#if defined(PIPE_ARCH_SSE)
#include "util/u_sse.h"
#endif
+#include "tgsi/tgsi_info.h"
#include "tgsi/tgsi_parse.h"
#include "tgsi/tgsi_util.h"
-#include "tgsi_exec.h"
-#include "tgsi_sse2.h"
+#include "tgsi/tgsi_dump.h"
+#include "tgsi/tgsi_exec.h"
+#include "tgsi/tgsi_sse2.h"
#include "rtasm/rtasm_x86sse.h"
/* 'vec' is the offset from the address register's value.
* We're loading CONST[ADDR+vec] into an xmm register.
*/
- struct x86_reg r0 = get_input_base();
- struct x86_reg r1 = get_output_base();
+ struct x86_reg r0 = get_immediate_base();
+ struct x86_reg r1 = get_coef_base();
uint i;
assert( indirectFile == TGSI_FILE_ADDRESS );
assert( indirectIndex == 0 );
+ assert( r0.mod == mod_REG );
+ assert( r1.mod == mod_REG );
x86_push( func, r0 );
x86_push( func, r1 );
const struct tgsi_full_src_register *reg,
const unsigned chan_index )
{
- unsigned swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
+ unsigned swizzle = tgsi_util_get_full_src_register_swizzle( reg, chan_index );
switch (swizzle) {
- case TGSI_EXTSWIZZLE_X:
- case TGSI_EXTSWIZZLE_Y:
- case TGSI_EXTSWIZZLE_Z:
- case TGSI_EXTSWIZZLE_W:
+ case TGSI_SWIZZLE_X:
+ case TGSI_SWIZZLE_Y:
+ case TGSI_SWIZZLE_Z:
+ case TGSI_SWIZZLE_W:
switch (reg->SrcRegister.File) {
case TGSI_FILE_CONSTANT:
emit_const(
}
break;
- case TGSI_EXTSWIZZLE_ZERO:
- emit_tempf(
- func,
- xmm,
- TGSI_EXEC_TEMP_00000000_I,
- TGSI_EXEC_TEMP_00000000_C );
- break;
-
- case TGSI_EXTSWIZZLE_ONE:
- emit_tempf(
- func,
- xmm,
- TEMP_ONE_I,
- TEMP_ONE_C );
- break;
-
default:
assert( 0 );
}
const struct tgsi_full_instruction *inst,
unsigned chan_index )
{
+ switch( inst->Instruction.Saturate ) {
+ case TGSI_SAT_NONE:
+ break;
+
+ case TGSI_SAT_ZERO_ONE:
+ sse_maxps(
+ func,
+ make_xmm( xmm ),
+ get_temp(
+ TGSI_EXEC_TEMP_00000000_I,
+ TGSI_EXEC_TEMP_00000000_C ) );
+
+ sse_minps(
+ func,
+ make_xmm( xmm ),
+ get_temp(
+ TGSI_EXEC_TEMP_ONE_I,
+ TGSI_EXEC_TEMP_ONE_C ) );
+ break;
+
+ case TGSI_SAT_MINUS_PLUS_ONE:
+ assert( 0 );
+ break;
+ }
+
+
switch( reg->DstRegister.File ) {
case TGSI_FILE_OUTPUT:
emit_output(
default:
assert( 0 );
}
-
- switch( inst->Instruction.Saturate ) {
- case TGSI_SAT_NONE:
- break;
-
- case TGSI_SAT_ZERO_ONE:
- /* assert( 0 ); */
- break;
-
- case TGSI_SAT_MINUS_PLUS_ONE:
- assert( 0 );
- break;
- }
}
#define STORE( FUNC, INST, XMM, INDEX, CHAN )\
{
float rgba[NUM_CHANNELS][QUAD_SIZE];
(*sampler)->get_samples(*sampler,
- &store[0],
- &store[4],
- &store[8],
- 0.0f, /*store[12], lodbias */
- rgba);
+ &store[0], /* s */
+ &store[4], /* t */
+ &store[8], /* r */
+ store[12], /* lodbias */
+ rgba); /* results */
memcpy( store, rgba, 16 * sizeof(float));
}
switch (inst->InstructionExtTexture.Texture) {
case TGSI_TEXTURE_1D:
- case TGSI_TEXTURE_SHADOW1D:
count = 1;
break;
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_RECT:
- case TGSI_TEXTURE_SHADOW2D:
- case TGSI_TEXTURE_SHADOWRECT:
count = 2;
break;
+ case TGSI_TEXTURE_SHADOW1D:
+ case TGSI_TEXTURE_SHADOW2D:
+ case TGSI_TEXTURE_SHADOWRECT:
case TGSI_TEXTURE_3D:
case TGSI_TEXTURE_CUBE:
count = 3;
/* This mask stores component bits that were already tested. Note that
