#include "pipe/p_compiler.h"
#include "pipe/p_state.h"
-#include "pipe/p_util.h"
#include "pipe/p_shader_tokens.h"
#include "tgsi/tgsi_parse.h"
#include "tgsi/tgsi_util.h"
#include "tgsi_exec.h"
+#include "util/u_memory.h"
+#include "util/u_math.h"
+
+#define FAST_MATH 1
#define TILE_TOP_LEFT 0
#define TILE_TOP_RIGHT 1
struct tgsi_exec_machine *mach,
const struct tgsi_token *tokens,
uint numSamplers,
- struct tgsi_sampler *samplers)
+ struct tgsi_sampler **samplers)
{
uint k;
struct tgsi_parse_context parse;
tgsi_dump(tokens, 0);
#endif
+ util_init_math();
+
mach->Tokens = tokens;
mach->Samplers = samplers;
union tgsi_exec_channel *dst,
const union tgsi_exec_channel *src)
{
+#if FAST_MATH
+ dst->f[0] = util_fast_exp2( src->f[0] );
+ dst->f[1] = util_fast_exp2( src->f[1] );
+ dst->f[2] = util_fast_exp2( src->f[2] );
+ dst->f[3] = util_fast_exp2( src->f[3] );
+#else
dst->f[0] = powf( 2.0f, src->f[0] );
dst->f[1] = powf( 2.0f, src->f[1] );
dst->f[2] = powf( 2.0f, src->f[2] );
dst->f[3] = powf( 2.0f, src->f[3] );
-}
-
-static void
-micro_f2it(
- union tgsi_exec_channel *dst,
- const union tgsi_exec_channel *src )
-{
- dst->i[0] = (int) src->f[0];
- dst->i[1] = (int) src->f[1];
- dst->i[2] = (int) src->f[2];
- dst->i[3] = (int) src->f[3];
+#endif
}
static void
union tgsi_exec_channel *dst,
const union tgsi_exec_channel *src )
{
+#if FAST_MATH
+ dst->f[0] = util_fast_log2( src->f[0] );
+ dst->f[1] = util_fast_log2( src->f[1] );
+ dst->f[2] = util_fast_log2( src->f[2] );
+ dst->f[3] = util_fast_log2( src->f[3] );
+#else
dst->f[0] = logf( src->f[0] ) * 1.442695f;
dst->f[1] = logf( src->f[1] ) * 1.442695f;
dst->f[2] = logf( src->f[2] ) * 1.442695f;
dst->f[3] = logf( src->f[3] ) * 1.442695f;
+#endif
}
static void
const union tgsi_exec_channel *src0,
const union tgsi_exec_channel *src1 )
{
+#if FAST_MATH
+ dst->f[0] = util_fast_pow( src0->f[0], src1->f[0] );
+ dst->f[1] = util_fast_pow( src0->f[1], src1->f[1] );
+ dst->f[2] = util_fast_pow( src0->f[2], src1->f[2] );
+ dst->f[3] = util_fast_pow( src0->f[3], src1->f[3] );
+#else
dst->f[0] = powf( src0->f[0], src1->f[0] );
dst->f[1] = powf( src0->f[1], src1->f[1] );
dst->f[2] = powf( src0->f[2], src1->f[2] );
dst->f[3] = powf( src0->f[3], src1->f[3] );
+#endif
}
static void
dst->f[3] = floorf( src->f[3] + 0.5f );
}
+static void
+micro_sgn(
+ union tgsi_exec_channel *dst,
+ const union tgsi_exec_channel *src )
+{
+ dst->f[0] = src->f[0] < 0.0f ? -1.0f : src->f[0] > 0.0f ? 1.0f : 0.0f;
+ dst->f[1] = src->f[1] < 0.0f ? -1.0f : src->f[1] > 0.0f ? 1.0f : 0.0f;
+ dst->f[2] = src->f[2] < 0.0f ? -1.0f : src->f[2] > 0.0f ? 1.0f : 0.0f;
+ dst->f[3] = src->f[3] < 0.0f ? -1.0f : src->f[3] > 0.0f ? 1.0f : 0.0f;
+}
+
static void
micro_shl(
union tgsi_exec_channel *dst,
case TGSI_EXTSWIZZLE_W:
switch( file ) {
case TGSI_FILE_CONSTANT:
- chan->f[0] = mach->Consts[index->i[0]][swizzle];
- chan->f[1] = mach->Consts[index->i[1]][swizzle];
- chan->f[2] = mach->Consts[index->i[2]][swizzle];
- chan->f[3] = mach->Consts[index->i[3]][swizzle];
+ assert(mach->Consts);
+ if (index->i[0] < 0)
+ chan->f[0] = 0.0f;
+ else
+ chan->f[0] = mach->Consts[index->i[0]][swizzle];
+ if (index->i[1] < 0)
+ chan->f[1] = 0.0f;
+ else
+ chan->f[1] = mach->Consts[index->i[1]][swizzle];
+ if (index->i[2] < 0)
+ chan->f[2] = 0.0f;
+ else
+ chan->f[2] = mach->Consts[index->i[2]][swizzle];
+ if (index->i[3] < 0)
+ chan->f[3] = 0.0f;
+ else
+ chan->f[3] = mach->Consts[index->i[3]][swizzle];
break;
case TGSI_FILE_INPUT:
union tgsi_exec_channel index;
uint swizzle;
+ /* We start with a direct index into a register file.
