diff options
Diffstat (limited to 'src/gallium/drivers/cell/ppu')
| -rw-r--r-- | src/gallium/drivers/cell/ppu/Makefile | 3 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_context.h | 1 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_gen_fp.c | 523 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_gen_fp.h | 42 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_gen_fragment.c | 870 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_gen_fragment.h | 38 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_state_emit.c | 93 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_state_per_fragment.c | 30 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_state_shader.c | 8 | ||||
| -rw-r--r-- | src/gallium/drivers/cell/ppu/cell_vertex_fetch.c | 5 |
10 files changed, 1547 insertions, 66 deletions
diff --git a/src/gallium/drivers/cell/ppu/Makefile b/src/gallium/drivers/cell/ppu/Makefile index 25473e200c..b28f4c5c31 100644 --- a/src/gallium/drivers/cell/ppu/Makefile +++ b/src/gallium/drivers/cell/ppu/Makefile @@ -25,9 +25,10 @@ SOURCES = \ cell_context.c \ cell_draw_arrays.c \ cell_flush.c \ + cell_gen_fragment.c \ + cell_gen_fp.c \ cell_state_derived.c \ cell_state_emit.c \ - cell_state_per_fragment.c \ cell_state_shader.c \ cell_pipe_state.c \ cell_screen.c \ diff --git a/src/gallium/drivers/cell/ppu/cell_context.h b/src/gallium/drivers/cell/ppu/cell_context.h index 8cec9f45b2..14914b9c6f 100644 --- a/src/gallium/drivers/cell/ppu/cell_context.h +++ b/src/gallium/drivers/cell/ppu/cell_context.h @@ -61,6 +61,7 @@ struct cell_fragment_shader_state { struct pipe_shader_state shader; struct tgsi_shader_info info; + struct spe_function code; void *data; }; diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.c b/src/gallium/drivers/cell/ppu/cell_gen_fp.c new file mode 100644 index 0000000000..6ffe94eb14 --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.c @@ -0,0 +1,523 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + + +/** + * Generate SPU fragment program/shader code. + * + * Note that we generate SOA-style code here. So each TGSI instruction + * operates on four pixels (and is translated into four SPU instructions, + * generally speaking). + * + * \author Brian Paul + */ + + +#include "pipe/p_defines.h" +#include "pipe/p_state.h" +#include "pipe/p_shader_tokens.h" +#include "tgsi/tgsi_parse.h" +#include "tgsi/tgsi_util.h" +#include "tgsi/tgsi_exec.h" +#include "tgsi/tgsi_dump.h" +#include "rtasm/rtasm_ppc_spe.h" +#include "util/u_memory.h" +#include "cell_context.h" +#include "cell_gen_fp.h" + + +/** Set to 1 to enable debug/disassembly printfs */ +#define DISASSEM 01 + + +/** + * Context needed during code generation. + */ +struct codegen +{ + int inputs_reg; /**< 1st function parameter */ + int outputs_reg; /**< 2nd function parameter */ + int constants_reg; /**< 3rd function parameter */ + int temp_regs[8][4]; /**< maps TGSI temps to SPE registers */ + + int one_reg; /**< register containing {1.0, 1.0, 1.0, 1.0} */ + + /** Per-instruction temps / intermediate temps */ + int num_itemps; + int itemps[3]; + + struct spe_function *f; + boolean error; +}; + + +/** + * Allocate an intermediate temporary register. + */ +static int +get_itemp(struct codegen *gen) +{ + int t = spe_allocate_available_register(gen->f); + assert(gen->num_itemps < Elements(gen->itemps)); + gen->itemps[gen->num_itemps++] = t; + return t; +} + +/** + * Free all intermediate temporary registers. To be called after each + * instruction has been emitted. + */ +static void +free_itemps(struct codegen *gen) +{ + int i; + for (i = 0; i < gen->num_itemps; i++) { + spe_release_register(gen->f, gen->itemps[i]); + } + gen->num_itemps = 0; +} + + +/** + * Return index of an SPE register containing {1.0, 1.0, 1.0, 1.0}. + * The register is allocated and initialized upon the first call. + */ +static int +get_const_one_reg(struct codegen *gen) +{ + if (gen->one_reg <= 0) { + gen->one_reg = spe_allocate_available_register(gen->f); + } + + /* one = {1.0, 1.0, 1.0, 1.0} */ + spe_load_float(gen->f, gen->one_reg, 1.0f); +#if DISASSEM + printf("il\tr%d, 1.0f\n", gen->one_reg); +#endif + + return gen->one_reg; +} + + +/** + * Return the index of the SPU temporary containing the named TGSI + * source register. If the TGSI register is a TGSI_FILE_TEMPORARY we + * just return the corresponding SPE register. If the TGIS register + * is TGSI_FILE_INPUT/CONSTANT/IMMEDIATE we allocate a new SPE register + * and emit an SPE load instruction. + */ +static int +get_src_reg(struct codegen *gen, + int channel, + const struct tgsi_full_src_register *src) +{ + int reg; + + /* XXX need to examine src swizzle info here. + * That will involve changing the channel var... + */ + + + switch (src->SrcRegister.File) { + case TGSI_FILE_TEMPORARY: + reg = gen->temp_regs[src->SrcRegister.Index][channel]; + break; + case TGSI_FILE_INPUT: + { + /* offset is measured in quadwords, not bytes */ + int offset = src->SrcRegister.Index * 4 + channel; + reg = get_itemp(gen); + /* Load: reg = memory[(machine_reg) + offset] */ + spe_lqd(gen->f, reg, gen->inputs_reg, offset); +#if DISASSEM + printf("lqd\tr%d, r%d + %d\n", reg, gen->inputs_reg, offset); +#endif + } + break; + case TGSI_FILE_IMMEDIATE: + /* xxx fall-through for now / fix */ + case TGSI_FILE_CONSTANT: + /* xxx fall-through for now / fix */ + default: + assert(0); + } + + return reg; +} + + +/** + * Return the index of an SPE register to use for the given TGSI register. + * If the TGSI register is TGSI_FILE_TEMPORARAY, the index of the + * corresponding SPE register is returned. If the TGSI register is + * TGSI_FILE_OUTPUT we allocate an intermediate temporary register. + * See store_dest_reg() below... + */ +static int +get_dst_reg(struct codegen *gen, + int channel, + const struct tgsi_full_dst_register *dest) +{ + int reg; + + switch (dest->DstRegister.File) { + case TGSI_FILE_TEMPORARY: + reg = gen->temp_regs[dest->DstRegister.Index][channel]; + break; + case TGSI_FILE_OUTPUT: + reg = get_itemp(gen); + break; + default: + assert(0); + } + + return reg; +} + + +/** + * When a TGSI instruction is writing to an output register, this + * function emits the SPE store instruction to store the value_reg. + * \param