* we test if the value is less than zero, so 1.0 and 0.0 need not to be
* tested. */
- uniquemask = (1 << TGSI_EXTSWIZZLE_ZERO) | (1 << TGSI_EXTSWIZZLE_ONE);
+ uniquemask = 0;
FOR_EACH_CHANNEL( chan_index ) {
unsigned swizzle;
/* unswizzle channel */
- swizzle = tgsi_util_get_full_src_register_extswizzle(
+ swizzle = tgsi_util_get_full_src_register_swizzle(
reg,
chan_index );
break;
case TGSI_OPCODE_MOV:
- case TGSI_OPCODE_SWZ:
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
- FETCH( func, *inst, 0, 0, chan_index );
- STORE( func, *inst, 0, 0, chan_index );
+ FETCH( func, *inst, 4 + chan_index, 0, chan_index );
+ }
+ FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+ STORE( func, *inst, 4 + chan_index, 0, chan_index );
}
break;
break;
case TGSI_OPCODE_RCP:
- /* TGSI_OPCODE_RECIP */
FETCH( func, *inst, 0, 0, CHAN_X );
emit_rcp( func, 0, 0 );
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
break;
case TGSI_OPCODE_RSQ:
- /* TGSI_OPCODE_RECIPSQRT */
FETCH( func, *inst, 0, 0, CHAN_X );
emit_abs( func, 0 );
emit_rsqrt( func, 1, 0 );
break;
case TGSI_OPCODE_DP3:
- /* TGSI_OPCODE_DOT3 */
FETCH( func, *inst, 0, 0, CHAN_X );
FETCH( func, *inst, 1, 1, CHAN_X );
emit_mul( func, 0, 1 );
break;
case TGSI_OPCODE_DP4:
- /* TGSI_OPCODE_DOT4 */
FETCH( func, *inst, 0, 0, CHAN_X );
FETCH( func, *inst, 1, 1, CHAN_X );
emit_mul( func, 0, 1 );
break;
case TGSI_OPCODE_SLT:
- /* TGSI_OPCODE_SETLT */
emit_setcc( func, inst, cc_LessThan );
break;
case TGSI_OPCODE_SGE:
- /* TGSI_OPCODE_SETGE */
emit_setcc( func, inst, cc_NotLessThan );
break;
case TGSI_OPCODE_MAD:
- /* TGSI_OPCODE_MADD */
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( func, *inst, 0, 0, chan_index );
FETCH( func, *inst, 1, 1, chan_index );
}
break;
- case TGSI_OPCODE_LERP:
- /* TGSI_OPCODE_LRP */
+ case TGSI_OPCODE_LRP:
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( func, *inst, 0, 0, chan_index );
FETCH( func, *inst, 1, 1, chan_index );
return 0;
break;
- case TGSI_OPCODE_CND0:
- return 0;
- break;
-
- case TGSI_OPCODE_DOT2ADD:
- /* TGSI_OPCODE_DP2A */
+ case TGSI_OPCODE_DP2A:
FETCH( func, *inst, 0, 0, CHAN_X ); /* xmm0 = src[0].x */
FETCH( func, *inst, 1, 1, CHAN_X ); /* xmm1 = src[1].x */
emit_mul( func, 0, 1 ); /* xmm0 = xmm0 * xmm1 */
}
break;
- case TGSI_OPCODE_INDEX:
- return 0;
- break;
-
- case TGSI_OPCODE_NEGATE:
- return 0;
- break;
-
- case TGSI_OPCODE_FRAC:
- /* TGSI_OPCODE_FRC */
+ case TGSI_OPCODE_FRC:
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( func, *inst, 0, 0, chan_index );
emit_frc( func, 0, 0 );
return 0;
break;
- case TGSI_OPCODE_FLOOR:
- /* TGSI_OPCODE_FLR */
+ case TGSI_OPCODE_FLR:
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( func, *inst, 0, 0, chan_index );
emit_flr( func, 0, 0 );
}
break;
- case TGSI_OPCODE_EXPBASE2:
- /* TGSI_OPCODE_EX2 */
+ case TGSI_OPCODE_EX2:
FETCH( func, *inst, 0, 0, CHAN_X );
emit_ex2( func, 0, 0 );
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
}
break;
- case TGSI_OPCODE_LOGBASE2:
- /* TGSI_OPCODE_LG2 */
+ case TGSI_OPCODE_LG2:
FETCH( func, *inst, 0, 0, CHAN_X );
