Mesa (gallium-0.2): gallium: ppc: use a src register cache to avoid redundant loads
Brian Paul
brianp at kemper.freedesktop.org
Tue Oct 28 00:25:52 UTC 2008
Module: Mesa
Branch: gallium-0.2
Commit: 604be5561f17042f61db42b31caf4d720cf66389
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=604be5561f17042f61db42b31caf4d720cf66389
Author: Brian Paul <brian.paul at tungstengraphics.com>
Date: Mon Oct 27 15:36:25 2008 -0600
gallium: ppc: use a src register cache to avoid redundant loads
---
src/gallium/auxiliary/tgsi/tgsi_ppc.c | 300 ++++++++++++++++++++++-----------
1 files changed, 204 insertions(+), 96 deletions(-)
diff --git a/src/gallium/auxiliary/tgsi/tgsi_ppc.c b/src/gallium/auxiliary/tgsi/tgsi_ppc.c
index 9ad7ecd..000ddba 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_ppc.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_ppc.c
@@ -72,11 +72,14 @@ const float ppc_builtin_constants[] ALIGN16_ATTRIB = {
#define CHAN_Z 2
#define CHAN_W 3
-#define TEMP_ONE_I TGSI_EXEC_TEMP_ONE_I
-#define TEMP_ONE_C TGSI_EXEC_TEMP_ONE_C
-#define TEMP_R0 TGSI_EXEC_TEMP_R0
-#define TEMP_ADDR TGSI_EXEC_TEMP_ADDR
+
+struct reg_chan_vec
+{
+ struct tgsi_full_src_register src;
+ uint chan;
+ uint vec;
+};
/**
@@ -94,6 +97,18 @@ struct gen_context
int one_vec; /**< vector register with {1.0, 1.0, 1.0, 1.0} */
int bit31_vec; /**< vector register with {1<<31, 1<<31, 1<<31, 1<<31} */
+
+
+ /**
+ * Cache of src registers.
+ * This is used to avoid redundant load instructions.
+ */
+ struct {
+ struct tgsi_full_src_register src;
+ uint chan;
+ uint vec;
+ } regs[12]; /* 3 src regs, 4 channels */
+ uint num_regs;
};
@@ -261,8 +276,83 @@ emit_fetch(struct gen_context *gen,
}
}
-#define FETCH( GEN, INST, DST_VEC, SRC_REG, CHAN ) \
- emit_fetch( GEN, DST_VEC, &(INST).FullSrcRegisters[SRC_REG], CHAN )
+
+
+/**
+ * Test if two TGSI src registers refer to the same memory location.
+ * We use this to avoid redundant register loads.
+ */
+static boolean
+equal_src_locs(const struct tgsi_full_src_register *a, uint chan_a,
+ const struct tgsi_full_src_register *b, uint chan_b)
+{
+ int swz_a, swz_b;
+ int sign_a, sign_b;
+ if (a->SrcRegister.File != b->SrcRegister.File)
+ return FALSE;
+ if (a->SrcRegister.Index != b->SrcRegister.Index)
+ return FALSE;
+ swz_a = tgsi_util_get_full_src_register_extswizzle(a, chan_a);
+ swz_b = tgsi_util_get_full_src_register_extswizzle(b, chan_b);
+ if (swz_a != swz_b)
+ return FALSE;
+ sign_a = tgsi_util_get_full_src_register_sign_mode(a, chan_a);
+ sign_b = tgsi_util_get_full_src_register_sign_mode(b, chan_b);
+ if (sign_a != sign_b)
+ return FALSE;
+ return TRUE;
+}
+
+
+/**
+ * Given a TGSI src register and channel index, return the PPC vector
+ * register containing the value. We use a cache to prevent re-loading
+ * the same register multiple times.