value_reg the SPE register containing the value to store. + * This would have been returned by get_dst_reg(). + */ +static void +store_dest_reg(struct codegen *gen, + int value_reg, int channel, + const struct tgsi_full_dst_register *dest) +{ + switch (dest->DstRegister.File) { + case TGSI_FILE_TEMPORARY: + /* no-op */ + break; + case TGSI_FILE_OUTPUT: + { + /* offset is measured in quadwords, not bytes */ + int offset = dest->DstRegister.Index * 4 + channel; + /* Store: memory[(machine_reg) + offset] = reg */ + spe_stqd(gen->f, value_reg, gen->outputs_reg, offset); +#if DISASSEM + printf("stqd\tr%d, r%d + %d\n", value_reg, gen->outputs_reg, offset); +#endif + } + break; + default: + assert(0); + } +} + + +static boolean +emit_MOV(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + int src_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + int dst_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + /* XXX we don't always need to actually emit a mov instruction here */ + spe_move(gen->f, dst_reg, src_reg); +#if DISASSEM + printf("mov\tr%d, r%d\n", dst_reg, src_reg); +#endif + store_dest_reg(gen, dst_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + return true; +} + + +/** + * Emit addition instructions. Recall that a single TGSI_OPCODE_ADD + * becomes (up to) four SPU "fa" instructions because we're doing SOA + * processing. + */ +static boolean +emit_ADD(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + /* Loop over Red/Green/Blue/Alpha channels */ + for (ch = 0; ch < 4; ch++) { + /* If the dest R, G, B or A writemask is enabled... */ + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + /* get indexes of the two src, one dest SPE registers */ + int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + + /* Emit actual SPE instruction: d = s1 + s2 */ + spe_fa(gen->f, d_reg, s1_reg, s2_reg); +#if DISASSEM + printf("fa\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg); +#endif + + /* Store the result (a no-op for TGSI_FILE_TEMPORARY dests) */ + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + /* Free any intermediate temps we allocated */ + free_itemps(gen); + } + } + return true; +} + + +/** + * Emit multiply. See emit_ADD for comments. + */ +static boolean +emit_MUL(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + /* d = s1 * s2 */ + spe_fm(gen->f, d_reg, s1_reg, s2_reg); +#if DISASSEM + printf("fm\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg); +#endif + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + return true; +} + + +/** + * Emit set-if-greater-than. + * Note that the SPE fcgt instruction produces 0x0 and 0xffffffff as + * the result but OpenGL/TGSI needs 0.0 and 1.0 results. + * We can easily convert 0x0/0xffffffff to 0.0/1.0 with a bitwise AND. + */ +static boolean +emit_SGT(struct codegen *gen, const struct tgsi_full_instruction *inst) +{ + int ch; + + for (ch = 0; ch < 4; ch++) { + if (inst->FullDstRegisters[0].DstRegister.WriteMask & (1 << ch)) { + int s1_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[0]); + int s2_reg = get_src_reg(gen, ch, &inst->FullSrcRegisters[1]); + int d_reg = get_dst_reg(gen, ch, &inst->FullDstRegisters[0]); + + /* d = (s1 > s2) */ + spe_fcgt(gen->f, d_reg, s1_reg, s2_reg); +#if DISASSEM + printf("fcgt\tr%d, r%d, r%d\n", d_reg, s1_reg, s2_reg); +#endif + + /* convert d from 0x0/0xffffffff to 0.0/1.0 */ + /* d = d & one_reg */ + spe_and(gen->f, d_reg, d_reg, get_const_one_reg(gen)); +#if DISASSEM + printf("and\tr%d, r%d, r%d\n", d_reg, d_reg, get_const_one_reg(gen)); +#endif + + store_dest_reg(gen, d_reg, ch, &inst->FullDstRegisters[0]); + free_itemps(gen); + } + } + + return true; +} + + +/** + * Emit END instruction. + * We just return from the shader function at this point. + * + * Note that there may be more code after this that would be + * called by TGSI_OPCODE_CALL. + */ +static boolean +emit_END(struct codegen *gen) +{ + /* return from function call */ + spe_bi(gen->f, SPE_REG_RA, 0, 0); +#if DISASSEM + printf("bi\trRA\n"); +#endif + return true; +} + + +/** + * Emit code for the given instruction. Just a big switch stmt. + */ +static boolean +emit_instruction(struct codegen *gen, + const struct tgsi_full_instruction *inst) +{ + switch (inst->Instruction.Opcode) { + case TGSI_OPCODE_MOV: + return emit_MOV(gen, inst); + case TGSI_OPCODE_MUL: + return emit_MUL(gen, inst); + case TGSI_OPCODE_ADD: + return emit_ADD(gen, inst); + case TGSI_OPCODE_SGT: + return emit_SGT(gen, inst); + case TGSI_OPCODE_END: + return emit_END(gen); + + /* XXX lots more cases to do... */ + + default: + return false; + } + + return true; +} + + + +/** + * Emit "code" for a TGSI declaration. + * We only care about TGSI TEMPORARY register declarations at this time. + * For each TGSI TEMPORARY we allocate four SPE registers. + */ +static void +emit_declaration(struct codegen *gen, const struct tgsi_full_declaration *decl) +{ + int i, ch; + + switch (decl->Declaration.File) { + case TGSI_FILE_TEMPORARY: +#if DISASSEM + printf("Declare temp reg %d .. %d\n", + decl->DeclarationRange.First, + decl->DeclarationRange.Last); +#endif + for (i = decl->DeclarationRange.First; + i <= decl->DeclarationRange.Last; + i++) { + for (ch = 0; ch < 4; ch++) { + gen->temp_regs[i][ch] = spe_allocate_available_register(gen->f); + } + + /* XXX if we run out of SPE registers, we need to spill + * to SPU memory. someday... + */ + +#if DISASSEM + printf(" SPE regs: %d %d %d %d\n", + gen->temp_regs[i][0], + gen->temp_regs[i][1], + gen->temp_regs[i][2], + gen->temp_regs[i][3]); +#endif + } + break; + default: + ; /* ignore */ + } +} + + +/** + * Translate TGSI shader code to SPE instructions. This is done when + * the state tracker gives us a new shader (via pipe->create_fs_state()). + * + * \param cell the rendering context (in) + * \param tokens the TGSI shader (in) + * \param f the generated function (out) + */ +boolean +cell_gen_fragment_program(struct cell_context *cell, + const struct tgsi_token *tokens, + struct spe_function *f) +{ + struct tgsi_parse_context parse; + struct codegen gen; + + memset(&gen, 0, sizeof(gen)); + gen.f = f; + + /* For SPE function calls: reg $3 = first param, $4 = second param, etc. */ + gen.inputs_reg = 3; /* pointer to inputs array */ + gen.outputs_reg = 4; /* pointer to outputs array */ + gen.constants_reg = 5; /* pointer to constants array */ + + spe_init_func(f, SPU_MAX_FRAGMENT_PROGRAM_INSTS * SPE_INST_SIZE); + spe_allocate_register(f, gen.inputs_reg); + spe_allocate_register(f, gen.outputs_reg); + spe_allocate_register(f, gen.constants_reg); + +#if DISASSEM + printf("Begin %s\n", __FUNCTION__); + tgsi_dump(tokens, 0); +#endif + + tgsi_parse_init(&parse, tokens); + + while (!tgsi_parse_end_of_tokens(&parse) && !gen.error) { + tgsi_parse_token(&parse); + + switch (parse.FullToken.Token.Type) { + case TGSI_TOKEN_TYPE_IMMEDIATE: +#if 0 + if (!note_immediate(&gen, &parse.FullToken.FullImmediate )) + goto fail; +#endif + break; + + case TGSI_TOKEN_TYPE_DECLARATION: + emit_declaration(&gen, &parse.FullToken.FullDeclaration); + break; + + case TGSI_TOKEN_TYPE_INSTRUCTION: + if (!emit_instruction(&gen, &parse.FullToken.FullInstruction )) { + gen.error = true; + } + break; + + default: + assert(0); + + } + } + + + if (gen.error) { + /* terminate the SPE code */ + return emit_END(&gen); + } + +#if DISASSEM + printf("cell_gen_fragment_program nr instructions: %d\n", f->num_inst); + printf("End %s\n", __FUNCTION__); +#endif + + tgsi_parse_free( &parse ); + + return !gen.error; +} diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fp.h b/src/gallium/drivers/cell/ppu/cell_gen_fp.h new file mode 100644 index 0000000000..99faea7046 --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_gen_fp.h @@ -0,0 +1,42 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + + +#ifndef CELL_GEN_FP_H +#define CELL_GEN_FP_H + + + +extern boolean +cell_gen_fragment_program(struct cell_context *cell, + const struct tgsi_token *tokens, + struct spe_function *f); + + +#endif /* CELL_GEN_FP_H */ + diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.c b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c new file mode 100644 index 0000000000..06219d4e98 --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.c @@ -0,0 +1,870 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + + +/** + * Generate SPU per-fragment code (actually per-quad code). + * \author Brian Paul + */ + + +#include "pipe/p_defines.h" +#include "pipe/p_state.h" +#include "rtasm/rtasm_ppc_spe.h" +#include "cell_context.h" +#include "cell_gen_fragment.h" + + + +/** Do extra optimizations? */ +#define OPTIMIZATIONS 1 + + +/** + * Generate SPE code to perform Z/depth testing. + * + * \param dsa Gallium depth/stencil/alpha state to gen code for + * \param f SPE function to append instruction onto. + * \param mask_reg register containing quad/pixel "alive" mask (in/out) + * \param ifragZ_reg register containing integer fragment Z values (in) + * \param ifbZ_reg register containing integer frame buffer Z values (in/out) + * \param zmask_reg register containing result of Z test/comparison (out) + */ +static void +gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa, + struct spe_function *f, + int mask_reg, int ifragZ_reg, int ifbZ_reg, int zmask_reg) +{ + ASSERT(dsa->depth.enabled); + + switch (dsa->depth.func) { + case PIPE_FUNC_EQUAL: + /* zmask = (ifragZ == ref) */ + spe_ceq(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & zmask) */ + spe_and(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_NOTEQUAL: + /* zmask = (ifragZ == ref) */ + spe_ceq(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & ~zmask) */ + spe_andc(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_GREATER: + /* zmask = (ifragZ > ref) */ + spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & zmask) */ + spe_and(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_LESS: + /* zmask = (ref > ifragZ) */ + spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); + /* mask = (mask & zmask) */ + spe_and(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_LEQUAL: + /* zmask = (ifragZ > ref) */ + spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg); + /* mask = (mask & ~zmask) */ + spe_andc(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_GEQUAL: + /* zmask = (ref > ifragZ) */ + spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg); + /* mask = (mask & ~zmask) */ + spe_andc(f, mask_reg, mask_reg, zmask_reg); + break; + + case PIPE_FUNC_NEVER: + spe_il(f, mask_reg, 0); /* mask = {0,0,0,0} */ + spe_move(f, zmask_reg, mask_reg); /* zmask = mask */ + break; + + case PIPE_FUNC_ALWAYS: + /* mask unchanged */ + spe_il(f, zmask_reg, ~0); /* zmask = {~0,~0,~0,~0} */ + break; + + default: + ASSERT(0); + break; + } + + if (dsa->depth.writemask) { + /* + * If (ztest passed) { + * framebufferZ = fragmentZ; + * } + * OR, + * framebufferZ = (ztest_passed ? fragmentZ : framebufferZ; + */ + spe_selb(f, ifbZ_reg, ifbZ_reg, ifragZ_reg, mask_reg); + } +} + + +/** + * Generate SPE code to perform alpha testing. + * + * \param dsa Gallium depth/stencil/alpha state to gen code for + * \param f SPE function to append instruction onto. + * \param mask_reg register containing quad/pixel "alive" mask (in/out) + * \param fragA_reg register containing four fragment alpha values (in) + */ +static void +gen_alpha_test(const struct pipe_depth_stencil_alpha_state *dsa, + struct spe_function *f, int mask_reg, int fragA_reg) +{ + int ref_reg = spe_allocate_available_register(f); + int amask_reg = spe_allocate_available_register(f); + + ASSERT(dsa->alpha.enabled); + + if ((dsa->alpha.func != PIPE_FUNC_NEVER) && + (dsa->alpha.func != PIPE_FUNC_ALWAYS)) { + /* load/splat the alpha reference float value */ + spe_load_float(f, ref_reg, dsa->alpha.ref); + } + + /* emit code to do the alpha comparison, updating 'mask' */ + switch (dsa->alpha.func) { + case PIPE_FUNC_EQUAL: + /* amask = (fragA == ref) */ + spe_fceq(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & amask) */ + spe_and(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_NOTEQUAL: + /* amask = (fragA == ref) */ + spe_fceq(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & ~amask) */ + spe_andc(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_GREATER: + /* amask = (fragA > ref) */ + spe_fcgt(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & amask) */ + spe_and(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_LESS: + /* amask = (ref > fragA) */ + spe_fcgt(f, amask_reg, ref_reg, fragA_reg); + /* mask = (mask & amask) */ + spe_and(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_LEQUAL: + /* amask = (fragA > ref) */ + spe_fcgt(f, amask_reg, fragA_reg, ref_reg); + /* mask = (mask & ~amask) */ + spe_andc(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_GEQUAL: + /* amask = (ref > fragA) */ + spe_fcgt(f, amask_reg, ref_reg, fragA_reg); + /* mask = (mask & ~amask) */ + spe_andc(f, mask_reg, mask_reg, amask_reg); + break; + + case PIPE_FUNC_NEVER: + spe_il(f, mask_reg, 0); /* mask = [0,0,0,0] */ + break; + + case PIPE_FUNC_ALWAYS: + /* no-op, mask unchanged */ + break; + + default: + ASSERT(0); + break; + } + +#if OPTIMIZATIONS + /* if mask == {0,0,0,0} we're all done, return */ + { + /* re-use amask reg here */ + int tmp_reg = amask_reg; + /* tmp[0] = (mask[0] | mask[1] | mask[2] | mask[3]) */ + spe_orx(f, tmp_reg, mask_reg); + /* if tmp[0] == 0 then return from function call */ + spe_biz(f, tmp_reg, SPE_REG_RA, 0, 0); + } +#endif + + spe_release_register(f, ref_reg); + spe_release_register(f, amask_reg); +} + + + +/** + * Generate SPE code to implement the given blend mode for a quad of pixels. + * \param f SPE function to append instruction onto. + * \param fragR_reg register with fragment red values (float) (in/out) + * \param fragG_reg register with fragment green values (float) (in/out) + * \param fragB_reg register with fragment blue values (float) (in/out) + * \param fragA_reg register with fragment alpha values (float) (in/out) + * \param fbRGBA_reg register with packed framebuffer colors (integer) (in) + */ +static void +gen_blend(const struct pipe_blend_state *blend, + struct spe_function *f, + enum pipe_format color_format, + int fragR_reg, int fragG_reg, int fragB_reg, int fragA_reg, + int fbRGBA_reg) +{ + int term1R_reg = spe_allocate_available_register(f); + int term1G_reg = spe_allocate_available_register(f); + int term1B_reg = spe_allocate_available_register(f); + int term1A_reg = spe_allocate_available_register(f); + + int term2R_reg = spe_allocate_available_register(f); + int term2G_reg = spe_allocate_available_register(f); + int term2B_reg = spe_allocate_available_register(f); + int term2A_reg = spe_allocate_available_register(f); + + int fbR_reg = spe_allocate_available_register(f); + int fbG_reg = spe_allocate_available_register(f); + int fbB_reg = spe_allocate_available_register(f); + int fbA_reg = spe_allocate_available_register(f); + + int one_reg = spe_allocate_available_register(f); + int tmp_reg = spe_allocate_available_register(f); + + boolean one_reg_set = false; /* avoid setting one_reg more than once */ + + ASSERT(blend->blend_enable); + + /* Unpack/convert framebuffer colors from four 32-bit packed colors + * (fbRGBA) to four float RGBA vectors (fbR, fbG, fbB, fbA). + * Each 8-bit color component is expanded into a float in [0.0, 1.0]. + */ + { + int mask_reg = spe_allocate_available_register(f); + + /* mask = {0x000000ff, 0x000000ff, 0x000000ff, 0x000000ff} */ + spe_load_int(f, mask_reg, 0xff); + + /* XXX there may be more clever ways to implement the following code */ + switch (color_format) { + case PIPE_FORMAT_A8R8G8B8_UNORM: + /* fbB = fbB & mask */ + spe_and(f, fbB_reg, fbRGBA_reg, mask_reg); + /* mask = mask << 8 */ + spe_roti(f, mask_reg, mask_reg, 8); + + /* fbG = fbRGBA & mask */ + spe_and(f, fbG_reg, fbRGBA_reg, mask_reg); + /* fbG = fbG >> 8 */ + spe_roti(f, fbG_reg, fbG_reg, -8); + /* mask = mask << 8 */ + spe_roti(f, mask_reg, mask_reg, 8); + + /* fbR = fbRGBA & mask */ + spe_and(f, fbR_reg, fbRGBA_reg, mask_reg); + /* fbR = fbR >> 16 */ + spe_roti(f, fbR_reg, fbR_reg, -16); + /* mask = mask << 8 */ + spe_roti(f, mask_reg, mask_reg, 8); + + /* fbA = fbRGBA & mask */ + spe_and(f, fbA_reg, fbRGBA_reg, mask_reg); + /* fbA = fbA >> 24 */ + spe_roti(f, fbA_reg, fbA_reg, -24); + break; + + case PIPE_FORMAT_B8G8R8A8_UNORM: + /* fbA = fbA & mask */ + spe_and(f, fbA_reg, fbRGBA_reg, mask_reg); + /* mask = mask << 8 */ + spe_roti(f, mask_reg, mask_reg, 8); + + /* fbR = fbRGBA & mask */ + spe_and(f, fbR_reg, fbRGBA_reg, mask_reg); + /* fbR = fbR >> 8 */ + spe_roti(f, fbR_reg, fbR_reg, -8); + /* mask = mask << 8 */ + spe_roti(f, mask_reg, mask_reg, 8); + + /* fbG = fbRGBA & mask */ + spe_and(f, fbG_reg, fbRGBA_reg, mask_reg); + /* fbG = fbG >> 16 */ + spe_roti(f, fbG_reg, fbG_reg, -16); + /* mask = mask << 8 */ + spe_roti(f, mask_reg, mask_reg, 8); + + /* fbB = fbRGBA & mask */ + spe_and(f, fbB_reg, fbRGBA_reg, mask_reg); + /* fbB = fbB >> 24 */ + spe_roti(f, fbB_reg, fbB_reg, -24); + break; + + default: + ASSERT(0); + } + + /* convert int[4] in [0,255] to float[4] in [0.0, 1.0] */ + spe_cuflt(f, fbR_reg, fbR_reg, 8); + spe_cuflt(f, fbG_reg, fbG_reg, 8); + spe_cuflt(f, fbB_reg, fbB_reg, 8); + spe_cuflt(f, fbA_reg, fbA_reg, 8); + + spe_release_register(f, mask_reg); + } + + + /* + * Compute Src RGB terms + */ + switch (blend->rgb_src_factor) { + case PIPE_BLENDFACTOR_ONE: + spe_move(f, term1R_reg, fragR_reg); + spe_move(f, term1G_reg, fragG_reg); + spe_move(f, term1B_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_ZERO: + spe_zero(f, term1R_reg); + spe_zero(f, term1G_reg); + spe_zero(f, term1B_reg); + break; + case PIPE_BLENDFACTOR_SRC_COLOR: + spe_fm(f, term1R_reg, fragR_reg, fragR_reg); + spe_fm(f, term1G_reg, fragG_reg, fragG_reg); + spe_fm(f, term1B_reg, fragB_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_SRC_ALPHA: + spe_fm(f, term1R_reg, fragR_reg, fragA_reg); + spe_fm(f, term1G_reg, fragG_reg, fragA_reg); + spe_fm(f, term1B_reg, fragB_reg, fragA_reg); + break; + /* XXX more cases */ + default: + ASSERT(0); + } + + /* + * Compute Src Alpha term + */ + switch (blend->alpha_src_factor) { + case PIPE_BLENDFACTOR_ONE: + spe_move(f, term1A_reg, fragA_reg); + break; + case PIPE_BLENDFACTOR_SRC_COLOR: + spe_fm(f, term1A_reg, fragA_reg, fragA_reg); + break; + case PIPE_BLENDFACTOR_SRC_ALPHA: + spe_fm(f, term1A_reg, fragA_reg, fragA_reg); + break; + /* XXX more cases */ + default: + ASSERT(0); + } + + /* + * Compute Dest RGB terms + */ + switch (blend->rgb_dst_factor) { + case PIPE_BLENDFACTOR_ONE: + spe_move(f, term2R_reg, fbR_reg); + spe_move(f, term2G_reg, fbG_reg); + spe_move(f, term2B_reg, fbB_reg); + break; + case