emit_lg2( func, 0, 0 );
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
}
break;
- case TGSI_OPCODE_POWER:
- /* TGSI_OPCODE_POW */
+ case TGSI_OPCODE_POW:
FETCH( func, *inst, 0, 0, CHAN_X );
FETCH( func, *inst, 1, 1, CHAN_X );
emit_pow( func, 0, 0, 0, 1 );
}
break;
- case TGSI_OPCODE_CROSSPRODUCT:
- /* TGSI_OPCODE_XPD */
+ case TGSI_OPCODE_XPD:
if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
FETCH( func, *inst, 1, 1, CHAN_Z );
}
break;
- case TGSI_OPCODE_MULTIPLYMATRIX:
- return 0;
- break;
-
case TGSI_OPCODE_ABS:
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( func, *inst, 0, 0, chan_index );
break;
case TGSI_OPCODE_SEQ:
- return 0;
+ emit_setcc( func, inst, cc_Equal );
break;
case TGSI_OPCODE_SFL:
break;
case TGSI_OPCODE_SGT:
- return 0;
+ emit_setcc( func, inst, cc_NotLessThanEqual );
break;
case TGSI_OPCODE_SIN:
break;
case TGSI_OPCODE_SLE:
- return 0;
+ emit_setcc( func, inst, cc_LessThanEqual );
break;
case TGSI_OPCODE_SNE:
- return 0;
+ emit_setcc( func, inst, cc_NotEqual );
break;
case TGSI_OPCODE_STR:
break;
case TGSI_OPCODE_SSG:
- /* TGSI_OPCODE_SGN */
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
FETCH( func, *inst, 0, 0, chan_index );
emit_sgn( func, 0, 0 );
return 0;
break;
- case TGSI_OPCODE_LOOP:
+ case TGSI_OPCODE_BGNFOR:
return 0;
break;
return 0;
break;
- case TGSI_OPCODE_ENDLOOP:
+ case TGSI_OPCODE_ENDFOR:
return 0;
break;
func,
get_coef_base(),
x86_fn_arg( func, 4 ) );
-
- x86_mov(
- func,
- get_sampler_base(),
- x86_make_disp( get_machine_base(),
- Offset( struct tgsi_exec_machine, Samplers ) ) );
}
+ x86_mov(
+ func,
+ get_sampler_base(),
+ x86_make_disp( get_machine_base(),
+ Offset( struct tgsi_exec_machine, Samplers ) ) );
+
while( !tgsi_parse_end_of_tokens( &parse ) && ok ) {
tgsi_parse_token( &parse );
&parse.FullToken.FullInstruction );
if (!ok) {
- debug_printf("failed to translate tgsi opcode %d to SSE (%s)\n",
- parse.FullToken.FullInstruction.Instruction.Opcode,
+ uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
+ debug_printf("failed to translate tgsi opcode %d (%s) to SSE (%s)\n",
+ opcode,
+ tgsi_get_opcode_name(opcode),
parse.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX ?
"vertex shader" : "fragment shader");
}
+
+ if (tgsi_check_soa_dependencies(&parse.FullToken.FullInstruction)) {
+ uint opcode = parse.FullToken.FullInstruction.Instruction.Opcode;
+
+ /* XXX: we only handle src/dst aliasing in a few opcodes
+ * currently. Need to use an additional temporay to hold
+ * the result in the cases where the code is too opaque to
+ * fix.
+ */
+ if (opcode != TGSI_OPCODE_MOV) {
+ debug_printf("Warning: src/dst aliasing in instruction"
+ " is not handled:\n");
+ tgsi_dump_instruction(&parse.FullToken.FullInstruction, 1);
+ }
+ }
break;
case TGSI_TOKEN_TYPE_IMMEDIATE:
assert(num_immediates < TGSI_EXEC_NUM_IMMEDIATES);
for( i = 0; i < size; i++ ) {
immediates[num_immediates][i] =
- parse.FullToken.FullImmediate.u.ImmediateFloat32[i].Float;
+ parse.FullToken.FullImmediate.u[i].Float;
}
#if 0
debug_printf("SSE FS immediate[%d] = %f %f %f %f\n",
projects / mesa.git / blobdiff