[fix] Fix block header decompression for inflate with indexing on asynchronous path (#803)

Co-authored-by: Meng Cao <[email protected]>

Author: abdelrahim-hentabli

doc/source/documentation/dev_guide_docs/c_use_cases/deflate/c_deflate_indexing_usage.rst

@@ -171,6 +171,12 @@ After the job completes, the returned CRC value (in job crc) should
 match the CRC value from the end of the mini-block, in this example 10.
 If it does not match, then the mini-block is not decompressed properly.
 
+.. note::
+
+    The state of decompression with mini-blocks is not saved from job
+    to job. Namely the :c:member:`qpl_job.total_out` will not store the
+    cumulative number of bytes written, but only that job's particular
+    mini-block's number of bytes written.
 
 Single Block Usage
 ==================

sources/c_api/legacy_hw_path/hardware_defs.h

@@ -88,6 +88,13 @@ extern "C" {
 #define IS_RND_ACCESS_BODY(flag) \
     (((flag) & (QPL_FLAG_RND_ACCESS | QPL_FLAG_FIRST)) == QPL_FLAG_RND_ACCESS)
 
+
+/**
+ * Random Access Header is FLAG_RND_ACCESS and FLAG_FIRST
+ */
+#define IS_RND_ACCESS_HDR(flag) \
+    (((flag) & (QPL_FLAG_RND_ACCESS | QPL_FLAG_FIRST)) == (QPL_FLAG_RND_ACCESS | QPL_FLAG_FIRST))
+
 /**
  * @todo
  */

sources/c_api/legacy_hw_path/qpl_hw_inflate_job.cpp

@@ -54,8 +54,13 @@ extern "C" qpl_status hw_submit_decompress_job(qpl_job *qpl_job_ptr,
     auto *const state_ptr = reinterpret_cast<qpl_hw_state *>(job::get_state(qpl_job_ptr));
 
     hw_iaa_analytics_descriptor *const desc_ptr = &state_ptr->desc_ptr;
-    hw_iaa_aecs_analytic    *const aecs_ptr = GET_DCFG(state_ptr);
+    hw_iaa_aecs_analytic        *aecs_ptr       = GET_DCFG(state_ptr);
 
+    if (IS_RND_ACCESS_HDR(qpl_job_ptr->flags)) {
+        aecs_ptr = (hw_iaa_aecs_analytic *) ((uint8_t *) (state_ptr->dcfg) + HW_AECS_FILTER_AND_DECOMPRESS_WA_HB);
+    }
+
+    HW_IMMEDIATELY_RET_NULLPTR(aecs_ptr);
     last_job = (available_in > MAX_BUF_SIZE) ? 0u : last_job;
 
     // Descriptor buffers set
@@ -79,7 +84,6 @@ extern "C" qpl_status hw_submit_decompress_job(qpl_job *qpl_job_ptr,
     const bool is_dictionary_mode = (qpl_job_ptr->flags & QPL_FLAG_FIRST ||
                                      !state_ptr->execution_history.first_job_has_been_submitted) &&
                                              qpl_job_ptr->dictionary != NULL;
-
     bool is_aecs_format2_expected = qpl::ml::util::are_iaa_gen_2_min_capabilities_present();
 
     if (state_ptr->execution_history.first_job_has_been_submitted) {
@@ -92,13 +96,6 @@ extern "C" qpl_status hw_submit_decompress_job(qpl_job *qpl_job_ptr,
             operation_flags |= ADOF_READ_SRC2(AD_RDSRC2_AECS);
         }
 
-        // Decompress random header
-        if (qpl_job_ptr->ignore_start_bits != 0u) {
-            operation_flags |= ADOF_READ_SRC2(AD_RDSRC2_AECS);
-            core_sw::util::set_zeros((uint8_t *) aecs_ptr, sizeof(hw_iaa_aecs_analytic));
-            aecs_ptr->inflate_options.decompress_state = DEF_STATE_HDR;
-        }
-
         // Decompress huffman only
         if (qpl_job_ptr->flags & QPL_FLAG_NO_HDRS) {
             operation_flags |= ADOF_READ_SRC2(AD_RDSRC2_AECS);
@@ -131,13 +128,14 @@ extern "C" qpl_status hw_submit_decompress_job(qpl_job *qpl_job_ptr,
 