+ *
+ * file[1],
+ * where:
+ * file = SrcRegister.File
+ * [1] = SrcRegister.Index
+ */
index.i[0] =
index.i[1] =
index.i[2] =
index.i[3] = reg->SrcRegister.Index;
+ /* There is an extra source register that indirectly subscripts
+ * a register file. The direct index now becomes an offset
+ * that is being added to the indirect register.
+ *
+ * file[ind[2].x+1],
+ * where:
+ * ind = SrcRegisterInd.File
+ * [2] = SrcRegisterInd.Index
+ * .x = SrcRegisterInd.SwizzleX
+ */
if (reg->SrcRegister.Indirect) {
union tgsi_exec_channel index2;
union tgsi_exec_channel indir_index;
+ const uint execmask = mach->ExecMask;
+ uint i;
+ /* which address register (always zero now) */
index2.i[0] =
index2.i[1] =
index2.i[2] =
index2.i[3] = reg->SrcRegisterInd.Index;
+ /* get current value of address register[swizzle] */
swizzle = tgsi_util_get_src_register_swizzle( ®->SrcRegisterInd, CHAN_X );
fetch_src_file_channel(
mach,
&index2,
&indir_index );
- index.i[0] += indir_index.i[0];
- index.i[1] += indir_index.i[1];
- index.i[2] += indir_index.i[2];
- index.i[3] += indir_index.i[3];
+ /* add value of address register to the offset */
+ index.i[0] += (int) indir_index.f[0];
+ index.i[1] += (int) indir_index.f[1];
+ index.i[2] += (int) indir_index.f[2];
+ index.i[3] += (int) indir_index.f[3];
+
+ /* for disabled execution channels, zero-out the index to
+ * avoid using a potential garbage value.
+ */
+ for (i = 0; i < QUAD_SIZE; i++) {
+ if ((execmask & (1 << i)) == 0)
+ index.i[i] = 0;
+ }
}
- if( reg->SrcRegister.Dimension ) {
- switch( reg->SrcRegister.File ) {
+ /* There is an extra source register that is a second
+ * subscript to a register file. Effectively it means that
+ * the register file is actually a 2D array of registers.
+ *
+ * file[1][3] == file[1*sizeof(file[1])+3],
+ * where:
+ * [3] = SrcRegisterDim.Index
+ */
+ if (reg->SrcRegister.Dimension) {
+ /* The size of the first-order array depends on the register file type.
+ * We need to multiply the index to the first array to get an effective,
+ * "flat" index that points to the beginning of the second-order array.
+ */
+ switch (reg->SrcRegister.File) {
case TGSI_FILE_INPUT:
- index.i[0] *= 17;
- index.i[1] *= 17;
- index.i[2] *= 17;
- index.i[3] *= 17;
+ index.i[0] *= TGSI_EXEC_MAX_INPUT_ATTRIBS;
+ index.i[1] *= TGSI_EXEC_MAX_INPUT_ATTRIBS;
+ index.i[2] *= TGSI_EXEC_MAX_INPUT_ATTRIBS;
+ index.i[3] *= TGSI_EXEC_MAX_INPUT_ATTRIBS;
break;
case TGSI_FILE_CONSTANT:
- index.i[0] *= 4096;
- index.i[1] *= 4096;
- index.i[2] *= 4096;
- index.i[3] *= 4096;
+ index.i[0] *= TGSI_EXEC_MAX_CONST_BUFFER;
+ index.i[1] *= TGSI_EXEC_MAX_CONST_BUFFER;
+ index.i[2] *= TGSI_EXEC_MAX_CONST_BUFFER;
+ index.i[3] *= TGSI_EXEC_MAX_CONST_BUFFER;
break;
default:
assert( 0 );
index.i[2] += reg->SrcRegisterDim.Index;
index.i[3] += reg->SrcRegisterDim.Index;
+ /* Again, the second subscript index can be addressed indirectly
+ * identically to the first one.
+ * Nothing stops us from indirectly addressing the indirect register,
+ * but there is no need for that, so we won't exercise it.