+ * \return index of PPC vector register with the desired src operand
+ */
+static int
+get_src_vec(struct gen_context *gen,
+ struct tgsi_full_instruction *inst, int src_reg, uint chan)
+{
+ const const struct tgsi_full_src_register *src =
+ &inst->FullSrcRegisters[src_reg];
+ int vec;
+ uint i;
+
+ /* check the cache */
+ for (i = 0; i < gen->num_regs; i++) {
+ if (equal_src_locs(&gen->regs[i].src, gen->regs[i].chan, src, chan)) {
+ /* cache hit */
+ return gen->regs[i].vec;
+ }
+ }
+
+ /* cache miss: allocate new vec reg and emit fetch/load code */
+ vec = ppc_allocate_vec_register(gen->f);
+ gen->regs[gen->num_regs].src = *src;
+ gen->regs[gen->num_regs].chan = chan;
+ gen->regs[gen->num_regs].vec = vec;
+ gen->num_regs++;
+ emit_fetch(gen, vec, src, chan);
+
+ assert(gen->num_regs <= Elements(gen->regs));
+
+ return vec;
+}
+
+
+/**
+ * Clear the src operand cache. To be called at the end of each emit function.
+ */
+static void
+release_src_vecs(struct gen_context *gen)
+{
+ uint i;
+ for (i = 0; i < gen->num_regs; i++) {
+ ppc_release_vec_register(gen->f, gen->regs[i].vec);
+ }
+ gen->num_regs = 0;
+}
@@ -333,11 +423,10 @@ emit_store(struct gen_context *gen,
static void
emit_scalar_unaryop(struct gen_context *gen, struct tgsi_full_instruction *inst)
{
- int v0 = ppc_allocate_vec_register(gen->f);
- int v1 = ppc_allocate_vec_register(gen->f);
+ int v0, v1 = ppc_allocate_vec_register(gen->f);
uint chan_index;
- FETCH(gen, *inst, v0, 0, CHAN_X);
+ v0 = get_src_vec(gen, inst, 0, CHAN_X);
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_RSQ:
@@ -355,7 +444,8 @@ emit_scalar_unaryop(struct gen_context *gen, struct tgsi_full_instruction *inst)
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
STORE(gen, *inst, v1, 0, chan_index);
}
- ppc_release_vec_register(gen->f, v0);
+
+ release_src_vecs(gen);
ppc_release_vec_register(gen->f, v1);
}
@@ -363,10 +453,9 @@ emit_scalar_unaryop(struct gen_context *gen, struct tgsi_full_instruction *inst)
static void
emit_unaryop(struct gen_context *gen, struct tgsi_full_instruction *inst)
{
- int v0 = ppc_allocate_vec_register(gen->f);
uint chan_index;
FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan_index) {
- FETCH(gen, *inst, 0, 0, chan_index); /* v0 = srcreg[0] */
+ int v0 = get_src_vec(gen, inst, 0, chan_index); /* v0 = srcreg[0] */
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_ABS:
/* turn off the most significant bit of each vector float word */
@@ -404,20 +493,30 @@ emit_unaryop(struct gen_context *gen, struct tgsi_full_instruction *inst)
}
STORE(gen, *inst, v0, 0, chan_index); /* store v0 */
}
- ppc_release_vec_register(gen->f, v0);
+
+ release_src_vecs(gen);
}
static void
emit_binop(struct gen_context *gen, struct tgsi_full_instruction *inst)
{
- int v0 = ppc_allocate_vec_register(gen->f);
- int v1 = ppc_allocate_vec_register(gen->f);
- int v2 = ppc_allocate_vec_register(gen->f);
- uint chan_index;
- FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan_index) {
- FETCH(gen, *inst, v0, 0, chan_index); /* v0 = srcreg[0] */
- FETCH(gen, *inst, v1, 1, chan_index); /* v1 = srcreg[1] */
+ int v2, zero_vec = -1;
+ uint chan;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_MUL) {
+ zero_vec = ppc_allocate_vec_register(gen->f);