     }
 
-    // Set the input accum alignment if we're doing random access
-    if ((qpl_job_ptr->flags & QPL_FLAG_RND_ACCESS)
-        && (hw_iaa_aecs_decompress_is_empty_input_accumulator(&aecs_ptr->inflate_options))) {
-        aecs_ptr->inflate_options.idx_bit_offset = 7u & qpl_job_ptr->ignore_start_bits;
-    }
 
     if (0u != qpl_job_ptr->ignore_start_bits) {
+        if (IS_RND_ACCESS_HDR(qpl_job_ptr->flags)) {
+            operation_flags |= ADOF_READ_SRC2(AD_RDSRC2_AECS);
+            core_sw::util::set_zeros(reinterpret_cast<uint8_t *>(aecs_ptr), HW_AECS_FILTER_AND_DECOMPRESS_WA_HB);
+            aecs_ptr->inflate_options.idx_bit_offset = 7u & qpl_job_ptr->ignore_start_bits;
+            aecs_ptr->inflate_options.decompress_state = DEF_STATE_HDR;
+        }
         auto status = hw_iaa_aecs_decompress_set_input_accumulator(&aecs_ptr->inflate_options,
                                                                    desc_ptr->src1_ptr,
                                                                    qpl_job_ptr->available_in,
@@ -149,10 +147,12 @@ extern "C" qpl_status hw_submit_decompress_job(qpl_job *qpl_job_ptr,
         desc_ptr->src1_ptr  = ++qpl_job_ptr->next_in_ptr;
         desc_ptr->src1_size = --qpl_job_ptr->available_in;
     }
-
     hw_iaa_aecs_decompress_state_set_aecs_format(&aecs_ptr->inflate_options, is_aecs_format2_expected);
 
     // AECS Write policy
+    if (IS_RND_ACCESS_HDR(qpl_job_ptr->flags)) {
+        operation_flags     |= ADOF_AECS_SEL;
+    }
     if (IS_RND_ACCESS_BODY(qpl_job_ptr->flags)) {
         decompression_flags |= ADDF_FLUSH_OUTPUT;
         hw_iaa_aecs_decompress_set_crc_seed(aecs_ptr, qpl_job_ptr->crc);

sources/c_api/legacy_hw_path/qpl_hw_submit_job.cpp

@@ -400,7 +400,7 @@ extern "C" qpl_status hw_submit_job (qpl_job * qpl_job_ptr) {
                                QPL_STS_OPERATION_ERR)
             return hw_submit_analytic_task(qpl_job_ptr);
 
-        case qpl_op_decompress:
+        case qpl_op_decompress: {
             if (qpl_job_ptr->dictionary != NULL && qpl_job_ptr->flags & QPL_FLAG_CANNED_MODE) {
                 // dictionary with canned mode
                 // TODO: remove once it's supported
@@ -416,9 +416,16 @@ extern "C" qpl_status hw_submit_job (qpl_job * qpl_job_ptr) {
                 break; // Workaround for header reading
             }
 
+            uint32_t crc = qpl_job_ptr->crc;
             job::reset<qpl_op_decompress>(qpl_job_ptr);
+
+            if (flags & QPL_FLAG_RND_ACCESS){
+                own_hw_state_reset(state_ptr);
+                qpl_job_ptr->crc = crc;
+            }
             state_ptr->aecs_size = HW_AECS_FILTER_AND_DECOMPRESS_WA_HB;
             return hw_submit_task(qpl_job_ptr);
+        }
         case qpl_op_compress:
             if (flags & QPL_FLAG_FIRST) {
                 job::reset<qpl_op_compress>(qpl_job_ptr);
@@ -483,9 +490,10 @@ extern "C" qpl_status hw_submit_job (qpl_job * qpl_job_ptr) {
     uint32_t source_size = qpl_job_ptr->available_in;
 
     qpl_buffer *const accumulator_ptr = &state_ptr->accumulation_buffer;
-    bool is_last_job = flags & QPL_FLAG_LAST;
+    bool is_last_job      = flags & QPL_FLAG_LAST;
+    bool is_indexing_mode = flags & QPL_FLAG_RND_ACCESS;
 
-    if ((!is_last_job)
+    if ((!is_last_job && !is_indexing_mode)
         && own_qpl_buffer_touch(accumulator_ptr, source_size)) {
         own_qpl_buffer_fill(accumulator_ptr, source_ptr, source_size);
         hw_iaa_completion_record_init_trivial_completion(&state_ptr->comp_ptr, source_size);

sources/core-iaa/sources/include/own_hw_definitions.h

@@ -113,6 +113,13 @@ extern "C" {
 #define IS_RND_ACCESS_BODY(flag) \
     (((flag) & (QPL_FLAG_RND_ACCESS | QPL_FLAG_FIRST)) == QPL_FLAG_RND_ACCESS)
 
+/**
+ * Random Access Header is FLAG_RND_ACCESS and FLAG_FIRST
+ */
+#define IS_RND_ACCESS_HDR(flag) \
+    (((flag) & (QPL_FLAG_RND_ACCESS | QPL_FLAG_FIRST)) == (QPL_FLAG_RND_ACCESS | QPL_FLAG_FIRST))
+
+
 #ifdef __GNUC__
 #define UNREFERENCED_PARAMETER(p) p __attribute__((unused)) /**< Unreferenced parameter - warning removal */
 #else