 0
> +
> + dst.z = src0.zw >= src1.zw ? \sim 0 : 0
> .. opcode:: DFRAC - Fraction
> @@ -1930,7 +1949,54 @@ This opcode is the inverse of :opcode:`DFRACEXP`.
> dst.zw = \sqrt{src.zw}
> +.. opcode:: F2D - Float to Double
> +
> +.. math::
> +
> + dst.xy = double(src0.x)
> +
> + dst.zw = double(src0.z)
> +
> +.. opcode:: D2F - Double to Float
> +
> +.. math::
> +
> + dst.x = float(src0.xy)
> +
> + dst.z = float(src0.zw)
> +
> +.. opcode:: I2D - Int to Double
> +
> +.. math::
> +
> + dst.xy = double(src0.x)
> +
> + dst.zw = double(src0.z)
> +
> +.. opcode:: D2I - Double to Int
> +
> +.. math::
> +
> + dst.x = int(src0.xy)
> +
> + dst.z = int(src0.zw)
> +
> +.. opcode:: U2D - Unsigned Int to Double
> +
> +.. math::
> +
> + dst.xy = double(src0.x)
> +
> + dst.zw = double(src0.z)
> +
> +.. opcode:: D2U - Double to Unsigned Int
> +
> +.. math::
> + dst.x = unsigned(src0.xy)
> +
> + dst.z = unsigned(src0.zw)
> +
> .. _samplingopcodes:
> Resource Sampling Opcodes
> diff --git a/src/gallium/include/pipe/p_shader_tokens.h
> b/src/gallium/include/pipe/p_shader_tokens.h
> index 442b67b..970e168 100644
> --- a/src/gallium/include/pipe/p_shader_tokens.h
> +++ b/src/gallium/include/pipe/p_shader_tokens.h
> @@ -494,7 +494,31 @@ struct tgsi_property_data {
> #define TGSI_OPCODE_INTERP_SAMPLE 193
> #define TGSI_OPCODE_INTERP_OFFSET 194
> -#define TGSI_OPCODE_LAST 195
> +/* sm5 marked opcodes are supported in D3D11 optionally - also DMOV,
> DMOVC */
> +#define TGSI_OPCODE_F2D 195 /* SM5 */
> +#define TGSI_OPCODE_D2F 196
> +#define TGSI_OPCODE_DABS 197
> +#define TGSI_OPCODE_DNEG 198 /* SM5 */
> +#define TGSI_OPCODE_DADD 199 /* SM5 */
> +#define TGSI_OPCODE_DDIV 200 /* SM5 */
> +#define TGSI_OPCODE_DMUL 201 /* SM5 */
> +#define TGSI_OPCODE_DMAX 202 /* SM5 */
> +#define TGSI_OPCODE_DMIN 203 /* SM5 */
> +#define TGSI_OPCODE_DSLT 204 /* SM5 */
> +#define TGSI_OPCODE_DSGE 205 /* SM5 */
> +#define TGSI_OPCODE_DSEQ 206 /* SM5 */
> +#define TGSI_OPCODE_DSNE 207 /* SM5 */
> +#define TGSI_OPCODE_DRCP 208 /* eg, cayman */
> +#define TGSI_OPCODE_DSQRT 209 /* eg, cayman also has DRSQ
> */
Adding DRSQ seems like a good idea, at least for graphics with floats its
more frequent than SQRT.
> +#define TGSI_OPCODE_DMAD 210 /* DFMA? */
> +#define TGSI_OPCODE_DFRAC 211 /* eg, cayman */
> +#define TGSI_OPCODE_DLDEXP 212 /* eg, cayman */
> +#define TGSI_OPCODE_DFRACEXP 213 /* eg, cayman */
> +#define TGSI_OPCODE_D2I 214
> +#define TGSI_OPCODE_I2D 215
> +#define TGSI_OPCODE_D2U 216
> +#define TGSI_OPCODE_U2D 217
> +#define TGSI_OPCODE_LAST 218
> #define TGSI_SAT_NONE 0 /* do not saturate */
> #define TGSI_SAT_ZERO_ONE 1 /* clamp to [0,1] */
More information about the mesa-dev
mailing list