+ *
+ * file[1][ind[4].y+3],
+ * where:
+ * ind = SrcRegisterDimInd.File
+ * [4] = SrcRegisterDimInd.Index
+ * .y = SrcRegisterDimInd.SwizzleX
+ */
if (reg->SrcRegisterDim.Indirect) {
union tgsi_exec_channel index2;
union tgsi_exec_channel indir_index;
+ const uint execmask = mach->ExecMask;
+ uint i;
index2.i[0] =
index2.i[1] =
&index2,
&indir_index );
- index.i[0] += indir_index.i[0];
- index.i[1] += indir_index.i[1];
- index.i[2] += indir_index.i[2];
- index.i[3] += indir_index.i[3];
+ index.i[0] += (int) indir_index.f[0];
+ index.i[1] += (int) indir_index.f[1];
+ index.i[2] += (int) indir_index.f[2];
+ index.i[3] += (int) indir_index.f[3];
+
+ /* for disabled execution channels, zero-out the index to
+ * avoid using a potential garbage value.
+ */
+ for (i = 0; i < QUAD_SIZE; i++) {
+ if ((execmask & (1 << i)) == 0)
+ index.i[i] = 0;
+ }
}
+
+ /* If by any chance there was a need for a 3D array of register
+ * files, we would have to check whether SrcRegisterDim is followed
+ * by a dimension register and continue the saga.
+ */
}
swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
/*
- * Fetch a texel using STR texture coordinates.
+ * Fetch a four texture samples using STR texture coordinates.
*/
static void
fetch_texel( struct tgsi_sampler *sampler,
boolean projected)
{
const uint unit = inst->FullSrcRegisters[1].SrcRegister.Index;
- union tgsi_exec_channel r[8];
+ union tgsi_exec_channel r[4];
uint chan_index;
float lodBias;
else
lodBias = 0.0;
- fetch_texel(&mach->Samplers[unit],
+ fetch_texel(mach->Samplers[unit],
&r[0], NULL, NULL, lodBias, /* S, T, P, BIAS */
&r[0], &r[1], &r[2], &r[3]); /* R, G, B, A */
break;
else
lodBias = 0.0;
- fetch_texel(&mach->Samplers[unit],
+ fetch_texel(mach->Samplers[unit],
&r[0], &r[1], &r[2], lodBias, /* inputs */
&r[0], &r[1], &r[2], &r[3]); /* outputs */
break;
else
lodBias = 0.0;
- fetch_texel(&mach->Samplers[unit],
+ fetch_texel(mach->Samplers[unit],
&r[0], &r[1], &r[2], lodBias,
&r[0], &r[1], &r[2], &r[3]);
break;
break;
default:
+ eval = NULL;
assert( 0 );
}
case TGSI_OPCODE_ARL:
FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( &r[0], 0, chan_index );
- micro_f2it( &r[0], &r[0] );
+ micro_trunc( &r[0], &r[0] );
STORE( &r[0], 0, chan_index );
}
break;
micro_flr( &r[1], &r[0] ); /* r1 = floor(r0) */
if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
micro_exp2( &r[2], &r[1] ); /* r2 = 2 ^ r1 */
- STORE( &r[2], 0, CHAN_X ); /* store r2 */
+ STORE( &r[2], 0, CHAN_X ); /* store r2 */
}
if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
micro_sub( &r[2], &r[0], &r[1] ); /* r2 = r0 - r1 */
- STORE( &r[2], 0, CHAN_Y ); /* store r2 */
+ STORE( &r[2], 0, CHAN_Y ); /* store r2 */
}
if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
micro_exp2( &r[2], &r[0] ); /* r2 = 2 ^ r0 */
- STORE( &r[2], 0, CHAN_Z ); /* store r2 */
+ STORE( &r[2], 0, CHAN_Z ); /* store r2 */
}
if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
- STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+ STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
}
break;
micro_lg2( &r[1], &r[2] ); /* r1 = lg2(r2) */
micro_flr( &r[0], &r[1] ); /* r0 = floor(r1) */
if (IS_CHANNEL_ENABLED( *inst, CHAN_X )) {
- STORE( &r[0], 0, CHAN_X );
+ STORE( &r[0], 0, CHAN_X );
}
if (IS_CHANNEL_ENABLED( *inst, CHAN_Y )) {
