i965/blorp: switch eu-emitter to use FS IR and fs_generator
authorTopi Pohjolainen <topi.pohjolainen@intel.com>
Tue, 10 Dec 2013 13:12:30 +0000 (15:12 +0200)
committerTopi Pohjolainen <topi.pohjolainen@intel.com>
Thu, 23 Jan 2014 06:47:12 +0000 (08:47 +0200)
No regressions on IVB (piglit quick + unit tests).

v2 (Paul):
  - no need to patch the unit tests anymore. Original logic
    was altered and unit tests updated to match the
    fs-generator
  - lrp emission moves from the blorp compiler core into the
    emitter here (previously there was a separate refactoring
    patch which is not really needed anymore as the lrp logic
    got refactored when the original lrp logic got fixed).
  - pass 'BRW_BLORP_RENDERBUFFER_BINDING_TABLE_INDEX' to the
    generator in fs_inst::target instead of hardcoding it

Signed-off-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
src/mesa/drivers/dri/i965/brw_blorp_blit.cpp
src/mesa/drivers/dri/i965/brw_blorp_blit_eu.cpp
src/mesa/drivers/dri/i965/brw_blorp_blit_eu.h

index 6454d2a7ede049b5a6addfbff6949010e0bee7a1..c4d1108bcdf110382aacc85d9c92d3e9005b697e 100644 (file)
@@ -654,11 +654,6 @@ private:
                        const sampler_message_arg *args, int num_args);
    void render_target_write();
 
-   void emit_lrp(const struct brw_reg &dst,
-                 const struct brw_reg &src1,
-                 const struct brw_reg &src2,
-                 const struct brw_reg &src3);
-
    /**
     * Base-2 logarithm of the maximum number of samples that can be blended.
     */
@@ -1584,21 +1579,6 @@ brw_blorp_blit_program::manual_blend_average(unsigned num_samples)
       emit_endif();
 }
 
-void
-brw_blorp_blit_program::emit_lrp(const struct brw_reg &dst,
-                                 const struct brw_reg &src1,
-                                 const struct brw_reg &src2,
-                                 const struct brw_reg &src3)
-{
-   brw_set_access_mode(&func, BRW_ALIGN_16);
-   brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
-   brw_LRP(&func, dst, src1, src2, src3);
-   brw_set_compression_control(&func, BRW_COMPRESSION_2NDHALF);
-   brw_LRP(&func, sechalf(dst), sechalf(src1), sechalf(src2), sechalf(src3));
-   brw_set_compression_control(&func, BRW_COMPRESSION_COMPRESSED);
-   brw_set_access_mode(&func, BRW_ALIGN_1);
-}
-
 void
 brw_blorp_blit_program::manual_blend_bilinear(unsigned num_samples)
 {
index 9b634589e7df38428b200dcecfc246cfae6a091a..5b652add1e4f382ab74f47b02efb43b67193d1ea 100644 (file)
 #include "brw_blorp.h"
 
 brw_blorp_eu_emitter::brw_blorp_eu_emitter(struct brw_context *brw)
-   : mem_ctx(ralloc_context(NULL))
+   : mem_ctx(ralloc_context(NULL)), c(rzalloc(mem_ctx, struct brw_wm_compile)),
+     generator(brw, c, NULL, NULL, false)
 {
-   brw_init_compile(brw, &func, mem_ctx);
-
-   /*
-    * By default everything is emitted as 16-wide with only a few expections
-    * handled explicitly either here in the compiler or by one of the specific
-    * code emission calls.
-    * It should be also noted that here in this file any alterations of the
-    * compression control settings are only used to affect the execution size
-    * of the instructions. The instruction template used to initialise all the
-    * instructions is effectively not altered -- the value stays at zero
-    * representing either GEN6_COMPRESSION_1Q or GEN6_COMPRESSION_1H depending
-    * on the context.
-    * If any other settings are used in the instruction headers, they are set
-    * elsewhere by the individual code emission calls.
-    */
-   brw_set_compression_control(&func, BRW_COMPRESSION_COMPRESSED);
 }
 