PIPE_BLENDFACTOR_ZERO: + spe_zero(f, term2R_reg); + spe_zero(f, term2G_reg); + spe_zero(f, term2B_reg); + break; + case PIPE_BLENDFACTOR_SRC_COLOR: + spe_fm(f, term2R_reg, fbR_reg, fragR_reg); + spe_fm(f, term2G_reg, fbG_reg, fragG_reg); + spe_fm(f, term2B_reg, fbB_reg, fragB_reg); + break; + case PIPE_BLENDFACTOR_SRC_ALPHA: + spe_fm(f, term2R_reg, fbR_reg, fragA_reg); + spe_fm(f, term2G_reg, fbG_reg, fragA_reg); + spe_fm(f, term2B_reg, fbB_reg, fragA_reg); + break; + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + /* one = {1.0, 1.0, 1.0, 1.0} */ + if (!one_reg_set) { + spe_load_float(f, one_reg, 1.0f); + one_reg_set = true; + } + /* tmp = one - fragA */ + spe_fs(f, tmp_reg, one_reg, fragA_reg); + /* term = fb * tmp */ + spe_fm(f, term2R_reg, fbR_reg, tmp_reg); + spe_fm(f, term2G_reg, fbG_reg, tmp_reg); + spe_fm(f, term2B_reg, fbB_reg, tmp_reg); + break; + /* XXX more cases */ + default: + ASSERT(0); + } + + /* + * Compute Dest Alpha term + */ + switch (blend->alpha_dst_factor) { + case PIPE_BLENDFACTOR_ONE: + spe_move(f, term2A_reg, fbA_reg); + break; + case PIPE_BLENDFACTOR_ZERO: + spe_zero(f, term2A_reg); + break; + case PIPE_BLENDFACTOR_SRC_ALPHA: + spe_fm(f, term2A_reg, fbA_reg, fragA_reg); + break; + case PIPE_BLENDFACTOR_INV_SRC_ALPHA: + /* one = {1.0, 1.0, 1.0, 1.0} */ + if (!one_reg_set) { + spe_load_float(f, one_reg, 1.0f); + one_reg_set = true; + } + /* tmp = one - fragA */ + spe_fs(f, tmp_reg, one_reg, fragA_reg); + /* termA = fbA * tmp */ + spe_fm(f, term2A_reg, fbA_reg, tmp_reg); + break; + /* XXX more cases */ + default: + ASSERT(0); + } + + /* + * Combine Src/Dest RGB terms + */ + switch (blend->rgb_func) { + case PIPE_BLEND_ADD: + spe_fa(f, fragR_reg, term1R_reg, term2R_reg); + spe_fa(f, fragG_reg, term1G_reg, term2G_reg); + spe_fa(f, fragB_reg, term1B_reg, term2B_reg); + break; + case PIPE_BLEND_SUBTRACT: + spe_fs(f, fragR_reg, term1R_reg, term2R_reg); + spe_fs(f, fragG_reg, term1G_reg, term2G_reg); + spe_fs(f, fragB_reg, term1B_reg, term2B_reg); + break; + /* XXX more cases */ + default: + ASSERT(0); + } + + /* + * Combine Src/Dest A term + */ + switch (blend->alpha_func) { + case PIPE_BLEND_ADD: + spe_fa(f, fragA_reg, term1A_reg, term2A_reg); + break; + case PIPE_BLEND_SUBTRACT: + spe_fs(f, fragA_reg, term1A_reg, term2A_reg); + break; + /* XXX more cases */ + default: + ASSERT(0); + } + + spe_release_register(f, term1R_reg); + spe_release_register(f, term1G_reg); + spe_release_register(f, term1B_reg); + spe_release_register(f, term1A_reg); + + spe_release_register(f, term2R_reg); + spe_release_register(f, term2G_reg); + spe_release_register(f, term2B_reg); + spe_release_register(f, term2A_reg); + + spe_release_register(f, fbR_reg); + spe_release_register(f, fbG_reg); + spe_release_register(f, fbB_reg); + spe_release_register(f, fbA_reg); + + spe_release_register(f, one_reg); + spe_release_register(f, tmp_reg); +} + + +static void +gen_logicop(const struct pipe_blend_state *blend, + struct spe_function *f, + int fragRGBA_reg, int fbRGBA_reg) +{ + /* XXX to-do */ + /* operate on 32-bit packed pixels, not float colors */ +} + + +static void +gen_colormask(uint colormask, + struct spe_function *f, + int fragRGBA_reg, int fbRGBA_reg) +{ + /* XXX to-do */ + /* operate on 32-bit packed pixels, not float colors */ +} + + + +/** + * Generate code to pack a quad of float colors into a four 32-bit integers. + * + * \param f SPE function to append instruction onto. + * \param color_format the dest color packing format + * \param r_reg register containing four red values (in/clobbered) + * \param g_reg register containing four green values (in/clobbered) + * \param b_reg register containing four blue values (in/clobbered) + * \param a_reg register containing four alpha values (in/clobbered) + * \param rgba_reg register to store the packed RGBA colors (out) + */ +static void +gen_pack_colors(struct spe_function *f, + enum pipe_format color_format, + int r_reg, int g_reg, int b_reg, int a_reg, + int rgba_reg) +{ + /* Convert float[4] in [0.0,1.0] to int[4] in [0,~0], with clamping */ + spe_cfltu(f, r_reg, r_reg, 32); + spe_cfltu(f, g_reg, g_reg, 32); + spe_cfltu(f, b_reg, b_reg, 32); + spe_cfltu(f, a_reg, a_reg, 32); + + /* Shift the most significant bytes to least the significant positions. + * I.e.: reg = reg >> 24 + */ + spe_rotmi(f, r_reg, r_reg, -24); + spe_rotmi(f, g_reg, g_reg, -24); + spe_rotmi(f, b_reg, b_reg, -24); + spe_rotmi(f, a_reg, a_reg, -24); + + /* Shift the color bytes according to the surface format */ + if (color_format == PIPE_FORMAT_A8R8G8B8_UNORM) { + spe_roti(f, g_reg, g_reg, 8); /* green <<= 8 */ + spe_roti(f, r_reg, r_reg, 16); /* red <<= 16 */ + spe_roti(f, a_reg, a_reg, 24); /* alpha <<= 24 */ + } + else if (color_format == PIPE_FORMAT_B8G8R8A8_UNORM) { + spe_roti(f, r_reg, r_reg, 8); /* red <<= 8 */ + spe_roti(f, g_reg, g_reg, 16); /* green <<= 16 */ + spe_roti(f, b_reg, b_reg, 24); /* blue <<= 24 */ + } + else { + ASSERT(0); + } + + /* Merge red, green, blue, alpha registers to make packed RGBA colors. + * Eg: after shifting according to color_format we might have: + * R = {0x00ff0000, 0x00110000, 0x00220000, 0x00330000} + * G = {0x0000ff00, 0x00004400, 0x00005500, 0x00006600} + * B = {0x000000ff, 0x00000077, 0x00000088, 0x00000099} + * A = {0xff000000, 0xaa000000, 0xbb000000, 0xcc000000} + * OR-ing all those together gives us four packed colors: + * RGBA = {0xffffffff, 0xaa114477, 0xbb225588, 0xcc336699} + */ + spe_or(f, rgba_reg, r_reg, g_reg); + spe_or(f, rgba_reg, rgba_reg, b_reg); + spe_or(f, rgba_reg, rgba_reg, a_reg); +} + + + + +/** + * Generate SPE code to implement the fragment operations (alpha test, + * depth test, stencil test, blending, colormask, and final + * framebuffer write) as specified by the current context state. + * + * Logically, this code will be called after running the fragment + * shader. But under some circumstances we could run some of this + * code before the fragment shader to cull fragments/quads that are + * totally occluded/discarded. + * + * XXX we only support PIPE_FORMAT_Z24S8_UNORM z/stencil buffer right now. + * + * See the spu_default_fragment_ops() function to see how the per-fragment + * operations would be done with ordinary C code. + * The code we generate here though has no branches, is SIMD, etc and + * should be much faster. + * + * \param cell the rendering context (in) + * \param f the generated function (out) + */ +void +cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f) +{ + const struct pipe_depth_stencil_alpha_state *dsa = + &cell->depth_stencil->base; + const struct pipe_blend_state *blend = &cell->blend->base; + const enum pipe_format color_format = cell->framebuffer.cbufs[0]->format; + + /* For SPE function calls: reg $3 = first param, $4 = second param, etc. */ + const int x_reg = 3; /* uint */ + const int y_reg = 4; /* uint */ + const int color_tile_reg = 5; /* tile_t * */ + const int depth_tile_reg = 6; /* tile_t * */ + const int fragZ_reg = 7; /* vector float */ + const int fragR_reg = 8; /* vector float */ + const int fragG_reg = 9; /* vector float */ + const int fragB_reg = 10; /* vector float */ + const int fragA_reg = 11; /* vector float */ + const int mask_reg = 12; /* vector uint */ + + /* offset of quad from start of tile + * XXX assuming 4-byte pixels for color AND Z/stencil!!!! + */ + int quad_offset_reg; + + int fbRGBA_reg; /**< framebuffer's RGBA colors for quad */ + int fbZS_reg; /**< framebuffer's combined z/stencil values for quad */ + + spe_init_func(f, SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + spe_allocate_register(f, x_reg); + spe_allocate_register(f, y_reg); + spe_allocate_register(f, color_tile_reg); + spe_allocate_register(f, depth_tile_reg); + spe_allocate_register(f, fragZ_reg); + spe_allocate_register(f, fragR_reg); + spe_allocate_register(f, fragG_reg); + spe_allocate_register(f, fragB_reg); + spe_allocate_register(f, fragA_reg); + spe_allocate_register(f, mask_reg); + + quad_offset_reg = spe_allocate_available_register(f); + fbRGBA_reg = spe_allocate_available_register(f); + fbZS_reg = spe_allocate_available_register(f); + + /* compute offset of quad from start of tile, in bytes */ + { + int x2_reg = spe_allocate_available_register(f); + int y2_reg = spe_allocate_available_register(f); + + ASSERT(TILE_SIZE == 32); + + spe_rotmi(f, x2_reg, x_reg, -1); /* x2 = x / 2 */ + spe_rotmi(f, y2_reg, y_reg, -1); /* y2 = y / 2 */ + spe_shli(f, y2_reg, y2_reg, 4); /* y2 *= 16 */ + spe_a(f, quad_offset_reg, y2_reg, x2_reg); /* offset = y2 + x2 */ + spe_shli(f, quad_offset_reg, quad_offset_reg, 4); /* offset *= 16 */ + + spe_release_register(f, x2_reg); + spe_release_register(f, y2_reg); + } + + + if (dsa->alpha.enabled) { + gen_alpha_test(dsa, f, mask_reg, fragA_reg); + } + + if (dsa->depth.enabled || dsa->stencil[0].enabled) { + const enum pipe_format zs_format = cell->framebuffer.zsbuf->format; + boolean write_depth_stencil; + + int fbZ_reg = spe_allocate_available_register(f); /* Z values */ + int fbS_reg = spe_allocate_available_register(f); /* Stencil values */ + + /* fetch quad of depth/stencil values from tile at (x,y) */ + /* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */ + spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + + if (dsa->depth.enabled) { + /* Extract Z bits from fbZS_reg into fbZ_reg */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + int mask_reg = spe_allocate_available_register(f); + spe_fsmbi(f, mask_reg, 0x7777); /* mask[0,1,2,3] = 0x00ffffff */ + spe_and(f, fbZ_reg, fbZS_reg, mask_reg); /* fbZ = fbZS & mask */ + spe_release_register(f, mask_reg); + /* OK, fbZ_reg has four 24-bit Z values now */ + } + else { + /* XXX handle other z/stencil formats */ + ASSERT(0); + } + + /* Convert fragZ values from float[4] to uint[4] */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM || + zs_format == PIPE_FORMAT_Z24S8_UNORM || + zs_format == PIPE_FORMAT_Z24X8_UNORM) { + /* 24-bit Z values */ + int scale_reg = spe_allocate_available_register(f); + + /* scale_reg[0,1,2,3] = float(2^24-1) */ + spe_load_float(f, scale_reg, (float) 0xffffff); + + /* XXX these two instructions might be combined */ + spe_fm(f, fragZ_reg, fragZ_reg, scale_reg); /* fragZ *= scale */ + spe_cfltu(f, fragZ_reg, fragZ_reg, 0); /* fragZ = (int) fragZ */ + + spe_release_register(f, scale_reg); + } + else { + /* XXX handle 16-bit Z format */ + ASSERT(0); + } + } + + if (dsa->stencil[0].enabled) { + /* Extract Stencil bit sfrom fbZS_reg into fbS_reg */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + /* XXX extract with a shift */ + ASSERT(0); + } + else if (zs_format == PIPE_FORMAT_Z24S8_UNORM || + zs_format == PIPE_FORMAT_Z24X8_UNORM) { + /* XXX extract with a mask */ + ASSERT(0); + } + } + + + if (dsa->stencil[0].enabled) { + /* XXX this may involve depth testing too */ + // gen_stencil_test(dsa, f, ... ); + ASSERT(0); + } + else if (dsa->depth.enabled) { + int zmask_reg = spe_allocate_available_register(f); + gen_depth_test(dsa, f, mask_reg, fragZ_reg, fbZ_reg, zmask_reg); + spe_release_register(f, zmask_reg); + } + + /* do we need to write Z and/or Stencil back into framebuffer? */ + write_depth_stencil = (dsa->depth.writemask | + dsa->stencil[0].write_mask | + dsa->stencil[1].write_mask); + + if (write_depth_stencil) { + /* Merge latest Z and Stencil values into fbZS_reg. + * fbZ_reg has four Z vals in bits [23..0] or bits [15..0]. + * fbS_reg has four 8-bit Z values in bits [7..0]. + */ + if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */ + spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */ + } + else if (zs_format == PIPE_FORMAT_S8Z24_UNORM || + zs_format == PIPE_FORMAT_X8Z24_UNORM) { + /* XXX to do */ + ASSERT(0); + } + else if (zs_format == PIPE_FORMAT_Z16_UNORM) { + /* XXX to do */ + ASSERT(0); + } + else if (zs_format == PIPE_FORMAT_S8_UNORM) { + /* XXX to do */ + ASSERT(0); + } + else { + /* bad zs_format */ + ASSERT(0); + } + + /* Store: memory[depth_tile_reg + quad_offset_reg] = fbZS */ + spe_stqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg); + } + + spe_release_register(f, fbZ_reg); + spe_release_register(f, fbS_reg); + } + + + /* Get framebuffer quad/colors. We'll need these for blending, + * color masking, and to obey the quad/pixel mask. + * Load: fbRGBA_reg = memory[color_tile + quad_offset] + * Note: if mask={~0,~0,~0,~0} and we're not blending or colormasking + * we could skip this load. + */ + spe_lqx(f, fbRGBA_reg, color_tile_reg, quad_offset_reg); + + + if (blend->blend_enable) { + gen_blend(blend, f, color_format, + fragR_reg, fragG_reg, fragB_reg, fragA_reg, fbRGBA_reg); + } + + /* + * Write fragment colors to framebuffer/tile. + * This involves converting the fragment colors from float[4] to the + * tile's specific format and obeying the quad/pixel mask. + */ + { + int rgba_reg = spe_allocate_available_register(f); + + /* Pack four float colors as four 32-bit int colors */ + gen_pack_colors(f, color_format, + fragR_reg, fragG_reg, fragB_reg, fragA_reg, + rgba_reg); + + if (blend->logicop_enable) { + gen_logicop(blend, f, rgba_reg, fbRGBA_reg); + } + + if (blend->colormask != 0xf) { + gen_colormask(blend->colormask, f, rgba_reg, fbRGBA_reg); + } + + + /* Mix fragment colors with framebuffer colors using the quad/pixel mask: + * if (mask[i]) + * rgba[i] = rgba[i]; + * else + * rgba[i] = framebuffer[i]; + */ + spe_selb(f, rgba_reg, fbRGBA_reg, rgba_reg, mask_reg); + + /* Store updated quad in tile: + * memory[color_tile + quad_offset] = rgba_reg; + */ + spe_stqx(f, rgba_reg, color_tile_reg, quad_offset_reg); + + spe_release_register(f, rgba_reg); + } + + //printf("gen_fragment_ops nr instructions: %u\n", f->num_inst); + + spe_bi(f, SPE_REG_RA, 0, 0); /* return from function call */ + + + spe_release_register(f, fbRGBA_reg); + spe_release_register(f, fbZS_reg); + spe_release_register(f, quad_offset_reg); +} + diff --git a/src/gallium/drivers/cell/ppu/cell_gen_fragment.h b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h new file mode 100644 index 0000000000..b59de198dc --- /dev/null +++ b/src/gallium/drivers/cell/ppu/cell_gen_fragment.h @@ -0,0 +1,38 @@ +/************************************************************************** + * + * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + + +#ifndef CELL_GEN_FRAGMENT_H +#define CELL_GEN_FRAGMENT_H + + +extern void +cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f); + + +#endif /* CELL_GEN_FRAGMENT_H */ + diff --git a/src/gallium/drivers/cell/ppu/cell_state_emit.c b/src/gallium/drivers/cell/ppu/cell_state_emit.c index f2feaa329a..2da3097983 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_emit.c +++ b/src/gallium/drivers/cell/ppu/cell_state_emit.c @@ -27,6 +27,7 @@ #include "util/u_memory.h" #include "cell_context.h" +#include "cell_gen_fragment.h" #include "cell_state.h" #include "cell_state_emit.h" #include "cell_state_per_fragment.h" @@ -54,23 +55,6 @@ emit_state_cmd(struct cell_context *cell, uint cmd, void cell_emit_state(struct cell_context *cell) { - if (cell->dirty & (CELL_NEW_FRAMEBUFFER | CELL_NEW_BLEND)) { - struct cell_command_logicop logicop; - - if (cell->logic_op.store != NULL) { - spe_release_func(& cell->logic_op); - } - - cell_generate_logic_op(& cell->logic_op, - & cell->blend->base, - cell->framebuffer.cbufs[0]); - - logicop.base = (intptr_t) cell->logic_op.store; - logicop.size = 64 * 4; - emit_state_cmd(cell, CELL_CMD_STATE_LOGICOP, &logicop, - sizeof(logicop)); - } - if (cell->dirty & CELL_NEW_FRAMEBUFFER) { struct pipe_surface *cbuf = cell->framebuffer.cbufs[0]; struct pipe_surface *zbuf = cell->framebuffer.zsbuf; @@ -83,44 +67,49 @@ cell_emit_state(struct cell_context *cell) fb->depth_format = zbuf ? zbuf->format : PIPE_FORMAT_NONE; fb->width = cell->framebuffer.width; fb->height = cell->framebuffer.height; +#if 0 + printf("EMIT color format %s\n", pf_name(fb->color_format)); + printf("EMIT depth format %s\n", pf_name(fb->depth_format)); +#endif } - if (cell->dirty & CELL_NEW_BLEND) { - struct cell_command_blend blend; - - if (cell->blend != NULL) { - blend.base = (intptr_t) cell->blend->code.store; - blend.size = (char *) cell->blend->code.csr - - (char *) cell->blend->code.store; - blend.read_fb = TRUE; - } - else { - blend.base = 0; - blend.size = 0; - blend.read_fb = FALSE; + if (cell->dirty & (CELL_NEW_FS)) { + /* Send new fragment program to SPUs */ + struct cell_command_fragment_program *fp + = cell_batch_alloc(cell, sizeof(*fp)); + fp->opcode = CELL_CMD_STATE_FRAGMENT_PROGRAM; + fp->num_inst = cell->fs->code.num_inst; + memcpy(&fp->code, cell->fs->code.store, + SPU_MAX_FRAGMENT_PROGRAM_INSTS * SPE_INST_SIZE); + if (0) { + int i; + printf("PPU Emit CELL_CMD_STATE_FRAGMENT_PROGRAM:\n"); + for (i = 0; i < fp->num_inst; i++) { + printf(" %3d: 0x%08x\n", i, fp->code[i]); + } } - - emit_state_cmd(cell, CELL_CMD_STATE_BLEND, &blend, sizeof(blend)); } - if (cell->dirty & CELL_NEW_DEPTH_STENCIL) { - struct cell_command_depth_stencil_alpha_test dsat; - - if (cell->depth_stencil != NULL) { - dsat.base = (intptr_t) cell->depth_stencil->code.store; - dsat.size = (char *) cell->depth_stencil->code.csr - - (char *) cell->depth_stencil->code.store; - dsat.read_depth = TRUE; - dsat.read_stencil = FALSE; - } - else { - dsat.base = 0; - dsat.size = 0; - dsat.read_depth = FALSE; - dsat.read_stencil = FALSE; - } - - emit_state_cmd(cell, CELL_CMD_STATE_DEPTH_STENCIL, &dsat, sizeof(dsat)); + if (cell->dirty & (CELL_NEW_FRAMEBUFFER | + CELL_NEW_DEPTH_STENCIL | + CELL_NEW_BLEND)) { + /* XXX we don't want to always do codegen here. We should have + * a hash/lookup table to cache previous results... + */ + struct cell_command_fragment_ops *fops + = cell_batch_alloc(cell, sizeof(*fops)); + struct spe_function spe_code; + + /* generate new code */ + cell_gen_fragment_function(cell, &spe_code); + /* put the new code into the batch buffer */ + fops->opcode = CELL_CMD_STATE_FRAGMENT_OPS; + memcpy(&fops->code, spe_code.store, + SPU_MAX_FRAGMENT_OPS_INSTS * SPE_INST_SIZE); + fops->dsa = cell->depth_stencil->base; + fops->blend = cell->blend->base; + /* free codegen buffer */ + spe_release_func(&spe_code); } if (cell->dirty & CELL_NEW_SAMPLER) { @@ -160,7 +149,8 @@ cell_emit_state(struct cell_context *cell) emit_state_cmd(cell, CELL_CMD_STATE_VERTEX_INFO, &cell->vertex_info, sizeof(struct vertex_info)); } - + +#if 0 if (cell->dirty & CELL_NEW_VS) { const struct draw_context *const draw = cell->draw; struct cell_shader_info info; @@ -175,4 +165,5 @@ cell_emit_state(struct cell_context *cell) emit_state_cmd(cell, CELL_CMD_STATE_BIND_VS, &info, sizeof(info)); } +#endif } diff --git a/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c b/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c