+ ppc_vzero(gen->f, zero_vec);
+ }
+
+ v2 = ppc_allocate_vec_register(gen->f);
+
+ FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan) {
+ /* fetch src operands */
+ int v0 = get_src_vec(gen, inst, 0, chan);
+ int v1 = get_src_vec(gen, inst, 1, chan);
+
+ /* emit binop */
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_ADD:
ppc_vaddfp(gen->f, v2, v0, v1);
@@ -426,8 +525,7 @@ emit_binop(struct gen_context *gen, struct tgsi_full_instruction *inst)
ppc_vsubfp(gen->f, v2, v0, v1);
break;
case TGSI_OPCODE_MUL:
- ppc_vxor(gen->f, v2, v2, v2); /* v2 = {0, 0, 0, 0} */
- ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v0 */
+ ppc_vmaddfp(gen->f, v2, v0, v1, zero_vec);
break;
case TGSI_OPCODE_MIN:
ppc_vminfp(gen->f, v2, v0, v1);
@@ -438,11 +536,54 @@ emit_binop(struct gen_context *gen, struct tgsi_full_instruction *inst)
default:
assert(0);
}
- STORE(gen, *inst, v2, 0, chan_index); /* store v2 */
+
+ /* store v2 */
+ STORE(gen, *inst, v2, 0, chan);
}
- ppc_release_vec_register(gen->f, v0);
- ppc_release_vec_register(gen->f, v1);
+
ppc_release_vec_register(gen->f, v2);
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_MUL)
+ ppc_release_vec_register(gen->f, zero_vec);
+
+ release_src_vecs(gen);
+}
+
+
+static void
+emit_triop(struct gen_context *gen, struct tgsi_full_instruction *inst)
+{
+ int v3;
+ uint chan;
+
+ v3 = ppc_allocate_vec_register(gen->f);
+
+ FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan) {
+ /* fetch src operands */
+ int v0 = get_src_vec(gen, inst, 0, chan);
+ int v1 = get_src_vec(gen, inst, 1, chan);
+ int v2 = get_src_vec(gen, inst, 2, chan);
+
+ /* emit ALU */
+ switch (inst->Instruction.Opcode) {
+ case TGSI_OPCODE_MAD:
+ ppc_vmaddfp(gen->f, v3, v0, v1, v2); /* v3 = v0 * v1 + v2 */
+ break;
+ case TGSI_OPCODE_LRP:
+ ppc_vsubfp(gen->f, v3, v1, v2); /* v3 = v1 - v2 */
+ ppc_vmaddfp(gen->f, v3, v0, v3, v2); /* v3 = v0 * v3 + v2 */
+ break;
+ default:
+ assert(0);
+ }
+
+ /* store v3 */
+ STORE(gen, *inst, v3, 0, chan);
+ }
+
+ ppc_release_vec_register(gen->f, v3);
+
+ release_src_vecs(gen);
}
@@ -452,16 +593,17 @@ emit_binop(struct gen_context *gen, struct tgsi_full_instruction *inst)
static void
emit_inequality(struct gen_context *gen, struct tgsi_full_instruction *inst)
{
- int v0 = ppc_allocate_vec_register(gen->f);
- int v1 = ppc_allocate_vec_register(gen->f);
- int v2 = ppc_allocate_vec_register(gen->f);
- uint chan_index;
- boolean complement = FALSE;
+ int v2;
+ uint chan;
int one_vec = gen_one_vec(gen);
- FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan_index) {
- FETCH(gen, *inst, v0, 0, chan_index); /* v0 = srcreg[0] */
- FETCH(gen, *inst, v1, 1, chan_index); /* v1 = srcreg[1] */
+ v2 = ppc_allocate_vec_register(gen->f);
+
+ FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan) {
+ /* fetch src operands */
+ int v0 = get_src_vec(gen, inst, 0, chan);
+ int v1 = get_src_vec(gen, inst, 1, chan);
+ boolean complement = FALSE;
switch (inst->Instruction.Opcode) {
case TGSI_OPCODE_SNE:
@@ -495,89 +637,58 @@ emit_inequality(struct gen_context *gen, struct tgsi_full_instruction *inst)
else
ppc_vand(gen->f, v2, one_vec, v2); /* v2 = one_vec & v2 */
- STORE(gen, *inst, v2, 0, chan_index); /* store v2 */
+ /* store v2 */