 brw_blorp_eu_emitter::~brw_blorp_eu_emitter()
@@ -54,15 +39,17 @@ brw_blorp_eu_emitter::~brw_blorp_eu_emitter()
 const unsigned *
 brw_blorp_eu_emitter::get_program(unsigned *program_size, FILE *dump_file)
 {
-   brw_set_uip_jip(&func);
+   const unsigned *res;
 
    if (unlikely(INTEL_DEBUG & DEBUG_BLORP)) {
       printf("Native code for BLORP blit:\n");
-      brw_dump_compile(&func, dump_file, 0, func.next_insn_offset);
+      res = generator.generate_assembly(NULL, &insts, program_size, dump_file);
       printf("\n");
+   } else {
+      res = generator.generate_assembly(NULL, &insts, program_size);
    }
 
-   return brw_get_program(&func, program_size);
+   return res;
 }
 
 /**
@@ -80,17 +67,15 @@ brw_blorp_eu_emitter::emit_kill_if_outside_rect(const struct brw_reg &x,
 {
    struct brw_reg f0 = brw_flag_reg(0, 0);
    struct brw_reg g1 = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
-   struct brw_reg null32 = vec16(retype(brw_null_reg(), BRW_REGISTER_TYPE_UD));
 
-   brw_CMP(&func, null32, BRW_CONDITIONAL_GE, x, dst_x0);
-   brw_CMP(&func, null32, BRW_CONDITIONAL_GE, y, dst_y0);
-   brw_CMP(&func, null32, BRW_CONDITIONAL_L, x, dst_x1);
-   brw_CMP(&func, null32, BRW_CONDITIONAL_L, y, dst_y1);
+   emit_cmp(BRW_CONDITIONAL_GE, x, dst_x0);
+   emit_cmp(BRW_CONDITIONAL_GE, y, dst_y0)->predicate = BRW_PREDICATE_NORMAL;
+   emit_cmp(BRW_CONDITIONAL_L, x, dst_x1)->predicate = BRW_PREDICATE_NORMAL;
+   emit_cmp(BRW_CONDITIONAL_L, y, dst_y1)->predicate = BRW_PREDICATE_NORMAL;
 
-   brw_set_predicate_control(&func, BRW_PREDICATE_NONE);
-
-   struct brw_instruction *inst = brw_AND(&func, g1, f0, g1);
-   inst->header.mask_control = BRW_MASK_DISABLE;
+   fs_inst *inst = new (mem_ctx) fs_inst(BRW_OPCODE_AND, g1, f0, g1);
+   inst->force_writemask_all = true;
+   insts.push_tail(inst);
 }
 
 void
@@ -99,40 +84,14 @@ brw_blorp_eu_emitter::emit_texture_lookup(const struct brw_reg &dst,
                                           unsigned base_mrf,
                                           unsigned msg_length)
 {
-   unsigned msg_type;
-
-   switch (op) {
-   case SHADER_OPCODE_TEX:
-      msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
-      break;
-   case SHADER_OPCODE_TXF:
-      msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
-      break;
-   case SHADER_OPCODE_TXF_CMS:
-      msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
-      break;
-   case SHADER_OPCODE_TXF_UMS:
-      msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS;
-      break;
-   case SHADER_OPCODE_TXF_MCS:
-      msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
-      break;
-   default:
-      assert(!"Unsupported texture lookup operation");
-   }
+   fs_inst *inst = new (mem_ctx) fs_inst(op, dst, brw_message_reg(base_mrf));
+
+   inst->base_mrf = base_mrf;
+   inst->mlen = msg_length;
+   inst->sampler = 0;
+   inst->header_present = false;
 