index 53ae3aa50e..78cb446c14 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c +++ b/src/gallium/drivers/cell/ppu/cell_state_per_fragment.c @@ -132,9 +132,9 @@ emit_alpha_test(struct pipe_depth_stencil_alpha_state *dsa, /** + * Generate code to perform Z testing. Four Z values are tested at once. * \param dsa Current depth-test state * \param f Function to which code should be appended - * \param m Mask of allocated / free SPE registers * \param mask Index of register to contain depth-pass mask * \param stored Index of register containing values from depth buffer * \param calculated Index of register containing per-fragment depth values @@ -198,6 +198,7 @@ emit_depth_test(struct pipe_depth_stencil_alpha_state *dsa, /** + * Generate code to apply the stencil operation (after testing). * \note Emits a maximum of 5 instructions. * * \warning @@ -222,9 +223,13 @@ emit_stencil_op(struct spe_function *f, spe_il(f, result, ref); break; case PIPE_STENCIL_OP_INCR: + /* clamp = [0xff, 0xff, 0xff, 0xff] */ spe_il(f, clamp, 0x0ff); + /* result[i] = in[i] + 1 */ spe_ai(f, result, in, 1); + /* clamp_mask[i] = (result[i] > 0xff) */ spe_clgti(f, clamp_mask, result, 0x0ff); + /* result[i] = clamp_mask[i] ? clamp[i] : result[i] */ spe_selb(f, result, result, clamp, clamp_mask); break; case PIPE_STENCIL_OP_DECR: @@ -259,10 +264,10 @@ emit_stencil_op(struct spe_function *f, /** + * Generate code to do stencil test. Four pixels are tested at once. * \param dsa Depth / stencil test state * \param face 0 for front face, 1 for back face * \param f Function to append instructions to - * \param reg_mask Mask of allocated registers * \param mask Register containing mask of fragments passing the * alpha test * \param depth_mask Register containing mask of fragments passing the @@ -310,13 +315,14 @@ emit_stencil_test(struct pipe_depth_stencil_alpha_state *dsa, switch (dsa->stencil[face].func) { case PIPE_FUNC_NEVER: - spe_il(f, stencil_mask, 0); + spe_il(f, stencil_mask, 0); /* stencil_mask[0..3] = [0,0,0,0] */ break; case PIPE_FUNC_NOTEQUAL: complement = TRUE; /* FALLTHROUGH */ case PIPE_FUNC_EQUAL: + /* stencil_mask[i] = (stored[i] == ref) */ spe_ceqi(f, stencil_mask, stored, ref); break; @@ -324,6 +330,8 @@ emit_stencil_test(struct pipe_depth_stencil_alpha_state *dsa, complement = TRUE; /* FALLTHROUGH */ case PIPE_FUNC_GREATER: + complement = TRUE; + /* stencil_mask[i] = (stored[i] > ref) */ spe_clgti(f, stencil_mask, stored, ref); break; @@ -331,8 +339,11 @@ emit_stencil_test(struct pipe_depth_stencil_alpha_state *dsa, complement = TRUE; /* FALLTHROUGH */ case PIPE_FUNC_GEQUAL: + /* stencil_mask[i] = (stored[i] > ref) */ spe_clgti(f, stencil_mask, stored, ref); + /* tmp[i] = (stored[i] == ref) */ spe_ceqi(f, tmp, stored, ref); + /* stencil_mask[i] = stencil_mask[i] | tmp[i] */ spe_or(f, stencil_mask, stencil_mask, tmp); break; @@ -461,7 +472,7 @@ cell_generate_depth_stencil_test(struct cell_depth_stencil_alpha_state *cdsa) * + 25 (front stencil) + 25 (back stencil) + 4 = 63 instructions. Round * up to 64 to make it a happy power-of-two. */ - spe_init_func(f, 4 * 64); + spe_init_func(f, SPE_INST_SIZE * 64); /* Allocate registers for the function's input parameters. Cleverly (and @@ -540,7 +551,7 @@ cell_generate_depth_stencil_test(struct cell_depth_stencil_alpha_state *cdsa) spe_selb(f, depth, depth, zvals, mask); } - spe_bi(f, 0, 0, 0); + spe_bi(f, 0, 0, 0); /* return from function call */ #if 0 @@ -956,7 +967,7 @@ cell_generate_alpha_blend(struct cell_blend_state *cb) * + 4 (fragment mask) + 1 (return) = 55 instlructions. Round up to 64 to * make it a happy power-of-two. */ - spe_init_func(f, 4 * 64); + spe_init_func(f, SPE_INST_SIZE * 64); const int frag[4] = { @@ -1144,9 +1155,10 @@ cell_generate_alpha_blend(struct cell_blend_state *cb) } -int PC_OFFSET(const struct spe_function *f, const void *d) +static int +PC_OFFSET(const struct spe_function *f, const void *d) { - const intptr_t pc = (intptr_t) f->csr; + const intptr_t pc = (intptr_t) &f->store[f->num_inst]; const intptr_t ea = ~0x0f & (intptr_t) d; return (ea - pc) >> 2; @@ -1178,7 +1190,7 @@ cell_generate_logic_op(struct spe_function *f, * bytes (equiv. to 8 instructions) are needed for data storage. Round up * to 64 to make it a happy power-of-two. */ - spe_init_func(f, 4 * 64); + spe_init_func(f, SPE_INST_SIZE * 64); /* Pixel colors in framebuffer format in AoS layout. diff --git a/src/gallium/drivers/cell/ppu/cell_state_shader.c b/src/gallium/drivers/cell/ppu/cell_state_shader.c index 97e44eeb1a..3a0d066da2 100644 --- a/src/gallium/drivers/cell/ppu/cell_state_shader.c +++ b/src/gallium/drivers/cell/ppu/cell_state_shader.c @@ -34,7 +34,7 @@ #include "cell_context.h" #include "cell_state.h" - +#include "cell_gen_fp.h" /** cast wrapper */ @@ -61,7 +61,7 @@ static void * cell_create_fs_state(struct pipe_context *pipe, const struct pipe_shader_state *templ) { - /*struct cell_context *cell = cell_context(pipe);*/ + struct cell_context *cell = cell_context(pipe); struct cell_fragment_shader_state *cfs; cfs = CALLOC_STRUCT(cell_fragment_shader_state); @@ -76,6 +76,8 @@ cell_create_fs_state(struct pipe_context *pipe, tgsi_scan_shader(templ->tokens, &cfs->info); + cell_gen_fragment_program(cell, cfs->shader.tokens, &cfs->code); + return cfs; } @@ -102,6 +104,8 @@ cell_delete_fs_state(struct pipe_context *pipe, void *fs) { struct cell_fragment_shader_state *cfs = cell_fragment_shader_state(fs); + spe_release_func(&cfs->code); + FREE((void *) cfs->shader.tokens); FREE(cfs); } diff --git a/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c b/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c index 2ece0250f6..566df7f59e 100644 --- a/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c +++ b/src/gallium/drivers/cell/ppu/cell_vertex_fetch.c @@ -297,10 +297,9 @@ void cell_update_vertex_fetch(struct draw_context *draw) /* Each fetch function can be a maximum of 34 instructions (note: this is - * actually a slight over-estimate). That means (34 * 4) = 136 bytes - * each maximum. + * actually a slight over-estimate). */ - spe_init_func(p, 136 * unique_attr_formats); + spe_init_func(p, 34 * SPE_INST_SIZE * unique_attr_formats); /* Allocate registers for the function's input parameters. |