+ STORE(gen, *inst, v2, 0, chan);
}
- ppc_release_vec_register(gen->f, v0);
- ppc_release_vec_register(gen->f, v1);
ppc_release_vec_register(gen->f, v2);
+
+ release_src_vecs(gen);
}
static void
emit_dotprod(struct gen_context *gen, struct tgsi_full_instruction *inst)
{
- int v0 = ppc_allocate_vec_register(gen->f);
- int v1 = ppc_allocate_vec_register(gen->f);
- int v2 = ppc_allocate_vec_register(gen->f);
+ int v0, v1, v2;
uint chan_index;
+ v2 = ppc_allocate_vec_register(gen->f);
+
ppc_vxor(gen->f, v2, v2, v2); /* v2 = {0, 0, 0, 0} */
- FETCH(gen, *inst, v0, 0, CHAN_X); /* v0 = src0.XXXX */
- FETCH(gen, *inst, v1, 1, CHAN_X); /* v1 = src1.XXXX */
+ v0 = get_src_vec(gen, inst, 0, CHAN_X); /* v0 = src0.XXXX */
+ v1 = get_src_vec(gen, inst, 1, CHAN_X); /* v1 = src1.XXXX */
ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */
- FETCH(gen, *inst, v0, 0, CHAN_Y); /* v0 = src0.YYYY */
- FETCH(gen, *inst, v1, 1, CHAN_Y); /* v1 = src1.YYYY */
+ v0 = get_src_vec(gen, inst, 0, CHAN_Y); /* v0 = src0.YYYY */
+ v1 = get_src_vec(gen, inst, 1, CHAN_Y); /* v1 = src1.YYYY */
ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */
- FETCH(gen, *inst, v0, 0, CHAN_Z); /* v0 = src0.ZZZZ */
- FETCH(gen, *inst, v1, 1, CHAN_Z); /* v1 = src1.ZZZZ */
+ v0 = get_src_vec(gen, inst, 0, CHAN_Z); /* v0 = src0.ZZZZ */
+ v1 = get_src_vec(gen, inst, 1, CHAN_Z); /* v1 = src1.ZZZZ */
ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */
if (inst->Instruction.Opcode == TGSI_OPCODE_DP4) {
- FETCH(gen, *inst, v0, 0, CHAN_W); /* v0 = src0.WWWW */
- FETCH(gen, *inst, v1, 1, CHAN_W); /* v1 = src1.WWWW */
- ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */
+ v0 = get_src_vec(gen, inst, 0, CHAN_W); /* v0 = src0.WWWW */
+ v1 = get_src_vec(gen, inst, 1, CHAN_W); /* v1 = src1.WWWW */
+ ppc_vmaddfp(gen->f, v2, v0, v1, v2); /* v2 = v0 * v1 + v2 */
}
else if (inst->Instruction.Opcode == TGSI_OPCODE_DPH) {
- FETCH(gen, *inst, v1, 1, CHAN_W); /* v1 = src1.WWWW */
- ppc_vaddfp(gen->f, v2, v2, v1); /* v2 = v2 + v1 */
+ v1 = get_src_vec(gen, inst, 1, CHAN_W); /* v1 = src1.WWWW */
+ ppc_vaddfp(gen->f, v2, v2, v1); /* v2 = v2 + v1 */
}
FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan_index) {
STORE(gen, *inst, v2, 0, chan_index); /* store v2 */
}
- ppc_release_vec_register(gen->f, v0);
- ppc_release_vec_register(gen->f, v1);
- ppc_release_vec_register(gen->f, v2);
-}
+ release_src_vecs(gen);
-static void
-emit_triop(struct gen_context *gen, struct tgsi_full_instruction *inst)
-{
- int v0 = ppc_allocate_vec_register(gen->f);
- int v1 = ppc_allocate_vec_register(gen->f);
- int v2 = ppc_allocate_vec_register(gen->f);
- int v3 = ppc_allocate_vec_register(gen->f);
- uint chan_index;
- FOR_EACH_DST0_ENABLED_CHANNEL(*inst, chan_index) {
- FETCH(gen, *inst, v0, 0, chan_index); /* v0 = srcreg[0] */
- FETCH(gen, *inst, v1, 1, chan_index); /* v1 = srcreg[1] */
- FETCH(gen, *inst, v2, 2, chan_index); /* v2 = srcreg[2] */
- switch (inst->Instruction.Opcode) {
- case TGSI_OPCODE_MAD:
- ppc_vmaddfp(gen->f, v3, v0, v1, v2); /* v3 = v0 * v1 + v2 */
- break;
- case TGSI_OPCODE_LRP:
- ppc_vsubfp(gen->f, v3, v1, v2); /* v3 = v1 - v2 */
- ppc_vmaddfp(gen->f, v3, v0, v3, v2); /* v3 = v0 * v3 + v2 */
- break;
- default:
- assert(0);
- }
- STORE(gen, *inst, v3, 0, chan_index); /* store v3 */