-   brw_SAMPLE(&func,
-              retype(dst, BRW_REGISTER_TYPE_UW) /* dest */,
-              base_mrf /* msg_reg_nr */,
-              brw_message_reg(base_mrf) /* src0 */,
-              BRW_BLORP_TEXTURE_BINDING_TABLE_INDEX,
-              0 /* sampler */,
-              msg_type,
-              8 /* response_length.  TODO: should be smaller for non-RGBA formats? */,
-              msg_length,
-              0 /* header_present */,
-              BRW_SAMPLER_SIMD_MODE_SIMD16,
-              BRW_SAMPLER_RETURN_FORMAT_FLOAT32);
+   insts.push_tail(inst);
 }
 
 void
@@ -141,16 +100,15 @@ brw_blorp_eu_emitter::emit_render_target_write(const struct brw_reg &src0,
                                                unsigned msg_length,
                                                bool use_header)
 {
-   brw_fb_WRITE(&func,
-                16 /* dispatch_width */,
-                msg_reg_nr,
-                src0,
-                BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE,
-                BRW_BLORP_RENDERBUFFER_BINDING_TABLE_INDEX,
-                msg_length,
-                0 /* response_length */,
-                true /* eot */,
-                use_header);
+   fs_inst *inst = new (mem_ctx) fs_inst(FS_OPCODE_BLORP_FB_WRITE);
+
+   inst->src[0] = src0;
+   inst->base_mrf = msg_reg_nr;
+   inst->mlen = msg_length;
+   inst->header_present = use_header;
+   inst->target = BRW_BLORP_RENDERBUFFER_BINDING_TABLE_INDEX;
+
+   insts.push_tail(inst);
 }
 
 void
@@ -161,8 +119,18 @@ brw_blorp_eu_emitter::emit_combine(enum opcode combine_opcode,
 {
    assert(combine_opcode == BRW_OPCODE_ADD || combine_opcode == BRW_OPCODE_AVG);
 
-   if (combine_opcode == BRW_OPCODE_ADD)
-      brw_ADD(&func, dst, src_1, src_2);
-   else
-      brw_AVG(&func, dst, src_1, src_2);
+   insts.push_tail(new (mem_ctx) fs_inst(combine_opcode, dst, src_1, src_2));
 }
+
+fs_inst *
+brw_blorp_eu_emitter::emit_cmp(int op,
+                               const struct brw_reg &x,
+                               const struct brw_reg &y)
+{
+   fs_inst *cmp = new (mem_ctx) fs_inst(BRW_OPCODE_CMP,
+                                        vec16(brw_null_reg()), x, y);
+   cmp->conditional_mod = op;
+   insts.push_tail(cmp);
+   return cmp;
+}
+
index 736f5b06ff3cb96e3ba62a31c9134ca693a7da7a..c10695e4b48a8f6729aad3b361b9c1077295e332 100644 (file)
@@ -25,7 +25,7 @@
 #define BRW_BLORP_BLIT_EU_H
 
 #include "brw_context.h"
-#include "brw_eu.h"
+#include "brw_fs.h"
 
 class brw_blorp_eu_emitter
 {
@@ -63,9 +63,11 @@ protected:
                              const struct brw_reg &dst,
                              const struct brw_reg &src)
    {
-      brw_CMP(&func, vec16(brw_null_reg()), op, x, y);
-      brw_MOV(&func, dst, src);
-      brw_set_predicate_control(&func, BRW_PREDICATE_NONE);
+      emit_cmp(op, x, y);
+
+      fs_inst *mv = new (mem_ctx) fs_inst(BRW_OPCODE_MOV, dst, src);
+      mv->predicate = BRW_PREDICATE_NORMAL;
+      insts.push_tail(mv);
    }
 
    inline void emit_if_eq_mov(const struct brw_reg &x, unsigned y,
@@ -74,101 +76,115 @@ protected:
       emit_cond_mov(x, brw_imm_d(y), BRW_CONDITIONAL_EQ, dst, brw_imm_d(src));
    }
 
