Boxis R700 User Manual download pdf (Page 67)

ATI R700 Technology

Scalar Operands 4-9

4.6.4.5 Dynamically-Indexed Constant Access (AR-relative, Constant Waterfalling)

To support DX9 vertex shaders, we provide dynamic indexing of constant-file

constants. This means that a GPR value is used as the index into the constant

file. Since the value comes from a GPR, it can be unique for each pixel. In the

worst case, it may take 64 times as long to execute this instruction, since up to

64 constant-file reads can be required.

Dynamic indexing requires two instructions:

• MOVA: Move one element of a GPR into the Address Register (AR) to be used

as the index value.

• <any ALU instruction>: Use the indices from the MOVA and perform

the indirect lookup.

There is a two-instruction delay slot between loading and using the GPR index

value. Hardware inserts delays if the kernel does not. The GPR indices loaded

by a MOVA instruction only persist for one clause; at the end of the clause they

are invalidated.

4.6.4.6 ALU Constant Buffer Sharing

ES, GS, and VS kernels can, on a per-ALU-clause basis, access their own

constant buffers or those of the other shader type.

ES/VS can use their own or use GS constant buffers, and GS can use its own

or ES/VS ones. This is provided for cases when the GS and VS shaders can be

merged into a single hardware shader stage.

This capability is activated by setting the ALT_CONSTS bit in the

SQ_CF_ALU_WORD1.

4.7 Scalar Operands

For each instruction, the operands src0, src1, and src2 are specified in the

instruction’s SRC*_SEL and SRC*_ELEM fields. GPR and constant-register

addresses can be relative-addressed, as specified in the SRC*_REL and

INDEX_MODE fields. In the OP2 microcode format, src2 is undefined.

4.7.1 Source Addresses

The data source address is specified in the SRC*_SEL field. This can refer to one

of the following.

• A GPR address, GPR[0, 127], with values [0, 127].

• A kcache constant in bank 0, kcache0[0, 31], with values [128, 159];

kcache0[16, 31] are accessible only if two cache lines have been locked.

• A kcache constant in bank 1, kcache1[0, 31], with values [160, 191];

kcache1[16, 31] are accessible only if two cache lines are locked.

• A constant-register address, c[0, 255], with values [256, 511].

1 2 ... 62 63 64 65 66 67 68 69 70 71 72 ... 391 392

No comments

Boxis R700 User Manual Page 67