micro_exp2( &r[0], &r[0] ); /* r0 = 2 ^ r0 */
micro_div( &r[0], &r[2], &r[0] ); /* r0 = r2 / r0 */
- STORE( &r[0], 0, CHAN_Y );
+ STORE( &r[0], 0, CHAN_Y );
}
if (IS_CHANNEL_ENABLED( *inst, CHAN_Z )) {
- STORE( &r[1], 0, CHAN_Z );
+ STORE( &r[1], 0, CHAN_Z );
}
if (IS_CHANNEL_ENABLED( *inst, CHAN_W )) {
- STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
+ STORE( &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], 0, CHAN_W );
}
break;
case TGSI_OPCODE_DOT2ADD:
/* TGSI_OPCODE_DP2A */
- assert (0);
+ FETCH( &r[0], 0, CHAN_X );
+ FETCH( &r[1], 1, CHAN_X );
+ micro_mul( &r[0], &r[0], &r[1] );
+
+ FETCH( &r[1], 0, CHAN_Y );
+ FETCH( &r[2], 1, CHAN_Y );
+ micro_mul( &r[1], &r[1], &r[2] );
+ micro_add( &r[0], &r[0], &r[1] );
+
+ FETCH( &r[2], 2, CHAN_X );
+ micro_add( &r[0], &r[0], &r[2] );
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( &r[0], 0, chan_index );
+ }
break;
case TGSI_OPCODE_INDEX:
break;
case TGSI_OPCODE_ROUND:
+ case TGSI_OPCODE_ARR:
FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( &r[0], 0, chan_index );
micro_rnd( &r[0], &r[0] );
/* TGSI_OPCODE_EX2 */
FETCH(&r[0], 0, CHAN_X);
+#if FAST_MATH
+ micro_exp2( &r[0], &r[0] );
+#else
micro_pow( &r[0], &mach->Temps[TEMP_2_I].xyzw[TEMP_2_C], &r[0] );
+#endif
FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
STORE( &r[0], 0, chan_index );
assert (0);
break;
- case TGSI_OPCODE_ARR:
- assert (0);
- break;
-
case TGSI_OPCODE_BRA:
assert (0);
break;
break;
case TGSI_OPCODE_SSG:
- assert (0);
+ /* TGSI_OPCODE_SGN */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ FETCH( &r[0], 0, chan_index );
+ micro_sgn( &r[0], &r[0] );
+ STORE( &r[0], 0, chan_index );
+ }
break;
case TGSI_OPCODE_CMP:
break;
case TGSI_OPCODE_NRM:
- assert (0);
+ /* 3-component vector normalize */
+ {
+ union tgsi_exec_channel tmp, dot;
+
+ /* tmp = dp3(src0, src0): */
+ FETCH( &r[0], 0, CHAN_X );
+ micro_mul( &tmp, &r[0], &r[0] );
+
+ FETCH( &r[1], 0, CHAN_Y );
+ micro_mul( &dot, &r[1], &r[1] );
+ micro_add( &tmp, &tmp, &dot );
+
+ FETCH( &r[2], 0, CHAN_Z );
+ micro_mul( &dot, &r[2], &r[2] );
+ micro_add( &tmp, &tmp, &dot );
+
+ /* tmp = 1 / sqrt(tmp) */
+ micro_sqrt( &tmp, &tmp );
+ micro_div( &tmp, &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &tmp );
+
+ /* note: w channel is undefined */
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ /* chan = chan * tmp */
+ micro_mul( &r[chan_index], &tmp, &r[chan_index] );
+ STORE( &r[chan_index], 0, chan_index );
+ }
+ }
+ break;
+
+ case TGSI_OPCODE_NRM4:
+ /* 4-component vector normalize */
+ {
+ union tgsi_exec_channel tmp, dot;
+
+ /* tmp = dp4(src0, src0): */
+ FETCH( &r[0], 0, CHAN_X );
+ micro_mul( &tmp, &r[0], &r[0] );
+
+ FETCH( &r[1], 0, CHAN_Y );
+ micro_mul( &dot, &r[1], &r[1] );
+ micro_add( &tmp, &tmp, &dot );
+
+ FETCH( &r[2], 0, CHAN_Z );
+ micro_mul( &dot, &r[2], &r[2] );
+ micro_add( &tmp, &tmp, &dot );
+
+ FETCH( &r[3], 0, CHAN_W );
+ micro_mul( &dot, &r[3], &r[3] );
+ micro_add( &tmp, &tmp, &dot );
+
+ /* tmp = 1 / sqrt(tmp) */
+ micro_sqrt( &tmp, &tmp );
+ micro_div( &tmp, &mach->Temps[TEMP_1_I].xyzw[TEMP_1_C], &tmp );
+
+ FOR_EACH_ENABLED_CHANNEL( *inst, chan_index ) {
+ /* chan = chan * tmp */
+ micro_mul( &r[chan_index], &tmp, &r[chan_index] );
+ STORE( &r[chan_index], 0, chan_index );
+ }
+ }
break;
case TGSI_OPCODE_DIV: