pipeline: add SWX instruction optimizer
authorCristian Dumitrescu <cristian.dumitrescu@intel.com>
Thu, 1 Oct 2020 10:19:56 +0000 (11:19 +0100)
committerDavid Marchand <david.marchand@redhat.com>
Thu, 1 Oct 2020 16:43:08 +0000 (18:43 +0200)
Instruction optimizer. Detects frequent patterns and replaces them
with some more powerful vector-like pipeline instructions without any
user effort. Executes at instruction translation, not at run-time.

Signed-off-by: Cristian Dumitrescu <cristian.dumitrescu@intel.com>
lib/librte_pipeline/rte_swx_pipeline.c

index d51fec8..77eae19 100644 (file)
@@ -5700,6 +5700,230 @@ instr_verify(struct rte_swx_pipeline *p __rte_unused,
        return 0;
 }
 
+static int
+instr_pattern_extract_many_detect(struct instruction *instr,
+                                 struct instruction_data *data,
+                                 uint32_t n_instr,
+                                 uint32_t *n_pattern_instr)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instr; i++) {
+               if (data[i].invalid)
+                       break;
+
+               if (instr[i].type != INSTR_HDR_EXTRACT)
+                       break;
+
+               if (i == RTE_DIM(instr->io.hdr.header_id))
+                       break;
+
+               if (i && data[i].n_users)
+                       break;
+       }
+
+       if (i < 2)
+               return 0;
+
+       *n_pattern_instr = i;
+       return 1;
+}
+
+static void
+instr_pattern_extract_many_optimize(struct instruction *instr,
+                                   struct instruction_data *data,
+                                   uint32_t n_instr)
+{
+       uint32_t i;
+
+       for (i = 1; i < n_instr; i++) {
+               instr[0].type++;
+               instr[0].io.hdr.header_id[i] = instr[i].io.hdr.header_id[0];
+               instr[0].io.hdr.struct_id[i] = instr[i].io.hdr.struct_id[0];
+               instr[0].io.hdr.n_bytes[i] = instr[i].io.hdr.n_bytes[0];
+
+               data[i].invalid = 1;
+       }
+}
+
+static int
+instr_pattern_emit_many_tx_detect(struct instruction *instr,
+                                 struct instruction_data *data,
+                                 uint32_t n_instr,
+                                 uint32_t *n_pattern_instr)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instr; i++) {
+               if (data[i].invalid)
+                       break;
+
+               if (instr[i].type != INSTR_HDR_EMIT)
+                       break;
+
+               if (i == RTE_DIM(instr->io.hdr.header_id))
+                       break;
+
+               if (i && data[i].n_users)
+                       break;
+       }
+
+       if (!i)
+               return 0;
+
+       if (instr[i].type != INSTR_TX)
+               return 0;
+
+       i++;
+
+       *n_pattern_instr = i;
+       return 1;
+}
+
+static void
+instr_pattern_emit_many_tx_optimize(struct instruction *instr,
+                                   struct instruction_data *data,
+                                   uint32_t n_instr)
+{
+       uint32_t i;
+
+       /* Any emit instruction in addition to the first one. */
+       for (i = 1; i < n_instr - 1; i++) {
+               instr[0].type++;
+               instr[0].io.hdr.header_id[i] = instr[i].io.hdr.header_id[0];
+               instr[0].io.hdr.struct_id[i] = instr[i].io.hdr.struct_id[0];
+               instr[0].io.hdr.n_bytes[i] = instr[i].io.hdr.n_bytes[0];
+
+               data[i].invalid = 1;
+       }
+
+       /* The TX instruction is the last one in the pattern. */
+       instr[0].type++;
+       instr[0].io.io.offset = instr[i].io.io.offset;
+       instr[0].io.io.n_bits = instr[i].io.io.n_bits;
+       data[i].invalid = 1;
+}
+
+static int
+instr_pattern_dma_many_detect(struct instruction *instr,
+                             struct instruction_data *data,
+                             uint32_t n_instr,
+                             uint32_t *n_pattern_instr)
+{
+       uint32_t i;
+
+       for (i = 0; i < n_instr; i++) {
+               if (data[i].invalid)
+                       break;
+
+               if (instr[i].type != INSTR_DMA_HT)
+                       break;
+
+               if (i == RTE_DIM(instr->dma.dst.header_id))
+                       break;
+
+               if (i && data[i].n_users)
+                       break;
+       }
+
+       if (i < 2)
+               return 0;
+
+       *n_pattern_instr = i;
+       return 1;
+}
+
+static void
+instr_pattern_dma_many_optimize(struct instruction *instr,
+                               struct instruction_data *data,
+                               uint32_t n_instr)
+{
+       uint32_t i;
+
+       for (i = 1; i < n_instr; i++) {
+               instr[0].type++;
+               instr[0].dma.dst.header_id[i] = instr[i].dma.dst.header_id[0];
+               instr[0].dma.dst.struct_id[i] = instr[i].dma.dst.struct_id[0];
+               instr[0].dma.src.offset[i] = instr[i].dma.src.offset[0];
+               instr[0].dma.n_bytes[i] = instr[i].dma.n_bytes[0];
+
+               data[i].invalid = 1;
+       }
+}
+
+static uint32_t
+instr_optimize(struct instruction *instructions,
+              struct instruction_data *instruction_data,
+              uint32_t n_instructions)
+{
+       uint32_t i, pos = 0;
+
+       for (i = 0; i < n_instructions; ) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+               uint32_t n_instr = 0;
+               int detected;
+
+               /* Extract many. */
+               detected = instr_pattern_extract_many_detect(instr,
+                                                            data,
+                                                            n_instructions - i,
+                                                            &n_instr);
+               if (detected) {
+                       instr_pattern_extract_many_optimize(instr,
+                                                           data,
+                                                           n_instr);
+                       i += n_instr;
+                       continue;
+               }
+
+               /* Emit many + TX. */
+               detected = instr_pattern_emit_many_tx_detect(instr,
+                                                            data,
+                                                            n_instructions - i,
+                                                            &n_instr);
+               if (detected) {
+                       instr_pattern_emit_many_tx_optimize(instr,
+                                                           data,
+                                                           n_instr);
+                       i += n_instr;
+                       continue;
+               }
+
+               /* DMA many. */
+               detected = instr_pattern_dma_many_detect(instr,
+                                                        data,
+                                                        n_instructions - i,
+                                                        &n_instr);
+               if (detected) {
+                       instr_pattern_dma_many_optimize(instr, data, n_instr);
+                       i += n_instr;
+                       continue;
+               }
+
+               /* No pattern starting at the current instruction. */
+               i++;
+       }
+
+       /* Eliminate the invalid instructions that have been optimized out. */
+       for (i = 0; i < n_instructions; i++) {
+               struct instruction *instr = &instructions[i];
+               struct instruction_data *data = &instruction_data[i];
+
+               if (data->invalid)
+                       continue;
+
+               if (i != pos) {
+                       memcpy(&instructions[pos], instr, sizeof(*instr));
+                       memcpy(&instruction_data[pos], data, sizeof(*data));
+               }
+
+               pos++;
+       }
+
+       return pos;
+}
+
 static int
 instruction_config(struct rte_swx_pipeline *p,
                   struct action *a,
@@ -5752,6 +5976,8 @@ instruction_config(struct rte_swx_pipeline *p,
        if (err)
                goto error;
 
+       n_instructions = instr_optimize(instr, data, n_instructions);
+
        err = instr_jmp_resolve(instr, data, n_instructions);
        if (err)
                goto error;