- }
- ppc_release_vec_register(gen->f, v0);
- ppc_release_vec_register(gen->f, v1);
ppc_release_vec_register(gen->f, v2);
- ppc_release_vec_register(gen->f, v3);
}
-
/** Approximation for vr = pow(va, vb) */
static void
ppc_vec_pow(struct ppc_function *f, int vr, int va, int vb)
@@ -610,10 +721,10 @@ emit_lit(struct gen_context *gen, struct tgsi_full_instruction *inst)
/* Compute Y, Z */
if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Y) ||
IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
- int x_vec = ppc_allocate_vec_register(gen->f);
+ int x_vec;
int zero_vec = ppc_allocate_vec_register(gen->f);
- FETCH(gen, *inst, x_vec, 0, CHAN_X); /* x_vec = src[0].x */
+ x_vec = get_src_vec(gen, inst, 0, CHAN_X); /* x_vec = src[0].x */
ppc_vzero(gen->f, zero_vec); /* zero = {0,0,0,0} */
ppc_vmaxfp(gen->f, x_vec, x_vec, zero_vec); /* x_vec = max(x_vec, 0) */
@@ -623,18 +734,16 @@ emit_lit(struct gen_context *gen, struct tgsi_full_instruction *inst)
}
if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_Z)) {
- int y_vec = ppc_allocate_vec_register(gen->f);
- int z_vec = ppc_allocate_vec_register(gen->f);
- int w_vec = ppc_allocate_vec_register(gen->f);
+ int y_vec, z_vec, w_vec;
int pow_vec = ppc_allocate_vec_register(gen->f);
int pos_vec = ppc_allocate_vec_register(gen->f);
int p128_vec = ppc_allocate_vec_register(gen->f);
int n128_vec = ppc_allocate_vec_register(gen->f);
- FETCH(gen, *inst, y_vec, 0, CHAN_Y); /* y_vec = src[0].y */
+ y_vec = get_src_vec(gen, inst, 0, CHAN_Y); /* y_vec = src[0].y */
ppc_vmaxfp(gen->f, y_vec, y_vec, zero_vec); /* y_vec = max(y_vec, 0) */
- FETCH(gen, *inst, w_vec, 0, CHAN_W); /* w_vec = src[0].w */
+ w_vec = get_src_vec(gen, inst, 0, CHAN_W); /* w_vec = src[0].w */
/* clamp Y to [-128, 128] */
load_constant_vec(gen, p128_vec, 128.0f);
@@ -653,16 +762,12 @@ emit_lit(struct gen_context *gen, struct tgsi_full_instruction *inst)
STORE(gen, *inst, z_vec, 0, CHAN_Z); /* store Z */
- ppc_release_vec_register(gen->f, y_vec);
- ppc_release_vec_register(gen->f, z_vec);
- ppc_release_vec_register(gen->f, w_vec);
ppc_release_vec_register(gen->f, pow_vec);
ppc_release_vec_register(gen->f, pos_vec);
ppc_release_vec_register(gen->f, p128_vec);
ppc_release_vec_register(gen->f, n128_vec);
}
- ppc_release_vec_register(gen->f, x_vec);
ppc_release_vec_register(gen->f, zero_vec);
}
@@ -670,6 +775,8 @@ emit_lit(struct gen_context *gen, struct tgsi_full_instruction *inst)
if (IS_DST0_CHANNEL_ENABLED(*inst, CHAN_W)) {
STORE(gen, *inst, one_vec, 0, CHAN_W);
}
+
+ release_src_vecs(gen);
}
@@ -723,11 +830,10 @@ emit_instruction(struct gen_context *gen,
default:
return 0;
}
-
-
return 1;
}
+
static void
emit_declaration(
struct ppc_function *func,
@@ -805,6 +911,7 @@ emit_epilogue(struct ppc_function *func)
{
ppc_return(func);
/* XXX restore prev stack frame */
+ debug_printf("PPC: Emitted %u instructions\n", func->num_inst);
}
@@ -839,6 +946,7 @@ tgsi_emit_ppc(const struct tgsi_token *tokens,
util_init_math();
+ memset(&gen, 0, sizeof(gen));
gen.f = func;
gen.inputs_reg = ppc_reserve_register(func, 3); /* first function param */
gen.outputs_reg = ppc_reserve_register(func, 4); /* second function param */
More information about the mesa-commit
mailing list