diff options
Diffstat (limited to 'src/mesa/drivers')
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp | 85 |
1 files changed, 58 insertions, 27 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp b/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp index c4ea7eca322..67693106fa3 100644 --- a/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp +++ b/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp @@ -36,13 +36,17 @@ extern "C" { namespace brw { +struct copy_entry { + src_reg *value[4]; + int saturatemask; +}; + static bool is_direct_copy(vec4_instruction *inst) { return (inst->opcode == BRW_OPCODE_MOV && !inst->predicate && inst->dst.file == GRF && - !inst->saturate && !inst->dst.reladdr && !inst->src[0].reladdr && inst->dst.type == inst->src[0].type); @@ -74,16 +78,16 @@ is_channel_updated(vec4_instruction *inst, src_reg *values[4], int ch) static bool try_constant_propagate(struct brw_context *brw, vec4_instruction *inst, - int arg, src_reg *values[4]) + int arg, struct copy_entry *entry) { /* For constant propagation, we only handle the same constant * across all 4 channels. Some day, we should handle the 8-bit * float vector format, which would let us constant propagate * vectors better. */ - src_reg value = *values[0]; + src_reg value = *entry->value[0]; for (int i = 1; i < 4; i++) { - if (!value.equals(*values[i])) + if (!value.equals(*entry->value[i])) return false; } @@ -213,22 +217,22 @@ is_logic_op(enum opcode opcode) static bool try_copy_propagate(struct brw_context *brw, vec4_instruction *inst, - int arg, src_reg *values[4]) + int arg, struct copy_entry *entry, int reg) { /* For constant propagation, we only handle the same constant * across all 4 channels. Some day, we should handle the 8-bit * float vector format, which would let us constant propagate * vectors better. */ - src_reg value = *values[0]; + src_reg value = *entry->value[0]; for (int i = 1; i < 4; i++) { /* This is equals() except we don't care about the swizzle. */ - if (value.file != values[i]->file || - value.reg != values[i]->reg || - value.reg_offset != values[i]->reg_offset || - value.type != values[i]->type || - value.negate != values[i]->negate || - value.abs != values[i]->abs) { + if (value.file != entry->value[i]->file || + value.reg != entry->value[i]->reg || + value.reg_offset != entry->value[i]->reg_offset || + value.type != entry->value[i]->type || + value.negate != entry->value[i]->negate || + value.abs != entry->value[i]->abs) { return false; } } @@ -239,7 +243,7 @@ try_copy_propagate(struct brw_context *brw, vec4_instruction *inst, */ int s[4]; for (int i = 0; i < 4; i++) { - s[i] = BRW_GET_SWZ(values[i]->swizzle, + s[i] = BRW_GET_SWZ(entry->value[i]->swizzle, BRW_GET_SWZ(inst->src[arg].swizzle, i)); } value.swizzle = BRW_SWIZZLE4(s[0], s[1], s[2], s[3]); @@ -300,6 +304,25 @@ try_copy_propagate(struct brw_context *brw, vec4_instruction *inst, if (value.equals(inst->src[arg])) return false; + /* Limit saturate propagation only to SEL with src1 bounded within 1.0 and 1.0 + * otherwise, skip copy propagate altogether + */ + if (entry->saturatemask & (1 << arg)) { + switch(inst->opcode) { + case BRW_OPCODE_SEL: + if (inst->src[1].file != IMM || + inst->src[1].fixed_hw_reg.dw1.f < 0.0 || + inst->src[1].fixed_hw_reg.dw1.f > 1.0) { + return false; + } + if (!inst->saturate) + inst->saturate = true; + break; + default: + return false; + } + } + value.type = inst->src[arg].type; inst->src[arg] = value; return true; @@ -309,9 +332,9 @@ bool vec4_visitor::opt_copy_propagation() { bool progress = false; - src_reg *cur_value[virtual_grf_reg_count][4]; + struct copy_entry entries[virtual_grf_reg_count]; - memset(&cur_value, 0, sizeof(cur_value)); + memset(&entries, 0, sizeof(entries)); foreach_in_list(vec4_instruction, inst, &instructions) { /* This pass only works on basic blocks. If there's flow @@ -322,7 +345,7 @@ vec4_visitor::opt_copy_propagation() * src/glsl/opt_copy_propagation.cpp to track available copies. */ if (!is_dominated_by_previous_instruction(inst)) { - memset(cur_value, 0, sizeof(cur_value)); + memset(&entries, 0, sizeof(entries)); continue; } @@ -343,33 +366,38 @@ vec4_visitor::opt_copy_propagation() /* Find the regs that each swizzle component came from. */ - src_reg *values[4]; + struct copy_entry entry; + memset(&entry, 0, sizeof(copy_entry)); int c; for (c = 0; c < 4; c++) { - values[c] = cur_value[reg][BRW_GET_SWZ(inst->src[i].swizzle, c)]; + int channel = BRW_GET_SWZ(inst->src[i].swizzle, c); + entry.value[c] = entries[reg].value[channel]; /* If there's no available copy for this channel, bail. * We could be more aggressive here -- some channels might * not get used based on the destination writemask. */ - if (!values[c]) + if (!entry.value[c]) break; + entry.saturatemask |= + (entries[reg].saturatemask & (1 << channel) ? 1 : 0) << c; + /* We'll only be able to copy propagate if the sources are * all from the same file -- there's no ability to swizzle * 0 or 1 constants in with source registers like in i915. */ - if (c > 0 && values[c - 1]->file != values[c]->file) + if (c > 0 && entry.value[c - 1]->file != entry.value[c]->file) break; } if (c != 4) continue; - if (try_constant_propagate(brw, inst, i, values)) + if (try_constant_propagate(brw, inst, i, &entry)) progress = true; - if (try_copy_propagate(brw, inst, i, values)) + if (try_copy_propagate(brw, inst, i, &entry, reg)) progress = true; } @@ -383,9 +411,11 @@ vec4_visitor::opt_copy_propagation() * the new value, so clear it. */ bool direct_copy = is_direct_copy(inst); + entries[reg].saturatemask = 0x0; for (int i = 0; i < 4; i++) { if (inst->dst.writemask & (1 << i)) { - cur_value[reg][i] = direct_copy ? &inst->src[0] : NULL; + entries[reg].value[i] = direct_copy ? &inst->src[0] : NULL; + entries[reg].saturatemask |= (((inst->saturate && direct_copy) ? 1 : 0) << i); } } @@ -393,13 +423,14 @@ vec4_visitor::opt_copy_propagation() * our destination's updated channels, as the two are no longer equal. */ if (inst->dst.reladdr) - memset(cur_value, 0, sizeof(cur_value)); + memset(&entries, 0, sizeof(entries)); else { for (int i = 0; i < virtual_grf_reg_count; i++) { for (int j = 0; j < 4; j++) { - if (is_channel_updated(inst, cur_value[i], j)){ - cur_value[i][j] = NULL; - } + if (is_channel_updated(inst, entries[i].value, j)){ + entries[i].value[j] = NULL; + entries[i].saturatemask &= ~(1 << j); + } } } } |