+   inline void emit_lrp(const struct brw_reg &dst,
+                        const struct brw_reg &src1,
+                        const struct brw_reg &src2,
+                        const struct brw_reg &src3)
+   {
+      insts.push_tail(
+         new (mem_ctx) fs_inst(BRW_OPCODE_LRP, dst, src1, src2, src3));
+   }
+
    inline void emit_mov(const struct brw_reg& dst, const struct brw_reg& src)
    {
-      brw_MOV(&func, dst, src);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_MOV, dst, src));
    }
 
    inline void emit_mov_8(const struct brw_reg& dst, const struct brw_reg& src)
    {
-      brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
-      emit_mov(dst, src);
-      brw_set_compression_control(&func, BRW_COMPRESSION_COMPRESSED);
+      fs_inst *mv = new (mem_ctx) fs_inst(BRW_OPCODE_MOV, dst, src);
+      mv->force_uncompressed = true;
+      insts.push_tail(mv);
    }
 
    inline void emit_and(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      brw_AND(&func, dst, src1, src2);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_AND, dst, src1, src2));
    }
 
    inline void emit_add(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      brw_ADD(&func, dst, src1, src2);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ADD, dst, src1, src2));
    }
 
    inline void emit_add_8(const struct brw_reg& dst,
                           const struct brw_reg& src1,
                           const struct brw_reg& src2)
    {
-      brw_set_compression_control(&func, BRW_COMPRESSION_NONE);
-      emit_add(dst, src1, src2);
-      brw_set_compression_control(&func, BRW_COMPRESSION_COMPRESSED);
+      fs_inst *add = new (mem_ctx) fs_inst(BRW_OPCODE_ADD, dst, src1, src2);
+      add->force_uncompressed = true;
+      insts.push_tail(add);
    }
 
    inline void emit_mul(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      brw_MUL(&func, dst, src1, src2);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_MUL, dst, src1, src2));
    }
 
    inline void emit_shr(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      brw_SHR(&func, dst, src1, src2);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_SHR, dst, src1, src2));
    }
 
    inline void emit_shl(const struct brw_reg& dst,
                         const struct brw_reg& src1,
                         const struct brw_reg& src2)
    {
-      brw_SHL(&func, dst, src1, src2);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_SHL, dst, src1, src2));
    }
 
    inline void emit_or(const struct brw_reg& dst,
                        const struct brw_reg& src1,
                        const struct brw_reg& src2)
    {
-      brw_OR(&func, dst, src1, src2);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_OR, dst, src1, src2));
    }
 
    inline void emit_frc(const struct brw_reg& dst,
                         const struct brw_reg& src)
    {
-      brw_FRC(&func, dst, src);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_FRC, dst, src));
    }
 
    inline void emit_rndd(const struct brw_reg& dst,
                          const struct brw_reg& src)
    {
-      brw_RNDD(&func, dst, src);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_RNDD, dst, src));
    }
 
    inline void emit_cmp_if(int op,
                            const struct brw_reg &x,
                            const struct brw_reg &y)
    {
-      brw_CMP(&func, vec16(brw_null_reg()), op, x, y);
-      brw_IF(&func, BRW_EXECUTE_16);
+      emit_cmp(op, x, y);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_IF));
    }
 
    inline void emit_else(void)
    {
-      brw_ELSE(&func);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ELSE));
    }
 
    inline void emit_endif(void)
    {
-      brw_ENDIF(&func);
+      insts.push_tail(new (mem_ctx) fs_inst(BRW_OPCODE_ENDIF));
    }
 
+private:
+   fs_inst *emit_cmp(int op, const struct brw_reg &x, const struct brw_reg &y);
+
    void *mem_ctx;
-   struct brw_compile func;
+   struct brw_wm_compile *c;
+   exec_list insts;
+   fs_generator generator;
 };
 
 #endif /* BRW_BLORP_BLIT_EU_H */